text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: |
We present an example of the gauge mediated SUSY breaking flipped $SU(5)$ model. The messengers of the SUSY breaking are either only colour triplets which belong to the minimal content of the scalar supermultiplets or together with triplets as a messengers emerge the ordinary Higgs doublets. In both cases the model predicts light gauginos in respect of the squarks and sleptons, which could be tested in the nearest LEP experiments.
In both cases ”all order“ solution of the doublet-triplet splitting problem is obtained, the $\mu $-term of the order of $100$ GeV is generated and the left handed neutrino masses are suppressed.
---
-2cm -1.5cm
[**Realistic GUT with Gauge Mediated\
Supersymmetry Breaking**]{}
[**Zurab Tavartkiladze [^1]** ]{}
[*INFN Sezione di Ferrara, I-44100 Ferrara, Italy\
and Institute of Physics of Georgian Academy of Sciences, 380077 Tbilisi, Georgia* ]{}
The supersymmetric theories suggest the elegant possibilities for solution of the gauge hierarchy problem. Non-renormalization theorems [@wess] in SUSY theories imply that certain ratios of coupling constants are non-renormalized in exact SUSY limit. This nice feature and also the successful prediction of the numerical value of $\sin^2 \theta_W$ [@lan] supported the idea of the SUSY Grand Unified Theories (GUT).
The most interesting question is the origin of the SUSY breaking. It is usually assumed that SUSY is broken in a ”hidden“ sector and by some interactions transmitted in the visible sector. The most famous scenarios are the supergravity theories [@barb] , in which the SUSY breaking in the visible sector transmitted by the gravity. In this case the soft SUSY breaking (SSB) terms are presented at the energies which correspond to the Planck scale - $M_P$ and even they have the universal form, they will renormalized between the $M_P$ and $M_{GUT}$ . At $M_{GUT}$ one has to integrate out the heavy particles and evolve again the SSB parameters from $M_{GUT}$ to $m_W$ scale with the RGEs of the MSSM. These processes violate the universalities (see [@pom] and references there) and lead to the flavour changing neutral currents (FCNC).
To another class of the SUSY breaking scenarios belong the gauge mediated supersymmetry breaking (GMSB) models [@dine; @fer] , in which the supersymmetry breaking is transmitted by the gauge interactions. Because the fact, that this models do not suffer from the FCNC problem the interest in models of this type was renewed recently [@nel; @amb] .
In this paper we present an example of the SUSY GUT in which the SUSY breaking occurs in the sector of scalar superfields which are used for the GUT symmetry breaking.
The main contribution to the soft masses to the squarks and sleptons comes from the nonzero $D$-term, which is just of the order of SUSY scale; while gauginos gain masses through the $SU(3)_C\times SU(2)_W\times U(1)_Y $ gauge interactions.
As a realistic model we consider the flipped $SU(5)$ theory which provides the natural solution of the doublet-triplet (DT) splitting problem through the missing partner mechanism . Crucial role in the SUSY breaking is played by the anomalous $U(1)_A$ symmetry.
The model predicts gauginos with masses in the range of $1$ GeV, while the soft masses of the scalar particles are in the region of $10^2-10^3$ GeV.
By the special implementation of the $U(1)_A$ charges of some superfields the model suggests two different sets of the messenger superfields. In first case in the role of the messenger superfields emerge the colour triplets and standard electroweak Higgs doublets. which could give the nonuniversal contributions tu the masses of the squarks and sleptons; However in the case considered in the present paper these contributions are strongly suppressed. In this case the masses of all gauginos are generated through the one loop diagrams. While in the second case in the role of the messengers we naturally have only color triplets and no mass term generated for wino. In both cases the desirable $\mu $-term is generated, left handed neutrinos are naturally light and proton decay through the $d=5$ operators is strongly suppressed.
Let us note, that some examples in which the standard Higgs doublets emerge as a messengers of the SUSY breaking also was considered in the recent work [@mes] .
The flipped version of the $SU(5)$ model provides the solution of the DT splitting problem through the missing partner mechanism by the most economical way [@flip2; @flip3] .
The gauge group is $SU(5)\times U(1)$ and the matter superfields transform under this group as: $(10_1+\bar 5_{-3}+1_5)_i$ ($i$ is a family index ) in which the ordinary quark and lepton superfields are compressed as:
$$10_1=(q, d^c, \nu_R )_1~,$$ $$\bar 5_{-3}=(u^c, l)_{-3}~,$$ $$1_5=e^c_5~.
\label{rep}$$ The $10_1$ contains $\nu_R$ additional state which is singlet under the $ G_{321}\equiv SU(3)_C\times SU(2)_W \times U(1)_Y $ group. The Higgs sector consists to the following superfields:
$$H \sim 10_1=(Q, D^c, N)_1~,~~~~~~
\bar H \sim \overline {10}_{-1}=(\bar Q, \bar D^c, \bar N)_{-1}~,$$ $$\phi \sim 5_{-2}=(T, h_d)_{-2}~,~~~~~~~
\overline {\phi } \sim \bar 5_{2}=(\bar T, h_u)_{2}~.
\label{screp}$$ The $h_u$ and $h_d$ fields are generate the masses of up and down quarks and leptons:
$$W^0_Y=10\cdot \bar 5 \cdot \overline {\phi } + 10\cdot 10 \cdot \phi
+1\cdot \bar 5 \cdot \phi ~.
\label{yuk1}$$
the first term generates the masses of up type quarks, while the second and third - masses of down quarks and leptons respectively.
The $H+\bar H$ pair is used for the GUT symmetry breaking. If $N+\bar
N$ from the set $H+\bar H$ develop VEVs of order $M_X\simeq 10^{16}$ GeV, then $SU(5)\times U(1)$ directly is broken to $G_{321} $ and $Q(3, 2)+\bar Q (\bar 3 , 2)$ from $H+\bar H$ are eaten Goldstone modes. The couplings between $H$ ($\bar H$) and $\phi $ ($\overline{\phi }$) superfields are described by the superpotential:
$$W_1^0=\lambda_1HH\phi+\lambda_2\bar H\bar H\overline {\phi }~.
\label{w1}$$
Substituting the VEVs of the $N$ and $\bar N$ fields the mass terms for the triplet components will get the form:
$$W_m=\lambda_1\langle N \rangle D^cT+
\lambda_2\langle \bar N \rangle \bar D^c\bar T~.
\label{wm}$$
So, after the GUT symmetry breaking the triplet states decouple. While $H$ and $\bar H$ do not contain doublet fragments, $h_u$ and $h_d$ remain naturally light.
Suppose, by some mechanism (which will be presented below) $N$ and $\bar N$ have F-terms with nonzero VEVs which magnitudes are of the order $\sim mM_X$ ($m$ is mass scale up to which the SUSY is switched on). This will cause the shift between the masses of the scalar and fermionic components from the triplet superfields by the value $\sim \sqrt{mM_X}$. While $mM_X\ll \langle N \rangle^2, \langle \bar N\rangle ^2$, masses of the scalar components will not changed (see (\[wm\])). So, we will not have the light triplet states in the spectra and the successful unification of the three gauge coupling constants will not be altered. While the colour-triplet fragments transform nontrivially under $SU(3)_C$ and $U(1)_Y$ gauge groups, the SUSY breaking will transferred from the $W_1^0$ sector by the gauge interactions. The gauginos get masses through one loop diagrams. For general set of messengers their masses are given by the formula [@nel]:
$$M_a=\frac{\alpha_a}{4\pi }\frac{F_i}{M}n_a(i)~,
\label{mgaug}$$
where the $M$ is the mass of the corresponding messenger, and $F_i$ - appropriate $F$-term. $n_a(i)$ is Dynkin index and for fundamental representation of $SU(N)$ equals to 1 and for $U(1)_Y$, $n_1=\frac{6}{5}Y^2$ ($Y=Q_{em}-T_3$). Index $a$ in (\[mgaug\]) corresponds to the gauge group and is $1$, $2$ and $3$ for $U(1)_Y$, $SU(2)_W$ and $SU(3)_C$ respectively. The (\[mgaug\]) is written for $F_i \ll M^2$ case.
In our case only gluinos and $U(1)_Y$ gauginos get masses and winos remain massless in the one loop level, since there is not doublets among the messengers.
In this situation one state of chargino is lighter then $W$ boson [^2]. According to refs. [@far] - [@alon] this case did not excluded and requires the low $\tan \beta $ regime .
The squarks and sleptons get masses through two loop diagrams [@dim] and are given by the following formula:
$$\tilde{m}^2=2\left(\frac{F}{M} \right)^2
\sum \left(\frac{\alpha_a}{4\pi } \right)^2C_an_a~,
\label{msc}$$
where $C_3=4/3$, $C_2=3/4$ and $C_1=3/5Y^2$.
Let us now describe how the SUSY breaking occurs in our model. For SUSY breaking we introduce the anomalous $U(1)_A$ gauge symmetry. As a source of the SUSY breaking $U(1)_A$ symmetry was used in the recent works [@ano]. The properties of the anomalous $U(1)_A$ symmetry also were used for explaining the problem of gauge hierarchy [@anDT; @anDT1] as well as for the understanding of the pattern of fermion masses and mixing [@anDT1; @ferm].
It is well known, that anomalous extra $U(1)$ factors appear in effective field theories from strings. The cancellation of its anomalies occurs by the Green-Schwarz mechanism [@green] . Because of the anomaly the Fayet-Iliopoulos term is always generated [@fayet] and in string theories it equals to [@wit]
$$\xi =\frac{g_A^2M_P^2}{192\pi^2}TrQ~.
\label{xi}$$
So, $D_A$ term will have the form:
$$\frac{g_A^2}{8}D_A^2=\frac{g_A^2}{8}
\left(\Sigma q_i|\varphi_i |^2+\xi \right)^2~.$$
In our model, which gauge group is $G=SU(5)\times U(1)\times U(1)_A$, the superfields have the following prescription of the $U(1)_A$ charges: $q_H=q_{\bar H}=-1$, $q_{\phi }=q_{\overline {\phi }}=2$. This choice of the charges will not change the form of the $W_1^0$ (see (\[w1\])). We also introduce the singlet superfields $X$, $Y$ and $\bar Y$ with $U(1)_A$ charges: $q_X=2$ and $q_Y=q_{\bar Y}=0$. The scalar superpotential
$$W_0=\lambda \frac{\bar Y Y}{M_P^2}X\bar HH
\label{w0}$$
is the most general under $G\times {\cal R}$ symmetry, where ${\cal R}$ symmetry acts on superfields $\varphi_i \to e^{{\rm i}R_{\varphi_i}}\varphi_i$ in such a way that $W\to e^{{\rm i}\alpha }W$. So, the ${\cal R}$ ‘charges’ of the superpotential and superfields are arranged as follows: $$R_W=\alpha ~,~~~~R_{\overline{\phi }}=\alpha -2R_{\bar H}~,~~~~
R_{\phi }=\alpha -2R_H~,$$ $$R_Y+R_{\bar Y}=\alpha -R_H-R_{\bar H }-R_X~,~~~~
\label{rs}$$ so, $W_1^0+W_0$ (see (\[w1\]) and (\[w0\])) can be the most general without fixing $R$ numbers of the all superfields. The potential builded from $F$ and $D$-terms will have the form:
$$V=\sum |F_i|^2+\tilde{g}^2\left(|H|^2-|\bar H|^2 \right)^2+
\frac{g^2_A}{8}\left(\xi - |H|^2-|\bar H|^2
+2|X|^2\right)^2~,
\label{v}$$
where $\tilde{g}^2=\frac{3}{10}g^2+\frac{1}{8}g_1^2$ ($g$ and $g_1$ are the $SU(5)$ and $U(1)$ coupling constants respectively). Supposing that $\xi >0$, one can easily see that there exists the SUSY conserving minima:
$$\bar YY=0~,~~~|H|=|\bar H|~,~~~$$ $$|H|^2+|\bar H|^2-2|X|^2=\xi ~;
\label{cl1}$$ So, for the scalar superpotential given by (\[w0\]) SUSY remains unbroken.
Let us imply the proposal of ref. [@ano] and suppose that the $\bar Y$ and $Y$ superfields transform nontrivially under the some gauge group which interaction becomes strong below some $\Lambda$ scale . The simplest case is the $SU(2)$ gauge group under which $Y$ and $\bar Y$ are the pair of doublet-antidoublet. Non-perturbative superpotential induced by the instanton effect have the form [@inst] :
$$W_{\rm inst}=\frac{\Lambda^5}{\bar Y Y}
\label{inst}$$
and whole scalar superpotential will be [^3] :
$$W=\lambda \frac{\bar YY}{M_P^2} X\bar H H
+ \frac{\Lambda^5}{\bar Y Y}~. \label{w}$$
The $F$ and $D$-terms will have the forms:
$$F_H=\lambda \frac{\bar YY}{M_P^2}X\bar H ~,
~~~~F_{\bar H}=\lambda \frac{\bar YY}{M_P^2}XH ~,
~~~~F_X=\lambda \frac{\bar YY}{M_P^2}\bar HH~,$$ $$F_Y=\lambda \frac{\bar Y}{M_P^2}X\bar HH-
\frac{\Lambda^5}{\bar YY^2}~,~~~
F_{\bar Y}=\lambda \frac{Y}{M_P^2}X\bar HH
-\frac{\Lambda^5}{\bar Y^2Y}$$ $$D=|H|^2-|\bar H|^2 ~,~~~
D_A=\xi - |H|^2-|\bar H|^2+2|X|^2~.
\label{fd}$$ It is easy to see that SUSY is broken because there is no solution with vanishing $F$ and $D$ terms.
Minimizing the potential, builded from the $F$ and $D$ terms (see (\[fd\] )), we can find that minimum can be obtained for the solutions:
$$H^2=\bar H^2 =\frac{3}{5}\xi ~,~~~
X^2=\frac{1}{10}\xi +\frac{5m^2}{g_A^2}~,$$ $$\bar Y^4=Y^4=\frac{25}{3}M_P^4\frac{m^2}{\xi }
\left(1+\frac{125}{6\sqrt{3}\lambda }
\frac{mM_P^2}{\xi \sqrt{\xi }}\right)~.
\label{sols}$$
where
$$m^2=\frac{2}{\sqrt{10}}\frac{\lambda \Lambda^5 }{M_P^2\sqrt {\xi }}
\label{pars}$$
and for $\Lambda \sim 10^{11}-10^{12}$ GeV, $\sqrt {\xi }\sim
10^{16}$ GeV and $M_P \sim 10^{18}$ GeV we obtain $m \sim 100$ GeV-$10$ TeV. For (\[sols\]) solutions taking into account (\[fd\]), (\[pars\])
$$F_X\sim F_H=F_{\bar H}=m\sqrt {\xi }~,~~~
F_Y=F_{\bar Y}\sim
mM_P\left(\frac{m}{\sqrt {\xi }} \right)^{1/2}~,$$ $$D=0~,~~~D_A=\frac{10m^2}{g_A^2}~.
\label{sol1}$$
As we see the SUSY in broken and nonzero $F_H$, $F_{\bar H}$-terms are the middle geometrical between the $M_X$ (GUT scale) and $\sim
m_W$ scales ; Also the nonvanishing $D_A$-term with magnitude $m^2$ is generated and the main contribution in the masses of the scalar components of the ordinary superfields comes from this term [^4] and equals to:
$$\tilde{m}^2_{\varphi_i }=\frac{5}{2}m^2q_i~,
\label{dmas}$$
where $q_i$ is the anomalous $U(1)_A$ charge of the appropriate $\varphi_i$ superfield. The non-universal contributions to the squark masses through the supergravity corrections $\sim m^2\epsilon_X^2 $ (where $\epsilon_X =M_X/M_P$) for $\epsilon_X \sim 10^{-2}$ will be negligible.
The upper bound of the soft masses (which are also proportional to the $m^2$) of the electroweak Higgs doublets could be obtained from the requirement of the electroweak symmetry breaking and related to the mass of the top quark. Namely for $m_t=175-180$ GeV upper bound on $m^2_{H_u}$ is $\sim (350~{\rm GeV})^2$ [@bar]. For this order of $m^2$ the mass of the gluino and also ’Majorana’ masses of wino and zino are of the order of $1$ GeV. The recent analyses of percentage exclusion of such a light gluino was presented in [@csik], where results of [@far1] were performed. Existence of light (or massless) wino leads to the one state of chargino lighter then $W$ boson. According refs. [@far; @alon] this case did not excluded and requires the low $\tan \beta $ regime. As far as the light bino concerned its phenomenological implications were described in refs. [@far] .
The GMSB example with light gauginos was presented in ref. [@moh] and with light wino in ref. [@alon]; While in the recent [@raby] works the models with a gluino as a lightest SUSY particle (LSP) were considered.
While the gaugino masses are generated through the nonzero F-terms (of the $N+\bar N$ components of the $H+\bar H$ pair) in the one loop level, there magnitude will be $\alpha_a /(4\pi ) m$. So, the model predicts the gauginos with low soft masses in respect to the soft masses of the squarks and sleptons.
From (\[dmas\]) we see that the matter superfields must have the positive $U(1)_A$ charges. This can easily obtained if $W_1^0$ (see (\[w1\])) will be rewritten to the form:
$$W_1=\lambda_1HH\phi+
\lambda_2 \frac{Z}{M_P}\bar H\bar H\overline {\phi }
\label{w11}$$
and the Yukawa superpotential for ordinary quarks and leptons will have the form:
$$W_Y=10\cdot \bar 5\cdot \overline{\phi }+
\frac{Z_1 }{M_P}10\cdot 10\cdot \phi +
\frac{Z_1 }{M_P}\bar 5\cdot 1\cdot \phi~,
\label{modyuk}$$
where we have introduced $Z $ and $Z_1 $ superfields with anomalous $q_Z$ and $q_{Z_1 }$ charges respectively and assumed that $Z$, $Z_1 $ share the VEVs with $H$, $\bar H$ and $X$ fields in the $D_A$ term. Therefore from (\[w11\]) and (\[modyuk\])
$$q_{\bar 5}=-1+\frac{1}{2}q_{Z_1}+q_Z~,~~~~~
q_{10}=-1-\frac{1}{2}q_{Z_1}~,$$ $$q_1=-1-\frac{3}{2}q_{Z_1}-q_Z~.
\label{muxts}$$ For $1-q_{Z_1}/2 <q_Z<-1-3q_{Z_1}/2$ and $q_{Z_1 }<-2$ charges of the all matter superfields will be positive.
Modification of the $W_1^0$ do not change the picture of the SUSY breaking because the ratio of the ‘effective’ $F$ term and the mass of the messengers will be unchanged. As far as the masses of one pair of triplet-antitriplet components are concerned, their magnitudes are $\sim M_X^2/M_P $. This threshold will not spoil the picture of unification and even suggests the possibility of the obtaining small value of the $\alpha_s $. Namely, for $\sin^2 \theta_W =0.2313 $ , $\alpha_s=0.11 $ is obtained [^5], which indeed coincides with the QCD sum analyses [@shif].
Until going to the fermion sector let us note that taking into account (\[w\]), (\[sols\]) and (\[w11\]) the nucleon decay parameter, which is the $(1,1)$ element of the inverse matrix of triplets- $(\hat{M_T}^{-1})_{11}$ have the magnitude $\sim mM_P/M_X^3$ and therefore nucleon decay through $d=5$ operator is strongly suppressed.
Turning to the fermion sector, we will see that $R$ charges can be arranged in such a way that the $\mu $-term will have the desirable magnitude.
One of the nice feature of the flipped $SU(5)$ theory is that in its framework, because $e^c$ is identified as a singlet state of $SU(5)$, there do not exists the dangerous relation $\hat{M_d}=\hat{M_e}$ which is concomitant to the minimal $SU(5)$ theories. However, one can see that $10\cdot \bar 5 \cdot \overline {\phi }$ coupling generate the large ”Dirac“ mass for the neutrino. To suppress this mass the ”Majorana“ mass should be generated for the $\nu_R $ state and then mass of the $\nu_L$ can be suppressed by the universal seesow mechanism [@seesow] .
Here we introduce the additional fermionic states $\Psi \sim 10_1$, $\overline{\Psi} \sim 10_{-1}$ and ${\cal N} \sim 1_0$ (for each generation). Let us also introduce the three scalar superfields $S$, $S_1$ and $S_2$, which carry the $U(1)_A$ charges. So they can also contribute in the $D_A$-term and can share the VEVs together with $H$, $\bar H$, $X$, $Z$ and $Z_1 $ states. So, if there do not exist for them couplings in the superpotential the abovepresented picture of the SUSY breaking will not changed. The transformation properties under the $SU(5)\times U(1)\times U(1)_A\times {\cal R}$ symmetry of all introduced superfields are presented in the Table 1.
$\begin{array}{|c|c|c|c|}
\hline
{\rm Fields} &SU(5)\times U(1) &U(1)_A &{\cal R} \\
\hline
&&&\\
H &10_1 &-1 & R_{\bar H} \\
&&&\\
\bar H &\overline {10}_{-1} &-1 & R_{\bar H} \\
&&&\\
\phi &5_{-2} &2 & \alpha -2 R_H \\
&&&\\
\overline {\phi } &\bar 5_2 &2-q_Z & \alpha -2R_{\bar H}-R_Z \\
&&&\\
X &1_0 &2 & R_X \\
&&&\\
Y &1_0 &0 & R_{Y} \\
&&&\\
\bar Y &1_0 &0 & \alpha -R_H-R_{\bar H}-R_X-R_{Y} \\
&&&\\
Z &1_0 &q_Z & R_Z \\
&&&\\
Z_1 &1_0 &q_{Z_1 } & R_{Z_1 } \\
&&&\\
S &1_0 &q &R_S \\
&&&\\
S_1 &1_0 &q_1 &R_1 \\
&&&\\
S_2 &1_0 &2(q-q_1)-q_{Z_1 } &-\alpha +2(R_H+R_{\bar H}+
R_S+R_X-R_1)-R_{Z_1 } \\
&&&\\
10_i &10_1 &-1-\frac{1}{2}q_{Z_1 } & R_H-\frac{1}{2}R_{Z_1 } \\
&&&\\
\bar 5_i &\bar 5_{-3} &-1+\frac{1}{2}q_{Z_1 }+q_Z &
-R_H+2R_{\bar H}+R_Z+\frac{1}{2}R_{Z_1 } \\
&&&\\
1_i &1_5 &-1-\frac{3}{2}q_{Z_1 }-q_Z &
3R_H-2R_{\bar H}-R_Z-\frac{3}{2}R_{Z_1 } \\
&&&\\
\Psi & 10_1&1-\frac{1}{2}q_{Z_1 }+q-q_1 &
R_H+R_X+R_S-R_1-\frac{1}{2}R_{Z_1 } \\
&&&\\
\overline{\Psi } &10_{-1} &-1+\frac{1}{2}q_{Z_1 }-q
&\alpha -R_H-R_X-R_S+\frac{1}{2}R_{Z_1 } \\
&&&\\
{\cal N} &1_0 &q_1+\frac{1}{2}q_{Z_1 }-q &
\alpha -R_H-R_{\bar H}-R_X-R_S-R_1+\frac{1}{2}R_{Z_1 } \\
&&&\\
\hline
\end{array}$
Under these assignments of charges the Yukawa superpotential which generates masses of the quarks and leptons and also neutrinos ”Dirac“ and ”Majorana“ masses have the form:
$${\cal W}_Y=W_Y+W_Y'~,
\label{bigyuk}$$
where $$W_Y'=A10\cdot \overline{\Psi }\cdot S\frac{X}{M_P}+
B\overline{\Psi }\cdot \Psi S_1 +C\Psi \cdot {\cal N}\cdot \bar H +
DS_2\cdot {\cal N}^2~.
\label{nu}$$ $A,\cdots ,D$ are the Yukawa matrices, which elements can be assumed to be of the order of one.
The neutrino mass matrix will have the form:
$$\begin{array}{ccccc}
& {\begin{array}{ccccc} \,\,\nu_L & \,\,~~ \nu_R &
\,\,~~N_{\overline{\Psi}}&
\,\,~~ N_{\Psi } &\,\,~~ N_{{\cal N}}
\end{array}}\\ \vspace{2mm}
\hat{m}_{\nu }= \begin{array}{c}
\nu_L \\ \nu_R \\ N_{\overline{\Psi}}
\\N_{\Psi } \\ N_{{\cal N}} \end{array}\!\!\!\!\!
&{\left(\begin{array}{ccccc}
\,\, 0 &\,\,~~h_u^0 &\,\,~~0 &\,\,~~0 &\,\,~~0 \\
\,\, h_u^0 &\,\,~~0 &\,\,~~\frac{XS}{M_P} &\,\,~~ 0 &\,\,~~ 0 \\
\,\,0&\,\,~~\frac{XS}{M_P} &\,\,~~0 &\,\,~~S_1 &\,\,~~0 \\
\,\,0&\,\,~~0 &\,\,~~S_1 &\,\,~~0 &\,\,~~\bar H \\
\,\,0&\,\,~~0 &\,\,~~0&\,\,~~\bar H & \,\,~~S_2 \end{array}\right)~.}
\end{array} \!\! ~~~~~
\label{matnu}$$
This is the seesaw mass matrix which results the light neutrino with a mass of order:
$$m_{\nu }\simeq M_P^2\frac{\langle h_u^0 \rangle ^2}{M_X^3}~.
\label{mnu}$$
Suppressing the neutrino masses the superpotential (\[nu\]) was used in which the $X$ superfield (with nonzero $F$-term) has couplings with superfields which are transforming nontrivially under the $G_{321}$ gauge group and could emerge as a messengers.
Some remarks about implications of this fact should be done.
For simplicity let us consider the case of the one generation. After integrating out the heavy states of $\overline{\Psi }+\Psi $, which masses are $\sim M_X$, the decoupled state of decuplet (let us denote it by $10_h$) will leave in $10$ by the weight $\frac{SX}{S_1M_P}\sim \epsilon_X $, while the light state (denoting it by $10_f$) contained in it approximately by the weight $1$. Taking into the account these facts the first two terms of the (\[nu\]) can be rewritten as follows:
$$W_Y'^{(2)}=\frac{SX}{M_P}\left(\epsilon 10_h\cdot \overline{\Psi }
+10_f\cdot \overline{\Psi }\right)+
S_1\cdot \overline{\Psi }\cdot 10_h~.
\label{nu1}$$
As we see the $\overline{\Psi }+10_h$ also emerge as a messengers but there contribution to the gaugino masses are strongly suppressed:
$$\delta M_a \sim \frac{\alpha_a}{4\pi }m\epsilon_X^2~.
\label{dgau}$$
Because there exists the matter coupling with messenger (see (\[nu1\])) the soft masses for scalar components of $10_f$ arise in the one loop level and as was shown in [@see] is of the order:
$$\delta m^2_{10} \sim \frac{1}{16\pi^2 }\epsilon_X^2
\frac{F^4}{M^6}~,
\label{dsq}$$
which is also negligible in comparison of (\[dmas\]).
Turning to the case of the three generations, without loosing the generality the last term of (\[nu1\]) could be taken diagonal, while the remaining terms in general are nondiagonal. As was shown in [@giu] because this fact the sparticles gain soft masses through the tadpole $D$-term of the $U(1)_Y$; However in our case this contribution is miserable:
$$\delta m^2_{\tilde{\phi_i}} \sim \frac{\alpha_1}{4\pi }
Y_{\tilde{\phi_i}}m^2\epsilon_X^4~.
\label{dsq1}$$
We have not fixed yet the anomalous $q$ and $q_1$ as well as $R_S$ and $R_1$ charges of the $S$ and $S_1$ superfields respectively. Two cases can be considered:
[**1. $q=-5$ and $R_S=-\frac{5}{2}R_X$**]{}
In this case the term which is permitted by the $U(1)_A$ and ${\cal R}$ symmetries is $$W_{\mu }=\overline{\phi } \phi
\frac{X^3Z Z_1 S_1^2S_2}{M_P^7}~.
\label{mu1}$$ Substituting VEVs $\langle S_1 \rangle \sim \langle S_2\rangle \sim
\langle X \rangle \sim M_X$ and taking into account that $M_X/M_P=\epsilon_X \sim 10^{-2}$ we will have $\mu =M_X\epsilon_X^7 \sim 100$ GeV. Also substituting the nonzero $F$-term of the $X$ superfield we will obtain the shift with the magnitude $mM_X\epsilon_X^7$ between the masses of the scalar and fermionic components from $\bar {\phi }$ and $\phi $ superfields , this means that the standard doublets also emerge as a messengers and wino can get mass through the one loop diagram [@nel]. In this case the masses of the gauginos will be:
$$M_3=
4\frac{\alpha_3}{4\pi }\frac{F_H}{\langle N \rangle}
=\frac{\alpha_3}{\pi }m~,
\label{g11}$$
$$M_2=
3\frac{\alpha_2}{4\pi }\frac{mM_X\epsilon_X^7}{\mu }
=3\frac{\alpha_2}{4\pi }m~,
\label{g12}$$
$$M_1
=\frac{\alpha_1}{4\pi }\left(
4\frac{F_H}{\langle N \rangle}
\frac{4}{30}+3\frac{mM_X\epsilon_X^7}{\mu }\frac{9}{30}
\right)
=\frac{43}{30}\frac{\alpha_1}{4\pi }m~.
\label{g13}$$
In (\[g11\]) the factor $4$ emerges because there are two pairs of the messengers and after the substitution of the $F$-terms in (\[wm\]) the combinator factor $2$ arise ; while in (\[g12\]) factor $3$ arise because in (\[mu1\]) the field $X$ (with nonzero $F$-term) is in the third power (the same arguments were taken into the account during the calculation of $M_1$).
So, without introducing the additional states of the messenger superfields we can obtain the desirable pattern of SUSY breaking in the framework of the flipped $SU(5)$ GUT, with successful DT splitting and with $\mu $-term of the order of $100$ GeV. All messengers are from the minimal content of the GUT supermultiplets. Interestingly in this case the standard Higgs doublets also belong to the messenger superfields. Coupling of messenger doublets with ordinary matter induce the nonuniversalities in the one loop level but will be miserable in respect of (\[dmas\]) [^6].
[**2. $q=-\frac{4}{5}$ and $R_S=-\frac{2}{5}R_X$**]{}
In this case the $\mu $-term generating coupling is:
$$W_{\mu }=\overline{\phi } \phi
\frac{S^3Z Z_1 S_1^2S_2}{M_P^7}~,
\label{mu2}$$
which also gives $\mu \sim 100$ GeV. In this case the doublets do not have couplings with superfields which $F$-terms have the nonzero VEVs.
Note, that in both cases the values of the $q_1$ and $q_Z$ charges still were undetermined. In order to insure the nonzero VEVs for $H$, $\bar H$ fields there charges and the $\xi $ term in the $D_A$ (see (\[xi\]), (\[fd\])) must have the different signs. Therefore ${\rm Tr}Q =-46-28q_1+8q_Z-21q_{Z_1}/2 > 0$ and from this condition we obtain $q_1 < (16q_Z-21q_{Z_1}-92)/56$.
Building our model we have assumed, that $\epsilon_X=\sqrt{\xi}/M_P \sim 10^{-2}$ which for $M_P\sim 10^{18}$ GeV gives $\sqrt{\xi}\sim 10^{16}$ GeV, this value was dictated from the scale of the grand unification. However, for the flipped $SU(5)$ model derived from strings [@flip2] the preferable value of $M_X$ is $10^{17}$ GeV without loosing the successful prediction of the $\sin^2 \theta_W $ [@unif]. Increasing the value of $M_X$ the picture of the abovepresented scenarios will not changed if for the values of $q$, $R_S$ for the two cases $-13$, $-13R_X/2$ and $-4/13$, $-2R_X/13$ will be taken respectively. Then the couplings generating the $\mu $-term will be $ X^{11}Z Z_1 S_1^2S_2/M_P^{15}\overline{\phi } \phi$ and $ S^{11}Z Z_1 S_1^2S_2/M_P^{15}\overline{\phi } \phi$ respectively, which for $\epsilon_X =\sqrt{\xi}/M_P \sim 10^{-1} $ still give $\mu \sim 100$ GeV.
[**[Acknowledgments]{}**]{}
I am indebted to D. Comelli, G. Dvali and I. Gogoladze for discussions and important comments.
[99]{}
J. Wess and B. Zumino, Phys. Lett. B 49 (1974) 52; M. Grisaru, W. Siegel and M. Rocek, Nucl. Phys. B 159 (1979) 429.
C. Giunti, C.W. Kim and U.W. Lee, Mod. Phys. Lett. A 6 (1991) 1745; P. Langacker and M. Luo, Phys. Rev. D 44 (1991) 817; U. Amaldi, W. de Boer and H. Furstenau, Phys. Lett. B 260 (1991) 447; J. Ellis, S. Kelley and D. Nanopoulos, Phys. Lett. B 260 (1991) 131.
R. Barbieri, S. Ferrara and C.A. Savoy , Phys. Lett. B 119 (1982) 343;\
A.H. Chamseddine, R. Arnowitt and P. Nath, Phys. Rev. D 49 (1982) 970; L.J. Hall, J. Lykken and S. Weinberg, Phys. Rev. D 27 (1983) 2359.
A. Pomarol and S. Dimopoulos, Nucl. Phys. B 453 (1995) 83.
M. Dine, W. Fischler and M. Srednichi, Nucl. Phys. B 189 (1981) 575; S. Dimopoulos and S. Raby, Nucl. Phys. B 192 (1981) 353; M. Dine and W. Fischler, Phys. Lett. B 110 (1982) 227; M. Dine and M. Srednichi, Nucl. Phys. B 202 (1982) 238; M. Dine and W. Fischler , Nucl. Phys. B 204 (1982) 346; L. Alvarez-Gaume, M. Claudson and M. Wise, Nucl. Phys. B 207 (1982) 96; C.R. Nappi and B.A. Ovrut, Phys. Lett. B 113 (1982) 175; S. Dimopoulos and S. Raby, Nucl. Phys. B 219 (1983) 479.
S. Ferrara, L. Girardello and F. Palumbo, Phys. Rev. D 20 (1979) 403.
M. Dine, A.E. Nelson and Y. Shirman, Phys. Rev. D 51 (1995) 1362;\
M. Dine, A.E. Nelson, Y. Nir and Y. Shirman, preprint SCIPP-95-32 (1982);\
S. Dimopoulos, M. Dine, S. Raby and S. Thomas, preprint SLAC-PUB-96-7104 (1996).
S. Ambrosanio, et al, Phys. Rev. Lett. 76 (1996) 3498; G. Dvali, G. F. Giudice and A. Pomarol, Nucl. Phys. B 478 (1996) 31; K.S. Babu, C. Kolda and F. Wilczek, hep-ph/9605408; I. Disgupta, B. Dobrescu and L. Randall, hep-ph/9607487; S.P. Martin, hep-ph/9608224; A. Riotto, N. Tornquist and R. N. Mohapatra, Phys. Lett. B 388 (1996) 599; S. Dimopoulos, S. Thomas and J. Wells, hep-ph/96094344; J. Bagger, D. Pierche, K. Matchev and R.-J Zhang, hep-ph/9609443; E. Poppitz and S. Trivedi, hep-ph/9609529; H. Baer et al, hep-ph/9610358; G. Bhattacharyya and A. Romanino, hep-ph/9611243; A. de Gouvea, T. Moroi and H. Murayama, hep-ph/9701244; N. Arkani-hamed, H. Murayama and J. March-Russell, hep-ph/9701286224; D. Dicus, B. Dutta and S. Nandi, hep-ph/9701341.
G. Dvali and M. Shifman, Phys. Lett. B 399 (1997) 60.
I. Antoniadis, J. Ellis, J.S. Hagelin and D.V. Nanopoulos, Phys. Lett. B 194 (1987) 231; B 205 (1988) 459; B 208 (1988) 209; B 231 (1989) 65.
G.K. Leontaris, J. Rizos and K. Tamvakis, Phys. Lett. B 243 (1990) 220; B 251 (1990) 83; I. Antoniadis, G.K. Leontaris and J. Rizos, Phys. Lett. B 245 (1990) 161; I. Antoniadis, J. Rizos and K. Tamvakis, Phys. Lett. B 278 (1992) 257; B 279 (1992) 281; J.L. Lopez and D.V. Nanopoulos, Nucl. Phys. B 338 (1990) 73; Phys. Lett. B 251 (1990) 73; D. Bailin and A. Love, Phys. Lett. B 280 (1992) 26.
G. Farrar, hep-ph/9608387; G. Farrar and A. Masiero, hep-ph/9410401; L. Clavelli, hep-ph/9210257.
G. Farrar, hep-ph/9707467.
A.E. Faraggi, Phys. Lett. B 387 (1996) 775.
S. Dimopoulos, G.F. Giudice and A. Pomarol, Phys. Lett. B 389 (1996) 37.
G. Dvali and A. Pomarol, Phys. Rev. Lett. 77 (1996) 3738; P. Binetruy and E. Dudas, Phys. Lett. B 389 (1996) 503.
G. Dvali and S. Pokorski, Phys. Rev. Lett. 78 (1997) 807.
Z. Berezhiani and Z. Tavartkiladze, Phys. Lett. B 396 (1997) 150, Phys. Lett. B 409 (1997) 220.
R.N. Mohapatra and A. Riotto, Phys. Rev. D 55 (1997) 1138; Phys. Rev. D 55 (1997) 4262; A.E. Nelson and D. Wright; hep-ph/9702359.
M. Green and J. Schwarz, Phys. Lett. B 149 (1984) 117.
E. Witten, Nucl. Phys. B 188 (1981) 513; W. Fischler et al., Rhys. Rev. Lett. 47 (1981) 657.
M. Dine, N. Seiberg and E. Witten, Nucl. Phys. B 289 (1987) 584; J. Atick, L. Dixon and A. Sen, Nucl. Phys. B 292 (1987) 109; M. Dine, I. Ichinose and N. Seiberg, Nucl. Phys. B 293 (1987) 253.
G. Dvali and A. Pomarol, hep-ph/9708364.
I. Affleck, M. Dine and N. Seiberg, Nucl. Phys. B 256 (1985) 557; A.I. Vainshtein, V.I. Zakharov and M.A. Shifman, Sov. Phys. Usp. 28 (1985) 709.
R. Barbieri and G.F. Giudice, Nucl. Phys. B 306 (1988) 63.
R.N. Mohapatra and S. Nandi, Phys. Rev. Lett. 79 (1997) 181.
F. Csikor and Z. Fodor, hep-ph/9712269.
S. Raby, Phys. Rev. D 56 (1997) 2852; hep-ph/9712254.
M. Shifman, Mod. Phys. Lett. A 10 (1995) 605.
M. Gell-Mann, P. Ramond and R. Slansky, in: Supergravity, eds. P. van Nieuwenhuizen and D. Freedman (North-Holland, Amsterdam, 1979) p. 315;\
T. Yanagida, Prog. Th. Phys. B 135 (1979) 66;\
R. Mohapatra and G. Senjanovic, Phys. Rev. Lett. 44 (1980) 912.
See $2^{{\rm nd}}$ ref. of [@amb].
$2^{{\rm nd}}$ ref. of [@amb] and S. Dimopoulos and G.F. Giudice, Phys. Lett. B 393 (1997) 72.
I. Antoniadis, J. Ellis, R. Lacaze and D.V. Nanopoulos; Phys. Lett. B 286 (1991) 188.
[^1]: E-mail address: [email protected]
[^2]: This happens if the $\mu $ term exists for the doublet components. If $\mu $-term is zero we will have two light states in the theory. However as will be shown later in our model the $\mu $ term with the desirable magnitude can be generated.
[^3]: Non-perturbative term can violate the [R]{} symmetry if the [R]{} symmetry is an anomalous.
[^4]: The model in which the soft masses for the matter particles are generated from the nonvanishing part of the anomalous $D$-term was considered in [@GA] .
[^5]: I thank I. Gogoladze for bringing my attention to this issue.
[^6]: See also the general arguments of ref. [@mes].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The introduction of Parikh matrices by Mateescu et al. in 2001 has sparked numerous new investigations in the theory of formal languages by various researchers, among whom is Şerbǎnuţǎ. Recently, a decade-old conjecture by Şerbǎnuţǎ on the $M$-ambiguity of words was disproved, leading to new possibilities in the study of such words. In this paper, we investigate how selective repeated duplications of letters in a word affect the $M$-ambiguity of the resulting words. The corresponding $M$-ambiguity of those words are then presented in sequences, which we term as $M$-ambiguity sequences. We show that nearly all patterns of $M$-ambiguity sequences are attainable. Finally, by employing certain algebraic approach and some underlying theory in integer programming, we show that repeated periodic duplications of letters of the same type in a word results in an $M$-ambiguity sequence that is eventually periodic.'
address:
- |
School of Mathematical Sciences\
Universiti Sains Malaysia\
11800 USMMalaysia
- |
School of Mathematics, Actuarial and Quantitative Studies\
Asia Pacific University of Technology & Innovation\
Technology Park Malaysia, Bukit Jalil\
57000 Kuala Lumpur, Malaysia
- |
School of Mathematical Sciences\
Universiti Sains Malaysia\
11800 USMMalaysia
author:
- Ghajendran Poovanandran
- Wen Chean Teh
title: '*M*-Ambiguity Sequences for Parikh Matrices and Their Periodicity Revisited'
---
Introduction
============
The classical Parikh Theorem [@rP66], which states that the Parikh vectors of all words from a context-free language form a semilinear set, established the Parikh mapping as a significant advancement in the theory of formal languages. The Parikh matrix mapping, introduced in [@MSSY01], is a canonical generalization of the Parikh mapping. On top of dealing with the number of occurrences of individual letters (as in the case of Parikh vectors), the Parikh matrix of a word stores information on the number of occurrences of certain subwords in that word as well. The introduction of Parikh matrices has led to various new studies in the combinatorial study of words (for example, see [@aA07; @AAP08; @vS09; @SS06; @wT16; @wT16b; @aS10; @SY10; @AT16; @GT16a; @GT17; @GT17b; @MBS17; @TSB18; @GT17c; @TK14]).
A word is $M$-ambiguous if and only if it shares the same Parikh matrix with another distinct word. In the pursuit of characterizing $M$-unambiguous words, Şerbǎnuţǎ proposed a conjecture in [@SS06] that the duplication of any letter in an $M$-ambiguous word will result in another $M$-ambiguous word. The conjecture was however overturned in [@GT17b] by a counterexample from the quaternary alphabet.
In this work, we will show that by duplicating certain letters in a word, it is possible to continuously change the $M$-ambiguity of the resulting words. In fact, we will see that such changes in the $M$-ambiguity of a word can occur in nearly any pattern. Given an infinite sequence of words, obtained by repeatedly duplicating certain letters in the first word, we present the corresponding of those words in what we term as an $M$-ambiguity sequence. This work also proposes an algebraic way to determine the $M$-ambiguity of a word. This algebraic approach is then used together with some underlying theory in integer linear programming to show that if we repeatedly duplicate—in a periodic manner—the letters of the same type in a word, the corresponding $M$-ambiguity sequence will be eventually periodic.
The remainder of this paper is structured as follows. Section 2 provides the basic terminology and preliminaries. Section 3 highlights some previous results pertaining to the overturn of Şerbǎnuţǎ’s conjecture and serves the main motivation of this paper. After that, the central notion of our study, namely the $M$-ambiguity sequences, is introduced. It is then shown that nearly any pattern of $M$-ambiguity sequence can be realized. Section 4 mainly studies the periodicity of $M$-ambiguity sequences. In relative to that, an algebraic analysis to determine the $M$-ambiguity of a word is illustrated. Certain theories pertaining to rational polyhedra are then used together with the algebraic approach to prove a main result on the periodicity of $M$-ambiguity sequences. Our conclusion follows after that.
Preliminaries
=============
We denote as follows—$\mathbb{R}$ is the set of real numbers, $\mathbb{Q}$ is the set of rational numbers, $\mathbb{Z}$ is the set of integers, $\mathbb{Z^+}$ is the set of positive integers and $\mathbb{Z}_{\ge 0}$ is the set of nonnegative integers.
Suppose $\Sigma$ is a finite nonempty alphabet. The set of all words over $\Sigma$ is denoted by $\Sigma^*$. The unique empty word is denoted by $\lambda$. Given two words $v,w\in\Sigma^*$, the concatenation of $v$ and $w$ is denoted by $vw$. An *ordered alphabet* is an alphabet $\Sigma=\{a_1, a_2, \dotsc,a_s\}$ with a total ordering on it. For example, if $a_1<a_2<\dotsb < a_s$, then we may write $\Sigma= \{a_1<a_2<\dotsb<a_s\}$. Conversely, if $\Sigma= \{a_1<a_2<\dotsb<a_s\}$, then $\{a_1, a_2, \dotsc,a_s\}$ is the *underlying alphabet*. Frequently, we will abuse notation and use $\Sigma$ to stand for both the ordered alphabet and its underlying alphabet. Suppose $\Gamma\subseteq\Sigma$. The projective morphism $\pi_{\Gamma}:\Sigma^*\rightarrow\Gamma^*$ is defined by $$\pi_{\Gamma}(a)=\begin{cases}
a, & \text{if } a\in\Gamma\\
\lambda, & \text{otherwise.}
\end{cases}$$
A word $v$ is a *scattered subword* (or simply *subword*) of $w\in \Sigma^*$ if and only if there exist $x_1,x_2,\dotsc, x_n$, $y_0, y_1, \dotsc,y_n\in \Sigma^*$ (possibly empty) such that $v=x_1x_2\dotsm x_n$ and $w=y_0x_1y_1\dotsm y_{n-1}x_ny_n$. The number of occurrences of a word $v$ as a subword of $w$ is denoted by $\vert w\vert_v$. Two occurrences of $v$ are considered different if and only if they differ by at least one position of some letter. For example, $\vert bcbcc\vert_{bc}=5$ and $\vert aabcbc\vert_{abc}=6$. By convention, $\vert w\vert_{\lambda}=1$ for all $w\in \Sigma^*$. For any integer $k\geq 2$, let $\mathcal{M}_k$ denote the multiplicative monoid of $k \times k$ upper triangular matrices with nonnegative integral entries and unit diagonal.
[@MSSY01] Suppose $\Sigma=\{a_1<a_2<\cdots <a_k\}$ is an ordered alphabet. The *Parikh matrix mapping* with respect to $\Sigma$, denoted by $\Psi_\Sigma$, is the morphism: $$\Psi_\Sigma:\Sigma^*\rightarrow\mathcal{M}_{k+1},$$ defined such that for every integer $1\le q\le k$, if $\Psi_\Sigma(a_q)=(m_{i,j})_{1\le i,j\le k+1}$, then
- $m_{i,i}=1$ for all $1\le i\le k+1$;
- $m_{q,q+1}=1$; and
- all other entries of the matrix $\Psi_\Sigma(a_q)$ are zero.
Matrices of the form $\Psi_\Sigma(w)$ for $w\in\Sigma^*$ are termed as *Parikh matrices*.
[@MSSY01]\[1206a\] Suppose $\Sigma=\{a_1<a_2< \dotsb<a_s\}$ is an ordered alphabet and $w\in \Sigma^*$. The matrix $\Psi_{\Sigma}(w)=(m_{i,j})_{1\leq i,j\leq s+1}$ has the following properties:
- $m_{i,i}=1$ for each $1\leq i \leq s+1$;
- $m_{i,j}=0$ for each $1\leq j<i\leq s+1$;
- $m_{i,j+1}=\vert w \vert_{a_ia_{i+1}\dotsm a_j }$ for each $1\leq i\leq j \leq s$.
Suppose $\Sigma=\{a<b<c<d\,\}$ and $w=abcdbc$. Then $$\begin{aligned}
\Psi_{\Sigma}(w)&=\Psi_{\Sigma}(a)\Psi_{\Sigma}(b)\Psi_{\Sigma}(c)\Psi_{\Sigma}(d)\Psi_{\Sigma}(b)\Psi_{\Sigma}(c)\\
&= \begin{pmatrix}
1 & 1 & 0 & 0 & 0\\
0 & 1 & 0 & 0 & 0\\
0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 1
\end{pmatrix}
\begin{pmatrix}
1 & 0 & 0 & 0 & 0\\
0 & 1 & 1 & 0 & 0\\
0 & 0 & 1 & 0 & 0\\
0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 1
\end{pmatrix}
\dotsm
\begin{pmatrix}
1 & 0 & 0 & 0 & 0\\
0 & 1 & 0 & 0 & 0\\
0 & 0 & 1 & 1 & 0\\
0 & 0 & 0 & 1 & 0\\
0 & 0 & 0 & 0 & 1
\end{pmatrix}\\
&= \begin{pmatrix}
1 & 1 & 2 & 3 & 1\\
0 & 1 & 2 & 3 & 1\\
0 & 0 & 1 & 2 & 1\\
0 & 0 & 0 & 1 & 1\\
0 & 0 & 0 & 0 & 1
\end{pmatrix}
=\begin{pmatrix}
1 & \vert w\vert_a & \vert w\vert_{ab} & \vert w\vert_{abc} & \vert w\vert_{abcd}\\
0 &1 & \vert w\vert_b & \vert w\vert_{bc} & \vert w\vert_{bcd}\\
0 & 0 & 1 & \vert w\vert_c & \vert w\vert_{cd}\\
0 & 0 & 0 & 1 & \vert w\vert_{d}\\
0 & 0 & 0 & 0 & 1
\end{pmatrix}.\end{aligned}$$
Suppose $\Sigma$ is an ordered alphabet. Two words $w,w'\in \Sigma^*$ are *$M$-equivalent*, denoted by $w\equiv_Mw'$, iff $\Psi_{\Sigma}(w)=\Psi_{\Sigma}(w')$. A word $w\in \Sigma^*$ is *$M$-ambiguous* iff it is $M$-equivalent to another distinct word. Otherwise, $w$ is . For any word $w\in\Sigma^*$, we denote by $C_w$ the set of all words that are $M$-equivalent to $w$.
The following is a simple equivalence relation which involves the most evident rewriting rules that preserve $M$-equivalence (see [@AAP08]).
\[1906a\] Suppose $\Sigma=\{a_1<a_2< \dotsb<a_s\}$ is an ordered alphabet. Two words $w,w'\in\Sigma^*$ are , denoted by $w\equiv_1w'$, iff $w'$ can be obtained from $w$ by applying finitely many rewriting rules of the following form: $$xa_ka_ly\rightarrow xa_la_k y \text{ where } x,y\in\Sigma^* \text{ and } \vert k-l\vert \geq 2.$$
[@TK14] Suppose $\Sigma$ is an alphabet and $v,w\in\Sigma^*$. The *v-core* of $w$, denoted by $\operatorname{core}_v(w)$, is the unique subword $w'$ of $w$ such that $w'$ is the subword of shortest length which satisfies $|w'|_v=|w|_v$.
\[CoreabcMiddle\][@GT16a] Suppose $\Sigma=\{a<b<c\}$ and $w\in\Sigma^*$ with $|w|_{abc}\ge 1$. Then, $w\equiv_1u\operatorname{core}_{abc}(w)v$ for some unique $u\in\{b,c\}^*$ and $v\in\{a,b\}^*$.
Attainable Patterns of *M*-ambiguity Sequences
==============================================
The following conjecture was proposed by Şerbǎnuţǎ in [@SS06] as an open pertaining to $M$-ambiguity of words.
\[SerbConj\] Suppose $\Sigma$ is an ordered alphabet. For any $u,v\in\Sigma^*$ and $a\in\Sigma$, if $uaav$ is $M$-unambiguous, then $uav$ is $M$-unambiguous as well. Equivalently, if $uav$ is $M$-ambiguous, then $uaav$ is also $M$-ambiguous.
The above conjecture holds for the binary and ternary alphabets. (For exhaustive lists of $M$-unambiguous binary and ternary words, readers are referred to [@MS04 Theorem 3] and [@SS06 Theorem A.1] respectively.) On the contrary, for the quarternary alphabet, it was shown in [@GT17b] that the conjecture is invalid. The counterexample given was the $M$-ambiguous word $cbcbabcdcbabcbc$ (which is to the word $bccabcbdbcbaccb$). The following result was then proven, thus overturning the conjecture.
[@GT17b]\[ConjectureOverturn\] The word $w=cbcbabc^ndcbabcbc$ is $M$-unambiguous with respect to $\Sigma=\{a<b<c<d\}$ for every integer $n>1$.
At this point, it is natural for one to ask Question \[MainQues\], which is in a more general setting.
[@SS06] Suppose $\Sigma$ is an alphabet and $w\in \Sigma^*$. Suppose $w=a^{p_1}_1a^{p_2}_2\cdots a^{p_n}_n$ such that $a_i\in\Sigma$ and $p_i>0$ for all $1\le i\le n$ with $a_i\neq a_{i+1}$ for all $1\le i\le n-1$. The *print* of $w$, denoted by $pr(w)$, is the word $a_1a_2\cdots a_n$.
Suppose $\Sigma$ is an ordered alphabet and $w,w'\in\Sigma^*$. We write $w\dashv w'$ iff $w=uav$ and $w'=uaav$ for some $u,v\in\Sigma^*$ and $a\in\Sigma$.
\[MainQues\] Suppose $\Sigma=\{a<b<c<d\}$. Consider an infinite sequence of words $w_i\in\Sigma^*,i\ge 0$ such that $pr(w_0)=w_0$ and $$w_0\dashv w_1\dashv w_2\dashv\dotsb.$$ In what patterns can the $M$-ambiguity of these words sequentially change?
In the spirit of answering the above question, we define the following notion.
Suppose $\Sigma$ is an ordered alphabet. Let $\varphi=\{w_i\}_{i\ge 0}$ be a sequence of words over $\Sigma$ such that for all integers $i\ge 0$, we have $w_i\dashv w_{i+1}$. We say that a sequence $\{m_i\}_{i\ge 0}$ is the *$M$-ambiguity sequence corresponding to $\varphi$*, denoted by $\Theta_\varphi$, if and only if for every integer $i\ge 0$, we have $m_i\in\{A,U\}$ such that if $m_i=A$, then $w_i$ is $M$-ambiguous; otherwise if $m_i=U$, then $w_i$ is $M$-unambiguous.
By the above definition, one can see that Question \[MainQues\] actually asks for the attainable patterns of $M$-ambiguity sequence, where the associated sequence of words starts with a print word. The following two examples, first presented in [@GT17e], provide a partial answer to this question.
For the remaining part of this section, we fix $\Sigma=\{a<b<c<d\}$. Whenever the of a word is mentioned, it is understood that it is with
\[Example1\] For each integer $n\ge 1$, let $w_{n,m}=c^nbcbabc^mdcbabcbc$. If $m=n$, then $w_{n,m}$ is as it is $M$-equivalent to the word $bc^{n+1}abc^nbdbcbaccb$. If $m=n+1$, then $w_{n,m}$ is $M$-unambiguous [@GT17e Theorem 3.6].
Therefore, if one wants to obtain a sequence $\varphi$ of words (where the first word is a print word) such that $\Theta_\varphi=A,U,A,U,A,U,\cdots$, the duplication of letters can be carried out in the following manner: $$w_{1,1},\, w_{1,2},\, w_{2,2},\, w_{2,3},\, w_{3,3},\, w_{3,4},\, \cdots .$$
\[Example2\] The words $cbabcdcbabc$ and $cbabcdcbabbc$ are $M$-unambiguous (computationally verified). By duplicating the first letter $b$ in that word, we obtain the $M$-ambiguous word $cbbabcdcbabbc$ (it is to the word $bcabcbdbcbacb$). For each integer $n\ge 1$, let $w_{n,m}=c^nbbabc^mdcbabbc$. If $m=n,$ then $w_{n,m}$ is as it is $M$-equivalent to the word $bc^nabc^nbdbcbacb$. If $m=n+1$, then $w_{n,m}$ is $M$-unambiguous [@GT17e Theorem 3.7].
Therefore, if one wants to obtain a sequence $\varphi$ of words (where the first word is a print word) such that $\Theta_\varphi=U,U,A,U,A,U,A,\cdots$, the duplication of letters can be carried out in the following manner: $$cbabcdcbabc, cbabcdcbabbc, w_{1,1},\, w_{1,2},\, w_{2,2},\, w_{2,3},\, w_{3,3},\, w_{3,4},\, \cdots .$$
\[RemReason\] Example \[Example1\] and Example \[Example2\] shows that $M$-ambiguity sequences with alternating $A$ and $U$ are attainable. In contrast to Example \[Example1\], the word $w_{1,1}$ in Example \[Example2\] is not a print word. That is why we needed the word $cbabcdcbabc$ to begin the sequence, followed by $cbabcdcbabbc$, before we reach $w_{1,1}$.
We now generalize the words used in Example \[Example1\] and Example \[Example2\] to provide a more nearly complete answer—almost any pattern of sequence is attainable. For that, we need the following observations and theorems as a basis.
\[BasisObs1\] For all positive integers $n$ and $p$, the word $c^nbcbabc^ndcbabcbc^p$ is as it is to the word $bc^{n+1}abc^nbdbcbaccbc^{p-1}$.
The proof of the following result closely resembles that of Theorem 3.6 in [@GT17e], yet we include it here for completeness.
\[BasisTheo1\] The word $w=c^nbcbabc^mdcbabcbc^p$ is $M$-unambiguous for all integers $n\geq 1$, $p\ge 1$ and $m\ge n+1$.
We argue by contradiction. Fix integers $n\geq 1$, $p\ge 1$ and $m\ge n+1$. Assume that $w$ is Then, for some $w'\in\Sigma^*$ such that $w'\neq w$. It follows that $\pi_{\{a,b\}}(w)\equiv_M\pi_{\{a,b\}}(w')$, $\pi_{\{b,c\}}(w)\equiv_M\pi_{\{b,c\}}(w')$ and $\pi_{\{c,d\}}(w)\equiv_M\pi_{\{c,d\}}(w')$. Note that $\pi_{\{c,d\}}(w)=c^{n+m+1}dc^{p+2}$ is , thus $\pi_{\{c,d\}}(w)=\pi_{\{c,d\}}(w')$. Meanwhile, $\pi_{\{a,b\}}(w)=bbabbabb$. Thus, $\pi_{\{a,b\}}(w')$ is either $bbabbabb$, $babbbbab$, $bbbaabbb$, or $abbbbbba$.
Write $w=\underbrace{c^nbcbabc^{m}}_{v_1}d\underbrace{cbabcbc^p}_{v_2}$ and $w'=v_1'dv_2'$, where $v_1',v_2'\in\{a,b,c\}^*$. (Note that $v_1$ and $v_2$ are both $M$-unambiguous[^1] as this fact will be needed later in this proof.) Since $|w|_x=|w'|_x$ for every $x\in\{abcd,bcd,cd\}$, it follows that $|v_1|_y=|v_1'|_y$ for every $y\in\{abc,bc,c\}$. Furthermore, since $|v_1|_c+|v_2|_c=|w|_c=|w'|_c=|v_1'|_c+|v_2'|_c$, we have $|v_2|_c=|v_2'|_c$.
Note that $|w'|_{bc}=|v_1'|_{bc}+|v_1'|_b|v_2'|_c+|v_2'|_{bc}=|v_1|_{bc}+|v_1'|_b\cdot (p+2)+|v_2'|_{bc}$. At the same time, $|w'|_{bc}=|w|_{bc}=|v_1|_{bc}+|v_1|_b|v_2|_c+|v_2|_{bc}=|v_1|_{bc}+3\cdot (p+2)+(3p+2)=|v_1|_{bc}+6p+8$. Thus, $$|v_1'|_b\cdot (p+2)+|v_2'|_{bc}=6p+8.\tag{$\star$}$$ Meanwhile, we have $|v_1'|_b\le |w'|_b=|w|_b=6$. If $|v_1'|_b=6$, then $|v_2'|_{bc}=-4$, which is impossible. Thus $|v_1'|_b\le 5$. Also, since $|v_1'|_{abc}=|v_1|_{abc}=m\ge n+1$, it follows that $|\operatorname{core}_{abc}(v_1')|_b\ge 1$.
$\pi_{\{a,b\}}(w')=bbabbabb$.\
Since $|v_1'|_b\le 5$, $|\operatorname{core}_{abc}(v_1')|_b\ge 1$ and $\pi_{\{a,b\}}(w')=\pi_{\{a,b\}}(v_1')\pi_{\{a,b\}}(v_2')$, it follows that $\pi_{\{a,b\}}(v_1')\in\{bbab,bbabb,bbabba,bbabbab\}$.
Assume $\pi_{\{a,b\}}(v_1')=bbabbab$. Then $|v_1'|_{ab}=4$. Furthermore, as $|v_1'|_b=5$, it holds by $(\star)$ that $|v_2'|_{bc}=p-2$. Note that $\pi_{\{a,b\}}(v_2')=b$, therefore $|v_2'|_a=0$ and consequently $|v_2'|_{abc}=0$. Thus $$\begin{aligned}
|w'|_{abc}={}&|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}+|v_2'|_{abc}\\
={}&|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}\\
={}&m+4\cdot (p+2)+2\cdot (p-2)\\
={}&m+4p+8+2p-4\\
={}&m+6p+4.\end{aligned}$$ That is to say, $|w'|_{abc}=m+6p+4<m+6p+5=|w|_{abc}$, which is a contradiction.
Assume $\pi_{\{a,b\}}(v_1')\in\{bbabb,bbabba\}$. Then, $|v_1'|_{ab}=2$. Furthermore, as $|v_1'|_b=4$, it holds by $(\star)$ that $|v_2'|_{bc}=2p$. Note that if $\pi_{\{a,b\}}(v_1')=bbabb$, then $\pi_{\{a,b\}}(v_2')=abb$. Consequently $|v_2'|_a=1$ and therefore $|v_2'|_{abc}=|v_2'|_{bc}$. Otherwise if $\pi_{\{a,b\}}(v_1')=bbabba$, then $\pi_{\{a,b\}}(v_2')=bb$. Consequently, $|v_2'|_a=0$ and therefore $|v_2'|_{abc}=0$ as well. In both cases, we have $|v_2'|_{abc}=|v_2'|_a|v_2'|_{bc}$. Thus $$\begin{aligned}
|w'|_{abc}={}&|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}+|v_2'|_{abc}\\
={}&|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}+|v_2'|_a|v_2'|_{bc}\\
={}&|v_1|_{abc}+|v_1'|_{ab}|v_2'|_c+(|v_1'|_a+|v_2'|_a)\cdot |v_2'|_{bc}\\
={}&m+2\cdot (p+2)+2\cdot 2p\\
={}&m+2p+4+4p\\
={}&m+6p+4.\end{aligned}$$ Similar to the case $\pi_{\{a,b\}}(v_1')=bbabbab$, we have $|w'|_{abc}=m+6p+4<m+6p+5=|w|_{abc}$, which is a contradiction.
Thus $\pi_{\{a,b\}}(v_1')=bbab$. We have $\pi_{\{a,b\}}(v_1')=\pi_{\{a,b\}}(v_1)$, therefore $|v_1'|_y=|v_1|_y$ for every $y\in\{a,b,ab\}$. As we already know that $|v_1'|_y=|v_1|_y$ for every $y\in\{abc,bc,c\}$, it follows that $v_1'\equiv_M v_1$ with respect to $\{a<b<c\}$. However, $v_1$ is , thus $v_1'=v_1$. Consequently, $v_2'\equiv_M v_2$ with respect to $\{a<b<c\}$ by the left invariance of $M$-equivalence. Similarly, $v_2$ is , thus $v_2'=v_2$. Therefore $w'=w$, which is a contradiction.
$\pi_{ab}(w')=babbbbab$.\
By similar reasoning as in Case 1, we have $\pi_{\{a,b\}}(v_1')=\{bab, babb, babbb, babbbb\}$. In all four cases, $|v_1'|_a=1$. Also, note that $|v_1'|_c=|v_1|_c=n+m+1$, $|v_1'|_{bc}=|v_1|_{bc}=3m+1$ and $|v_1'|_{abc}=|v_1|_{abc}=m$.
By Proposition \[CoreabcMiddle\], it holds that $v_1'\equiv_1u_1\operatorname{core}_{abc}(v_1')u_2$ for some unique $u_1\in\{b,c\}^*$ and $u_2\in\{a,b\}^*$. Since $\pi_{ab}(v_1')\in\{bab,babb,babbb,babbbb\}$ and $a$ is a prefix of $\operatorname{core}_{abc}(v_1')$, it follows that $\vert u_1\vert_b=1$. Also, note that $|v_1'|_{bc}=|u_1|_{bc}+|u_1|_b|\operatorname{core}_{abc}(v_1')|_c+|\operatorname{core}_{abc}(v_1')|_{bc}$. Since $|\operatorname{core}_{abc}(v_1')|_{abc}=\vert v_1'\vert_{abc}=m$, $a$ is a prefix of $\operatorname{core}_{abc}(v_1')$, and that is the only $a$ in $\operatorname{core}_{abc}(v_1')$, it follows that $|\operatorname{core}_{abc}(v_1')|_{bc}=m$. Additionally, since $|u_1|_{bc}+ |u_1|_{cb}=|u_1|_b|u_1|_c$, it follows that $|u_1|_{bc} \le |u_1|_b|u_1|_c= |u_1|_c$. Therefore, $|v_1'|_{bc}
\le |u_1|_{c}+|\operatorname{core}_{abc}(v_1')|_c+m
=|v_1'|_c+m
=n+m+1+m
=n+2m+1$. Consequently, $3m+1=|v_1|_{bc}=|v_1'|_{bc}\le n+2m+1$, which reduces to $m\le n$. Thus a contradiction occurs.
$\pi_{ab}(w')=bbbaabbb$.\
This case is impossible. Observe that $|v_1'|_b=3+|\operatorname{core}_{abc}(v_1')|_b$. Since $|v_1'|_b\le 5$ and $|\operatorname{core}_{abc}(v_1')|_b\ge 1$, it follows that $\pi_{ab}(v_1')\in\{bbbaab,bbbaabb\}$.
If $\pi_{ab}(v_1')=bbbaabb$, then $|v_1'|_b=5$ and consequently $|v_2'|_{bc}=p-2$ due to $(\star)$. Correspondingly, we have $|w'|_{abc}
=|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}+|v_2'|_{abc}
=m+4\cdot (p+2)+2\cdot (p-2)+0=m+6p+4$. On the other hand, if $\pi_{ab}(v_1')=bbbaab$, then $|v_1'|_b=4$ and consequently $|v_2'|_{bc}=2p$ due to $(\star)$. Correspondingly, we have $|w'|_{abc}
=|v_1'|_{abc}+|v_1'|_{ab}|v_2'|_c+|v_1'|_a|v_2'|_{bc}+|v_2'|_{abc}
=m+2\cdot (p+2)+2\cdot (2p)+0=m+6p+4$ as well. In both cases, $|w'|_{abc}<m+6p+5=|w|_{abc}$, which is a contradiction.
$\pi_{ab}(w')=abbbbbba$.\
This case is trivially impossible. Note that $|v_1'|_{bc}=|v_1|_{bc}=3m+1$. Consequently, $|v_1'|_{abc}=1\cdot|v_1'|_{bc}=3m+1$. However, $|v_1'|_{abc}=|v_1|_{abc}=m$, thus a contradiction.
Observation \[BasisObs1\] and Theorem \[BasisTheo1\] allow us to generate sequences of words (starting with a print word) such that the first word is and the of the remaining words sequentially change in an arbitrary pattern. This is illustrated by the following example.
\[ExFibo\] Consider the Fibonacci sequence $0,1,1,2,3,5,8,13,\cdots$. Suppose we want to generate a sequence of words realizing the $M$-ambiguity sequence where the number of terms $U$ between two consecutive terms $A$ follows the Fibonacci sequence—i.e, $A,A,U,A,U,A,U,U,A,\cdots$.
For all integers $n,m,p\ge 1$, let $w_{n,m,p}=c^nbcbabc^mdcbabcbc^p$. By Observation \[BasisObs1\], if $m=n$, then $w_{n,m,p}$ is $M$-ambiguous for any $p\ge 1$. By Theorem \[BasisTheo1\], if $m=n+1$, then $w_{n,m,p}$ is $M$-unambiguous for any $p\ge 1$. Thus, it remains to duplicate the letters in the following manner: $$w_{1,1,1},w_{1,1,2},w_{1,2,1},w_{2,2,1},w_{2,3,1},w_{3,3,1},w_{3,4,1},w_{3,4,2},w_{4,4,2},\cdots$$ (Notice that whenever we need to retain the preceding term, we increase the power $p$ by one—that is to duplicate the last letter $c$.)
On the other hand, to generate similar sequences of words such that the first word is , we need the following observation and result.
\[BasisObs2\] For all positive integers $n$ and $p$, the word $c^nbbabc^ndcbabbc^p$ is as it is to the word $bc^nabc^nbdbcbacbc^{p-1}$.
\[BasisTheo2\] The word $c^nbbabc^mdcbabbc^p$ is for all integers $n\geq 1$, $p\ge 1$ and $m\ge n+1$.
Argue similarly as in the proof of Theorem \[BasisTheo1\].
\[RemDrawback\] When $n\!=\!m\!=\!p\!=\!1$, in contrast to the word in the word in Theorem \[BasisTheo2\] is not a print word. Thus, similarly as in Example \[Example2\], we need the $M$-unambiguous words $cbabcdcbabc$ and $cbabcdcbabbc$ on top of Observation \[BasisObs2\] and Theorem \[BasisTheo2\] to realize $M$-ambiguity sequences starting with $U$. However, this forces the first three terms to be $U$, $U$, and $A$ before we can change the terms arbitrarily.
Periodicity of *M*-ambiguity Sequences
======================================
Consider the word $cbcbab\underline{c}dcbabcbc$ over the ordered alphabet $\{a<b<c<d\}$. By Theorem \[ConjectureOverturn\], it holds that every duplication of the underlined letter $c$ in that word gives rise to an $M$-unambiguous word. Thus for the sequence of words $\varphi=\{w_i\}_{i\ge 1}$ such that $w_i=cbcbabc^idcbabcbc$, we have $\Theta_{\varphi}=A,U,U,U,\cdots$.
We see that the sequence $\Theta_{\varphi}$ is eventually periodic with its period being one. Thus we seek to know whether the periodicity of an $M$-ambiguity sequence is a trait in the case of duplicating a single letter in a word. We formulate this question formally as follows.
\[QuesMotivation\] Suppose $\Sigma$ is an ordered alphabet. Let $\varphi=\{w_i\}_{i\ge 1}$ be a sequence of words over $\Sigma$ such that for every integer $k\ge 1$, we have $w_k=xa^ky$ for some $x,y\in\Sigma^*$ and $a\in\Sigma$. Is the sequence $\Theta_{\varphi}$ eventually periodic?
In the spirit of answering the above question, we first present a way to determine the $M$-ambiguity of a word—by transforming it to a problem of solving of linear equalities. To illustrate this, we analyze the word considered in Theorem \[ConjectureOverturn\] and deduce that it is $M$-unambiguous for every integer $n>1$.
Let $\Sigma=\{a<b<c<d\}$ and consider the word $w=cbcbabc^ndcbabcbc$, where $n$ is a nonnegative integer. If a word $w'\in\Sigma^*$ is $M$-equivalent to $w$, then $\pi_{\{a,b,d\}}(w')\equiv_M\pi_{\{a,b,d\}}(w)$ with respect to . Since $\pi_{\{a,b,d\}}(w)=bbabdbabb$, it follows that for such a word $w'$, the projection $\pi_{\{a,b,d\}}(w')$ must be one of the following: $$\tag{$\ast\ast$}
\begin{aligned}
&dbbabbabb,bdbabbabb,bbdabbabb,bbadbbabb,bbabdbabb,bbabbdabb,\\
&bbabbadbb,bbabbabdb,bbabbabbd,dbabbbbab,bdabbbbab,badbbbbab,\\
&babdbbbab,babbdbbab,babbbdbab,babbbbdab,babbbbadb,babbbbabd,\\
&dbbbaabbb,bdbbaabbb,bbdbaabbb,bbbdaabbb,bbbadabbb,bbbaadbbb,\\
&bbbaabdbb,bbbaabbdb,bbbaabbbd,dabbbbbba,adbbbbbba,abdbbbbba\\
&abbdbbbba,abbbdbbba,abbbbdbba,abbbbbdba,abbbbbbda,abbbbbbad.
\end{aligned}$$
Consider the scenario $\pi_{\{a,b,d\}}(w')=\pi_{\{a,b,d\}}(w)=bbabdbabb$. Then $$w'=c^{x_1}\boldsymbol{b}c^{x_2}\boldsymbol{b}c^{x_3}\boldsymbol{a}c^{x_4}\boldsymbol{b}c^{x_5}\boldsymbol{d}c^{x_6}\boldsymbol{b}c^{x_7}\boldsymbol{a}c^{x_8}\boldsymbol{b}c^{x_9}\boldsymbol{b}c^{x_{10}}$$ for some nonnegative integers $x_i\,(1\le i\le 10)$. Since $w'\equiv_Mw$, it follows that
-------------------------------------------------- ----- --------------- ----- -------------- ----- ----------
$x_1+x_2+x_3+x_4+x_5+x_6+x_7+x_8+x_9+x_{10}$ $=$ $|w'|_c$ $=$ $|w|_c$ $=$ $n+5$,
$x_2+2x_3+2x_4+3x_5+3x_6+4x_7+4x_8+5x_9+6x_{10}$ $=$ $|w'|_{bc}$ $=$ $|w|_{bc}$ $=$ $3n+15$,
$x_1+x_2+x_3+x_4+x_5$ $=$ $|w'|_{cd}$ $=$ $|w|_{cd}$ $=$ $n+2$,
$x_5+x_6+2x_7+2x_8+4x_9+6x_{10}$ $=$ $|w'|_{abc}$ $=$ $|w|_{abc}$ $=$ $n+11$,
$x_2+2x_3+2x_4+3x_5$ $=$ $|w'|_{bcd}$ $=$ $|w|_{bcd}$ $=$ $3n+1$,
$x_5$ $=$ $|w'|_{abcd}$ $=$ $|w|_{abcd}$ $=$ $n$.
-------------------------------------------------- ----- --------------- ----- -------------- ----- ----------
Solving the above system of linear equalities, we obtain the solution set $$\begin{split}
x_1 & = 1+x_3+x_4, \\
x_2 & = 1-2x_3-2x_4,\\
x_5 & = n,\\
x_6 & = 1,\\
x_7 & = -1-x_8+x_{10},\\
x_9 & = 3-2x_{10}.
\end{split}$$
By imposing the constraints $x_i\ge 0\,(1\le i\le 10)$, we now have the system of linear inequalities $$\tag{$\ast\ast\ast$}
\begin{split}
x_3+x_4&\ge -1, \\
2x_3+2x_4&\le 1,\\
n&\ge 0,\\
x_8-x_{10}&\le -1,\\
2x_{10}&\le 3\\
x_3,x_4,x_8,x_{10}&\ge 0.
\end{split}$$
From the above system of linear inequalities, notice that the only possible value of $x_{10}$ is 1 and therefore $x_8=0$. Also, observe that it can only be the case that $x_3=x_4=0$. By the system of linear equations before that, it follows that $x_1=x_2=x_6=x_9=1$, $x_5=n$ and $x_7=0$. As a result, we have $w'=cbcbabc^ndcbabcbc$. However, notice that $w'=w$, thus this scenario does not imply that $w$ is for every nonnegative
Next, consider the scenario $\pi_{\{a,b,d\}}(w')=babbbdbbab$. Analyzing similarly as above, we obtain the solution set $$\begin{split}
x_1 & = 1-n, \\
x_2 & = 1-x_3+x_5+n,\\
x_4 & = -2x_5+n,\\
x_6 & = -2+x_8+x_9,\\
x_7 & = 5-2x_8-2x_9,\\
x_{10} & = 0.
\end{split}$$ and the system of linear inequalities $$\begin{split}
n&\le 1, \\
x_3-x_5&\le 1,\\
2x_5-n&\le 0,\\
x_8+x_9&\ge 2,\\
2x_8+2x_9&\le 5\\
x_3,x_5,x_8,x_9&\ge 0.
\end{split}$$
By some simple analysis, one can see that for $n=0$ or $n=1$, integral solutions exist for the above system—each of them gives rise to a word $w'$ that is distinct from $w$. This implies that when $n=0$ or $n=1$, the word $w$ is . However, when $n>1$, there are no integral solutions, with such $n$, satisfying the system.
Arguing like this, one can see that each possibility of $\pi_{\{a,b,d\}}(w')$ in $(\ast\ast)$ leads to a system of linear equations and inequalities. Every such system can then be analyzed similarly as in above (thus we omit the details of the remaining computations). In our case here, when $n>1$, all the remaining $34$ systems lead to no solutions. Thus, we conclude that the word $w=cbcbabc^ndcbabcbc$ is $M$-unambiguous for all integers $n>1$.
\[RemForProof\] Suppose $\Sigma$ is an ordered alphabet. For a general word $$x_1a^{k}x_2a^{k}\cdots x_{j-1}a^{k}x_j$$ where $x_1,x_2,\cdots , x_j\in\Sigma^*$, $a\in\Sigma$, and $k$ is a positive integer, the above algebraic can be used to determine the values of $k$ such that the word is . The corresponding (finitely many) systems of linear equalities and are rational. We will need this observation for the proof of Theorem \[TheoEventPeriod1\] later.
Next, we need the following notion and known result, which in turn will be used to prove a lemma necessary for our purpose.
\[DefPoly\] Suppose $n$ is a positive integer. A set $P\subseteq\mathbb{R}^n$ is a *rational polyhedron* if and only if $P=\{\boldsymbol{x}\in\mathbb{R}^n\,|\,A\boldsymbol{x}\ge \boldsymbol{b}\}$ for some matrix $A\in\mathbb{Q}^{m\times n}$ and vector $\boldsymbol{b}\in\mathbb{Q}^m$, where $m$ is a positive integer.
The following result was deduced as Equation 19 in Chapter 16 of [@aS98]. We do not state the underlying details that lead to this result here as they are not essential for our purpose.
[@aS98]\[DecompTheo\] Suppose $n$ is a positive integer. For any rational polyhedra $P\subseteq\mathbb{R}^n$, there exist vectors $\boldsymbol{x_1},\boldsymbol{x_2},\cdots ,\boldsymbol{x_r},\boldsymbol{y_1},\boldsymbol{y_2},\cdots ,\boldsymbol{y_s}\in\mathbb{Z}^n$ such that $$\begin{aligned}
\{\boldsymbol{x}\in P\cap\mathbb{Z}^n\}=&\{\lambda_1\boldsymbol{x_1}+\cdots +\lambda_r\boldsymbol{x_r}+\mu_1\boldsymbol{y_1}+\cdots +\mu_s\boldsymbol{y_s}\,|\,\lambda_1,\cdots ,\lambda_r,\mu_1,\cdots ,\mu_s \text{ are }\\
&\text{ nonnegative integers with }\lambda_1+\cdots+\lambda_r=1\}.
\end{aligned}$$
We are now ready to prove our main lemma.
\[LemMain\] Suppose $n$ is a positive integer. Let $P=\{\boldsymbol{x}\in\mathbb{R}^n\,|\,A\boldsymbol{x}\ge \boldsymbol{b}\}$ for some matrix $A\in\mathbb{Q}^{m\times n}$ and vector $\boldsymbol{b}\in\mathbb{Q}^{m}$ where $m$ is a positive integer. Choose an arbitary integer $1\le k\le n$. Let $$P_k=\{p\in\mathbb{Z}^+\,|\, p \text{ is the $k^{th}$\! component of some } \boldsymbol{x}\in P\cap\mathbb{Z}_{\ge 0}^n\}.$$ Suppose the set $P_k$ is infinite. Then, for some positive integer $d$ and nonempty set $T\subseteq[0,d)\cap\mathbb{Z}$, there exists a positive integer $N$ such that $$\{p\in P_k\,|\,p\ge N\}=\{p\in\mathbb{Z}^+\,|\,p\ge N \text{ and } p=dq+t \text{ for some } t\in T \text{ and integers }q\}.$$
Clearly, by Definition \[DefPoly\], $P$ is a rational polyhedron. Therefore, by Theorem \[DecompTheo\], there exist vectors $\boldsymbol{x_1},\boldsymbol{x_2},\cdots ,\boldsymbol{x_r},\boldsymbol{y_1},\boldsymbol{y_2},\cdots ,\boldsymbol{y_s}\in\mathbb{Z}^n$ such that $$\begin{aligned}
\{\boldsymbol{x}\in P\cap\mathbb{Z}^n\}=\{&\lambda_1\boldsymbol{x_1}+\cdots +\lambda_r\boldsymbol{x_r}+\mu_1\boldsymbol{y_1}+\cdots +\mu_s\boldsymbol{y_s}\,|\,\lambda_1,\cdots ,\lambda_r,\mu_1,\cdots ,\mu_s \text{ are }\\
&\text{nonnegative integers with }\lambda_1+\cdots+\lambda_r=1\}.
\end{aligned}$$
Fix an arbitary integer $1\le k\le n$. Suppose the set $P_k$ is infinite. Let $\boldsymbol{x}[i]$ denote the $i^{th}$ component of a vector $\boldsymbol{x}$. Then, for an arbitrary $p\in P_k$, it holds that $$\tag{4.5.1}
\begin{aligned}
&p=\lambda_{1}\boldsymbol{x_1}[k]+\cdots +\lambda_{r}\boldsymbol{x_r}[k]+\mu_{1}\boldsymbol{y_1}[k]+\cdots +\mu_{s}\boldsymbol{y_s}[k] \text{ for some}\\
&\text{nonnegative integers }\lambda_{1},\cdots ,\lambda_{r},\mu_{1},\cdots ,\mu_{s}\text{ with }\lambda_{1}+\cdots+\lambda_{r}=1.
\end{aligned}$$
Assume $\boldsymbol{y_{i}}[k]$ is nonpositive for all integers $1\le i\le s$. Since $p$ is a positive integer and $\lambda_{1},\cdots ,\lambda_{r}$ are nonnegative integers with $\lambda_{1}+\cdots+\lambda_{r}=1$, it holds that $0<p\le \max\{\boldsymbol{x_i}[k]\,|\,1\le i\le r\}$. However, such values of integers $p$ are only finitely many, which is a contradiction as the set $P_k$ is infinite. Thus $\boldsymbol{y_{i}}[k]$ is positive for some integers $1\le i\le s$.
Choose an integer $I$ such that $\boldsymbol{y_{I}}[k]$ is positive. Let $$T=\{t\in[0,\boldsymbol{y_{I}}[k])\cap\mathbb{Z}\,|\,t=p-\boldsymbol{y_{I}}[k]\cdot q \text{ for some }p\in P_k \text{ and integer } q\}.$$ For every $t\in T$, let $$p_t^*=\min\{p\in P_k\,|\,p=\boldsymbol{y_{I}}[k]\cdot q+t \text{ for some integer }q\}.$$ Then, by (4.5.1), it follows that $$\begin{gathered}
\tag{4.5.2}
\text{for every $t\in T$ and integer $j\ge 0$, we have $p_t^*+j\cdot\boldsymbol{y_{I}}[k]\in P_k$.}\end{gathered}$$
Let $N=\max\{p_t^*\,|\,t\in T\}$ and $d=\boldsymbol{y_{I}}[k]$. Then the forward inclusion clearly holds by the definition of $N,d$ and $T$. To show that the backward inclusion holds, fix an arbitrary $p\in\mathbb{Z}^+$ with $p\ge N$ such that $p=\boldsymbol{y_{I}}[k]\cdot q+t$ for some $t\in T$ and integer $q$. Let $q^*_t$ be the integer such that $p^*_{t}=\boldsymbol{y_{I}}[k]\cdot q^*_{t}+t$. Note that $p^*_{t}\le N\le p$, thus $q^*_{t}\le q$. Notice that $$\begin{aligned}
p&=\boldsymbol{y_{I}}[k]\cdot q+t\\
&=\boldsymbol{y_{I}}[k]\cdot (q+q_t^*-q_t^*)+t\\
&=\boldsymbol{y_{I}}[k]\cdot (q^*_{t}+(q-q_t^*))+t\\
&=\boldsymbol{y_{I}}[k]\cdot q^*_{t} + t + \boldsymbol{y_{I}}[k]\cdot (q-q_t^*)\\
&=p_t^*+\boldsymbol{y_{I}}[k]\cdot (q-q_t^*).
\end{aligned}$$ It remains to see that since $q-q^*_{t}\ge 0$, by (4.5.2), it holds that $p\in P_k$. Thus our conclusion holds.
\[TheoEventPeriod1\] Suppose $\Sigma$ is an ordered alphabet. Let $\varphi=\{w_k\}_{k\ge 1}$ be a sequence of words over $\Sigma$ such that for every integer $k\ge 1$, we have $$w_k=x_1a^kx_2a^k\cdots x_{j-1}a^kx_j$$ for some $x_1,x_2,\cdots x_j\in\Sigma^*$ and $a\in\Sigma$. Then, $\Theta_\varphi$ is eventually periodic.
In Remark \[RemForProof\], we observe that for a word $w_k=x_1a^kx_2a^k\cdots x_{j-1}a^kx_j$ (as in the hypothesis) where $k$ is a positive integer, the algebraic analysis presented in the beginning of this section can be used to determine the values of $k$ such that $w_k$ is $M$-ambiguous. For the completeness of this proof, we will reiterate certain parts of the aforementioned analysis.
Let $\Gamma=\Sigma\backslash\{a\}$. Write $w_k$ in the form $a^{\gamma_1}\beta_1a^{\gamma_2}\beta_2\cdots a^{\gamma_{n}}\beta_{n}a^{\gamma_{n+1}}$ for some positive integer $n$, integers $\gamma_i\ge 0\,(1\le i\le n+1)$ and $\beta_i\in\Gamma \,(1\le i\le n)$. Note that $\beta_1\beta_2\cdots\beta_n=\pi_{\Gamma}(w_k)$. Suppose there exists $w'\in\Sigma^*$ such that $w'\equiv_Mw_k$. Then $\pi_{\Gamma}(w')\equiv_M\pi_{\Gamma}(w_k)$. Each possibility of the projection $\pi_{\Gamma}(w')$ gives rise to a rational system of linear inequalities as in $(\ast\ast\ast)$, with $k$ being a variable in it (due to the constraint $k\ge 1$). Each such system, when solved for nonnegative integral solutions, contains the values of $k$ such that $w_k$ is $M$-equivalent to $w'$ with that projection.
Assume $\pi_{\Gamma}(w')=\pi_{\Gamma}(w_k)=\beta_1\beta_2\cdots\beta_n$. Then $w'=a^{y_1}\beta_1a^{y_2}\beta_2\cdots a^{y_{n}}\beta_{n}a^{y_{n+1}}$ for some integers $y_i\ge 0\,(1\le i\le n+1)$. If $y_i=\gamma_i$ for every integer $1\le i\le n+1$, then $w'=w$. To avoid this, we impose the condition $y_i<\gamma_i$ or $y_i>\gamma_i$ for some integer $1\le i\le n+1$. Thus for every integer $1\le i\le n+1$, we consider two distinct systems of linear inequalities, each of them consisting of the ones obtained as in $(\ast\ast\ast)$, together with one of the conditions $y_i<\gamma_i$ or $y_i>\gamma_i$ —this gives a total of $2(n+1)$ systems of linear inequalities. On the other hand, if $\pi_{\Gamma}(w')\neq\pi_{\Gamma}(w_k)$, then it is impossible for $w'$ to be the same word as $w$. Thus, for each such possibility of $\pi_{\Gamma}(w')$, it suffices to consider the system of linear inequalities obtained as in $(\ast\ast\ast)$ —this gives a total of $|C_{\pi_{\Gamma}(w_k)}-1|$ systems.
Let $N=2(n+1)+|C_{\pi_{\Gamma}(w_k)}-1|$. Let integers $1\le i\le N$ enumerate the systems of linear inequalities that we have and write each of them in the form $A_i\boldsymbol{y}\ge \boldsymbol{b_i}$ for some matrix $A_i\in\mathbb{Q}^{r\times s}$ and vector $\boldsymbol{b_i}\in\mathbb{Q}^{r}$, where $r$ is a positive integer and $s=n+2$. For every integer $1\le i\le N$, let $P_i=\{\boldsymbol{y}\in\mathbb{R}^q\,|\,A_i\boldsymbol{y}\ge \boldsymbol{b_i}\}$ and let $\tau_i$ be the index such that the ${\tau_i}^{th}$ component of $\boldsymbol{y}$ corresponds to the variable $k$. Also, for every integer $1\le i\le N$, define the set $$P_i^*=\{p\in\mathbb{Z}^+\,|\, p \text{ is the ${\tau_i}^{th}$\! component of some } \boldsymbol{y}\in P_i\cap\mathbb{Z}_{\ge 0}^q\}.$$ Notice that $$\tag{4.6.1}
\text{the word } w_k \text{ is M\!-ambiguous if and only if } k\in\underset{1\le i\le N}{\bigcup} P_i^*.$$
The set $P_i^*$ is finite for every integer $1\le i\le N$.\
Then the set $\underset{1\le i\le N}{\bigcup} P_i^*$ is finite as well. By (4.6.1), the word $w_k$ is $M$-ambiguous for only finitely many values of $k$. For every integer $k>\!\max\{k\,|\,w_k \text{ is $M$\!-ambiguous}\}$, the word $w_k$ is $M$-unambiguous. Therefore, $\Theta_\varphi$ is eventually periodic (with its period being one).
The set $P_i^*$ is infinite for some integer $1\le i\le N$.\
Let $I=\{1\le i\le N\,|\,\text{the set }P_i^* \text{ is infinite}\}$. For every integer $i\in I$, by Lemma \[LemMain\], it follows that for some positive integer $d_i$ and nonempty set $T_i\subseteq[0,d_i)\cap\mathbb{Z}$, there exists a positive integer $M_i$ such that $$\{p\in P_i^*\,|\,p\ge M_i\}\!=\{p\in\mathbb{Z}^+\,|\,p\ge M_i \text{ and } p=d_iq+t \text{ for some } t\in T_i \text{ and integer }q\}.$$ Let $M'=\max(\{M_i\,|\,i\in I\}\cup\{p\in P_i^*\,|\,P_i^* \text{ is finite}\})$. Then, by (4.6.1), it follows that $$\tag{4.6.2}
\begin{aligned}
&\text{for every integer } k\ge M', \text{ the word } w_k \text{ is } M\!\text{-ambiguous if and only if}\\
&\text{there exists } i\in I \text{ such that } k=d_iq+t \text{ for some } t\in T_i \text{ and integer } q.
\end{aligned}$$
Let $d'=\underset{i\in I}{\prod} d$. By some simple argument, one can see that for any $i\in I$ and integer $k$, we have $k=d_iq+t$ for some $t\in T_i$ and integer $q$ if and only if $k+d'=d_iq+t$ for some $t\in T_i$ and integer $q$. Therefore, by (4.6.2), it holds that for every integer $k\ge M'$, the $M$-ambiguity of the words $w_{k+d'}$ and $w_k$ are the same. That is to say, the sequence $\Theta_\varphi$ is eventually periodic.
In both cases, our conclusion holds.
Finally, the following generalization holds as a consequence of the above theorem.
Suppose $\Sigma$ is an ordered alphabet. Let $\varphi=\{w_n\}_{n\ge 0}$ be a sequence of words over $\Sigma$ such that for every integer $n\ge 0$, we have $$w_n=x_1a^{k^{(1)}_n}\!x_2a^{k^{(2)}_n}\cdots x_{j}a^{k^{(j)}_n}\!x_{j+1}$$ for some $x_1,x_2,\cdots x_{j+1}\in\Sigma^*$ and $a\in\Sigma$ where
- $k^{(i)}_0=1$ for every integer $1\le i\le j$;
- for every integer $1\le i\le j$, let $e_i$ denote the $j$-tuple with $1$ in the $i^{th}$ coordinate and $0$ elsewhere, and for every integer $n\ge 1$, let $\alpha_n\in\{e_i\,|\,1\le i\le j\}$ and $$(k^{(1)}_n,k^{(2)}_n,\cdots, k^{(j)}_n)=(k^{(1)}_{n-1},k^{(2)}_{n-1},\cdots, k^{(j)}_{n-1})+\alpha_n;$$
If the sequence $\{\alpha_n\}_{n\ge 1}$ is periodic, then the sequence $\Theta_\varphi$ is eventually periodic.
Suppose the sequence $\{\alpha_n\}_{n\ge 1}$ is periodic, with a period $p$. Then for all integers $1\le n\le p$ and $m\ge 0$, we have $\alpha_{n+mp}=\alpha_n$. Let integers $d_i,(1\le i\le j)$ be such that $(d_1,d_2,\cdots ,d_j)=\sum\limits_{n=1}^{p}\alpha_n$. Next, we need the following observation. (The validity of the following claim can be easily verified by the reader, thus we omit its technical proof.)
\[Claim1\] For every integer $1\le n\le p$, let $\alpha^*_n$ be the $j$-tuple such that $\alpha^*_n=\sum\limits_{i=1}^{n}\alpha_i$ (the addition of tuples is defined element-wise). For all integers $1\le n\le p$ and $1\le i\le j$, let $\mu_{n,i}$ be the value in the $i^{th}$ coordinate of $\alpha^*_n$. Then, for all integers $1\le n\le p$, $1\le i\le j$ and $m\ge 0$, we have $k^{(i)}_{n+mp}=d_im+\mu_{n,i}+1$.
For all integers $1\le n\le p$ and $m\ge 0$, we have $$\tag{4.9.1}
\begin{aligned}
w_{n+mp}&=x_1a^{k^{(1)}_{n+mp}}x_2a^{k^{(2)}_{n+mp}}\cdots x_{j}a^{k^{(j)}_{n+mp}}x_{j+1}\\
&=x_1a^{d_1m+\mu_{n,1}+1}x_2a^{d_2m+\mu_{n,2}+1}\cdots x_{j}a^{d_jm+\mu_{n,j}+1}x_{j+1}\\
&=x_1\underbrace{a^m\cdots a^m}_\text{$d_1$ times}a^{\mu_{n,1}}ax_2\underbrace{a^m\cdots a^m}_\text{$d_2$ times}a^{\mu_{n,2}}a\cdots x_{j}\underbrace{a^m\cdots a^m}_\text{$d_j$ times}a^{\mu_{n,j}}ax_{j+1}
\end{aligned}$$ where the second equality holds by Claim \[Claim1\].
For all integers $0\le n<p$, define the sequence of words $\varphi_n=\{w_{n+mp}\}_{m\ge 0}$. Then, for every integer $0\le n<p$, it follows by $(4.9.1)$ and Theorem \[TheoEventPeriod1\] that the corresponding $M$-ambiguity sequence $\Theta_{\varphi_n}=\{\theta_{n,t}\}_{t\ge 0}$ is eventually periodic. That is to say, for every integer $0\le n<p$, there exists positive integers $T_n$ and $P_n$ such that for all integers $t\ge T_n$ and $m\ge 0$, we have $\theta_{n,t+mP_n}=\theta_{n,t}$. Let $T=\max\{T_n\,|\,0\le n<p\}$, then clearly $$\tag{4.9.2}
\text{for all integers } 0\le n<p,\, t\ge T \text{ and } m\ge 0,\text{ we have } \theta_{n,t+mP_n}=\theta_{n,t}.$$
Let $P=p\,\cdot\prod\limits_{n=1}^{p}P_n$. To see that the sequence $\Theta_\varphi=\{\vartheta_t\}_{t\ge 0}$ is eventually periodic, we show that for every integer $t\ge T$, we have $\vartheta_{t+P}=\vartheta_t$. Fix an arbitrary integer $t\ge T$. Let integers $q$ and $0\le r<p$ be such that $t=pq+r$. Then, it can be verified that $\vartheta_t=\theta_{r,q}$, and therefore $\vartheta_{t+P}=\theta_{r,q+\frac{P}{p}}$. It remains to see that since $\frac{P}{p}=\prod\limits_{n=1}^{p}P_n$, it follows by $(4.9.2)$ that $\vartheta_{t+P}=\theta_{r,q+\frac{P}{p}}=\theta_{r,q}=\vartheta_{t}$. Thus our conclusion holds.
Conclusion
==========
Unlike the case of binary and ternary alphabets, for larger alphabets, duplication of letters in a word can continuously alter the $M$-ambiguity of the resulting words. In fact, by using the main observations and results in Section 3, we have seen that nearly any pattern of $M$-ambiguity sequence is attainable.
As implied in Remark \[RemDrawback\], we are yet to find a print word such that selective repeated duplications of letters in that word could give rise to arbitrary sequences starting with the term $U$. We believe that by further investigation, this would be achievable as well. However, we leave it as an open problem.
The final result in Section 4 shows that repeated duplications of letters of the same type in a word, when done in a periodic manner, give rise to a periodic sequence. It remains to see if periodic duplications of different types of letters in a word would lead to the same conclusion. The main complexity would be that the associated systems consist of nonlinear equations and inequalities.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors gratefully acknowledge support for this research by a Research University Grant No. 1001/PMATHS/8011019 of Universiti Sains Malaysia. This study is an extension of the work in [@GT17e].
[10]{}
A. Atanasiu. Binary amiable words. , 18(2):387–400, 2007.
A. Atanasiu, R. Atanasiu, and I. Petre. Parikh matrices and amiable words. , 390(1):102–109, 2008.
A. Atanasiu and W. C. Teh. A new operator over [P]{}arikh languages. , 27(06):757–769, 2016.
K. Mahalingam, S. Bera, and K. G. Subramanian. Properties of [P]{}arikh matrices of words obtained by an extension of a restricted shuffle operator. , 29(3):403–3413, 2018.
A. Mateescu and A. Salomaa. Matrix indicators for subword occurrences and ambiguity. , 15(2):277–292, 2004.
A. Mateescu, A. Salomaa, K. Salomaa, and S. Yu. A sharpening of the [P]{}arikh mapping. , 35(6):551–564, 2001.
R. J. Parikh. On context-free languages. , 13:570–581, 1966.
G. Poovanandran and W. C. Teh. Strong $2\cdot t$ and strong $3\cdot t$ transformations for strong [M]{}-equivalence. .
G. Poovanandran and W. C. Teh. Elementary matrix equivalence and core transformation graphs for [P]{}arikh matrices. , 251:276–289, 2018.
G. Poovanandran and W. C. Teh. On [M]{}-equivalence and strong [M]{}-equivalence for [P]{}arikh matrices. , 29(01):123–137, 2018.
G. Poovanandran and W. C. Teh. Parikh matrices and [M]{}-ambiguity sequence. , 1132:012012, 2018.
A. Salomaa. Criteria for the matrix equivalence of words. , 411(16): 1818–1827, 2010.
A. Salomaa and S. Yu. Subword occurrences, [P]{}arikh matrices and [L]{}yndon images. , 21(1):91–111, 2010.
A. Schrijver. . John Wiley & Sons, 1998.
V. N. [Ş]{}erb[ă]{}nu[ţ]{}[ă]{}. On [P]{}arikh matrices, ambiguity, and prints. , 20(1):151–165, 2009.
V. N. [Ş]{}erb[ă]{}nu[ţ]{}[ă]{} and T. F. [Ş]{}erb[ă]{}nu[ţ]{}[ă]{}. Injectivity of the [P]{}arikh matrix mappings revisited. , 73(1):265–283, 2006.
W. C. Teh. arikh matrices and [P]{}arikh rewriting systems. , 146:305–320, 2016.
W. C. Teh and A. Atanasiu. On a conjecture about [P]{}arikh matrices. , 628:30–39, 2016.
W. C. Teh, A. Atanasiu, and G. Poovanandran. On strongly [M]{}-unambiguous prints and [[Ş]{}erb[ă]{}nu[ţ]{}[ă]{}]{}’s conjecture for [P]{}arikh matrices. , 719:86–93, 2018.
W. C. Teh and K. H. Kwa. Core words and [P]{}arikh matrices. , 582:60–69, 2015.
W. C. Teh, K. G. Subramanian, and S. Bera. Order of weak [M]{}-relation and [P]{}arikh matrices. , 743:83–92, 2018.
[^1]: See Theorem A.1 in [@SS06].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, the elliptic PDE-constrained optimization problem with box constraints on the control is studied. To numerically solve the problem, we apply the *‘optimize-discretize-optimize’* strategy. Specifically, the alternating direction method of multipliers (ADMM) algorithm is applied in function space first, then the standard piecewise linear finite element approach is employed to discretize the subproblems in each iteration. Finally, some efficient numerical methods are applied to solve the discretized subproblems based on their structures. Motivated by the idea of the multi-level strategy, instead of fixing the mesh size before the computation process, we propose the strategy of gradually refining the grid. Moreover, the subproblems in each iteration are solved inexactly. Based on the strategies above, an efficient convergent multi-level ADMM (mADMM) algorithm is proposed. We present the convergence analysis and the iteration complexity results $o(1/k)$ of the proposed algorithm for the PDE-constrained optimization problems. Numerical results show the high efficiency of the mADMM algorithm.'
author:
- 'Xiaotong Chen[^1]'
- 'Xiaoliang Song[^2]'
- Zixuan Chen
- Bo Yu
title: 'A multi-level ADMM algorithm for elliptic PDE-constrained optimization problems '
---
PDE-constrained optimization, ADMM, multi-level, convergence analysis
49N05, 49M25, 65N12, 68W15
Introduction {#intro}
============
In this paper, we consider the elliptic PDE-constrained optimization problem with box constraints on the control: $$\label{eqn:orginal problems}
\begin{aligned}
\min \limits_{(y,u)\in Y\times U}^{}\ \ &J(y,u)=\frac{1}{2}\|y-y_d\|_{L^2(\Omega)}^{2}+\frac{\alpha}{2}\|u\|_{L^2(\Omega)}^{2} \\
{\rm s.t.}\ \ \ \ \ \ &Ly=u+y_{r}\ \ \ \ \ \mathrm{in} \ \Omega, \\
&\ \ y=0 \quad \ \ \ \ \ \ \ \ \ \mathrm{on} \ \partial\Omega,\\
\qquad \qquad\quad &\ \ u\in U_{ad}=\{v(x)|a\leq v(x)\leq b, {\rm a.e }\ \mathrm{on}\ \Omega\}\subseteq U,
\end{aligned} \tag{$\mathrm{P}$}$$ where $Y:=H_{0}^{1}(\Omega), U:=L^{2}(\Omega), \Omega\subseteq\mathbb{R}^{n}(n=2,3)$ is a convex, open and bounded domain with $C^{1,1}$- or polygonal boundary; the desired state $y_{d}\in L^{2}(\Omega)$ and the source term $y_{r}\in L^{2}(\Omega)$ are given; parameters $\alpha>0$, $-\infty<a<b<+\infty$; $L$ is the uniformly elliptic differential operater given by $$Ly:=-\sum\limits_{i,j=1}^{n}(a_{ij}y_{x_{i}})_{{x_{j}}}+c_{0}y,$$ where $a_{ij},c_{0}\in L^{\infty}(\Omega),c_{0}\geqslant 0,a_{ij}=a_{ji},\sum\limits_{i,j=1}^{n}a_{ij}(x)\xi_{i}\xi_{j}\geqslant\theta\|\xi\|^{2}, \ \rm a.a. \ x\in \Omega, \forall \xi\in \mathbb{R}^{n}.$
By the classical conclusion in the PDE theory, for a given $y_{r}\in L^{2}(\Omega)$ and every $u\in L^{2}(\Omega)$, elliptic PDEs involved in the (\[eqn:orginal problems\]) $$\label{eqn:pde}
\begin{aligned}
Ly&=u+y_{r}\quad \mathrm{in}\ \Omega, \\
y&=0\quad \quad \quad \ \mathrm{on}\ \partial\Omega,\\
\end{aligned}$$ has a unique weak solution $y=y(u):=S(u+y_{r})$, where $S:L^{2}(\Omega)\rightarrow H_{0}^{1}(\Omega)$ denotes the solution operator. It is well-defined and is a continuous linear injective operator[@Kinderlehrer1980An Theorem B.4].
As is known to all, there are two possible approaches to tackle optimization problems with PDE constraints numerically. One is *First discretize, then optimize*, another is *First optimize, then discretize* [@Hinze2009Optimization]. In the first approach, the discretization is applied to the original PDE-constrained optimization problem, while in the second one, the discretization is applied to the KKT system of the PDE-constrained optimization problem. Different from the strategies mentioned above, instead of applying discretized concept to problem (\[eqn:orginal problems\]) or its KKT system directly, *‘optimize-discretize-optimize’* strategy was proposed by Song in [@Song2018]. First the optimization algorithm is given in the sense of continuous function space, then the subproblems in each iteration are discretized by the standard piecewise linear finite element approach. Finally, numerical optimization methods are applied to solve the discretized subproblems numerically. The advantage of this method is that from the optimization algorithm in function space, we can have a better knowledge of the structure of the PDE-constrained optimization problem, which is important for choosing an appropriate discretization format to discretize the subproblems. Moreover, this strategy gives the freedom to discretize the subproblems differently, which makes the subproblems can be solved more effectively.
In this paper, we focus on the alternating direction method of multipliers (ADMM) method, which was originally proposed by Glowinski et al. and Gabay et al. in [@Gabay1976A; @Glowinski1975] and has been broadly used in many areas. Motivated by the success of ADMM in solving finite dimensional large scale optimization problem [@Chen2017; @Fazel2013; @Li2015; @Li2016], ADMM-type method has been used to solve PDE-constrained optimization problems in function space in [@Chen2018; @li2018efficient; @Song2018A; @zhang2017alternating]. While these works all focus on *First discretize, then optimize* method, and to the best of our knowledge, very little work has been done to apply *First optimize, then discretize* to solve PDE-constrained optimization problems by ADMM-type method. We apply ADMM as the outer optimization algorithm of the *‘optimize-discretize-optimize’* strategy, propose a new algorithm to solve PDE-constrained optimization problems efficiently and give the convergence analysis.
First, we briefly give the iterative format of the classical ADMM for the 2-block convex optimization problem with linear constraints: $$\begin{aligned}
\min \limits_{(x,y)\in X\times Y}^{}&f(x)+g(y) \\
{\rm s.t.}\ \ \ \ &Ax+By=c,
\end{aligned}$$ where $X$,$Y$,$\Lambda$ are finite dimensional real Euclidean spaces, $f(x):X\rightarrow (-\infty, +\infty]$, $g(y):Y\rightarrow (-\infty, +\infty]$ are convex functions (maybe nonsmooth), $A: X\rightarrow \Lambda$, $B: Y\rightarrow \Lambda$ are linear mappings, $c \in \Lambda$ is given. The augmented Lagrangian function is given by the following form: $$\mathcal{L}_{\rho}(x,y,\lambda;\rho)=f(x)+g(y)+(\lambda,Ax+By-c)+\frac{\rho}{2}\|Ax+By-c\|^2,$$ where $\lambda \in \Lambda$ denotes the Lagrange multiplier, $\rho>0$ is the penalty parameter.
Given $(x^0,y^0,\lambda^0)\in {\rm dom}f\times{\rm dom}g\times \Lambda$, penalty parameter $\rho>0$ and the step size parameter $\tau\in \left(0,\frac{\sqrt{5}+1}{2}\right)$, then the iterative format of the classical ADMM is as follows: $$\left\{\begin{aligned}
x^{k+1}&={\text{argmin}}\mathcal{L}_{\rho}(x,y^k,\lambda^k;\rho),\\
y^{k+1}&={\text{argmin}} \mathcal{L}_{\rho}(x^{k+1},y,\lambda^k;\rho),\\
\lambda^{k+1}&=\lambda^k+\tau\rho(Ax^{k+1}+By^{k+1}-c).
\end{aligned} \right.$$
The advantage of ADMM is that it separates $f(x)$ and $g(y)$ into two subproblems. In each subproblem, there is only one variable, the other one is fixed. Thus each subproblem could be solved easily and efficiently. For the classical ADMM, the convergence analysis was first conducted by [@Fortin1983Chapter; @Gabay1976A; @Glowinski1980Lectures], and for the recent interesting new developments on the convergence analysis of ADMM-type method, see [@han2017linear; @Sun2014A; @2018arXiv180904249Y].
To apply ADMM-type method to solve the problem (\[eqn:orginal problems\]), we use the solution operator $S$ and introduce an artificial variable $z$, then we equivalently rewrite problem (\[eqn:orginal problems\]) as the following reduced form: $$\label{eqn:reduced problem with linear constraint}
\begin{aligned}
\min \limits_{(u,z)\in U\times Z}^{}\ \ \ &\hat{J}(u)+\delta_{U_{ad}}(z) \\
{\rm s.t.}\qquad \ &u=z,
\end{aligned} \tag{RP}$$ where $\hat{J}(u):=\dfrac{1}{2}\|S(u+y_{c})-y_{d}\|_{L^{2}(\Omega)}^{2}+\dfrac{\alpha}{2}\|u\|_{L^{2}(\Omega)}^{2}$ denotes the reduced cost function, $\delta_{U_{ad}}(z)$ denotes the indicator function of $U_{ad}$, $$\delta_{U_{ad}}(z)=\left\{\begin{array}{ll}{1,} & {z \in U_{ad}}, \\ {\infty,} & {z \notin U_{ad}}.\end{array}\right.$$ The augmented Lagrangian function of (\[eqn:reduced problem with linear constraint\]) is defined as follows: $$\label{eqn:lagrangian function}
L_{\sigma}(u,z,\lambda;\sigma)=\hat{J}(u)+\delta_{U_{ad}}(z)+\langle\lambda,u-z\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u-z\|_{L^{2}(\Omega)}^{2},$$ where $\lambda \in L^{2}(\Omega)$ denotes the Lagrangian multiplier, $\sigma>0$ is the penalty parameter.
The classical ADMM in finite dimensional spaces can be extended directly in Hilbert space for problem (\[eqn:reduced problem with linear constraint\]). Given $(u^{0},z^{0},{\lambda}^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, parameters $\sigma>0$, $\tau \in \left(0,\dfrac{1+\sqrt{5}}{2}\right)$, the iterative scheme is presented as follows:
$$\left\{\begin{aligned}
\bar u^{k+1}&={\rm argmin} \hat{J}(u)+\langle \bar \lambda^{k},u-\bar z^{k}\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u-\bar z^{k}\|_{L^{2}(\Omega)}^{2},\\
\bar z^{k+1}&={\rm argmin} \delta_{U_{ad}}(z)+\langle \bar \lambda^{k},\bar u^{k+1}-z\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|\bar u^{k+1}-z\|_{L^{2}(\Omega)}^{2},\\
\bar \lambda^{k+1}&=\bar \lambda^{k}+\sigma(\bar u^{k+1}-\bar z^{k+1}).
\end{aligned} \right.$$
Usually, it is expensive and unnecessary to exactly compute the solution of each subproblem even if it is feasible. Thus it is natural to use some iterative methods such as Krylov-based methods to solve the subproblems which are equivalent to large scale or ill-conditioned linear systems. The inexact ADMM algorithm in finite dimension space and its convergence results under certain error criterion have been studied extensively recently (see [@Chen2017; @Li2015]). Taking the inexactness of the solutions in the function space into account, Song et al. applied the inexact ADMM algorithm to Hilbert space for PDE-constrained optimization problems and presented the convergence results in [@Song2017Fe]. Given $(u^{0},z^{0},{\lambda}^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, parameters $\sigma>0$, $\tau \in \left(0,\dfrac{1+\sqrt{5}}{2}\right)$. Let$\lbrace\epsilon_{k}\rbrace_{k=0}^{\infty}$ be a sequence satisfying $\lbrace\epsilon_{k}\rbrace_{k=0}^{\infty}\subseteq[0,+\infty)$ and $\sum_{k=0}^{\infty}\epsilon_{k}<\infty$, the iterative scheme of the inexact ADMM in function space for (\[eqn:reduced problem with linear constraint\]) is as follows: $$\left\{\begin{aligned}
\text{Step 1}:\ &\text{Compute an approximation solution $u^{k+1}$ of}\\
&\min_{u} \hat{J}(u)+\langle \lambda^{k},u-z^{k}\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u-z^{k}\|_{L^{2}(\Omega)}^{2}\\
&\text{such that the error vector}\ \delta_{u}^{k+1}:=\nabla \hat{J}(u^{k+1})+\lambda^{k}+\sigma(u^{k+1}-z^{k}) \ \text{satisfies } \|\delta_{u}^{k+1}\|_{L^{2}(\Omega)}\leq \epsilon_{k}.\\
\text{Step 2}: \ &z^{k+1}={\rm argmin} \delta_{U_{ad}}(z)+\langle \lambda^{k},u^{k+1}-z\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u^{k+1}-z\|_{L^{2}(\Omega)}^{2}.\\
\text{Step 3}: \ &\lambda^{k+1}=\lambda^{k}+\tau \sigma(u^{k+1}-z^{k+1}).
\end{aligned} \right.$$
In [@Song2017Fe], the discretized problem is considered and the level of discretization is fixed. In this paper, instead of considering the discretized problem, the finite element method is employed to discretize the subproblems in each iteration of the inexact ADMM algorithm. This strategy gives the freedom and flexibility to discretize the subproblems by different discretization schemes. The total error of the proposed algorithm for (\[eqn:reduced problem with linear constraint\]) is consisted of two parts: the discretization error and the iteration error resulted from inexactly solving the discretized subproblems. For these two errors, our algorithm can be considered as an approximation of exact ADMM in function space, thus we regard it as an inexact ADMM algorithm in function space. In order to guarantee the convergence behavior of our algorithm, we consider controlling the mesh size and the inexactness of solving the subproblems.
In the classical finite element based ADMM-type algorithm to solve the PDE constrained optimization problem, the subproblems are always discretized on a fixed mesh size, which results in large scale optimization problems. Thus it is important to consider reducing the computation cost. Multi-grid method is a modern field of research starting with the works of Brandt [@Brandt1977Multi] and Hackbusch [@Hackbusch1985; @Hackbusch1978On]. It is well known that multi-grid method solves elliptic problems with optimal computational complexity. Motivated by the idea of applying multi-grid method to tackle infinite dimension problems by Newton method in [@Deuflhard2011newton; @Hackbusch1985], we apply the multi-level strategy to our algorithm. It is important to point out that in the initial stage of the algorithm, the iteration precision is required to be relatively low, which means using coarse mesh is sufficient. While as the iteration process proceeds, the iteration precision is supposed to be higher and higher. In this case, using finer mesh is necessary. It is obvious that the strategy of gradually refining the grid can strongly reduce the computation cost and make the algorithm faster than computing the problem on a fixed mesh size.
The main contribution of this paper is that we give the reasonable strategies of gradually refining the grid and inexactly solving the subproblems, and propose an efficient convergent multi-level ADMM (mADMM) algorithm. Specifically, we apply the ADMM algorithm in function space, then employ the standard piecewise linear finite element approach and implement the strategy of gradually refining the grid to the related subproblems appearing in each iteration. For the discretized subproblems, we use inexact strategies to solve them. For example, to solve the smooth subproblems, we can use some iterative methods such as Krylov-based methods. Thanks to the above strategies, we can solve the problem more efficiently and reduce the computation cost significantly. Moreover, we give the convergence analysis as well as the iteration complexity results $o(1/k)$ for the mADMM algorithm.
The paper is organized as follows. In section 2, we apply the inexact ADMM algorithm in function space first, then use finite element method to discretize the associated subproblems and propose the multi-level ADMM (mADMM) algorithm. Section 3 gives the convergence analysis and the iteration complexity of the proposed mADMM algorithm. The numerical results are given in section 4. Section 5 contains a brief summary of this paper.
A multi-level ADMM algorithm {#sec:2}
============================
In this section, we apply the *‘optimize-discretize-optimize’* strategy and propose an efficient convergent multi-level ADMM (mADMM) algorithm in Algorithm \[alg:multi-level ADMM algorithm\]. The strategies of gradually refining the grid and inexactly solving the subproblems are given to guarantee the convergence and efficiency of the mADMM algorithm.
The mADMM algorithm {#sec:2.1}
-------------------
Based on the inexact ADMM in function space, to numerically solve problem (\[eqn:reduced problem with linear constraint\]), we consider the finite element method. Specifically, piecewise linear functions are utilized to discretize the variables in the related subproblems appearing in each iteration of the inexact ADMM in function space. We first consider a family of regular and quasi-uniform triangulations $\lbrace \mathcal{T}_{h}\rbrace$ of $\bar \Omega$, i.e. $\bar \Omega=\bigcup_{T\in \mathcal{T}_{h}}\bar T$. With each element $T\in \mathcal{T}_{h}$, we define the diameter of the set $T$ by $\rho_{T}:={\rm diam} \ T$ and let $\sigma_{T}$ denotes the diameter of the largest ball contained in $T$. The mesh size of the grid is defined by $h:={\rm max}_{T\in \mathcal{T}_{h}}\ \rho_{T}$. We suppose the following standard assumption in the context of error estimates holds (see[@Hinze2010Variational; @Hinze2009Optimization]).
(Regular and quasi-uniform triangulations) There exist two positive constants $\kappa$ and $\tau$ such that $$\frac{\rho_{T}}{\sigma_{T}}\leq \kappa, \quad \frac{h}{\rho_{T}}\leq \tau$$ hold for all $T\in \mathcal{T}_{h}$ and all $h>0$. Moreover, let us define $\bar \Omega_{h}=\bigcup_{T\in \mathcal{T}_{h}}\bar T$ and let $\Omega_{h}\subseteq \Omega$ and $\Gamma_{h}$ denote its interior and its boundary, respectively. In the case that $\Omega$ is a convex polyhedral domain, we have $\Omega=\Omega_{h}$. In the case that $\Omega$ is a domain with a $C^{1,1}$- boundary $\Gamma$, we assume that $\bar \Omega_{h}$ is convex and that all boundary vertices of $\bar \Omega_{h}$ are contained in $\Gamma$, such that $$|\Omega\backslash\Omega_{h}|\leq ch^{2},$$ where $|\cdot|$ denotes the measure of the set, and $c>0$ is a constant.
For the state variable $y$ and the control variable $u$, we choose the same discretized state space and discretized control space defined by $$\begin{aligned}
&Y_{h}:=\{y_{h}\in C( \bar{\Omega})|y_{{h}|T}\in \mathcal{P}_{1}, \forall{T\in \mathcal{T}_{h},y_{h}=0 \ {\rm in} \ \bar{\Omega} \setminus\Omega_{h}}\},\\
&U_{h}:=\{u_{h}\in C( \bar{\Omega})|u_{{h}|T}\in \mathcal{P}_{1}, \forall{T\in \mathcal{T}_{h},u_{h}=0 \ {\rm in} \ \bar{\Omega} \setminus\Omega_{h}}\},\end{aligned}$$ where $\mathcal{P}_{1}$ denotes the space of polynomials of degree less than or equal to 1. For given source term $y_{r}$ and $u\in L^{2}(\Omega)$, let the discretized state associated with $u$ denoted by $y_{h}(u)$, which is defined as the unique solution for the discretized weak formulation of the state equation (\[eqn:pde\]): $$\label{eqn:weak formulation}
\int_{\Omega_{h}}\left(\sum_{i,j=1}^{n}a_{ij}(y_{h})_{x_{i}}(v_{h})_{x_{j}}+c_{0}y_{h}v_{h}\right)dx=\int_{\Omega_{h}}(u+y_{r})v_{h}dx, \ \ \forall v_{h}\in Y.$$ Moreover, $y_{h}$ can be expressed by $y_{h}(u)=S_{h}(u+y_{r})$, where $S_{h}$ denotes the discretized version of the solution operator $S$. Then we have the following well-known conclusion about error estimates.
\[lem:error estimates\][([@Ciarlet2002The], Thm. 4.4.6)]{} For a given $u\in L^{2}(\Omega)$, let $y$ be the unique weak solution of the state equation (\[eqn:pde\]) and $y_{h}$ be the unique solution of (\[eqn:weak formulation\]). Then there exists a constant $c>0$ independent of $h$, $u$ and $y_{r}$ such that $$\|y-y_{h}\|_{L^{2}(\Omega)}+h\|\nabla y-\nabla y_{h}\|_{L^{2}(\Omega)}\leq ch^{2}(\|u\|_{L^{2}(\Omega)}+\|y_{r}\|_{L^{2}(\Omega)}).$$ In particular, this implies $\|S-S_{h}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\leq ch^{2}$ and $\|S-S_{h}\|_{L^{2}(\Omega)\rightarrow H^{1}(\Omega)}\leq ch$.
For the given triangulation $\mathcal{T}_{h}$ with nodes $\{x_{i}\}_{i=1}^{N_{h}}$, let $\{\phi_{i}(x)\}_{i=1}^{N_{h}}$ be a set of nodal basis functions associated with nodes $\{x_{i}\}_{i=1}^{N_{h}}$ which spans $Y_{h}$, $U_{h}$ and satisfies the properties: $$\phi_{i}(x)\geqslant0,\ \ \|\phi_{i}(x)\|_{\infty}=1\ \ \ \forall{i=1,...,N_{h}}, \ \ \sum\limits_{i=1}^{N_{h}}\phi_{i}(x)=1.$$ Then $y_{h}\in Y_{h}$, $u_{h}\in U_{h}$ can be represented as $y_{h}=\sum\limits_{i=1}^{N_{h}}y_{i}\phi_{i},$ $u_{h}=\sum\limits_{i=1}^{N_{h}}u_{i}\phi_{i}$ and we have $y_{h}(x_{i})=y_{i}$, $u_{h}(x_{i})=u_{i}$. The other variables and operators in the subproblems of the inexact ADMM in function space are all discretized by piecewise linear functions similarly.
Before we give the proposed algorithm to solve the problem (\[eqn:reduced problem with linear constraint\]), we first introduce the definition of node interpolation operator $I_{h}$ for the following convergence analysis.
For a given regular and quasi-uniform triangulation $\mathcal{T}_{h}$ of $\Omega$ with nodes $\{x_{i}\}_{i=1}^{N_{h}}$, let $\{\phi_{i}(x)\}_{i=1}^{N_{h}}$ denotes a set of associated nodal basis functions. Then the interpolation operator is defined as $$I_{h}w(x):=\sum_{i=1}^{N_{h}}w(x_{i})\phi_{i}(x) \ \ {\rm for \ any} \ w \in L^{2}(\Omega).$$
Moreover, about the interpolation error estimate, we have the following result.
\[lem:interpolation error estimate\][([@Ciarlet2002The], Thm. 3.1.6)]{} For all $w \in W^{k+1,p}(\Omega), k\geq 0, p, q\in[0,+\infty)$ and $0\leq m \leq k+1$, we have $$\|w-I_{h}w\|_{W^{m,q}(\Omega)}\leq c_{I}h^{k+1-m}\|w\|_{W^{k+1,p}(\Omega)},$$ where $c_{I}$ is a constant which is independent of the mesh size $h$.
In the classical finite element based algorithm, the discretization mesh size is fixed in advance. When computing on the finer mesh, the scale of the discretized problem becomes larger and the computation cost becomes larger. So it is important to consider reducing the computation cost. In the first several iterations of the algorithm, the iteration precision is not that good, which means using coarse mesh will not make the precision worse, but reduce the computation amount. While as the iteration process proceeds, the iteration precision becomes higher and higher. In this case, using finer mesh is necessary. Thus we introduce the idea of gradually refining the grid. Specifically, in the initial iteration we choose coarse mesh and obtain a solution first, then use the interpolation operator to project the obtained solution to the finer mesh. Finally, we apply appropriate numerical methods to solve the subproblems in the finer mesh and obtain a more precise solution, so on and so forth.
Let the mesh size of the $k$th iterate denotes by $h_{k}, k\in \mathbb{Z}, k\geq 1$, then the discretized reduced cost function is defined as follows, $${\hat J_{h_{k+1}}}(u_{h_{k+1}}):=\dfrac{1}{2}\|S_{h_{k+1}}(u_{h_{k+1}}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}\|^{2}_{L^{2}(\Omega)}+\dfrac{\alpha}{2}\|u_{h_{k+1}}\|^{2}_{L^{2}(\Omega)}.$$ Let $U_{ad, h_{k+1}}$ denotes the discretized feasible set, $$U_{a d, h_{k+1}} :=U_{h_{k+1}} \cap U_{a d}=\left\{z_{h_{k+1}}=\sum_{i=1}^{N_{h_{k+1}}} z_{i} \phi_{i}(x) | a \leq z_{i} \leq b, \forall i=1, \cdots, N_{h_{k+1}}\right\} \subset U_{a d}.$$ Moreover, we define $\lambda_{h_{k+1}}^{k}:=I_{h_{k+1}}\lambda_{h_{k}}^{k}$ and $z_{h_{k+1}}^{k}:=I_{h_{k+1}}z_{h_{k}}^{k}$, where $I_{h_{k+1}}$ denotes the node interpolation operator. Based on the above representations, we present the iterative scheme of the multi-level ADMM algorithm (mADMM) in Algorithm \[alg:multi-level ADMM algorithm\].
$(u^{0},z^{0},{\lambda}^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, parameters $\sigma>0$, $\tau \in \left(0,\dfrac{1+\sqrt{5}}{2}\right)$. Let$\lbrace\xi_{k+1}\rbrace_{k=0}^{\infty}$ be a sequence satisfying $\lbrace\xi_{k+1}\rbrace_{k=0}^{\infty}\subseteq[0,+\infty)$ and $\sum_{k=0}^{\infty}\xi_{k+1}<\infty$, mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $\sum_{k=0}^{\infty}h_{k+1}< \infty.$ Set $k=0$. $u^{k}_{h_{k}},z^{k}_{h_{k}},{\lambda}^{k}_{h_{k}}$. Compute an approximation solution of
\_[u\_[h\_[k+1]{}]{}]{} [J\_[h\_[k+1]{}]{}]{}(u\_[h\_[k+1]{}]{})+\_[h\_[k+1]{}]{}\^[k]{},u\_[h\_[k+1]{}]{}-[z]{}\_[h\_[k+1]{}]{}\^[k]{}\_[L\^[2]{}()]{} +u\_[h\_[k+1]{}]{}-[z]{}\_[h\_[k+1]{}]{}\^[k]{}\^[2]{}\_[L\^[2]{}()]{}
such that the error vector $\delta_{u,h_{k+1}}^{k+1}:=\nabla {\hat J_{h_{k+1}}}(u_{h_{k+1}})+{\lambda}_{h_{k+1}}^{k}+\sigma(u_{h_{k+1}}^{k+1}-{z}_{h_{k+1}}^{k})$ satisfies\
$\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq \xi_{k+1}.$ Compute $z^{k+1}_{h_{k+1}}$ as follows: $$z^{k+1}_{h_{k+1}}={\rm argmin} \delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}})+\langle {\lambda}_{h_{k+1}}^{k},u^{k+1}_{h_{k+1}}-z_{h_{k+1}}\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u^{k+1}_{h_{k+1}}-z_{h_{k+1}}\|^{2}_{L^{2}(\Omega)}.$$ Compute $$\lambda^{k+1}_{h_{k+1}}=I_{h_{k+1}}\lambda^{k}_{h_{k}}+\tau \sigma(u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}).$$ If a termination criterion is met, stop; else, set $k:=k+1$ and go to Step 1.
In order to guarantee the sequence $\lbrace\xi_{k+1}\rbrace_{k=0}^{\infty}\subseteq[0,+\infty)$ satisfies $\sum_{k=0}^{\infty}\xi_{k+1}<\infty$. In the numerical experiment, we can choose $\xi_{k+1}=\frac{1}{(k+1)^{2}}$ as an example.
In order to guarantee the mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $\sum_{k=0}^{\infty}h_{k+1}< \infty.$ In the numerical experiment, we choose $h_{k}=2^{-(k+4)}$ in Example \[ex1\], $h_{k}=\sqrt{2}/2^{(k+3)}$ in Example \[ex2\].
Numerical computation of the subproblems in Algorithm \[alg:multi-level ADMM algorithm\]
----------------------------------------------------------------------------------------
For $u=\sum_{i=1}^{N_{h}}u_{i}\phi_{i}\in U_{h}$, $y=\sum_{i=1}^{N_{h}}y_{i}\phi_{i}\in Y_{h}$, let $\boldsymbol{\rm u}=(u_{1},...,u_{N_{h}})$, $\boldsymbol{\rm y}=(y_{1},...,y_{N_{h}})$ be the relative coefficient vectors respectively. Let the $L^{2}$-projections of $y_{r}$ and $y_{d}$ onto $Y_{h}$ be $y_{r,h}:=\sum_{i=1}^{N_{h}}y_{r}^{i}\phi_{i}(x)$ and $y_{d,h}:=\sum_{i=1}^{N_{h}}y_{d}^{i}\phi_{i}(x)$, respectively. Similarly, $\boldsymbol{\rm y}_{r}=(y_{r}^{1},y_{r}^{2}, ...,y_{r}^{N_{h}})$ and $\boldsymbol{\rm y}_{d}=(y_{d}^{1},y_{d}^{2}, ...,y_{d}^{N_{h}})$ denote their coefficient vectors. The stiffness matrix and the mass matrix are defined as: $$K_{h}:=(a(\phi_{i},\phi_{j}))_{i,j=1}^{N_{h}}, \ M_{h}:=\left( \int_{\Omega_{h}}\phi_{i}\phi_{j}dx \right)_{i,j=1}^{N_{h}}.$$ Furthermore, we define $$\begin{aligned}
f(\boldsymbol{\rm u})&:=\dfrac{1}{2}\|K_{h}^{-1}M_{h}(\boldsymbol{\rm u}+\boldsymbol{\rm y}_{r})-\boldsymbol{\rm y}_{d}\|_{M_{h}}^{2}+\dfrac{\alpha}{2}\|\boldsymbol{\rm u}\|_{M_{h}}^{2}, \\
g(\boldsymbol{\rm z})&:=\delta_{[a,b]^{N_{h}}}(\boldsymbol{\rm z}).\end{aligned}$$ Then we can obtain the matrix-vector form of Algorithm \[alg:multi-level ADMM algorithm\] in Algorithm \[alg:matrix-vector form\].
$(\boldsymbol{\rm u}^{0},\boldsymbol{\rm z}^{0},\boldsymbol{\rm \lambda}^{0})\in \mathbb{R}^{N_{h}}\times{[a,b]^{N_{h}}}\times\mathbb{R}^{N_{h}}$, parameters $\sigma>0$, $\tau \in \left(0,\dfrac{1+\sqrt{5}}{2}\right)$. Let$\lbrace\xi_{k+1}\rbrace_{k=0}^{\infty}$ be a sequence satisfying $\lbrace\xi_{k+1}\rbrace_{k=0}^{\infty}\subseteq[0,+\infty)$ and $\sum_{k=0}^{\infty}\xi_{k+1}<\infty$, mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $O(\sum_{k=0}^{\infty}h_{k+1})\leq \infty.$ Set $k=0$. $\boldsymbol{\rm u}^{k},\boldsymbol{\rm z}^{k},\boldsymbol{\rm \lambda}^{k}$. Compute an approximation solution of $$\min_{\boldsymbol{\rm u}} f(\boldsymbol{\rm u})+\langle \boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}_{h_{k}}^{k},M_{h_{k+1}}(\boldsymbol{\rm u}-\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm z}^{k}_{h_{k}})\rangle+\dfrac{\sigma}{2}\|\boldsymbol{\rm u}-\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm z}^{k}_{h_{k}}\|_{M_{h_{k+1}}}^{2},$$ such that the error vector $$\boldsymbol{\delta}_{{\boldsymbol u},h_{k+1}}^{k}:=\nabla f(\boldsymbol{\rm u}^{k+1}_{h_{k+1}})+M_{h_{k+1}}^{T}\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}^{k}_{h_{k}}+\sigma M_{h_{k+1}}(\boldsymbol{\rm u}^{k+1}_{h_{k+1}}-\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm z}^{k}_{h_{k}})$$ satisfies $\|\boldsymbol{\delta}_{{\boldsymbol u},h_{k+1}}^{k}\|_{2}\leq \xi_{k+1}$. Compute $\boldsymbol{\rm z}^{k}$ as follows: $$\boldsymbol{\rm z}^{k+1}_{h_{k+1}}={\rm argmin} \ g(\boldsymbol{\rm z})+\langle \boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}^{k}_{h_{k}},M_{h_{k+1}}(\boldsymbol{\rm z}-\boldsymbol{\rm u}^{k+1}_{h_{k+1}})\rangle+\dfrac{\sigma}{2}\|\boldsymbol{\rm u}^{k+1}_{h_{k+1}}-\boldsymbol{\rm z}\|_{M_{h_{k+1}}}^{2}.$$ Compute $$\boldsymbol{\rm \lambda}^{k+1}_{h_{k+1}}=\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}^{k}_{h_{k}}+\tau\sigma(\boldsymbol{\rm u}^{k+1}_{h_{k+1}}-\boldsymbol{\rm z}^{k+1}_{h_{k+1}}).$$ If a termination criterion is met, stop; else, set $k:=k+1$ and go to Step 1.
Let $\boldsymbol{\rm y}^{k+1}:=K_{h}^{-1}M_{h}(\boldsymbol{\rm u}^{k+1}+\boldsymbol{\rm y}_{r}),\
\boldsymbol{\rm p}^{k+1}:=K_{h}^{-1}M_{h}(\boldsymbol{\rm y}_{d}-\boldsymbol{\rm y}^{k+1})$ denotes the discretized state and the discretized adjoint state respectively. Then the $\boldsymbol {\rm u}$-subproblem at the $k$-th iterate is equivalent to solving the following linear system: $$\left[
\begin{array}{ccc}
M_{h} & 0 & K_{h}\\
0 & (\alpha+\sigma)M_{h} & -M_{h}\\
K_{h} & -M_{h} & 0\\
\end{array}
\right]
\begin{array}{ccc}
\left[
\begin{array}{ccc}
\boldsymbol{\rm y}^{k+1} \\
\boldsymbol{\rm u}^{k+1}\\
\boldsymbol{\rm p}^{k+1}\\
\end{array}
\right]
\end{array}
=
\left[
\begin{array}{ccc}
M_{h}\boldsymbol{\rm y}_{d} \\
M_{h}(-\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}^{k}+\sigma \boldsymbol{\rm I}_{h_{k+1}}{\boldsymbol{\rm z}}^{k})\\
M_{h}\boldsymbol{\rm y}_{r}\\
\end{array}
\right].$$ Since $\boldsymbol{\rm p}^{k+1}=(\alpha+\sigma)\boldsymbol{\rm u}^{k+1}+\boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm \lambda}^{k}-\sigma \boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm z}^{k}$, we eliminate the variable $\boldsymbol{\rm p}$, then the above problem can be rewritten in the following reduced form without any computational cost: $$\label{equ:linear}
\left[
\begin{array}{ccc}
M_{h} & K_{h}(\alpha+\sigma) \\
-K_{h} & M_{h} \\
\end{array}
\right]
\begin{array}{ccc}
\left[
\begin{array}{ccc}
\boldsymbol{\rm y}^{k+1} \\
\boldsymbol{\rm u}^{k+1}\\
\end{array}
\right]
\end{array}
=
\left[
\begin{array}{ccc}
M_{h}\boldsymbol{\rm y_{d}}-K_{h}\boldsymbol{\rm I}_{h_{k+1}}(\boldsymbol{\rm \lambda}^{k}+\sigma \boldsymbol{\rm I}_{h_{k+1}}\boldsymbol{\rm z}^{k}))\\
M_{h}\boldsymbol{\rm y}_{r}
\end{array}
\right].$$ The equivalent linear system (\[equ:linear\]) can be solved inexactly by some Krylov-based method.
Finally, we give a terminal condition of Algorithm \[alg:matrix-vector form\]. Let $\epsilon$ be a given accuracy tolerance, we terminate the algorithm when $\eta<\epsilon$, where the accuracy of a numerical solution is measured by the residual $\eta:={\rm max}\lbrace\eta_{1},\eta_{2},\eta_{3},\eta_{4},\eta_{5}\rbrace,$ where $$\begin{aligned}
\eta_{1}&=\frac{\|K_{h}\boldsymbol{\rm y}-M_{h}\boldsymbol{\rm u}-M_{h}\boldsymbol{\rm y_{r}}\|}{1+\|M_{h}{\boldsymbol{\rm y}}_{r}\|}, \
\eta_{2}=\frac{\|M_{h}(\boldsymbol{\rm u}-\boldsymbol{\rm z})\|}{1+\|\boldsymbol{\rm u}\|},\
\eta_{3}=\frac{\|M_{h}(\boldsymbol{\rm y}-{\boldsymbol{\rm y}}_{d})+K_{h}\boldsymbol{\rm p}\|}{1+\|M_{h}\boldsymbol{\rm y}_{d}\|},\\
\eta_{4}&=\frac{\|\alpha M_{h}\boldsymbol{\rm u}-M_{h}\boldsymbol{\rm p}+M_{h}\boldsymbol{\rm \lambda}\|}{1+\|\boldsymbol{\rm u}\|},\
\eta_{5}=\frac{\|\boldsymbol{\rm z}-\Pi_{[a,b]^{N_{h}}}(\boldsymbol{\rm z}+M_{h}\boldsymbol{\rm \lambda})\|}{1+\|\boldsymbol{\rm z}\|}.
\end{aligned}$$
Convergence analysis {#sec:3}
====================
In this section, based on the convergence results of the inexact ADMM in function space in Theorem \[convergence functional space\], we give the convergence analysis and the iteration complexity of the proposed mADMM algorithm.
[([@Song2017Fe], Thm. 2.5)]{} \[convergence functional space\] Suppose that the operator $L$ is uniformly elliptic. Let $(y^{\ast},u^{\ast},z^{\ast},p^{\ast},\lambda^{\ast})$ be the KKT point of (P), the sequence $\lbrace(u^{k},z^{k},\lambda^{k})\rbrace$ is generated by the inexact ADMM algorithm for (\[eqn:reduced problem with linear constraint\]) with the associated state $\lbrace y_{k}\rbrace$ and adjoint state $\lbrace p_{k}\rbrace$, then we have $$\begin{aligned}
&\lim \limits_{k\rightarrow \infty}\lbrace \|u^{k}-u^{\ast}\|_{L^{2}(\Omega)}+\|z^{k}-z^{\ast}\|_{L^{2}(\Omega)}+\|\lambda^{k}-\lambda^{\ast}\|_{L^{2}(\Omega)}\rbrace=0,\\
&\lim \limits_{k\rightarrow \infty}\lbrace \|y^{k}-y^{\ast}\|_{H_{0}^{1}(\Omega)}+\|p^{k}-p^{\ast}\|_{H_{0}^{1}(\Omega)}\rbrace=0.\end{aligned}$$ Moreover, there exists a constant $C$ only depends on the initial point $(u^{0}, z^{0}, \lambda^{0})$ and the optimal solution $(u^{\ast}, z^{\ast}, \lambda^{\ast})$ such that for $k\geq 1$, $$\min_{1\leq i\leq k}{R(u^{i}, z^{i}, \lambda^{i})}\leq \frac{C}{k}, \ \ \lim \limits_{k\rightarrow \infty}(k\times \min_{1\leq i\leq k}{R(u^{i}, z^{i}, \lambda^{i})})=0,$$ where the function $R: (u, z, \lambda)\rightarrow [0,\infty)$ defined as $$R(u, z, \lambda):= \|\nabla {\hat J}(u)+\lambda\|^{2}_{L^{2}(\Omega)}+{\rm dist}^{2}(0, -\lambda+\partial \delta_{U_{ad}}(z))+\|u-z\|^{2}_{L^{2}(\Omega)}.$$
For the convenience of proving the convergence results, we give the iterative scheme of the multi-level discretized ADMM for (\[eqn:reduced problem with linear constraint\]) and present a lemma to measure the gap between the solution sequences obtained by the ADMM in function space and the multi-level discretized ADMM.
Given $(u^{0},z^{0},{\lambda}^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, parameters $\sigma>0$, $\tau \in \left(0,\dfrac{1+\sqrt{5}}{2}\right)$. The mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $\sum_{k=0}^{\infty}h_{k+1}< \infty.$ Then the iterative format of the multi-level discretized ADMM algorithm is as follows: $$\left\{\begin{aligned}
\bar u^{k+1}_{h_{k+1}}&={\rm argmin}\hat J_{h_{k+1}}(u_{h_{k+1}})+\langle\bar{\lambda}_{h_{k+1}}^{k},u_{h_{k+1}}-\bar{z}_{h_{k+1}}^{k}\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|u_{h_{k+1}}-\bar{z}_{h_{k+1}}^{k}\|_{L^{2}(\Omega)}^{2},\\
\bar z^{k+1}_{h_{k+1}}&={\rm argmin} \delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}})+\langle {\bar \lambda}_{h_{k+1}}^{k},\bar u^{k+1}_{h_{k+1}}-z_{h_{k+1}}\rangle_{L^{2}(\Omega)}+\dfrac{\sigma}{2}\|\bar u^{k+1}_{h_{k+1}}-z_{h_{k+1}}\|_{L^{2}(\Omega)}^{2},\\
\bar \lambda^{k+1}_{h_{k+1}}&=I_{h_{k+1}}\bar \lambda^{k}_{h_{k}}+\sigma(\bar u^{k+1}_{h_{k+1}}-\bar z^{k+1}_{h_{k+1}}).
\end{aligned} \right.$$
\[lem:continuous exact and discretized exact\] Let the initial point be $(u^{0},z^{0};\lambda^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, then $\|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=\|\bar{z}^{k+1}-\bar{z}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=\|\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}=O(h_{k+1})$, $\forall k\geqslant 1,$ and $$\sum_{k=0}^{\infty}\|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=O\left(\sum_{k=0}^{\infty}h_{k+1}\right).$$
We employ the mathematical induction to prove the conclusion. While $k=1$, with the definition of $\lambda_{h_{k+1}}^{k}:=I_{h_{k+1}}\lambda_{h_{k}}^{k}$ and the interpolation error estimate in Lemma \[lem:interpolation error estimate\], we have $$\|\bar{\lambda}^{0}-\bar{\lambda}^{0}_{h_{1}}\|_{L^{2}(\Omega)}=\|\lambda^{0}-I_{h_{1}}\lambda^{0}\|_{L^{2}(\Omega)}=O(h_{1}),$$ Then we can easily obtain that $\|\bar{u}^{1}-\bar{u}^{1}_{h_{1}}\|_{L^{2}(\Omega)}=O(h_{1})$. The proof is similar to the case $k>1$, here we omit it.
While $k>1$, we assume for $\forall j\leq k,$ we have $\|\bar{u}^{j}-\bar{u}^{j}_{h_{j}}\|_{L^{2}(\Omega)}=\|\bar{\lambda}^{j-1}-\bar{\lambda}^{j-1}_{h_{j}}\|_{L^{2}(\Omega)}=O(h_{j})$. Then for $z$-subproblems in exact ADMM in function space and the multi-level discretized ADMM, we know $\bar{z}^{k}$ and $\bar{z}^{k}_{h_{k}}$ satisfy the following optimality conditions: $$\bar{z}^{k}=\Pi_{[a,b]}\left(\bar{u}^{k}+\frac{\bar{\lambda}^{k-1}}{\sigma}\right),\
\bar{z}^{k}_{h_{k}}=\Pi_{[a,b]}\left(\bar{u}^{k}_{h_{k}}+\frac{\bar{\lambda}^{k-1}_{h_{k}}}{\sigma}\right).$$ Then subtracting the above two equalities we obtain $$\begin{aligned}
\|\bar{z}^{k}-\bar{z}^{k}_{h_{k}}\|_{L^{2}(\Omega)}
&=\left\|\Pi_{[a,b]}\left(\bar{u}^{k}+\dfrac{\bar{\lambda}^{k-1}}{\sigma}\right)-\Pi_{[a,b]}\left(\bar{u}^{k}_{h_{k}}+\dfrac{\bar{\lambda}^{k-1}_{h_{k}}}{\sigma}\right)\right\|_{L^{2}(\Omega)}
\\
&\leq \left\|\bar{u}^{k}+\dfrac{\bar{\lambda}^{k-1}}{\sigma}-\bar{u}^{k}_{h_{k}}-\dfrac{\bar{\lambda}^{k-1}_{h_{k}}}{\sigma}\right\|_{L^{2}(\Omega)}\\
&\leq \|\bar{u}^{k}-\bar{u}^{k}_{h_{k}}\|_{L^{2}(\Omega)}+\dfrac{1}{\sigma}\|\bar{\lambda}^{k-1}-\bar{\lambda}^{k-1}_{h_{k}}\|_{L^{2}(\Omega)}\\
&=O(h_{k}).
\end{aligned}$$ For the multiplier $\bar{\lambda}^{k}$ and $\bar{\lambda}^{k}_{h_{k}}$, $$\bar{\lambda}^{k}=\bar{\lambda}^{k-1}+\sigma(\bar{u}^{k}-\bar{z}^{k}), \
\bar{\lambda}^{k}_{h_{k}}=\bar{\lambda}^{k-1}_{h_{k}}+\sigma(\bar{u}^{k}_{h_{k}}-\bar{z}^{k}_{h_{k}}),$$ we can get the estimate $$\label{equ:lambda}
\begin{aligned}
\|\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k}}\|_{L^{2}(\Omega)}
&=\|\bar{\lambda}^{k-1}-\bar{\lambda}^{k-1}_{h_{k}}+\sigma(\bar{u}^{k}-\bar{u}^{k}_{h_{k}})-\sigma(\bar{z}^{k}-\bar{z}^{k}_{h_{k}})\|_{L^{2}(\Omega)}\\
&\leq \|\bar{\lambda}^{k-1}-\bar{\lambda}^{k-1}_{h_{k}}\|_{L^{2}(\Omega)}+ \sigma\|\bar{u}^{k}-\bar{u}^{k}_{h_{k}}\|_{L^{2}(\Omega)}+\sigma\|\bar{z}^{k}-\bar{z}^{k}_{h_{k}}\|_{L^{2}(\Omega)}\\
&=O(h_{k}).
\end{aligned}$$ For $u$-subproblems, $\bar{u}^{k+1}$ and $\bar{u}^{k+1}_{h_{k+1}}$ satisfy the following optimality conditions respectively, $$\begin{aligned}
&S^{\ast}[S(\bar{u}^{k}+y_{r})-y_{d}]+\alpha\bar{u}^{k+1}+\bar{\lambda}^{k}+\sigma(\bar{u}^{k+1}-\bar{z}^{k})=0,\\
&S^{\ast}_{h_{k+1}}[S_{h_{k+1}}(\bar{u}_{h_{k+1}}^{k+1}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}]+\alpha\bar{u}^{k+1}_{h_{k+1}}+\bar{\lambda}^{k}_{h_{k+1}}+\sigma(\bar{u}^{k+1}_{h_{k+1}}-\bar{z}^{k}_{h_{k+1}})=0.\end{aligned}$$ Then we know from the above two equalities that $$\begin{aligned}
0=&
S^{\ast}S\bar{u}^{k+1}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}_{h_{k+1}}+S^{\ast}Sy_{r}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}I_{h_{k+1}}y_{r}\\&-S^{\ast}y_{d}+S^{\ast}_{h_{k+1}}I_{h_{k+1}}y_{d}+(\alpha+\sigma)(\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}})+\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}}-\sigma(\bar{z}^{k}-\bar{z}^{k}_{h_{k+1}}),
\end{aligned}$$ so $$\begin{aligned}
\label{equality}
(\alpha+\sigma)(\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}})=&
-(S^{\ast}S\bar{u}^{k+1}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}_{h_{k+1}}+S^{\ast}Sy_{r}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}I_{h_{k+1}}y_{r}
\\&-S^{\ast}y_{d}+S^{\ast}_{h_{k+1}}I_{h_{k+1}}y_{d}+\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}}-\sigma(\bar{z}^{k}-\bar{z}^{k}_{h_{k+1}}))\\
=&-(E_{1}+E_{2}+E_{3}+E_{4}+E_{5}),
\end{aligned}$$ where we define $$\begin{aligned}
E_{1}&:=S^{\ast}S\bar{u}^{k+1}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}_{h_{k+1}},\\
E_{2}&:=S^{\ast}Sy_{r}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}I_{h_{k+1}}y_{r},\\
E_{3}&:=-S^{\ast}y_{d}+S^{\ast}_{h_{k+1}}I_{h_{k+1}}y_{d},\\
E_{4}&:=\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}},\\
E_{5}&:=-\sigma(\bar{z}^{k}-\bar{z}^{k}_{h_{k+1}}).
\end{aligned}$$
For the term $E_{1}$, we make use of the decomposition, $$\begin{aligned}
\|E_{1}\|_{L^{2}(\Omega)}&=\|S^{\ast}S\bar{u}^{k+1}-S^{\ast}S_{h_{k+1}}\bar{u}^{k+1}+S^{\ast}S_{h_{k+1}}\bar{u}^{k+1}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}+S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}\\
&\leq\|S^{\ast}(S-S_{h_{k+1}})\bar{u}^{k+1}\|_{L^{2}(\Omega)}+\|(S^{\ast}-S^{\ast}_{h_{k+1}})S_{h_{k+1}}\bar{u}^{k+1}\|_{L^{2}(\Omega)}+\|S^{\ast}_{h_{k+1}}S_{h_{k+1}}(\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}})\|_{L^{2}(\Omega)}.
\end{aligned}$$ From the well known error estimate $\|S-S_{h}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}=O(h^{2})$ in lemma \[lem:error estimates\] and the property that $S^{\ast},S_{h_{k+1}}$ are bounded linear operators, we have $$\begin{aligned}
&\|S^{\ast}(S-S_{h_{k+1}})\bar{u}^{k+1}\|_{L^{2}(\Omega)}\leq
\|S^{\ast}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|S-S_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|\bar{u}^{k+1}\|_{L^{2}(\Omega)}=O(h_{k+1}^{2}),\\
&\|(S^{\ast}-S^{\ast}_{h_{k+1}})S_{h_{k+1}}\bar{u}^{k+1}\|_{L^{2}(\Omega)}\leq\|S^{\ast}-S^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|S_{h_{k+1}}\bar{u}^{k+1}\|_{L^{2}(\Omega)}=O(h_{k+1}^{2}).\end{aligned}$$ Hence, there exists a constant $\hat C$ such that $$\|E_{1}\|_{L^{2}(\Omega)}\leq \hat Ch_{k+1}^{2}+\|S^{\ast}_{h_{k+1}}S_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}.$$
Similarly, based on the property of the projection operator $\|y_{r}-I_{h_{k+1}}y_{r}\|_{L^{2}(\Omega)}=\|y_{d}-I_{h_{k+1}}y_{d}\|_{L^{2}(\Omega)}=O(h_{k+1})$ in Lemma \[lem:interpolation error estimate\], for the term $E_{2}$, we have $$\begin{aligned}
\|E_{2}\|_{L^{2}(\Omega)}=&\|S^{\ast}Sy_{r}-S^{\ast}_{h_{k+1}}Sy_{r}+S^{\ast}_{h_{k+1}}Sy_{r}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}y_{r}+S^{\ast}_{h_{k+1}}S_{h_{k+1}}y_{r}-S^{\ast}_{h_{k+1}}S_{h_{k+1}}I_{h_{k+1}}y_{r}\|_{L^{2}(\Omega)}\\
\leq& \|(S^{\ast}-S^{\ast}_{h_{k+1}})Sy_{r}\|_{L^{2}(\Omega)}+\|S^{\ast}_{h_{k+1}}(S-S_{h_{k+1}})y_{r}\|_{L^{2}(\Omega)}+\|S^{\ast}_{h_{k+1}}S_{h_{k+1}}(y_{r}-I_{h_{k+1}}y_{r})\|_{L^{2}(\Omega)}\\
\leq & \|S^{\ast}-S^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|Sy_{r}\|_{L^{2}(\Omega)}+\|S^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|S-S_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|y_{r}\|_{L^{2}(\Omega)}\\
&+\|S^{\ast}_{h_{k+1}}S_{h_{k+1}}\|_{L^{2}(\Omega)\rightarrow L^{2}(\Omega)}\|y_{r}-I_{h_{k+1}}y_{r}\|_{L^{2}(\Omega)}\\
=&O(h_{k+1}),
\end{aligned}$$ and $$\begin{aligned}
\|E_{3}\|_{L^{2}(\Omega)}&=\|-S^{\ast}y_{d}+S^{\ast}I_{h_{k+1}}y_{d}-S^{\ast}I_{h_{k+1}}y_{d}+S^{\ast}_{h_{k+1}}I_{h_{k+1}}y_{d}\|_{L^{2}(\Omega)}\\
&\leq \|S^{\ast}(y_{d}-I_{h_{k+1}}y_{d})\|_{L^{2}(\Omega)}+\|(S^{\ast}-S^{\ast}_{h_{k+1}})I_{h_{k+1}}y_{d}\|_{L^{2}(\Omega)}\\
&=O(h_{k+1}).
\end{aligned}$$
For the term $E_{4}$, $$\begin{aligned}
\|E_{4}\|_{L^{2}(\Omega)}
&=\|\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k}}+\bar{\lambda}^{k}_{h_{k}}-I_{h_{k+1}}\bar{\lambda}^{k}_{h_{k}}\|_{L^{2}(\Omega)}\\
&\leq \|\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k}}\|_{L^{2}(\Omega)}+\|\bar{\lambda}^{k}_{h_{k}}-I_{h_{k+1}}\bar{\lambda}^{k}_{h_{k}}\|_{L^{2}(\Omega)}\\
&=O(h_{k}+h_{k+1}),
\end{aligned}$$ where we used (\[equ:lambda\]), the property of the projection operator $\|\bar{\lambda}^{k}_{h_{k}}-I_{h_{k+1}}\bar{\lambda}^{k}_{h_{k}}\|_{L^{2}(\Omega)}=O(h_{k+1})$ and the property of the mesh size $h_{k}>h_{k+1}$. Moreover, as the mesh sizes satisfy $\sum_{k=0}^{\infty}h_{k+1}< \infty,$ there exists a constant $C_{k+1}$ such that $h_{k}<C_{k+1}h_{k+1}$, then we have $$\|E_{4}\|_{L^{2}(\Omega)}=\|\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}=O(h_{k+1}).$$
Similarly, $$\begin{aligned}
\|E_{5}\|_{L^{2}(\Omega)}
&=\sigma\|\bar{z}^{k}-\bar{z}^{k}_{h_{k}}+\bar{z}^{k}_{h_{k}}-\bar{z}^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}\\
&\leq \sigma\|\bar{z}^{k}-\bar{z}^{k}_{h_{k}}\|_{L^{2}(\Omega)}+\sigma\|\bar{z}^{k}_{h_{k}}-\bar{z}^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}\\
&=O(h_{k+1}).
\end{aligned}$$
Then with the fact that operators $S^{\ast}_{h_{k+1}}, S_{h_{k+1}}$ are bounded linear operators, we know from the equality (\[equality\]) and the estimations of $L^{2}$ norms of $\lbrace E_{i}\rbrace_{i=1}^{5}$ above that $$\|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=O(h_{k+1}).$$ Moreover, we have $$\begin{aligned}
\|\bar{z}^{k+1}-\bar{z}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}
&=\left\|\Pi_{[a,b]}\left(\bar{u}^{k+1}+\dfrac{\bar{\lambda}^{k}}{\sigma}\right)-\Pi_{[a,b]}\left(\bar{u}^{k+1}_{h_{k+1}}+\dfrac{\bar{\lambda}^{k}_{h_{k+1}}}{\sigma}\right)\right\|_{L^{2}(\Omega)}
\\
&\leq \|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}+\dfrac{1}{\sigma}\|(\bar{\lambda}^{k}-\bar{\lambda}^{k}_{h_{k+1}})\|_{L^{2}(\Omega)}\\
&=O(h_{k+1}).
\end{aligned}$$ Hence the conclusion holds for the case $k+1$ and we can get the assertion $$\sum_{k=0}^{\infty}\|\bar{u}^{k+1}-\bar{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=O\left(\sum_{k=0}^{\infty}h_{k+1}\right).$$
Similar to the Lemma \[lem:continuous exact and discretized exact\], we have the following lemma.
\[lem:continuous inexact and discretized inexact\] Let the initial point be $(u^{0},z^{0};\lambda^{0})\in L^{2}(\Omega)\times {\rm dom}(\delta_{U_{ad}}(\cdot))\times L^{2}(\Omega)$, then $\|u^{k+1}-{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=\|z^{k+1}-{z}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=\|{\lambda}^{k}-{\lambda}^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}=O(h_{k+1}+\delta_{u,h_{k+1}}^{k+1})$, $\forall k\geqslant 1,$ and $$\sum_{k=0}^{\infty}\|{u}^{k+1}-{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}=O\left(\sum_{k=0}^{\infty}\left(h_{k+1}+\delta_{u,h_{k+1}}^{k+1}\right)\right).$$
We employ the mathematical induction to prove the conclusion. The proof is similar to Lemma \[lem:continuous exact and discretized exact\], here we do not talk about it in detail.
To prove the convergence of the mADMM algorithm, let $(\tilde{u}_{h_{k+1}}^{k+1},\tilde{z}^{k+1}_{h_{k+1}})$ represents the exact solutions of the $(k+1)$th iteration of Algorithm \[alg:multi-level ADMM algorithm\]: $$\begin{aligned}
\label{tildeu}
&\tilde{u}_{h_{k+1}}^{k+1}:={\rm argmin}{\hat J}(u_{h_{k+1}})+\langle\lambda_{h_{k+1}}^{k},u_{h_{k+1}}-z_{h_{k+1}}^{k}\rangle+\dfrac{\sigma}{2}\|u_{h_{k+1}}-z_{h_{k+1}}^{k}\|_{L^{2}(\Omega)}^{2},\\
\label{tildez}
&\tilde{z}^{k+1}_{h_{k+1}}:={\rm argmin}\delta_{U_{ad,h}}(z_{h_{k+1}})+\langle{\lambda}^{k}_{h_{k+1}},\tilde{u}^{k+1}_{h_{k+1}}-z_{h_{k+1}}\rangle+\dfrac{\sigma}{2}\|\tilde{u}^{k+1}_{h_{k+1}}-z_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}.\end{aligned}$$ The following lemma gives the gap between $(\tilde{u}_{h_{k+1}}^{k+1}, \tilde{z}_{h_{k+1}}^{k+1})$ and $({u}_{h_{k+1}}^{k+1}, {z}_{h_{k+1}}^{k+1})$.
([@Song2017Fe], Lemma 4.4) \[lmm:tildeu\] For any $k\geqslant 0$, we have $$\begin{aligned}
&\|\tilde{u}_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq\rho\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)},\\
&\|\tilde{z}_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq\rho\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)},\end{aligned}$$ where $\rho:=\|[S^{\ast}_{h_{k+1}}S_{h_{k+1}}+(\alpha+\sigma)I]^{-1}\|_{{L^{2}(\Omega)}\rightarrow {L^{2}(\Omega)}}$.
For the convenience of analyzing the non-ergodic iteration complexity, let $(u^{\ast}_{h_{k+1}}, z^{\ast}_{h_{k+1}}, \lambda^{\ast}_{h_{k+1}})$ denotes the KKT point of the discretized reduced problem with the mesh size $h_{k+1}$ $$\label{discretized reduced problem}
\begin{aligned}
\min \limits_{u_{h_{k+1}},z_{h_{k+1}}}
&\frac{1}{2}\|S_{h_{k+1}}(u_{h_{k+1}}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_d\|_{L^{2}(\Omega_{h})}^{2}
+\frac{\alpha}{2}\|u_{h_{k+1}}\|_{{L^{2}(\Omega_{h})}}^{2}+\delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}})\\
{\rm s.t.} \ \ \ \ \ & u_{h_{k+1}}=z_{h_{k+1}}.
\end{aligned} \tag{DRP}$$ Moreover, we provide a lemma and two propositions which are essential for analyzing the iteration complexity of our mADMM.
([@Chen2017], Lemma 6.1) \[lmm:complexity1\] If a sequence $\left\{a_{i}\right\} \in \mathbb{R}$ satisfies the following conditions: $$a_{i} \geq 0 \ for \ any \ i \geq 0 \ \ and \ \ \sum_{i=0}^{\infty} a_{i}=\bar a<\infty .$$ Then we have $\min_{i=1,2, \cdots, k}\left\{a_{i}\right\} \leq \frac{\overline{a}}{k}$ and $\lim_{k \rightarrow \infty}\left\{k \cdot \min _{i=1,2, \cdots, k}\left\{a_{i}\right\}\right\}=0$.
\[lmm:complexity2\] Let $\left\{\left(u^{k}_{h_{k}}, z^{k}_{h_{k}}, \lambda^{k}_{h_{k}}\right)\right\}$ be the sequence generated by Algorithm \[alg:multi-level ADMM algorithm\] and $\left\{\left(u^{\ast}_{h_{k}}, z^{\ast}_{h_{k}}, \lambda^{\ast}_{h_{k}}\right)\right\}$ denotes the KKT point of the discretized reduced problem. Then for $k\geq 0$ we have $$\begin{aligned}
&\left\langle\delta^{k+1}_{u,h_{k+1}}, u^{k+1}_{h_{k+1}}-u^{\ast}_{h_{k+1}}\right\rangle+\frac{1}{2\tau\sigma}\|\lambda^{\ast}_{h_{k+1}}-\lambda^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{\sigma}{2}\|z^{\ast}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\\
&-\frac{1}{2\tau\sigma}\|\lambda^{\ast}_{h_{k+1}}-\lambda^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}-\frac{\sigma}{2}\|z^{\ast}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\\
\geq &\alpha\|u^{k+1}_{h_{k+1}}-u^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{(3-\tau)\sigma}{2}\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{\sigma}{2}\|u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}.\end{aligned}$$
First, for any $f_{1}, f_{1}^{\prime}, f_{2}, f_{2}^{\prime}\in L^{2}(\Omega)$, we have the following two important equalities hold $$\begin{aligned}
\label{equ1}
&\langle f_{1}, f_{2}\rangle_{L^{2}(\Omega)}=\frac{1}{2}\left(\|f_{1}\|_{L^{2}(\Omega)}^{2}+\|f_{2}\|_{L^{2}(\Omega)}^{2}-\|f_{1}-f_{2}\|_{L^{2}(\Omega)}^{2}\right)
=\frac{1}{2}\left(\|f_{1}+f_{2}\|_{L^{2}(\Omega)}^{2}-\|f_{1}\|_{L^{2}(\Omega)}^{2}-\|f_{2}\|_{L^{2}(\Omega)}^{2}\right),\\
\label{equ2}
&\left\langle f_{1}-f_{1}^{\prime}, f_{2}-f_{2}^{\prime}\right\rangle_{L^{2}(\Omega)}=\frac{1}{2}\left(\|f_{1}+f_{2}\|_{L^{2}(\Omega)}^{2}+\left\|f_{1}^{\prime}+f_{2}^{\prime}\right\|_{L^{2}(\Omega)}^{2}-\left\|f_{1}+f_{2}^{\prime}\right\|_{L^{2}(\Omega)}^{2}-\left\|f_{1}^{\prime}+f_{2}\right\|_{L^{2}(\Omega)}^{2}\right).\end{aligned}$$ The proof of the above two equalities can be easily obtained by the definition of $L^{2}-$norm.
By the optimality conditions of the $u$-subproblem and $z$-subproblem corresponding to $u_{h_{k+1}}^{k+1}$ and $z_{h_{k+1}}^{k+1}$, we have $$\begin{aligned}
\label{optimality1}
&\nabla {\hat J_{h_{k+1}}}(u_{h_{k+1}}^{k+1})=S^{\ast}_{h_{k+1}}[S_{h_{k+1}}(u_{h_{k+1}}^{k+1}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}]+\alpha u^{k+1}_{h_{k+1}}=\delta_{u,h_{k+1}}^{k+1}-(\lambda^{k}_{h_{k+1}}+\sigma(u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}})),\\
&\lambda^{k}_{h_{k+1}}+\sigma(u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}))\in \partial
\delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}}^{k+1}).\end{aligned}$$ Moreover, $\left\{\left(u^{\ast}_{h_{k}}, z^{\ast}_{h_{k}}, \lambda^{\ast}_{h_{k}}\right)\right\}$ denotes the KKT point of the discretized reduced problem, so it satisfies the following KKT system $$\begin{aligned}
\label{optimality2}
&\nabla {\hat J_{h_{k+1}}}(u_{h_{k+1}}^{\ast})=S^{\ast}_{h_{k+1}}[S_{h_{k+1}}(u_{h_{k+1}}^{\ast}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}]+\alpha u^{\ast}_{h_{k+1}}=-\lambda^{\ast}_{h_{k+1}},\\
&\lambda^{\ast}_{h_{k+1}}\in \partial
\delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}}^{\ast}),\\
&u^{\ast}_{h_{k+1}}=z^{\ast}_{h_{k+1}}.\end{aligned}$$ Then by combining (\[optimality1\]) and (\[optimality2\]), we obtain $$\begin{aligned}
&\left\langle \nabla {\hat J_{h_{k+1}}}(u_{h_{k+1}}^{k+1})-\nabla {\hat J_{h_{k+1}}}(u_{h_{k+1}}^{\ast}), u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast} \right\rangle\\
=&\left\langle S^{\ast}_{h_{k+1}}S_{h_{k+1}}(u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast})+\alpha(u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}), u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast} \right\rangle\\
=&\|S_{h_{k+1}}(u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast})\|_{L^{2}(\Omega)}^{2}+\alpha\|u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}\|_{L^{2}(\Omega)}^{2}.
\end{aligned}$$ Moreover, the subdifferential operator $\partial
\delta_{U_{ad,h_{k+1}}}(z)$ is a maximal monotone operator, so the following inequality holds, $$\left\langle \partial
\delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}}^{k+1})-\partial
\delta_{U_{ad,h_{k+1}}}(z_{h_{k+1}}^{\ast}), z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{\ast} \right\rangle\geq 0.$$ For the convenience of analyzing, we define $r_{h_{k+1}}^{k+1}=u_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k+1}$. Therefore, we can derive that $$\begin{aligned}
&\left\langle \delta_{u,h_{k+1}}^{k+1}-(\tilde \lambda_{h_{k+1}}^{k}+\sigma(z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k}))+\lambda_{h_{k+1}}^{\ast}, u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}\right\rangle\geq \alpha\|u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}\|_{L^{2}(\Omega)}^{2},\\
&\left\langle \lambda_{h_{k+1}}^{k}+\sigma r_{h_{k+1}}^{k+1}, z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{\ast} \right\rangle\geq 0.\end{aligned}$$ Then adding the above two equalities we obtain $$\label{estimate}
\left\langle \delta_{u,h_{k+1}}^{k+1}, u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}\right\rangle-\left\langle \lambda_{h_{k+1}}^{k}+\sigma r_{h_{k+1}}^{k+1}-\lambda_{h_{k+1}}^{\ast}, r_{h_{k+1}}^{k+1}\right\rangle-\sigma\left\langle z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k}, u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast} \right\rangle \geq \alpha\|u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{\ast}\|_{L^{2}(\Omega)}^{2}.$$ Next, we estimate the last two terms on the left side separately, $$\begin{aligned}
\label{estimate1}
&\left\langle \lambda_{h_{k+1}}^{\ast}-(\lambda_{h_{k+1}}^{k}+\sigma r_{h_{k+1}}^{k+1}), r_{h_{k+1}}^{k+1}\right\rangle
\\
=&\frac{1}{\tau\sigma}\left\langle \lambda_{h_{k+1}}^{\ast}-\lambda_{h_{k+1}}^{k}, \lambda_{h_{k+1}}^{k+1}-\lambda_{h_{k+1}}^{k} \right\rangle-\sigma\|r_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}\\
=&\frac{1}{2\tau\sigma}(\|\lambda_{h_{k+1}}^{\ast}-\lambda_{h_{k+1}}^{k}\|_{L^{2}(\Omega)}^{2}+\|\lambda_{h_{k+1}}^{k+1}-\lambda_{h_{k+1}}^{k}\|_{L^{2}(\Omega)}^{2}-\|\lambda_{h_{k+1}}^{\ast}-\lambda_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2})-\sigma\|r_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}\\
=&\frac{1}{2\tau\sigma}(\|\lambda_{h_{k+1}}^{\ast}-\lambda_{h_{k+1}}^{k}\|_{L^{2}(\Omega)}^{2}-\|\lambda_{h_{k+1}}^{\ast}-\lambda_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2})+\frac{(\tau-2)\sigma}{2}\|r_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2},
\end{aligned}$$ where we used the equality (\[equ1\]).
Moreover, by employing the equality (\[equ2\]) and using $u^{\ast}_{h_{k+1}}=z^{\ast}_{h_{k+1}}$, we have $$\begin{aligned}
\label{estimate2}
&\sigma\left\langle z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k}, u_{h_{k+1}}^{\ast}-u_{h_{k+1}}^{k+1} \right\rangle \\
=&\sigma\left\langle z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{k}, -u_{h_{k+1}}^{k+1}-(-z_{h_{k+1}}^{\ast}) \right\rangle \\
=&\frac{\sigma}{2}(\|r_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}+\|z_{h_{k+1}}^{k}-z_{h_{k+1}}^{\ast}\|_{L^{2}(\Omega)}^{2}-\|z_{h_{k+1}}^{k+1}-z_{h_{k+1}}^{\ast}\|_{L^{2}(\Omega)}^{2}
-\|z_{h_{k+1}}^{k}-u_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}).
\end{aligned}$$ Then, substituting (\[estimate1\]), (\[estimate2\]) into (\[estimate\]), we can get the assertion of Proposition \[lmm:complexity2\].
\[lmm:complexity3\] Let $\left\{\left(u^{k}_{h_{k}}, z^{k}_{h_{k}}, \lambda^{k}_{h_{k}}\right)\right\}$ be the sequence generated by Algorithm \[alg:multi-level ADMM algorithm\], $\left\{\left(u^{\ast}_{h_{k}}, z^{\ast}_{h_{k}}, \lambda^{\ast}_{h_{k}}\right)\right\}$ denotes the KKT point of the discretized reduced problem and $\tilde{u}_{h_{k+1}}^{k+1}$, $\tilde{z}^{k+1}_{h_{k+1}}$ defined in (\[tildeu\]), (\[tildez\]), respectively. Then for $k\geq 0$ we have $$\begin{aligned}
&\frac{1}{2\tau\sigma}\|\lambda^{\ast}_{h_{k+1}}-\lambda^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{\sigma}{2}\|z^{\ast}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}-\frac{1}{2\tau\sigma}\|\lambda^{\ast}_{h_{k+1}}-\tilde \lambda^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}-\frac{\sigma}{2}\|z^{\ast}_{h_{k+1}}-\tilde z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\\
\geq & \alpha\|\tilde u^{k+1}_{h_{k+1}}-u^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{(3-\tau)\sigma}{2}\|\tilde u^{k+1}_{h_{k+1}}-\tilde z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{\sigma}{2}\|\tilde u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}.\end{aligned}$$
For the proof of Proposition \[lmm:complexity3\], by substituting $\tilde u^{k+1}_{h_{k+1}}$ and $\tilde z^{k+1}_{h_{k+1}}$ for $u^{k+1}_{h_{k+1}}$ and $z^{k+1}_{h_{k+1}}$ in the proof of Proposition \[lmm:complexity2\], we can get the assertion.
Finally, based on the above results, the convergence results of Algorithm \[alg:multi-level ADMM algorithm\] is given by the following theorem.
\[convergence theorem\] Suppose that the operator $L$ is uniformly elliptic. Let $(y^{\ast},u^{\ast},z^{\ast},p^{\ast},\lambda^{\ast})$ be the KKT point of [(\[eqn:orginal problems\])]{}, $(u^{k}_{h_{k}},z^{k}_{h_{k}},\lambda^{k}_{h_{k}})$ is obtained in the $k$th iterate of Algorithm \[alg:multi-level ADMM algorithm\], where we suppose the mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $\sum_{k=0}^{\infty}h_{k+1}<\infty,$ and the error vector $\delta_{u,h_{k+1}}^{k+1}$ satisfies $\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq \xi_{k+1}$, $\sum_{k=0}^{\infty}\xi_{k+1}<\infty.$ Then we have $$\begin{aligned}
&\lim \limits_{k\rightarrow \infty}\lbrace \|u^{k}_{h_{k}}-u^{\ast}\|_{L^{2}(\Omega)}+\|z^{k}_{h_{k}}-z^{\ast}\|_{L^{2}(\Omega}+\|\lambda^{k}_{h_{k}}-\lambda^{\ast}\|_{L^{2}(\Omega)}\rbrace=0,\\
&\lim \limits_{k\rightarrow \infty}\lbrace \|y^{k}_{h_{k}}-y^{\ast}\|_{H_{0}^{1}(\Omega)}+\|p^{k}_{h_{k}}-p^{\ast}\|_{H_{0}^{1}(\Omega)}\rbrace=0.\end{aligned}$$ Moreover, there exists a constant $\tilde C$ only depends on the initial point $(u^{0}, z^{0}, \lambda^{0})$ and the optimal solution $(u^{\ast}, z^{\ast}, \lambda^{\ast})$ such that for $k\geq 1$, $$\min_{1\leq i\leq k}{R_{h_{i}}}(u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}})\leq \frac{\tilde C}{k}, \ \ \lim \limits_{k\rightarrow \infty}(k\times \min_{1\leq i\leq k}{R_{h_{i}}}(u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}}))=0,$$ where $R_{h_{i}}: (u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}})\rightarrow [0,\infty)$ is defined as $$R_{h_{i}}(u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}}):=\|\nabla {\hat J_{h_{i}}}(u^{i}_{h_{i}})+\lambda_{h_{i}}^{i-1}\|^{2}_{L^{2}(\Omega)}+{\rm dist}^{2}(0,-\lambda_{h_{i}}^{i-1}+\partial \delta_{U_{ad,h_{i}}}(z^{i}_{h_{i}}))+\|u^{i}_{h_{i}}-z^{i}_{h_{i}}\|^{2}_{L^{2}(\Omega)}.$$
By the optimality condition of the $u$-subproblem in the ADMM in function space, we have $$\begin{aligned}
S^{\ast}[S(\bar{u}^{k+1}+y_{r})-y_{d}]+\alpha\bar{u}^{k+1}+\bar \lambda^{k}+\sigma(\bar{u}^{k+1}-\bar z^{k})=0.\end{aligned}$$ As we know, the error between the inexact solution and exact solution contains two parts, error from gradually refining the grid and error from the inexactly solving the subproblems. We take them into consideration together as a total error, let $u_{h_{k+1}}^{k+1}$ represents the inexact solution of the $(k+1)$th iteration, then from the optimality condition of the $u$-subproblem, we have $$S^{\ast}[S(u_{h_{k+1}}^{k+1}+y_{r})-y_{d}]+\alpha u_{h_{k+1}}^{k+1}+\lambda^{k}+\sigma(u_{h_{k+1}}^{k+1}-z^{k})=\delta_{u}^{k+1}.$$ Moreover, by the optimality conditions of the $u-$subproblem corresponding to $u_{h_{k+1}}^{k+1}$ and $\bar u_{h_{k+1}}^{k+1}$ in Algorithm \[alg:multi-level ADMM algorithm\] and multi-level discretized ADMM, we have $$\begin{aligned}
&S^{\ast}_{h_{k+1}}[S_{h_{k+1}}(u_{h_{k+1}}^{k+1}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}]+\alpha u_{h_{k+1}}^{k+1}+\lambda^{k}_{h_{k+1}}+\sigma(u_{h_{k+1}}^{k+1}-z^{k}_{h_{k+1}})=\delta_{u,h_{k+1}}^{k+1},\\
&S^{\ast}_{h_{k+1}}[S_{h_{k+1}}(\bar u_{h_{k+1}}^{k+1}+I_{h_{k+1}}y_{r})-I_{h_{k+1}}y_{d}]+\alpha \bar u_{h_{k+1}}^{k+1}+\bar \lambda^{k}_{h_{k+1}}+\sigma(\bar u_{h_{k+1}}^{k+1}-\bar z^{k}_{h_{k+1}})=0.\end{aligned}$$
Then we know from the four equalities above that $$\begin{aligned}
\delta_{u}^{k+1}=&\delta_{u}^{k+1}-\delta_{u,h_{k+1}}^{k+1}+\delta_{u,h_{k+1}}^{k+1}\\
=&\delta_{u,h_{k+1}}^{k+1}+S^{\ast}S(u_{h_{k+1}}^{k+1}-\bar u^{k+1})+(\alpha+\sigma)(u_{h_{k+1}}^{k+1}-\bar u^{k+1})+(\lambda^{k}-\bar \lambda^{k})-\sigma(z^{k}-\bar z^{k})\\
&+S^{\ast}_{h_{k+1}}S_{h_{k+1}}(\bar u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{k+1})+(\alpha+\sigma)(\bar u_{h_{k+1}}^{k+1}-u_{h_{k+1}}^{k+1})+(\bar \lambda^{k}_{h_{k+1}}-\lambda^{k}_{h_{k+1}})-\sigma(\bar z^{k}_{h_{k+1}}-z^{k}_{h_{k+1}})\\
=&\delta_{u,h_{k+1}}^{k+1}+(\alpha+\sigma)(\bar u_{h_{k+1}}^{k+1}-\bar u^{k+1})+
(\lambda^{k}-\lambda^{k}_{h_{k+1}})-\sigma(z^{k}-z^{k}_{h_{k+1}})+(\bar \lambda^{k}_{h_{k+1}}-\bar \lambda^{k})-\sigma(\bar z^{k}_{h_{k+1}}-\bar z^{k})\\
&+(S^{\ast}S-S^{\ast}_{h_{k+1}}S_{h_{k+1}})(u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1})+S^{\ast}S(\bar u_{h_{k+1}}^{k+1}-\bar u^{k+1}).
\end{aligned}$$ Moreover, we have the estimate $$\begin{aligned}
\label{equ:estimate}
&\|(S^{\ast}S-S^{\ast}_{h_{k+1}}S_{h_{k+1}})(u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1})\|_{L^{2}(\Omega)}\\
=&\|(S^{\ast}S-S^{\ast}S_{h_{k+1}})(u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1})\|_{L^{2}(\Omega)}
+\|(S^{\ast}S_{h_{k+1}}-S^{\ast}_{h_{k+1}}S_{h_{k+1}})(u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1})\|_{L^{2}(\Omega)}\\
\leq &\|S^{\ast}\|\|S-S_{h_{k+1}}\|\|u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}+\|S^{\ast}-S^{\ast}_{h_{k+1}}\|\|S_{h_{k+1}}\|\|u_{h_{k+1}}^{k+1}-\bar u_{h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\\
= &O(h_{k+1}^{2}).
\end{aligned}$$
Then we know from (\[equ:estimate\]), Lemma \[lem:continuous exact and discretized exact\] and Lemma \[lem:continuous inexact and discretized inexact\] that there exists constant C such that $$\label{equ: delta}
\begin{aligned}
\sum_{k=0}^{\infty}\|\delta_{u}^{k+1}\|_{L^{2}(\Omega)}
\leq \sum_{k=0}^{\infty}\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}+C\sum_{k=0}^{\infty}\left(h_{k+1}+\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\right).
\end{aligned}$$ We know from Algorithm \[alg:multi-level ADMM algorithm\] that the mesh sizes $\lbrace h_{k} \rbrace_{k=0}^{\infty}$ of each iteration satisfy $\sum_{k=0}^{\infty}h_{k+1}<\infty.$ The error vector of the multi-level ADMM satisfy $\sum_{k=0}^{\infty}\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq\sum_{k=0}^{\infty}\xi_{k+1}<\infty,$ where $\xi_{k+1}$ is the upper bound of $\delta_{u,h_{k+1}}^{k+1}$, i.e. $\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}\leq \xi_{k+1}$. Thus we have $$\sum_{k=0}^{\infty}\|\delta_{u}^{k+1}\|_{L^{2}(\Omega)}
< \infty.$$ For the discretization error and the iteration error, Algorithm \[alg:multi-level ADMM algorithm\] can be considered as an inexact ADMM algorithm in function space, then we know from Theorem \[convergence functional space\] that the convergence of Algorithm \[alg:multi-level ADMM algorithm\] is guaranteed.
At last, we establish the proof of the iteration complexity results for the sequence generated by the mADMM. First, by the optimality condition for $(u^{k+1}_{h_{k+1}},z^{k+1}_{h_{k+1}})$, we have $$\begin{aligned}
&\delta_{u,h_{k+1}}^{k+1}-[\lambda^{k}_{h_{k+1}}+\sigma(u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}})]=\nabla \hat J_{h_{k+1}}(u^{k+1}_{h_{k+1}}),\\
&\lambda^{k}_{h_{k+1}}+\sigma(u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}})\in \partial \delta_{U_{ad},h_{k+1}}(z^{k+1}_{h_{k+1}}).
\end{aligned}$$ Then by the definition of $R_{h_{k}}$, we derive $$\label{R inequality}
\begin{aligned}
R_{h_{k+1}}(u^{k+1}_{h_{k+1}},z^{k+1}_{h_{k+1}},\lambda^{k+1}_{h_{k+1}})=&\|\nabla \hat J_{h_{k+1}}(u^{k+1}_{h_{k+1}})+\lambda^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+{\rm dist}^{2}(0,-\lambda^{k}_{h_{k+1}}+\partial \delta_{U_{ad},h_{k+1}}(z^{k+1}_{h_{k+1}}))
\\&
+\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\\
\leq &\|\delta_{u,h_{k+1}}^{k+1}-\sigma(u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}})\|_{L^{2}(\Omega)}^{2}+(1+\sigma^{2})\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\\
\leq & 2\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}+2\sigma^{2}\|u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+(1+\sigma^{2})\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}.
\end{aligned}$$
Next, for the convenience of giving an upper bound of $R_{h_{k+1}}(u^{k+1}_{h_{k+1}},z^{k+1}_{h_{k+1}},\lambda^{k+1}_{h_{k+1}})$, we define the following sequence $\theta_{k}$, $\bar \theta_{k}$ and $\tilde \theta_{k}$ with: $$\begin{aligned}
&\theta^{k}=\left(\frac{1}{\sqrt{2 \tau \sigma}} \left( \lambda^{k}_{h_{k+1}}-\lambda^{\ast}_{h_{k+1}}\right), \sqrt{\frac{\sigma}{2}} \left(z^{k}_{h_{k+1}}-z^{\ast}_{h_{k+1}}\right)\right),\\
&\bar \theta^{k}=\left(\frac{1}{\sqrt{2 \tau \sigma}} \left( \lambda^{k}_{h_{k}}-\lambda^{\ast}_{h_{k}}\right), \sqrt{\frac{\sigma}{2}} \left(z^{k}_{h_{k}}-z^{\ast}_{h_{k}}\right)\right),\\
&\tilde \theta^{k}=\left(\frac{1}{\sqrt{2 \tau \sigma}} \left(\tilde \lambda^{k}_{h_{k}}-\lambda^{\ast}_{h_{k}}\right), \sqrt{\frac{\sigma}{2}} \left(\tilde z^{k}_{h_{k}}-z^{\ast}_{h_{k}}\right)\right).\end{aligned}$$ First, we give an upper bound of $\theta^{k}$. We know from Lemma \[lmm:complexity3\] that $\|\tilde \theta^{k+1}\|_{L^{2}(\Omega)}\leq \|\theta^{k}\|_{L^{2}(\Omega)},$ so $$\begin{aligned}
\|\theta^{k+1}\|_{L^{2}(\Omega)}&\leq \|\tilde \theta^{k+1}\|_{L^{2}(\Omega)}+\|\tilde \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)},\\
&\leq \|\theta^{k}\|_{L^{2}(\Omega)}+\|\tilde \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)}.
\end{aligned}$$ We know from the definition of $\tilde \theta^{k+1}$, $\theta^{k+1}$ and Lemma \[lmm:tildeu\] that $$\label{theta}
\begin{aligned}
\|\tilde \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)}=&\frac{1}{2 \tau \sigma}\left\|\tilde{\lambda}^{k+1}_{h_{k+1}}-\lambda^{k+1}_{h_{k+2}}\right\|_{L^{2}(\Omega)}+\frac{\sigma}{2}\left\|\tilde{z}^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+2}}\right\|_{L^{2}(\Omega)}\\
\leq& \frac{1}{2 \tau \sigma}\left\|\tilde{\lambda}^{k+1}_{h_{k+1}}-\lambda^{k+1}_{h_{k+1}}\right\|_{L^{2}(\Omega)}+\frac{1}{2 \tau \sigma}\left\|\lambda^{k+1}_{h_{k+1}}-I_{h_{k+2}}\lambda^{k+1}_{h_{k+1}}\right\|_{L^{2}(\Omega)}\\
&+\frac{\sigma}{2}\left\|\tilde{z}^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\right\|_{L^{2}(\Omega)}+\frac{\sigma}{2}\left\|z^{k+1}_{h_{k+1}}-I_{h_{k+2}}z^{k+1}_{h_{k+1}}\right\|_{L^{2}(\Omega)}\\
\leq& \frac{1}{2 \tau \sigma}\|\tilde{u}^{k+1}_{h_{k+1}}-{u}^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}+\frac{\sigma}{2}\left\|\tilde{z}^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\right\|_{L^{2}(\Omega)}+{\tilde C_{1}}h_{k+2}\\
\leq& (\frac{1}{\tau \sigma}+\frac{\sigma}{2})\rho\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}+{\tilde C_{1}}h_{k+2},
\end{aligned}$$ where $\tilde C_{1}$ is a constant. So there exists a constant $C'_{1}$ such that for every $k$ $$\begin{aligned}
\label{es1}
\|\theta^{k+1}\|_{L^{2}(\Omega)}&\leq \|\theta^{k}\|_{L^{2}(\Omega)}+\|\tilde \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)}\\
&\leq \|\theta^{k}\|_{L^{2}(\Omega)}+\left(\frac{1}{\tau \sigma}+\frac{\sigma}{2}\right)\rho\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}+{\tilde C_{1}}h_{k+2},\\
&\leq \|\theta^{k-1}\|_{L^{2}(\Omega)}+\left(\frac{1}{\tau \sigma}+\frac{\sigma}{2}\right)\rho\|\delta_{u,h_{k}}^{k}\|_{L^{2}(\Omega)}+{\tilde C_{1}}h_{k+1}+\left(\frac{1}{\tau \sigma}+\frac{\sigma}{2}\right)\rho\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}+{\tilde C_{1}}h_{k+2}\\
&\leq \cdots\\
&\leq\|\theta^{0}\|_{L^{2}(\Omega)}+\left(\frac{1}{\tau \sigma}+\frac{\sigma}{2}\right)\rho\sum_{i=0}^{k}\|\delta_{u,h_{i+1}}^{i+1}\|_{L^{2}(\Omega)}+{\tilde C_{1}}\sum_{i=0}^{k}h_{i+2}\\
&<C'_{1}.
\end{aligned}$$ By Lemma \[lmm:complexity2\], we have $$\label{es2}
\|\bar \theta^{k+1}\|_{L^{2}(\Omega)}^{2}\leq \|\theta^{k}\|_{L^{2}(\Omega)}^{2}+ \bar \eta \|\delta^{k}_{u,h_{k+1}}\|_{L^{2}(\Omega)}\leq (C'_{1})^{2}+\bar \eta \|\delta^{k+1}_{u,h_{k+1}}\|_{L^{2}(\Omega)},$$ so there exists a constant $C'_{2}$ such that $\|\bar \theta^{k+1}\|_{L^{2}(\Omega)}\leq C'_{2}.$ Hence, $$\left\|\theta^{k+1}+\bar \theta^{k+1}\right\|_{L^{2}(\Omega)}\leq \left\|\theta^{k+1}\right\|_{L^{2}(\Omega)}+\left\|\bar \theta^{k+1}\right\|_{L^{2}(\Omega)}<C'_{1}+C'_{2}.$$ By the definition of $\theta_{k}$ and $\bar \theta_{k}$, we have $$\| \theta^{k+1}-\bar \theta^{k+1}\|_{L^{2}(\Omega)}^{2}=
\|\left(\frac{1}{\sqrt{2 \tau \sigma}}\left(\lambda^{k+1}_{h_{k+1}}-\lambda^{k+1}_{h_{k+2}}\right), \sqrt\frac{\sigma}{2}\left(z^{k+1}_{h_{k+2}}-z^{k+1}_{h_{k+1}}\right)\right)\|_{L^{2}(\Omega)}^{2}=O(h_{k+2}^{2}),$$ thus $\| \theta^{k+1}-\bar \theta^{k+1}\|_{L^{2}(\Omega)}=O(h_{k+2}).$
Moreover, we have the estimate that $$\begin{aligned}
\left\langle\delta^{k+1}_{u,h_{k+1}}, u^{k+1}_{h_{k+1}}-u^{\ast}_{h_{k+1}}\right\rangle &\leq \left\langle\delta^{k+1}_{u,h_{k+1}}, u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}+z^{k+1}_{h_{k+1}}-z^{\ast}_{h_{k+1}}\right\rangle\\
&=\left\langle\delta^{k+1}_{u,h_{k+1}}, \frac{1}{\tau\sigma}(\lambda^{k+1}_{h_{k+1}}-\lambda^{\ast}_{h_{k+1}}+\lambda^{\ast}_{h_{k+1}}-\lambda^{k}_{h_{k+1}})\right\rangle-\left\langle\delta^{k+1}_{u,h_{k+1}}, z^{k+1}_{h_{k+1}}-z^{\ast}_{h_{k+1}}\right\rangle\\
&\leq \frac{1}{\tau\sigma}\|\delta^{k+1}_{u,h_{k+1}}\|_{L^{2}(\Omega)}\left(\|\lambda^{k+1}_{h_{k+1}}-\lambda^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)}+\|\lambda^{\ast}_{h_{k+1}}-\lambda^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}+\|z^{k+1}_{h_{k+1}}-z^{\ast}_{h_{k+1}}\|_{L^{2}(\Omega)}\right)\\
&\leq \bar \eta \|\delta^{k+1}_{u,h_{k+1}}\|_{L^{2}(\Omega)}.
\end{aligned}$$ where $\bar \eta:=\sqrt{\frac{2C_{1}'}{\tau\sigma}}+\sqrt{\left(1+\frac{1}{\tau}\right)\frac{2C_{2}'}{\sigma}}$ is a constant, we used (\[es1\]), (\[es2\]) and the property $u^{\ast}_{h_{k+1}}=z^{\ast}_{h_{k+1}}$.
Then we know from Lemma \[lmm:complexity2\] that $$\begin{aligned}
&\sum_{k=0}^{\infty}\left(\frac{\sigma}{2}\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}+\frac{\sigma}{2}\|u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\right)\\
\leq& \sum_{k=0}^{\infty}\left(\left\|\theta^{k}\right\|^{2}_{L^{2}(\Omega)}-\left\|\theta^{k+1}\right\|^{2}_{L^{2}(\Omega)}+\left\|\theta^{k+1}\right\|^{2}_{L^{2}(\Omega)}-\left\|\bar \theta^{k+1}\right\|^{2}_{L^{2}(\Omega)}\right)+\sum_{k=0}^{\infty}\left\langle\delta^{k+1}_{u,h_{k+1}}, u^{k+1}_{h_{k+1}}-u^{\ast}_{h_{k+1}}\right\rangle
\\
\leq & \left\|\theta^{0}\right\|^{2}_{L^{2}(\Omega)}+\bar \eta \sum_{k=0}^{\infty}\|\delta^{k+1}_{u,h_{k+1}}\|_{L^{2}(\Omega)}+\sum_{k=0}^{\infty}(C'_{1}+C'_{2})\cdot{\rm max}_{k}(\|\bar \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)})
\\
< &\infty,
\end{aligned}$$ where we used the property $\left\|\theta^{k+1}\right\|^{2}_{L^{2}(\Omega)}-\left\|\bar \theta^{k+1}\right\|^{2}_{L^{2}(\Omega)}\leq \|\bar \theta^{k+1}+\theta^{k+1}\|_{L^{2}(\Omega)}\cdot\|\bar \theta^{k+1}-\theta^{k+1}\|_{L^{2}(\Omega)}$ and (\[theta\]). Hence, there exist constants $\tilde C_{1}, \tilde C_{2}$ such that $$\label{equ:bounded}
\sum_{k=0}^{\infty}\|u^{k+1}_{h_{k+1}}-z^{k+1}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\leq \tilde C_{1},\ \ \sum_{k=0}^{\infty}\|u^{k+1}_{h_{k+1}}-z^{k}_{h_{k+1}}\|_{L^{2}(\Omega)}^{2}\leq \tilde C_{2}.$$
Finally, by substituting (\[equ:bounded\]) to (\[R inequality\]), there exists a constant $\tilde C$, $$\sum_{k=0}^{\infty}R_{h_{k+1}}(u^{k+1}_{h_{k+1}},z^{k+1}_{h_{k+1}},\lambda^{k+1}_{h_{k+1}})\leq 2
\sum_{k=0}^{\infty}\|\delta_{u,h_{k+1}}^{k+1}\|_{L^{2}(\Omega)}^{2}+2\sigma^{2}\tilde C_{1}+(1+\sigma^{2})\tilde C_{2}=\tilde C<\infty.$$ Thus, by Lemma \[lmm:complexity1\], we know that $$\min_{1\leq i\leq k}{R_{h_{i}}}(u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}})\leq \frac{\tilde C}{k}, \ \ \lim \limits_{k\rightarrow \infty}(k\times \min_{1\leq i\leq k}{R_{h_{i}}}(u^{i}_{h_{i}}, z^{i}_{h_{i}}, \lambda^{i}_{h_{i}}))=0$$ holds. Therefore, combining the obtained global convergence results, we complete the whole proof of Theorem \[convergence theorem\].
Numerical experiments {#sec:4}
=====================
In this section we illustrate the numerical performance of the proposed multi-level ADMM algorithm for PDE-constrained optimization problems. All our computational results are obtained by MATLAB R2017b running on a computer with 64-bit Windows 7.0 operation system, Intel(R) Core(TM) i7-6700U CPU (3.40 GHz), and 32 GB of memory.
First, we introduce the algorithmic details that are common to all examples. The discretization was carried out by using the standard piecewise linear finite element approach. To present numerical results, it is convenient to introduce the experimental order of convergence (EOC), which for some positive error function $E(h):=\|u-u_{h}\|_{L^{2}(\Omega)}, h>0$ is defined by $${\rm EOC}:=\frac{{\rm log}E(h_{1})-{\rm log}E(h_{2})}{{\rm log}h_{1}-{\rm log}h_{2}}.$$ We note that if $E(h)=O(h^{\beta})$, then $EOC\approx \beta$. In numerical experiments, we measure the accuracy of an approximate optimal solution by using the corresponding KKT residual error for each algorithm. For the purpose of showing the efficiency of our mADMM, we report the numerical results obtained by running the ihADMM (see [@Song2017Fe] for details) and the classical ADMM method to compare with the results obtained by the mADMM. In this case, we terminate all the algorithms when $\eta < 10^{-6}$ with the maximum number of iterations set to 500. For all numerical examples and all algorithms, we choose zeros as the initial values and the penalty parameter $\sigma$ was chosen as $\sigma=0.1\alpha$. About the step length $\tau$, we choose $\tau = 1.618$.
\[ex1\] ([@Hinze2009Optimization], Example 3.3) Consider $$\begin{aligned}
\min \limits_{(y,u)\in H_{0}^{1}(\Omega)\times L^{2}(\Omega)}^{}\ \ J(y,u)&=\frac{1}{2}\|y-y_d\|_{L^2(\Omega)}^{2}+\frac{\alpha}{2}\|u\|_{L^2(\Omega)}^{2} \\
{\rm s.t.}\qquad \quad \quad \ -\Delta y&=u\ \ \ \mathrm{in}\ \Omega, \\
\ \ \ \ \ \ y&=0\quad \mathrm{on}\ \partial\Omega,\\
\qquad \qquad\quad u&\in U_{ad}=\{v(x)|a\leq v(x)\leq b, {\rm a.e }\ \mathrm{on}\ \Omega\},
\end{aligned}$$ where the domain is the unit circle $\Omega=B_{1}(0)\subseteq \mathbb{R}^{2}$. Set the desired state $y_{d}=(1-(x_{1}^{2}+x_{2}^{2}))x_{1}$, the parameters $\alpha=0.1, a=-0.2, b=0.2$.
In this example, the exact solutions of the problem are unknown in advance. Instead we use the numerical solutions computed on the grid with $h=2^{-10}$ as reference solutions. As an example, the discretized optimal control on the grid with $h=2^{-7}$ is presented in Figure \[fig:1\].
![Discretized optimal control solution for Example \[ex1\] on the grid with $h = 2^{-7}$.[]{data-label="fig:1"}](ex1level4.eps)
The error of the control $u$ [w.r.t.]{} the $L^{2}$-norm, the EOC for the control, the numerical results for the accuracy of solution, the CPU time and the number of iterations obtained by our mADMM, the ihADMM and the classical ADMM are shown in Table \[tab:1\]. We can see from Table \[tab:1\] that our mADMM is highly efficient in obtaining an approximate solution compared with the ihADMM and the classical ADMM in terms of the CPU time, especially when the discretization is in a fine level. Furthermore, it should be specially mentioned that the numerical results in terms of iterations illustrate the mesh-independent performance of the mADMM and the ihADMM. However, iterations of the classical ADMM will increase with the refinement of the discretization.
[lllllllll]{} h & $\#\rm dofs$ & $E$ &EOC & Index & mADMM & ihADMM & classical ADMM\
$2^{-5}$ & 635 &0.00168 & - & residual $\eta$& 8.46e-07 & 9.29e-07 & 8.32e-07\
& & & & CPU times/s & 0.21 & 0.61 & 0.95\
& & & & $\#$iter & 14 & 27 & 63\
$2^{-6}$ & 2629 &5.57e-04 & 1.5927 & residual $\eta$& 6.90e-07 & 8.78e-07 & 2.47e-07\
& & & & CPU times/s & 0.43 & 1.73 & 2.13\
& & & & $\#$iter & 15 & 26 & 30\
$2^{-7}$ & 10697 &2.05e-04 & 1.4420& residual $\eta$& 7.14e-07 & 8.83e-07 & 9.03e-07\
& & & & CPU times/s & 1.23 & 7.40 & 30.46\
& & & & $\#$iter & 13 & 25 & 54\
$2^{-8}$ & 43153 &1.13e-04 & 0.8593& residual $\eta$& 4.29e-07 & 7.16e-07 & 6.93e-07\
& & & & CPU times/s & 4.45 & 43.13 & 755.84\
& & & & $\#$iter & 13 & 25 & 119\
$2^{-9}$ & 173345 &6.68e-05 & 0.7584& residual $\eta$& 1.22e-07 & 8.69e-07 & 2.68e-07\
& & & & CPU times/s & 39.15 & 384.39 & 39646.84\
& & & & $\#$iter & 14 & 24 & 380\
$2^{-10}$ & 694849 & - & - & residual $\eta$& 8.7e-08 & 9.29e-07 & **3.58e-05**\
& & & & CPU times/s & 553.37 & 6451.98 & 289753.72\
& & & & $\#$iter & 15 & 23 & **500**\
\[ex2\] ([@Hinze2012The], Example 4.1) Consider $$\begin{aligned}
\min \limits_{(y,u)\in H_{0}^{1}(\Omega)\times L^{2}(\Omega)}^{}\ \ J(y,u)&=\frac{1}{2}\|y-y_d\|_{L^2(\Omega)}^{2}+\frac{\alpha}{2}\|u\|_{L^2(\Omega)}^{2} \\
{\rm s.t.}\qquad \quad \quad \ -\Delta y&=u\ \ \ \mathrm{in}\ \Omega, \\
y&=0\quad \mathrm{on}\ \partial\Omega,\\
\qquad \qquad\quad u&\in U_{ad}=\{v(x)|a\leq v(x)\leq b, {\rm a.e }\ \mathrm{on}\ \Omega\}.
\end{aligned}$$ where $\Omega=(0,1)^{2}$, the upper bound is $a=0.3$, the lower bound is $b=1$, and the regularization parameter is $\alpha=0.001$. We choose $y_{d}=-4\pi^{2} \alpha \sin(\pi x)\sin(\pi y)+Sr$, where $r=\min(1,\max(0.3,2\sin(\pi x)\sin(\pi y)))$, $S$ denotes the solution operator. In addition, from the choice of parameters, it is easy to know that $u=r$ is the unique solution of the continuous problem.
The exact control and the discretized optimal control on the grid with $h = \sqrt{2}/2^{7}$ are presented in Figure \[fig:2\].
[3in]{} ![Exact control solution (left) and discretized optimal control solution (right) for Example \[ex2\] on the grid with $h = \sqrt{2}/2^{7}$.[]{data-label="fig:2"}](ex2exactu.eps "fig:"){width="3.4in"}
[3in]{} ![Exact control solution (left) and discretized optimal control solution (right) for Example \[ex2\] on the grid with $h = \sqrt{2}/2^{7}$.[]{data-label="fig:2"}](ex2level5.eps "fig:"){width="3.4in"}
The error of the control $u$ [w.r.t.]{} the $L^{2}$- norm, the EOC for the control, the numerical results for the accuracy of solution, the CPU time and the number of iterations obtained by our mADMM, the ihADMM and the classical ADMM are shown in Table \[tab:2\]. Experiment results show that our mADMM has evident advantage on CPU time over the ihADMM and the classical ADMM. Furthermore, we also notice that the numerical results in terms of iteration numbers illustrate the mesh-independent performance of the mADMM.
[lllllllll]{} h & $\#\rm dofs$ & $E$ &EOC & Index & mADMM & ihADMM & classical ADMM\
$\sqrt{2}/2^{4}$ & 225 &0.0172 &- & residual $\eta$& 9.26e-07 & 9.88e-07 & 9.98e-07\
&& & & CPU times/s & 0.31 & 0.28 & 0.43\
& & & & $\#$iter & 22 & 25 & 120\
$\sqrt{2}/2^{5}$ & 961 &6.71e-03 &1.3580 & residual $\eta$& 7.83e-07 & 9.19e-07 & 7.25e-07\
& & & & CPU times/s & 0.85 & 0.69 & 0.87\
& & & & $\#$iter & 23 & 26 & 32\
$\sqrt{2}/2^{6}$ & 3969 &2.11e-03 &1.6691& residual $\eta$& 8.62e-07 & 7.10e-07 & 3.34e-07\
& & & & CPU times/s & 2.97 & 3.85 & 4.45\
& & & & $\#$iter & 24 & 30 & 32\
$\sqrt{2}/2^{7}$ & 16129 &8.02e-04 &1.3956& residual $\eta$& 9.80e-08 & 7.80e-08 & 8.86e-08\
&& & & CPU times/s & 14.92 & 37.39 & 91.15\
& & & & $\#$iter & 23 & 28 & 78\
$\sqrt{2}/2^{8}$ & 65025 &3.58e-04 &1.1636& residual $\eta$& 9.04e-07 & 9.62e-07 & 7.05e-07\
& & & & CPU times/s & 22.80 & 151.57 & 2457.02\
& & & & $\#$iter & 21 & 28 & 183\
$\sqrt{2}/2^{9}$ & 261121 &1.81e-04 &0.9840 & residual $\eta$& 3.09e-07 & 8.43e-07 & 5.56e-07\
&& & & CPU times/s & 168.58 & 1469.13 & 40739.35\
& & & & $\#$iter & 22 & 30 & 283\
Conclusion {#sec:5}
==========
In this paper, we employ a multi-level ADMM algorithm to solve optimization problems with PDE constraints. Instead of solving the discretized problems, we apply the *‘optimize-discretize-optimize’* strategy. Such approach has the flexibility that allows us to discretize the subproblems of the inexact ADMM algorithm by different discretization schemes. Motivated by the multi-level strategy, we propose the proper strategy of gradually refining the grid and the strategy of solving the subproblems inexactly. We designed the convergent multi-level ADMM (mADMM) algorithm, which can significantly reduce the computation cost and make the algorithm faster. The convergence analysis and the iteration complexity results $o(1/k)$ is presented. Numerical results demonstrated the efficiency of the proposed mADMM algorithm.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank Prof. Long Chen very much for his FEM package iFEM [@Chen2009] in Matlab.
, [Multi-level adaptive solutions to boundary-value problems]{}. [*Math. Comp.*]{} [**31** (1977)]{} [333-390]{}.
, [An efficient inexact symmetric Gauss-Seidel based majorized ADMM for high-dimensional convex composite conic programming]{}. [*Math. Program.*]{} [**161** (2017)]{} [237-270]{}.
, [iFEM: An Integrated Finite Element Methods Package in MATLAB]{}. [Technical Report.]{} [University of California at Irvine, Irvine (2009)]{}.
, [A FE-ADMM algorithm for Lavrentiev-regularized state-constrained elliptic control problem]{}. [*ESAIM: COCV*]{} [**25**]{} [(2019)]{} [E5]{}.
, [The Finite Element Method for Elliptic Problems]{}. [Volume 40 of *Classics in Applied Mathematics.*]{} [Society for Industrial and Applied Mathematics]{}, [Philadelphia (2002)]{}.
, [Newton Methods for Nonlinear Problems: Affine Invariance and Adaptive Algorithms]{}. [Springer]{}, [Berlin (2011)]{}.
, [Hankel matrix rank minimization with applications to system identification and realization]{}. [*SIAM J. Matrix Anal. Appl.*]{} [**34**]{} [(2013)]{} [946-977]{}.
, [On decomposition-coordination methods using an augmented Lagrangian]{}. [Augmented Lagrangian Methods: Applications to the Solution of Boundary Problems]{}. Elsevier, [Amsterdam (1983)]{}.
, [A dual algorithm for the solution of nonlinear variational problems via finite element approximation]{}. [*Comput. Math. Appl.*]{} [**2** (1976)]{} [17-40]{}.
, [Lectures on numerical methods for nonlinear variational problems]{}. [Springer]{}, [Berlin (1980)]{}.
, [Sur l’approximation, par éléments finis d’ordre un, et la résolution, par pénalisation-dualité d’une classe de problèmes de Dirichlet non linéaires]{}. [*Analyse Numérique*]{} [**9** (1975)]{} [41-76]{}.
, [Multi-Grid Methods and Applications]{}. [Springer]{}, [Berlin (1985)]{}.
, [On the multi-grid method applied to difference equations]{}. [*Computing*]{} [**20** (1978)]{}, [291-306]{}.
, [Linear rate convergence of the alternating direction method of multipliers for convex composite programming]{}. [*Math. Oper. Res.*]{} [**43** (2017)]{} [622-637]{}.
, [Variational discretization of Lavrentiev-regularized state constrained elliptic optimal control problems]{}. [*Comput. Optim. Appl.*]{} [**46** (2010)]{} [487-510]{}.
, [Optimization with PDE Constraints]{}. [Springer]{}, [Berlin (2009)]{}.
, [The semi-smooth Newton method for variationally discretized control constrained elliptic optimal control problems; implementation, convergence and globalization]{}. [*Optim. Methods Softw.*]{} [**27** (2012)]{} [933-950]{}.
, [An Introduction to Variational Inequalities and their Applications]{}. [Academic Press]{}, [New York (1980)]{}.
, [An efficient alternating direction method of multipliers for optimal control problems constrained by random Helmholtz equations]{}. [*Numer. Algorithms*]{} [**78** (2018)]{} [161-191]{}.
, [QSDPNAL: A two-phase Newton-CG proximal augmented Lagrangian method for convex quadratic semidefinite programming problems]{}. arXiv preprint [*arXiv:1512.08872* (2015)]{}.
, [A Schur complement based semi-proximal ADMM for convex quadratic conic programming and extensions]{}. [*Math. Program.*]{} [**155** (2016)]{} [333-373]{}.
, [Some alternating direction iteration methods for solving PDE-constrained optimization problems]{}. [PhD thesis]{}, [Dalian University of Technology]{}, [Dalian (2018)]{}.
, [A two phase strategy for control constrained elliptic optimal control problems]{}. [*Numer. Linear Algebra Appl.*]{} [**25** (2018)]{} [e2138]{}.
, [A FE-inexact heterogeneous [ADMM]{} algorithm for elliptic optimal control problems with $L^{1}$-control cost]{}. [*J. Syst. Sci. Complex.*]{} [**31** (2017)]{} [1659-1697]{}.
, [A convergent 3-block semiproximal alternating direction method of multipliers for conic programming with 4-type constraints]{}. [*SIAM J. Optim.*]{} [**25** (2015)]{} [882-915]{}.
, [A fast globally linearly convergent algorithm for the computation of Wasserstein Barycenters]{}. arXiv preprint [*arXiv:1809.04249* (2018)]{}.
, [An alternating direction method of multipliers for elliptic equation constrained optimization problem]{}. [*Sci. China Math.*]{} [**60** (2017)]{} [361-378]{}.
[^1]: School of Mathematical Sciences, Dalian University of Technology, Dalian, Liaoning 116025, China. ([[email protected], [email protected], [email protected]]{}).
[^2]: Department of Applied Mathematics, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China. ([[email protected]]{}).
| {
"pile_set_name": "ArXiv"
} |
, , and
We have used the Simbad database and VizieR catalogue access tools to construct the observational color–absolute magnitude diagrams of nearby K–M dwarfs with precise [*Hipparcos*]{} parallaxes ($\sigma_\pi/\pi\leq 0.05$). Particular attention has been paid to removing unresolved double/multiple stars and variables. In addition to archival data, we have made use of nearly 2000 new radial-velocity measurements of K–M dwarfs to identify spectroscopic binary candidates. The main sequences, cleaned from unresolved binaries, variable stars, and old population stars which can also widen the sequence due to their presumably lower metallicity, were compared to available solar-metallicity models. Significant offsets of most of the model main-sequence lines are seen with respect to observational data, especially for the lower-mass stars. Only the location and slope of the Victoria-Regina and, partly, BaSTI isochrones match the data quite well.
astronomical databases: miscellaneous – stars: late type – C–M diagrams – solar neighborhood
The validation of stellar theoretical models relies heavily upon the accurate determination of the location and properties of the sequences of stars in the color–luminosity diagrams. For main-sequence stars with masses above solar, most of the theoretical models are generally in line with observations. The situation is much worse when we discuss the domain of low-mass ($M<0.8\,M_\odot$) stars. On one hand, the complexity of their spectra with molecular features dominating complicates both the calculation of stellar models and the color-temperature transformations, what leads to difficulties in determining good theoretical main-sequence lines. On the other hand, comparisons of models to data require comprehensive observations from which it would be possible to accurately locate the observational ridge lines of different metallicity. Because of the inefficiency of photometric methods to determine metallicities of M-type dwarfs, the local samples suitable for such comparisons are too modest in size. Therefore, the more generally accepted approach has been to use instead of field stars the Galactic star clusters. With this approach, however, the advantage in having the lower main sequence of homogeneous chemical composition is often reduced by uncertainties coming from fitting procedure (distances, etc.) and the scatter within the fainter portion of the sequence due to observational limitations.
In recent years, large photometric surveys such as the Sloan Digital Sky Survey (SDSS) and the Two Micron All Sky Survey (2MASS) have proved to be superb data resources for statistical investigations of late-type dwarfs. However, accurate parallaxes are not available for the overwhelming majority of faint stars that could probe the lower-main sequence. Thus, until [*Gaia*]{} space mission becomes a reality, nearby [*Hipparcos*]{} stars still remain demanding objects.
In the paper by Kotoneva et al. (2002), [*Hipparcos*]{}-based absolute magnitudes of K dwarfs were compared with a set of theoretical isochrones for three metallicity ranges – solar, subsolar (\[Fe/H\] between –0.30 and –0.50) and super-solar (\[Fe/H\]=0.18–0.30). The metallicities were derived by photometric method using [*Geneva*]{} $b_1$ and Johnson-Cousins $(R-I)_{\rm C}$ or Str[ö]{}mgren $m_1$ and $(R-I)_{\rm C}$ colors. They found a tight relationship between luminosity, color and metallicity for K dwarfs. However, none of the isochrones tested by them in the $M_V$,$B$–$V$ plane fitted the observational K-dwarf sequences.
Just & Jahrei[ß]{} (2008) investigated the properties of the main sequence of [*Hipparcos*]{} F–K stars in the Catalogue of Nearby Stars (CNS), using the Johnson-Cousins $BV(RI)_{\rm C}$ data transformed to the SDSS $ugriz$ filter system. Systematic differences in the shape and location of the main sequences in the $ugriz$ system were found with respect to the theoretical isochrones of Padova and Dartmouth models. More recently, Bochanski et al. (2010) have derived color – absolute magnitude relations for low-mass stars using new yet unpublished $(ugriz)^\prime$ observations and 2MASS $JHK_{\rm S}$ photometry of nearby dwarfs with known trigonometric parallax measurements. These relations have been used by the SDSS team to estimate absolute magnitudes and distances to all faint stars in their huge sample.
The aim of this paper is to make a comparison of the sequence of nearby K–M dwarfs in the ($M_V$, color) diagrams, based on the data available at the CDS, with theoretical solar-metallicity isochrones of various stellar models. We accept that the ridge lines, defined in the ($M_V$, color) diagram by stars with kinematics typical of the young to intermediate age disk, may adequately represent the locus of solar-metallicity stars. The absolute magnitudes of K–M dwarfs will be solely based on [*Hipparcos*]{} parallaxes. Our analysis in this paper is restricted to Johnson-Cousins $BV(RI)_{\rm
C}$ photometry as in this particular system we can find the majority of accurate observational data on nearby stars and the majority of model isochrones calculated. We focus attention on the main contributors to the scatter of the observational sequence – photometric variability, unresolved binarity and measurement errors. While for information on variability and astrometric multiplicity of nearby stars we can address the [*Hipparcos*]{} survey and expanding datasets of supplementary observations, the census of spectroscopic binaries, which requires long-term radial-velocity programs, is far from being complete. Therefore, to identify new unresolved binaries among nearby K–M dwarfs, we have made use of nearly 2000 radial-velocity measurements obtained over the past decade within our CORAVEL program (Upgren, Sperauskas & Boyle 2002).
In the following section we will describe the selection of stars, refining of the samples and observational data used. Comparisons of the data to theoretical isochrones of different stellar models are demonstrated in Section 3.
We started with the selection of stars with precise trigonometric parallaxes from [*Hipparcos*]{} catalog (van Leeuwen 2007), which fall within the typical color and luminosity range for K and M dwarfs: $B$–$V$$>$0.80 and $M_V \geq 5.5$ mag. A limit on parallax accuracy of $\sigma_\pi/\pi\leq 0.05$ was chosen to ensure the absolute magnitudes to be accurate to within $\sim 0.1$ mag, with a negligible bias introduced by the Lutz-Kelker effect. We have 1815 stars that satisfy the above criteria (Figure 1). However, for nearly half of these stars, no homogeneous and accurate photometry or supplementary data can be found in the existing databases. To have the luminosities and colors based on rigorous observational basis, we have finally chosen for further analysis two separate samples of [*Hipparcos*]{} K–M dwarfs:
- The McCormick sample (hereafter MCC), which constitutes stars north of declination $-30^\circ$, selected years ago spectroscopically at the McCormick Observatory by A.N. Vyssotsky and his colleagues (for a review of their survey, see Upgren & Weis 1989). For all of these stars, the $BV(RI)_{\rm C}$ data are thoroughly collected at the ARI Database for Nearby Stars (CNS3, Gliese & Jahreiss (1991), with its more updated version of 1998).
- The sample of stars south of declination $+26^\circ$ from the recently published catalog of homogeneous and standardized $UBV(RI)_{\rm C}JHK$ photometry by Koen, Kilkenny, van Wyk & Marang (2010) (hereafter KKWM).
The coverage of color – luminosity diagram by the MCC and KKWM stars with respect to the remaining [*Hipparcos*]{} stars with $\sigma_\pi/\pi\leq 0.05$ is shown in Figure 1. There was no need to apply the extinction and reddening corrections, since the stars in both samples are at distances smaller than 50 pc (the KKWM stars are within 30 pc).
All stars in the MCC sample and part of northernmost stars in the KKWM sample, which had no high-quality or any radial-velocity measurements, were targets of our decadal program of radial-velocity observations (Upgren, Sperauskas & Boyle 2002). However, 126 stars of the selected KKWM sample are still lacking good radial-velocity measurements and thus were rejected from further analysis. Thus we have 478 stars in the MCC sample and 416 stars in the KKWM sample, which satisfy the $\sigma_\pi/\pi\leq 0.05$ criterion and have radial-velocity data of satisfactory quality. Of these, 173 stars are found to be common to both samples.
-1mm
-1mm
In Figures 2–4 we present ($M_V$, color) diagrams for the two samples. To demonstrate what effect the addition of stars known or suspected to be variables, binaries and belonging to presumably metal-deficient population has on the width of the observational sequence, we plotted all these groups of stars by different symbols. We applied the so-called kinematical age parameter $f(U,V,W)$, introduced by Grenon (1987), $$f = 1/C(a_1U^2+a_2V^2+a_3W^2)\,,$$ to divide the stars into different populations: young- to intermediate-age thin disk ($f\leq 0.20$), old thin disk ($0.20<f\leq0.35$) and thick disk ($f>0.35$). Here the space velocity components $U$, $V$ and $W$ are computed with the values of the Sun’s motion relative the LSR $(U_0,V_0,W_0)=(-10.0,12.2,7.2)$kms$^{-1}$ adopted from Sch[ö]{}nrich et al. (2010). The normalization constant $C$ and the coefficients $a_i$ are adopted such that for stars on nearly circular galactic orbits the parameter $f$ equals to orbital eccentricity. As can be seen in the figures, the stars with the thick disk kinematics are seen to lie below the sequence defined by the rest of the stars, and this is an indication of their having lower metallicity. The location of the sequence of stars with the old-thin-disk kinematics is nearly identical to that of the younger thin disk stars but differ in its larger (by 0.1 mag) scatter.
As one would expect, the known and suspected multiple stars tend to lie above the sequence of their single counterparts at the same color. The fraction of such stars, including those with the [*Hipparcos*]{} multiplicity flags, amounts to 30%. The spectroscopic binaries (SB) and radial-velocity variables (SB candidates) comprise 13% of the MCC sample for which our radial-velocity program is nearing completion. Contrary to the general trend, noted in the literature (e.g. Lada 2006), that the binary fraction is likely to steadily decline with spectral type, we do not see from our preliminary analysis of radial-velocity measurements a smaller fraction of SB candidates among M-type dwarfs than among K-type dwarfs (in both cases, around 10%). Based on our radial-velocity observations we report 24 new SB candidates among the MCC stars considered, in addition to those for which radial-velocity variability was noted or suspected in the literature.
The stars showing any signs of photometric variability, i.e. those found in the GCVS database (Samus et al. 2007–2011) or flagged in the [*Hipparcos*]{} catalog, constitute one third of the two samples (in Figures 2–4 indicated by crosses).
[r]{}\[0pt\][72mm]{} -5mm
For comparison with theoretical isochrones, we removed stars which may contribute increased scatter or broadening of the observational sequence: known and probable multiple stars, variables, and stars with $f$$>$$0.20$ which may belong to populations of subsolar metallicity. After refining the MCC and KKWM samples in this way, we are finally left with 160 and 105 stars in each sample, respectively. The ridge lines of the stellar loci in the absolute magnitude versus color planes, defined using each refined sample separately, are tabulated in Table 1 (in the column headings, superscripts (1) and (2) refer to MCC and KKWM, respectively). We note that over most of the color range the ridge lines of the KKWM sequences are systematically more luminous than those of the MCC stars, with average offsets of 0.1 mag in $M_V$, $M_{R_{\rm C}}$ and $M_{I_{\rm C}}$ (see Figure 5). Since systematic differences in $BV(RI)_{\rm C}$ magnitudes between the MCC and KKWM data sets are found to be of 0.02–0.03 mag (in the sense that KKWM magnitudes are brighter), the much larger offsets in the color-absolute magnitude space can be explained by a larger fraction of unrecognized binaries left in the KKWM sample, which may lead to a shift of the ridge lines toward brighter absolute magnitudes. Indeed, most of the KKWM stars were not targets for our long-term radial-velocity program and the fraction of known/suspected SB among them (7%) is much less significant than that among the MCC stars (13%) covered entirely by our program.
In Figure 6 we compare a set of solar composition isochrones to the sequences of K–M dwarfs with $f\leq0.20$, cleaned of variable and binary stars and SB candidates. The imposed cut in the kinematical age parameter $f$ ensures that the stars belong to the young and intermediate age population and are, on average, likely to be of solar metallicity. We have tested 5 Gyr isochrones (the effect of age is minor in the K–M dwarf range) of the following stellar evolution models:
- Dartmouth (Dotter et al. 2008);
- Padova (Marigo et al. 2008);
- BaSTI (Cordier et al. 2007);
- Victoria-Regina (VandenBerg et al. 2006);
- Yonsei-Yale ${\rm Y}^2$ (Demarque et al. 2004) with the older (GDK, Green et al. 1987) magnitude/color transformations;
- Geneva (Lejeune & Schaerer 2001);
- Siess et al. (2000).
In the $M_V$,$B$–$V$ diagram of Figure 6, we also show the empirical relation from the calibrations by Schmidt-Kaler (1982), and, in the $M_V$,$V$–$I_{\rm C}$ diagram, the fiducial line of the solar-metallicity open cluster M67, taken from Sandquist (2004) but extended to the redder colors ($V$–$I_{\rm C}\geq1.6$) using photometric data from the WEBDA Data Base. To aid visualization, we have plotted in the figure the sequences of the MCC and KKWM stars rather than their ridge lines. Note that photometry for 108 MCC stars not common to the KKWM sample comes from the CNS3 database, while for 52 stars common to both samples and for the rest 53 KKWM stars is taken from KKWM. Therefore, plotted in the figure are comparable numbers of stars representing the two sources of $BV(RI)_{\rm C}$ photometry.
We conclude by Figure 6 that, with a few exceptions, most of the isochrones provide a satisfactory fit to the dwarfs of spectral types earlier than K7 ($B$–$V$$<$1.3), but fail to reproduce the stellar locus of the less luminous dwarfs. In the region of M stars ($B$–$V$$\geq$1.4), some isochrones diverge from the main sequence by more than 1 mag. Most appropriate are the Victoria-Regina isochrones, which provide a good fit to the shape of the main sequence down to the limit of their models (around M5), and the BaSTI isochrones, matching the locus of K dwarfs well, especially in the $M_V$,$B$–$V$ plane. The BaSTI colors $V$–$R_{\rm C}$ and $V$–$I_{\rm C}$ at the reddest end (around K8–M0), however, do not correctly match the data. In the $M_V$ versus $V$–$R_{\rm C}$ and $V$–$I_{\rm C}$ planes, the Dartmouth models and isochrones by Siess et al. (2000) also fit the data, but only down to spectral type K8. There is also a general agreement between the sequence of field K–M dwarfs and the fiducial line of M67.
In summary, the samples of nearby K–M dwarfs show that there is a clear problem in fitting most of the stellar models to low mass stars. The Victoria-Regina isochrones (VandenBerg et al. 2006) with empirically constrained color-temperature relations by VandenBerg & Clem (2003) appear to be the only models which are entirely consistent with the observed loci of lower main-sequence stars in the $BV(RI)_{\rm C}$ system.
[^1]
Bochanski J. J., Hawley S. L., Covey K. R. et al. 2010, AJ, 139, 2679
Cordier D., Pietrinferni A., Cassisi S., Salaris M. 2007, AJ, 133, 468\
http://albione.oa-teramo.inaf.it/
Demarque P., Woo J.-H., Kim Y.-Ch., Yi S. K. 2004, ApJS, 155, 667\
http://www.astro.yale.edu/demarque/yyiso.html
Dotter A., Chaboyer B., Jevremovi[ć]{} D. et al. 2008, ApJS, 178, 89\
http://stellar.dartmouth.edu/ models/isolf.html
Gliese W., Jahreiss H. 1991. Preliminary Version of the Third Catalogue of Nearby Stars (CNS3); update at http://www.ari.uni-heidelberg.de/datenbanken/aricns/
Green E. M., Demarque P., King C. R. 1987, The Revised Yale Isochrones and Luminosity Functions, New Haven, Yale Univ. Obs. (GDK)
Grenon M. 1987, Journ. Astron. & Astroph., 8, 123
Just A., Jahrei[ß]{} H. 2008, Astron. Nachr., 329, 790
Koen C., Kilkenny D., van Wyk F., Marang F. 2010, MNRAS, 403, 1949 (KKWM)
Kotoneva E., Flynn C., Jimenez R. 2002, MNRAS, 335, 1147
Lada C. J. 2006, ApJ, 640, L63
Lejeune T., Schaerer D. 2001, A&A, 366, 468\
http://webast.ast.obs-mip.fr/equipe/stellar
Marigo P., Girardi L., Bressan A. et al. 2008, A&A, 482, 883\
http://stev.oapd.inaf.it/cmd
Samus N. N., Durlevich O. V., Kazarovets E. V. et al. 2007–2011, General Catalogue of Variable Stars (GCVS database, Version 2011Jan), CDS B/gcvs
Sandquist E. L. 2004, MNRAS, 347, 101
Schmidt-Kaler T. 1982, in Landolt-B[ö]{}rnstein New Series, Group VI, vol. 2b, eds. K. Schaifers & H. H. Voigt, Springer-Verlag, Berlin
Sch[ö]{}nrich R., Binney J., Dehnen W. 2010, MNRAS, 403, 1829
Siess L., Dufour E., Forestini M. 2000, A&A, 358, 593\
http://www.astro.ulb.ac.be/ siess/WWWTools/Isochrone
Upgren A. R., Weis E. W. 1989, in [*Star catalogues: a centennial tribute to A. N. Vyssotsky*]{}, eds. A. G. D. Philip & A. R. Upgren, L. Davis Press, Schenectady, p. 19
Upgren A. R., Sperauskas J., Boyle R. P. 2002, Baltic Astronomy, 11, 91
VandenBerg D. A., Clem J. L. 2003, AJ, 126, 778
VandenBerg D. A., Bergbusch P. A., Dowler P. D. 2006, ApJS, 162, 375; http://\
www1.cadc-ccda.hia-iha.nrc-cnrc.gc.ca/community/VictoriaReginaModels/
van Leeuwen F. 2007, A&A, 474, 653
[^1]: This research has made use of the SIMBAD data base and the VizieR catalogue access tool, CDS, Strasbourg. The ARICNS data base, Astronomisches Rechen-Institut, Heidelberg, and the WEBDA database, Viena, have also been our data source. Direct access to the BaSTI, Dartmouth, Geneva, Padova, Victoria-Regina and Yonsei-Yale databases and the isochrone browse tools by Lionel Siess are greatly acknowledged. The work was supported by the Research Council of Lithuania under the grant No. MIP-132/2010.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A self-consistent system of interaction nonlinear spinor and scalar fields within the scope of a BI cosmological model filled with perfect fluid is considered. The role of spinor field in the evolution of the Universe is studied. It is shown that the spinor field nonlinearity can generate a negative effective pressure, which can be seen as an alternative source for late time acceleration of the Universe.'
author:
- Bijan Saha
title: Spinor field and accelerated regimes in cosmology
---
Introduction
============
The accelerated mode of expansion of the present day Universe encourages many researchers to introduce different kind of sources that is able to explain this. Among them most popular is the dark energy given by a $\Lambda$ term [@PRpadma; @sahni; @lambda], quintessence [@caldwell; @starobinsky; @zlatev; @pfdenr], Chaplygin gas [@kamen; @pfden]. Recently cosmological models with spinor field have been extensively studied by a number of authors in a series of papers [@sgrg; @sjmp; @smpla; @sprd; @bited; @green]. The principal motive of the papers [@sgrg; @sjmp; @smpla; @sprd; @bited] was to find out the regular solutions of the corresponding field equations. In some special cases, namely with a cosmological constant ($\Lambda$ term) that plays the role of an additional gravitation field, we indeed find singularity-free solutions. It was also found that the introduction of nonlinear spinor field results in a rapid growth of the Universe. This allows us to consider the spinor field as a possible candidate to explain the accelerated mode of expansion. Note that similar attempt is made in a recent paper by Kremer [*et. al.*]{} [@kremer1]. In this paper we study the role of a spinor field in the late-time acceleration of the Universe. To avoid lengthy calculations, we mainly confine ourselves to the study of master equation describing the evolution of BI Universe. We here give the solutions to the spinor and scalar field equations symbolically, for details one can consult [@sprd; @bited].
Basic equations: a brief journey
================================
We consider a self consistent system of nonlinear spinor and scalar fields within the scope of a Bianchi type-I gravitational field filled with a perfect fluid. The spinor and the scalar field is given by the Lagrangian $${{\mathcal}L}= \frac{i}{2} \biggl[{\bar \psi}\gamma^{\mu} \nabla_{\mu} \psi-
\nabla_{\mu} \bar \psi \gamma^{\mu} \psi \biggr] - m{\bar \psi}\psi + F +
\frac{1}{2} (1 + \lambda_1 F_1)\,{\varphi}_{,\alpha}{\varphi}^{,\alpha},
\label{lag}$$ where $\lambda$ is the coupling constant and $F$ and $F_1$ are some arbitrary functions of invariants generated from the real bilinear forms of a spinor field. Here we assume $F = F (I,J)$ and $F_1 = F_1 (I,J)$ with $I = S^2$, $ S = {\bar \psi}\psi$, $J = P^2$ and $P = i \bar \psi \gamma^5 \psi$.
The gravitational field is chosen in the form $$ds^2 = dt^2 - a_1^2 dx_1^2 - a_2^2 dx_2^2 - a_3^2 dx_3^2,
\label{BI1}$$ where $a_i$ are the functions of $t$ only and the speed of light is taken to be unity. We also define $$\tau = a_1 a_2 a_3. \label{taudef}$$
We consider the spinor and scalar field to be space independent. In that case for the spinor, scalar and metric functions we find the following expressions [@bited].
For $F=F(I)$ we find $S=C_0/\tau$ with $C_0$ being an integration constant. The components of the spinor field in this case read $$\psi_{1,2}(t) = (C_{1,2}/\sqrt{\tau}) e^{-i\beta}, \quad
\psi_{3,4}(t) = (C_{3,4}/\sqrt{\tau}) e^{i\beta}, \label{psiI}$$ with the integration constants obeying $C_0$ as $C_0 = C_{1}^{2} +
C_{2}^{2} - C_{3}^{2} - C_{4}^{2}.$ Here $\beta = \int(m -
{{{\mathcal}D}})dt $.
For $F=F(J)$ in case of massless spinor field we find $P=D_0/\tau$. The corresponding components of the spinor field in this case read: with $$\begin{aligned}
\psi_{1,2} &=& \bigl(D_{1,2} e^{i \sigma} + iD_{3,4}
e^{-i\sigma}\bigr)/\sqrt{\tau},\nonumber \\ \label{psiJ}\\ \psi_{3,4} &=&
\bigl(iD_{1,2} e^{i \sigma} + D_{3,4} e^{-i
\sigma}\bigr)/\sqrt{\tau},\nonumber\end{aligned}$$ with $D_0=2\,(D_{1}^{2} + D_{2}^{2} - D_{3}^{2} -D_{4}^{2}).$
For the scalar field we find $${\varphi}= C \int \frac{dt}{\tau (1 + 2\lambda_1 F_1)}, \quad C = {\rm
const.} \label{sfsol}$$
Solving the Einstein equation for the metric functions we find $$a_i(t) = D_{i} [\tau(t)]^{1/3} \exp \bigl[X_i \int\limits_{0}^{t}
[\tau (t')]^{-1}dt' \bigr], \label{a_i}$$ with the integration constants obeying $$D_1 D_2 D_3 = 1, \qquad
X_1 + X_2 + X_3 = 0,$$
As one sees, the spinor, scalar and metric functions are in some functional dependence of $\tau$. It should be noted that besides these, other physical quantities such as spin-current, charge etc. and invariant of space-time are too expressed via $\tau$ [@sprd; @bited]. It should be noted that at any space-time points where $\tau = 0$ the spinor, scalar and gravitational fields become infinity, hence the space-time becomes singular at this point [@bited]. So it is very important to study the equation for $\tau$ (which can be viewed as master equation) in details, exactly what we shall do in the section to follow. In doing so we analyze the role of spinor field in the character of evolution.
Evolution of BI universe and role of spinor field
=================================================
The equation for $\tau$ is found from the Einstein one: $$R_\mu^\nu - \frac{1}{2} \delta_\mu^\nu R = \kappa T_\mu^\nu +
\delta_\mu^\nu \Lambda. \label{ee}$$ The details can be found in [@sprd]. This equation indeed describes the evolution of the universe and has the following general form: $$\frac{\ddot \tau}{\tau}= \frac{3}{2}\kappa
\Bigl(T_{1}^{1}+T_{0}^{0}\Bigr) + 3 \Lambda, \label{dtau}$$ where $\Lambda$ is the cosmological constant, $T_\mu^\nu$ is the energy-momentum tensor. Note also that here a positive $\Lambda$ corresponds to the universal repulsive force, while a negative one gives an additional gravitational force. Note that a positive $\Lambda$ is often considered to be a form of dark energy. Though our main object is to verify the role of spinor field in the evolution of the Universe, we include the $\Lambda$ term in order to explain some results obtained later. For this purpose we recall that the Bianchi identity $G_{\mu;\nu}^{\nu}= 0$ gives $${\dot T}_{0}^{0} = - \frac{\dot \tau}{\tau}\bigl(T_{0}^{0} -
T_{1}^{1}\bigr). \label{conservds}$$ After a little manipulation from and one finds the following expression for $T_0^0$:
$$\kappa T_0^0 = 3 H^2 - \Lambda - C_{00}/\tau^2, \label{kt00}$$
where the definition of the generalized Hubble constant $H$ as $$H = \frac{1}{3}{\dot \tau}/\tau, \label{HC}$$ Let us now stop here for a while. Consider the case when $\Lambda
= 0$. At the moment when expansion rate is zero (it might be at a time prior to the “Big Bang”, or sometimes in the far future when the universe cease to expand we have $H = 0$. Then the nonnegativity of $T_0^0$ suggests that $C_{00} \le 0$. Let us now consider another case when $\tau$ is large enough for the term $1/\tau^2$ to be omitted. As we know $T_0^0$ (the energy density), decreases with the increase of $\tau$. If $\tau$ is big enough for $T_0^0$ to be neglected, from we find $$3 H^2 -
\Lambda \to 0.$$ It means for $\tau$ to be infinitely large, $\Lambda \ge 0.$ In case of $\Lambda = 0$ we find that beginning from some value of $\tau$ the rate of expansion of the Universe becomes trivial, that is the universe does not expand with time. Whereas, for $\Lambda > 0$ the expansion process continues forever. As far as negative $\Lambda$ is concerned, its presence imposes some restriction on the energy density $T_0^0$, namely, $T_0^0$ can never be small enough to be ignored. It imposes some restrictions on $\tau$, precisely, there exists some upper limit for $\tau$ (note that $\tau$ is essentially nonnegative, i.e. bound from below). Thus we see that a negative $\Lambda$, depending on the choice of parameters can give rise to an oscillatory mode of expansion. Thus we come to the following conclusion:
[*Let $T_\mu^\nu$ be the source of the Einstein field equation; $T_0^0$ is the energy density and $T_1^1,\,T_2^2,\,T_3^3$ are the principal pressure and $T_1^1 = T_2^2 = T_3^3$. An ever-expanding BI Universe may be obtained if and only if the $\Lambda$ term is positive (describes a repulsive force and can be viewed as a form of dark energy) and is introduced into the system as in .*]{}
It should be noted that the other types of dark energy such as quintessence, Chaplygin gas enters into the system as a part of $T_\mu^\nu$ and corresponding energy density decreases with the increase of the Universe, hence cannot be considered as source for ever-expanding Universe.
Let us now go back to the Eq. . The components of the energy-momentum tensor read: $$\begin{aligned}
T_{0}^{0} &=& mS - F + \frac{1}{2} ( 1 + 2 \lambda_1
F_1) {\dot {\varphi}}^2 + {\varepsilon}_{pf}, \nonumber\\ \label{total}
\\
T_{1}^{1} &=& T_{2}^{2} = T_{3}^{3} = {{{\mathcal}D}} S + {{{\mathcal}G}} P - F -
\frac{1}{2} ( 1 + 2 \lambda_1 F_1) {\dot {\varphi}}^2 - p_{pf},\nonumber\end{aligned}$$ where, ${{{\mathcal}D}} = 2 S dF/dI + \lambda_1 S {\dot {\varphi}}^2 dF_1/dI$ and ${{{\mathcal}G}} = 2 P dF/dJ + \lambda_1 P {\dot {\varphi}}^2 dF_1/dJ.$ In ${\varepsilon}_{pf}$ and $p_{pf}$ are the energy density and pressure of the perfect fluid, respectively and related by the equation of state $p_{pf} = \zeta {\varepsilon}_{pf}$, where $\zeta \in
[0,\,1]$.
Let us now study the equation for $\tau$ in details and clarify the role of material field in the evolution of the Universe. For simplicity we consider the case when both $F$ and $F_1$ are the functions of $I\, (S)$ only. For simplicity we set $C = 1$ and $C_0 = 1$. Note that from the Bianchi identity for ${\varepsilon}_{pf}$ and $p_{pf}$ we find ${\varepsilon}_{pf} = {\varepsilon}_0/\tau^{1+\zeta}$ and $p_{pf} = \zeta_0
{\varepsilon}_0/\tau^{1+\zeta}$. Further we set ${\varepsilon}_0 = 1$. Assuming that $F = \lambda S^q$ and $F_1 = S^r$, for the effective energy density and effective pressure we find $$\begin{aligned}
T_{0}^{0} &=& \frac{m}{\tau} - \frac{\lambda}{\tau^q} +
\frac{\tau^{r-2}}{2(2 \lambda_1 + \tau^r)} +
\frac{1}{\tau^{1 + \zeta}} + \Lambda \equiv {\varepsilon}\, \nonumber\\ \label{total0}
\\
T_{1}^{1} &=& \frac{(q -1)\lambda}{\tau^q} - \frac{[(2 - r) \lambda_1 +
\tau^r ] \tau^{r-2}}{2(2 \lambda_1 + \tau^r)^2} - \frac{\zeta}{\tau^{1+\zeta}} -
\Lambda \equiv p.\nonumber\end{aligned}$$ Taking into account that $T_0^0$ and $T_1^1$ are the functions of $\tau$, only, the Eq. can now be presented as $$\ddot \tau = {{\mathcal}F}(q_1,\tau), \label{newtd}$$ where we define $${{\mathcal}F}(q_1,\tau) = (3/2) \kappa \Bigl(m + \lambda (q -2)\tau^{1
- q} + \lambda_1 r \tau^{r-1}/2(2 \lambda_1 + \tau^r)^2 +
(1-\zeta)/\tau^\zeta \Bigr) + 3 \Lambda \tau, \label{force}$$ where $q_1 = \{\kappa,m,\lambda,\lambda_1,q,r,\zeta\}$ is the set of problem parameters. The En. allows the following first integral: $$\dot \tau = \sqrt{2[E - {{\mathcal}U}(q_1,\tau)]} \label{1stint}$$ where we denote $${{\mathcal}U}(q_1,\tau) = - \frac{3}{2}\Bigl[\kappa\Bigl(m \tau -
\lambda /\tau^{q - 2} - \lambda_1 /2(2 \lambda_1 + \tau^r) +
\tau^{1-\zeta}\Bigr) - \Lambda \tau^2\Bigr]. \label{poten}$$ From a mechanical point of view Eq. can be interpreted as an equation of motion of a single particle with unit mass under the force $\mathcal F(q_1,\tau)$. In $E$ is the integration constant which can be treated as energy level, and ${{\mathcal}U}(q_1,\tau)$ is the potential of the force $\mathcal F(q_1, \tau)$. We solve the Eq. numerically using Runge-Kutta method. The initial value of $\tau$ is taken to be a reasonably small one, while the corresponding first derivative $\dot \tau$ is evaluated from for a given $E$. As one sees, the positivity of the radical imposes some restriction on the value of $\tau$, namely in case of $\lambda > 0$ and $q \ge 2$ the value of $\tau$ cannot be too close to zero at any space-time point. It is clearly seen from the graphical view of the potential \[cf. Fig. \[potenf\]\]. Thus we can conclude that for some special choice of problem parameters the introduction of nonlinear spinor field given by a self-action provides singularity-free solutions. For numerical solutions we set $\kappa = 1$, spinor mass $m = 1$, the power of nonlinearity we choose as $q = 4$, $ r = 4$ and for perfect fluid we set $\zeta = 1/3$ that corresponds to a radiation. Here, in the figures we use the following notations:\
1 corresponds to the case with self-action and interaction, i.e., $\lambda = 1$, $\lambda_1 = 1$;\
2 corresponds to the case with self-action only, i.e., $\lambda = 1$, $\lambda_1 = 0$;\
3 corresponds to the case with interaction only, i.e., $\lambda = 0$, $\lambda_1 = 1$.
1 cm
As one sees from Fig. \[potenf\], in presence of a self-action of the spinor field, there occurs an infinitely high barrier as $\tau \to 0$, it means that in the case considered here $\tau$ cannot be trivial \[if treated classically, the Universe cannot approach to a point unless it stays at an infinitely high energy level\]. Thus we see, the nonlinearity of the spinor field provided by the self-action generates singularity-free evolution of the Universe. But, as it was shown in [@sprd], this regularity can be achieved only at the expense of dominant energy condition in Hawking-Penrose theorem. It is also clear that if the nonlinearity is induced by a scalar field, $\tau$ may be trivial as well, thus giving rise to space-time singularity. It should be noted that introduction of a positive $\Lambda$ just accelerates the speed of expansion, whereas, a negative $\Lambda$ depending of the choice of $E$ generates oscillatory or non-periodic mode of evolution. These cases are thoroughly studied in [@sprd; @bited]. As it was shown in [@bited] the regular solution obtained my means of a negative $\Lambda$ is case of interaction does not result in broken dominant energy condition. In Fig. \[enpr\] we plot the effective energy density and effective pressure of the matter field. In case of self-action pressure is initially positive, but with the expansion of the Universe it becomes negative. In case of interaction field the pressure is always negative. It means, the models with nonlinear spinor field and interacting spinor and scalar fields can to some extent explain the late time acceleration of the Universe. As one sees, the corresponding quantities (potential, energy density and pressure) differs only at the initial stage depending on the type of nonlinearity.
1.5 cm
In Fig. \[H\] we graphically justified our assumptions about the role of $\Lambda$ term, namely, in absence of the cosmological constant, $H \to 0$ as the Universe expands. Finally, in Fig. \[tauf\] we illustrate the evolution of the Universe. As one sees, the character of evolution differs only at the initial stage depending on the choice of nonlinearity.
Finally we would like to emphasize that here we restrict within three cases only. Cases with nontrivial $\Lambda$ term is not considered, since they were thoroughly studied in previous papers [@lambda; @sprd; @bited]. Our main aim here was to emphasize the new role of spinor field to explain the late time acceleration of the Universe.
conclusion
==========
We considered a system of interaction nonlinear spinor and scalar fields within the scope of a BI cosmological model filled with perfect fluid. It is shown that the spinor field nonlinearity can generate a negative effective pressure, which can be seen as an alternative source for late time acceleration of the Universe. Here, beside spinor and scalar fields, we consider usual perfect fluid obeying $p_{pf} = \zeta {\varepsilon}_{pf}$. We plan to consider a few other fluids in near future that can provide an initial inflation as well.
[99]{}
Padmanabhan, T., Phys. Rep. [**380**]{} (2003) 235.
Sahni, V., [*Dark Matter and Dark Energy*]{} arXiv (2004) astro-ph/0403324.
Saha, Bijan, [*Anisotropic cosmological models with a perfect fluid and a $\Lambda$ term*]{} (accepted for publication in Astrophysics and space science) arXiv (2004) [gr-qc/0411080]{}[http://xxx.lanl.gov/abs/gr-qc/0411080]{}.
Cladwell, R.R., Dave, R., and Steinhardt, P.J., Phys. Rev. Lett. [**80**]{} (1998) 1582.
Sahni, V. and Starobinsky, A.A., Int. J. Mod. Phys. D [**9**]{} (2000) 373.
Zlatev, I., Wang, L., and Steinhardt, P.J., Phys. Rev. Lett. [**82**]{} (1999) 896.
Saha, Bijan, [*Anisotropic cosmological models with perfect fluid and dark energy revisited*]{} arXiv (2005) [gr-qc/0501067]{}[http://xxx.lanl.gov/abs/gr-qc/0501067]{}.
Kamenshchik, A.Yu., Moschella, U., and Pasquier, V., Phys. Lett. B. [**511**]{} (2001) 265.
Saha, Bijan, [*Anisotropic cosmological models with perfect fluid and dark energy*]{} (accepted for publication in Chinese Journal of Physics) arXiv (2004) [gr-qc/0412078]{}[http://xxx.lanl.gov/abs/gr-qc/0412078]{}.
Saha, B. and Shikin, G.N., J. Math. Phys. [**38**]{} (1997) [5305]{}[http://www.jinr.ru/ bijan/my\_papers/JMP05305.pdf]{}.
Saha, B. and Shikin, G.N., Gen. Relativ. Gravit. [**29**]{} (1997) [1099]{}[http://www.jinr.ru/ bijan/my\_papers/grg97\_1099.pdf]{}.
Saha, Bijan, Mod. Phys. Lett. A [**16**]{} (2001), [1287]{}[http://www.jinr.ru/ bijan/my\_papers/mpla01\_1287.pdf]{}.
Saha, Bijan, Phys. Rev. D [**64**]{} (2001) [123501]{}[http://www.jinr.ru/ bijan/my\_papers/PRD23501.pdf]{}.
Saha, Bijan and Boyadjiev, T., Phys. Rev. D [**69**]{} (2004) [124010]{}[http://www.jinr.ru/ bijan/my\_papers/PRD24010.pdf]{}.
Armend$\acute a$riz-Pic$\acute o$n, C. and Greene, P.B., Gen. Relativ. Gravit. [**35**]{} (2003) 1637.
Ribas, M.O., Devecchi, F.P., and Kremer, G.M., [*Fermions as sources of accelerated regimes in cosmology*]{} ArXiv (2005) gr-qc/0511099.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate a novel boundary condition for the $bc$ system with central charge $c=-2$. Its boundary state is constructed and tested in detail. It appears to give rise to the first example of a local logarithmic boundary sector within a bulk theory whose Virasoro zero modes are diagonalizable. 0.1cm -0.3cm PACS numbers: 11.25. HF; 11.25.-w; 11.25. Sq.'
author:
- 'Thomas Creutzig, Thomas Quella, Volker Schomerus'
date: December 2006
title: 'New boundary conditions for the $c=-2$ ghost system'
---
[DESY 06-219 NSF-KITP-06-122]{}\
[KCL-MTH-06-15 hep-th/0612040]{}\
Introduction
============
The $bc$ system with Virasoro central charge $c=-2$, and the closely related symplectic fermion model, have been studied extensively in the past, both in the bulk and on the boundary (see e.g. [@Kausch:2000fu; @Gaberdiel:2001tr; @Flohr:2001zs; @Gaberdiel:2006pp] and references therein). Most of the past work was driven by formal questions in the context of logarithmic conformal field theory. Recently, it was pointed out [@Schomerus:2005bf; @Gotz:2006qp; @Quella:2007] that the $bc$ system at $c=-2$ also plays a crucial role for the solution of WZNW models on supergroups. In fact, it enters through a Kac-Wakimoto type representation of such theories. The latter reduces the solution of the WZNW model on superspaces to that on the corresponding bosonic base. In order to extend such a free fermion representation to the boundary sector, we have to impose boundary conditions on the $bc$ system. Surprisingly, it turns out that the relevant Neumann-type boundary theory has not been discussed in the existing literature. We shall fill this gap below.
As the name indicates, the $bc$ system involves two sets of chiral bulk fields $c,\bar c$ and $b,\bar b$ of conformal dimension $h_c = 0$ and $h_b = 1$, respectively. In the conventional setup, we would glue $c$ to $\bar c$ and $b$ to $\bar b$ along the boundary [@Callan:1987px]. But for $c=-2$ there exists another possibility: namely, to glue $b$ to a derivative of $\bar c$ and vice versa. More precisely, we can demand that $$\label{bc}
b(z) \, = \, \mu \bar \partial \bar c(\bar z) \ \ , \ \
\bar b(\bar z) \, = \, - \mu \partial c(z) \ \ \text{ for }
\ z = \bar z\ \ .$$ These relations guarantee trivial gluing conditions for the energy momentum tensor $T=-b\partial c$. It is not difficult to check that the action of the $bc$ system is invariant under variations respecting (\[bc\]) provided we add an appropriate boundary term, $$\label{act}
S \ = \ \frac{1}{2\pi} \int d^2z \left[ b\,\bar \partial c +
\bar b\,\partial\bar c \right]
- \frac{i\mu}{4\pi} \int du \ c\,\partial_u \bar c \ \ .$$ Our aim here is to solve the theory that is defined by the action (\[act\]) and the boundary condition (\[bc\]). We shall set $\mu =1$ throughout our discussion. Formulas for the general case are easily obtained from the ones we display below.
Solution of the boundary theory
===============================
In order to construct the state space and the fields explicitly, we introduce an algebra that is generated by the modes $c_n, b_n$ and two additional zero modes ${\xi^b}_0, {\xi^c}_0$ subject to the conditions $$\begin{aligned}
\label{comm}
\{ c_n,b_m\} & = & n\,\delta_{n,-m} \ \ ,\\[2mm]
\{ {\xi^c}_0,b_0 \} \ = \ 1 \ \ \ \ & , & \ \ \ \ \{{\xi^b}_0 , c_0 \} \ = \ 1 \ \ .\end{aligned}$$ All other anti-commutators in the theory are assumed to vanish. The state space of our boundary theory is generated from a ground state with the properties $$\label{vac}
c_n | 0\rangle \ = \ b_n |0\rangle\ = \ 0
\ \ \ \text{ for } \ \ \ n \ \geq \ 0$$ by application of ‘raising operators’, including the zero modes ${\xi^b}_0$ and ${\xi^c}_0$. On this space we can introduce the local fields $c,\bar c,
b,\bar b$ through the prescription $$\begin{aligned}
\label{b6}
b(z) &= & \ \sum_{n\in{Z}} b_n z^{-n-1} \\[2mm]
c(z) & = & \sum_{n\neq0} \frac{c_n}{n} z^{-n} +
c_0 \ln z + {\xi^c}_0\\[2mm]
\bar b(\bar z) & = & \sum_{n\neq 0} c_n \bar z^{-n-1} - c_0 \bar
z^{-1}
\\[2mm]
\bar c(\bar z) & = & - \sum_{n\neq 0} \frac{b_n}{n} \bar z^{-n} +
b_0 \ln \bar z - {\xi^b}_0 \label{bc9} \end{aligned}$$ It is not difficult to check with the help of eqs. (\[comm\]) that these fields satisfy the correct local anti-commutation relations $$\bigl\{ b(z), c(w) \bigr\} \ = \ \delta(z-w) \ \ , \ \
\bigl\{ \bar b(\bar z),\bar c(\bar w) \bigr\} \ = \
\delta(\bar z- \bar w)\ $$ in the interior of the upper half plane. Needless to stress that they also fulfill our boundary conditions (\[bc\]) with $\mu = 1$.
For later use let us also spell out the construction of the Virasoro generators in terms of fermionic modes, $$L_n \ = \ \sum_{m \neq 0} \, :b_{n-m} c_m : - b_n
c_0\ \ .$$ It is important to stress that – due to the term $c_0 b_0$ – the element $L_0$ satisfies $L_0 {\xi^c}_0{\xi^b}_0 |0\rangle =
|0\rangle$. Since $L_0$ vanishes on all other ground states, it is non-diagonalizable. In other words, our boundary theory is an example of a logarithmic conformal field theory. The logarithms in this model, however, are restricted to the boundary sector since the Hamiltonian of the bulk theory is diagonalizable (see below).
Computations of correlation functions in our boundary theory, require to introduce a dual vacuum with the properties $$\begin{aligned}
\label{cav}
\langle 0| c_n \ = \ \langle 0| b_n & = & 0
\ \ \ \text{ for } \ \ \ n \ \leq \ 0 \\[2mm]
\langle 0 | {\xi^c}_0 {\xi^b}_0 |0\rangle
& = & 1 \ \ \ \ . \label{cavn} \end{aligned}$$ For the c=-2 ghost system the ground states $|0\rangle$ and $\langle
0|$ which are annihilated by the zero modes $b_0$ and $c_0$ are at the same time $SL(2,C)$ invariant vacua. Consistency with the commutation relations requires $\langle 0|0\rangle = \langle 0|\{b_0,{\xi^c}_0\}
|0\rangle =0$. Therefore, the simplest non-vanishing quantity is $\langle 0|{\xi^c}_0{\xi^b}_0|0\rangle$ for our boundary theory (see also [@Flohr:2001zs] for a more detailed discussion).
Finally, we would like to display the boundary state $|N\rangle$ for our new boundary condition. Before we provide explicit formulas let us briefly recall that the bulk fields are obtained as $$c(z) \ = \ {\xi^c}_0 + \sum_{n \neq 0} \, \frac{c_n}{n} \, z^{-n}
\ \ \ , \ \ \ b(z) \ = \ \sum_{n\in Z} \, b_n \, z^{-n-1}$$ and similarly for their anti-holomorphic counterparts. Note that there are no modes $c_0,\bar c_0$ and ${\xi^b}_0,\bar{\xi^b}_0$ in the bulk of our $bc$ ghost system. This feature distinguishes the $c=-2$ ghosts from the closely related symplectic fermions. According to the standard rules, the boundary state for our boundary theory must satisfy the following Ishibashi conditions [@Ishi] $$\label{Ish1}
(b_n - \bar c_{-n}) |N\rangle \ = \ 0 \ \ , \ \ (c_n + \bar
b_{-n}) |N\rangle \ = \ 0 \ $$ for $n\neq 0$ and $b_0 |N\rangle = \bar b_0 |N\rangle = 0$. As one may easily check, the unique solution to these conditions is given by $$\label{BS}
\frac{|N\rangle}{\sqrt{2\pi}} \, = \,
\exp\left(-\sum_{m=1}^\infty\,
\frac{1}{m} ( c_{-m} \bar c_{-m} +
b_{-m} \bar b_{-m}) \right) |0\rangle$$ where $|0\rangle$ is a state in the bulk theory that satisfies conditions of the form (\[vac\]) for both chiral and anti-chiral modes. There also exists a dual boundary state $\langle N|$, satisfying the conditions $$\langle N| (b_n + \bar c_{-n}) \ = \ 0 \ \ , \ \ \langle N | (c_n
- \bar b_{-n}) \ = \ 0 \ $$ for $n\neq 0$ and $\langle N| b_0 = \langle N| \bar b_0 = 0$. These linear relations are related to eqs. (\[Ish1\]) by conjugation using that $c_n^* = - c_{-n}$ for $n \neq 0$ and $b^*_n = b_{-n}$ etc. The dual boundary state is given by the following explicit formula $$\label{dBS}
\frac{\langle N|}{\sqrt{2\pi}} \, = \,
\langle 0|\, \exp\left(\sum_{m=1}^\infty \,
\frac{1}{m}\, ( c_{m} \bar c_{m} +
b_{m} \bar b_{m}) \right)$$ involving a dual closed string ground state $\langle 0|$ that obeys conditions of the form (\[cav\]) for modes of chiral and anti-chiral fields and that is normalized by $\langle 0 |
{\xi^c}_0 \bar {\xi^c}_0 |0\rangle=1$ (see comments after eq. \[cavn\]).
Cardy consistency conditions
============================
Having constructed our new boundary theory, and in particular its boundary state, we would now like to perform two Cardy-like consistency tests. To begin with, let us verify that $|N\rangle$ satisfies world-sheet duality. We stress that in this note we consider a theory in which bulk and boundary theory consist of Ramond sectors only, a choice that we shall comment in more detail below. In such a model, world-sheet duality relates quantities that are periodic in both world-sheet space and time. The simplest such quantity in our boundary theory would be ${{\text{tr}}}[q^{L_0 + 1/12} (-1)^F]$ which vanishes since bosonic and fermionic states come in pairs on each level of the state space. The same is certainly true for $\langle N| \tilde q^{L_0 +
1/12} (-1)^{F} |N\rangle$, in agreement with world-sheet duality. In order to probe finer details of the theory, we need to consider quantities with additional insertions of fields or zero modes. Here, we shall establish the relation $$\begin{aligned}
\nonumber
& & {{\text{tr}}}\left(q^{H^o} (-1)^F c(z) \bar c(\bar z) \right) \ = \
\\[2mm] & & \hspace*{1cm} \ = \ \langle N|
\tilde q^{\frac12 H^c} (-1)^{\frac12 F^c} c(\xi) \bar
c(\bar \xi) |N\rangle \label{test1}\end{aligned}$$ where $H^o = L_0 + 1/12, q = \exp(2\pi i \tau), \xi =
\exp(-\frac{1}{\tau}\ln z)$ and $F^c = F + \bar F$, as usual. The closed string Hamiltonian is given by $$H^c \ = \ \sum_{m\in{Z}}\Bigl[:b_{-m} c_m: + :\bar b_{-m}
\bar c_m:\Bigr] + 1/6 \ \ .$$ Validity of eq. (\[test1\]) is required by the definition of boundary states (see e.g. [@Recknagel:1997sb]). Starting with the left hand side, it is rather easy to see that $$\begin{aligned}
{{\text{tr}}}\left(q^{L_0 + 1/12} (-1)^F c(z) \bar c(\bar z)
\right)\!\!&=&\!\!-{{\text{tr}}}\left(q^{L_0 + 1/12} (-1)^F {\xi^c}_0 {\xi^b}_0
\right) \nonumber \\[2mm] & & \hspace*{-3cm} \ = \
2 \pi i \tau \eta(q)^2 \ = \ - 2 \pi \eta(\tilde q)^2\ \ . \end{aligned}$$ In the computation we split off the term $c_0 b_0$ from $H^o$ and use it to saturate the fermionic zero modes. The rest is then straightforward. We can reproduce the same result if we insert our explicit formulas for the boundary states $|N\rangle$ and $\langle N|$ into the right hand side of eq. (\[test1\]).
It is possible to perform another similar test of our boundary theory using the usual trivial boundary conditions of the ghost system. In this case, the field $c(z)$ is identified with its own anti-holomorphic partner $\bar
c (\bar z)$ along the boundary and likewise for the pair $b$ and $\bar b$. Let us recall that the boundary state $|{{\text{id}}}\rangle $ and its dual $\langle {{\text{id}}}|$ take the form [@Callan:1987px] $$\begin{aligned}
\nonumber
|{{\text{id}}}\rangle & = &
\text{exp}\Biggl(\sum_{m=1}^\infty \frac{1}{m} \Bigl(c_{-m}\bar{b}_{-m} +
\bar{c}_{-m} b_{-m}\Bigr)\Biggr)({\xi^c}_0-\xi^{\bar c}_0)|0\rangle
\\[2mm] \langle {{\text{id}}}|&=&i\langle0|({\xi^c}_0-\xi^{\bar c}_0)\,\text{exp}
\Biggl(\sum_{m=1}^\infty\, \frac{1}{m} \Bigl(\bar{b}_{m}c_{m}+b_{m}
\bar{c}_{m}\Bigr)\Biggr)\end{aligned}$$ where we use the same notations as before. For the exchange of closed string modes between $|N\rangle$ and $\langle {{\text{id}}}|$ the above formulas imply $$\begin{aligned}
\langle{{\text{id}}}|\, \tilde q^{\frac12 H^c }\, (-1)^{\frac12 F^c}\,
c(\xi) \, |N\rangle & = & \langle{{\text{id}}}|\,
\tilde q^{\frac12 H^c }\, (-1)^{\frac12 F^c}\, {\xi^c}_0 \, |N\rangle
\nonumber \\[2mm]
& & \hspace*{-4cm} \ = \
\sqrt{2\pi}\, \tilde{q}^{\frac{1}{12}}\prod_{n=1}^\infty
\bigl(1+\tilde{q}^{2n}\bigr)
\ = \
\sqrt{\frac{\pi \theta_2(2\tilde{\tau})}{\eta(2\tilde{\tau})}}\ \ .
\label{res1}\end{aligned}$$ Once more we had to insert the field $c(z)$ in order to get a non-vanishing result. For comparison with a world-sheet dual, we need to quantize the ghost system on the upper half plane with trivial boundary conditions on the positive real axis and our non-trivial ones on the other half. A moment of reflection reveals that the following combinations $$\begin{aligned}
\chi^+(z) & = & \frac{1}{\sqrt{2}}\bigl(b(z)+i\partial c(z)\bigr) \ \ ,
\\[2mm]
\chi^-(z) & = & \frac{1}{\sqrt{2}} \bigl(ib(z)+\partial c(z)\bigr)\end{aligned}$$ diagonalize the monodromy, i.e. they obey the following simple periodicity relations $\chi^\pm(e^{2\pi i} z) = \pm i
\chi^\pm(z)$. Hence, they take the following form [@Kausch:2000fu] $$\begin{aligned}
\chi^\pm(z) &= &\sum_{r\in{Z}\mp\frac{1}{4}}\chi_r^\pm\,z^{-r-1}\ \ .\end{aligned}$$ The modes $\chi^\pm_r$ satisfy the same canonical commutation relations, $\{\chi^+_r,\chi^-_s\}=r \delta_{r,-s}$, as before. Formulas for the Virasoro generators can easily be worked out. For us, it suffices to display the zero mode $\tilde L_0$, $$\tilde L_0 \ = \ -\sum_{r\in{Z}-\frac{1}{4}}:\chi^+_r\chi^-_{-r}:-
\frac{3}{32}\ \ .$$ The constant shift by $3/32$ is needed in order to obtain standard Virasoro relations with the other generators (see also [@Saleur:1991hk] for a closely related analysis of twisted sectors in the bulk theory). The state space of our boundary theory contains two ground states $|\Omega_\pm
\rangle$ which are related to each other by the action of a zero mode ${\xi^c}$. On this space we can introduce the field $c$ through $$c(z) \ = \ \sqrt{\pi} {\xi^c}+ \frac{i}{\sqrt2}
\sum _{r\in{Z}-\frac{1}{4}}\frac{\chi_r^+}{r} \,z^{-r}
- \frac{1}{\sqrt2} \sum _{r\in{Z}+\frac{1}{4}}\frac{\chi_r^-}{r}
\,z^{-r}\ .$$ From the construction of the state space and our formula for $\tilde H^o = \tilde L_0 + 1/12$ we infer the following expression for the mixed open string amplitude, $$\begin{aligned}
{{\text{tr}}}\Bigl(q^{\tilde H^o} (-1)^F c(z) \Bigr) & = &
\nonumber \\[2mm]
& & \hspace*{-3cm} \ = \ \sqrt{\pi}\, q^{-\frac{1}{96}}
\prod_{n=0}^\infty\Bigl(1-q^{\frac{1}{2}(n+1/2)}\Bigr)\ =\
\sqrt{\frac{\pi \theta_4(\tau/2)}{\eta(\tau/2)}},
\nonumber\end{aligned}$$ which reproduces exactly the previous result (\[res1\]) upon modular transformation and concludes our investigation of the new boundary theory. Let us remark that the same partition function was found recently in [@Jacobsen:2006bn] with the help of boundary loop models.
Conclusions and outlook
=======================
The choice of our new gluing condition for the $bc$ system was motivated by the interest in branes on supergroups. As we shall discuss in a forthcoming paper, maximally symmetric branes in a WZW model on a supergroup turn out to satisfy Neumann-type boundary conditions in the fermionic coordinates. This implies that all fermionic zero modes must act non-trivially on the space of open string states. In our toy model, the role of the fermionic coordinates is played by $c$ and $\bar c$. Hence, we needed to find boundary conditions with a four-fold degeneracy of ground states. For the standard boundary conditions of the $bc$ system, $c = \bar c$ along the boundary and hence only one fermionic zero mode survives, giving rise to a 2-dimensional space of ground states. In this sense, the usual boundary conditions of the $bc$ systems are localized in one of the fermionic directions. Our boundary conditions come with two non-vanishing zero modes ${\xi^b}_0$ and ${\xi^c}_0$ (and their dual momenta $c_0$ and $b_0$). This property makes them a good model for maximally symmetric branes on supergroups.
There exist various extensions of our theory that we want to briefly comment about. In our analysis we focused on the RR sector of the $bc$ ghost system in the bulk. It is certainly straightforward to include a NSNS sector in case this is required by the application. Furthermore, we can also replace the bulk theory by its logarithmic cousin, the symplectic fermion model. Since the formulas and results are very similar, we refrain from giving more details. Boundary theories for symplectic fermion theories have been studied extensively in the past (see e.g. [@Moghimi-Araghi:2000cx; @Kogan:2000fa; @Ishimoto:2001jv; @Kawai:2001ur; @Bredthauer:2002ct; @Kawai:2002fu; @Bredthauer:2002xb; @Gaberdiel:2006pp]). We would like to stress, however, that our boundary condition seems to be new, also in the context of symplectic fermions.
Let us be a bit more specific and relate our constructions to the results in [@Gaberdiel:2006pp]. A comparison of the gluing conditions shows that our state $|{{\text{id}}}\rangle$ is a close relative of the $(N,\pm)$ boundary condition of [@Gaberdiel:2006pp]. In fact, all boundary theories considered in [@Gaberdiel:2006pp] glue $\partial c $ to $\bar \partial\bar c$ and $b$ to $\bar b$, with different choices of signs. None of these models displays any enhancement of zero-modes in the boundary spectrum. In this sense, we would prefer to consider them all as being of the same type (mixed Dirichlet-Neumann in the context of the $bc$ system). Using the notations of [@Gaberdiel:2006pp], our new boundary theories arise when we glue $\chi^+$ to $\bar
\chi^-$ and vice versa. Such a choice gives rise to a non-trivial gluing automorphism on the so-called triplet algebra and therefore it was excluded from the analysis in [@Gaberdiel:2006pp].
In the case of the $bc$ ghost system, the boundary state $|N\rangle$ has a rather novel feature: it describes a logarithmic boundary theory in a non-logarithmic bulk. Put differently, the $bc$ ghost system possesses a diagonalizable bulk Hamiltonian $H^c$. Nevertheless, the Hamiltonian $H^o$ of our new boundary theory is non-diagonalizable. Hence, logarithmic singularities can appear, but [*only*]{} when two boundary fields approach each other. To the best of our knowledge, such a behavior has never been encountered before. It shows that conformal field theories may be logarithmic even if none of its correlators on the sphere contain logarithms.
We thank Matthias Gaberdiel, Andreas Recknagel, Sylvain Ribault, Ingo Runkel and Hubert Saleur for their useful comments and critical remarks. TQ acknowledges the warm hospitality at the KITP in Santa Barbara during the completion of this article. This work was partially supported by the EU Research Training Network grants “Euclid”, contract number HPRN-CT-2002-00325, “Superstring Theory", contract number MRTN-CT-2004-512194, and “ForcesUniverse”, contract number MRTN-CT-2004-005104, as well as by the PPARC rolling grant PP/C507145/1 and by the National Science Foundation under Grant No. PHY99-07949. Until September 2006 TQ has been funded by a PPARC postdoctoral fellowship under reference PPA/P/S/2002/00370.
[99]{}
H. G. Kausch, [*Symplectic fermions*]{}, Nucl. Phys. B [**583**]{} (2000) 513. M. R. Gaberdiel, [*An algebraic approach to logarithmic conformal field theory*]{}, Int. J. Mod. Phys. A [**18**]{} (2003) 4593. M. Flohr, [*Bits and pieces in logarithmic conformal field theory*]{}, Int. J. Mod. Phys. A [**18**]{} (2003) 4497. M. R. Gaberdiel and I. Runkel, [*The logarithmic triplet theory with boundary*]{}, hep-th/0608184.
V. Schomerus and H. Saleur, [*The GL(1$|$1) WZW model: From supergeometry to logarithmic CFT*]{}, Nucl. Phys. B [**734**]{} (2006) 221. G. Götz, T. Quella and V. Schomerus, [*The WZNW model on PSU(1,1$|$2)*]{}, arXiv:hep-th/0610070. T. Quella and V. Schomerus, in preparation
C. G. . Callan, C. Lovelace, C. R. Nappi and S. A. Yost, [*Adding holes and crosscaps to the superstring*]{}, Nucl. Phys. B [**293**]{} (1987) 83. N.Ishibashi, [*The boundary and crosscap states in conformal field theories*]{}, Mod. Phys. Lett. [**A 4**]{} (1989) 251.
A. Recknagel and V. Schomerus, [*D-branes in Gepner models*]{}, Nucl. Phys. B [**531**]{} (1998) 185 \[arXiv:hep-th/9712186\]. H. Saleur, [*Polymers and percolation in two-dimensions and twisted N=2 supersymmetry*]{}, Nucl. Phys. B [**382**]{} (1992) 486 \[arXiv:hep-th/9111007\]. J. L. Jacobsen and H. Saleur, [*Conformal boundary loop models*]{}, arXiv:math-ph/0611078. S. Moghimi-Araghi and S. Rouhani, [*Logarithmic conformal field theories near a boundary*]{}, Lett. Math. Phys. [**53**]{} (2000) 49 \[arXiv:hep-th/0002142\]. I. I. Kogan and J. F. Wheater, [*Boundary logarithmic conformal field theory*]{}, Phys. Lett. B [**486**]{}, 353 (2000) \[arXiv:hep-th/0003184\]. Y. Ishimoto, [*Boundary states in boundary logarithmic CFT*]{}, Nucl. Phys. B [**619**]{}, 415 (2001) \[arXiv:hep-th/0103064\]. S. Kawai and J. F. Wheater, [*Modular transformation and boundary states in logarithmic conformal field theory*]{}, Phys. Lett. B [**508**]{}, 203 (2001) \[arXiv:hep-th/0103197\]. A. Bredthauer and M. Flohr, [*Boundary states in c = -2 logarithmic conformal field theory*]{}, Nucl. Phys. B [**639**]{}, 450 (2002) \[arXiv:hep-th/0204154\]. S. Kawai, [*Logarithmic conformal field theory with boundary*]{}, Int. J. Mod. Phys. A [**18**]{}, 4655 (2003) \[arXiv:hep-th/0204169\]. A. Bredthauer, [*Boundary states and symplectic fermions*]{}, Phys. Lett. B [**551**]{} (2003) 378 \[arXiv:hep-th/0207181\].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Static flexoelectric effect in a finite sample of a solid is addressed in terms of phenomenological theory for the case of a thin plate subjected to bending. It has been shown that despite an explicit asymmetry inherent to the bulk constitutive electromechanical equations which take into account the flexoelectric coupling, the electromechanical response for a finite sample is “symmetric”. “Symmetric” means that if a sensor and an actuator are made of a flexoelectric element, performance of such devices can be characterized by the same effective piezoelectric coefficient. This behavior is consistent with the thermodynamic arguments offered earlier, being in conflict with the current point of view on the matter in literature. This result was obtained using standard mechanical boundary conditions valid for the case where the polarization vanishes at the surface. It was shown that, for the case where there is the polarization is nonzero at the surface, the aforementioned symmetry of electromechanical response may be violated if standard mechanical boundary conditions are used, leading to a conflict with the thermodynamic arguments. It was argued that this conflict may be resolved when using modified mechanical boundary conditions. It was also shown that the contribution of surface piezoelectricity to the flexoelectric response of a finite sample is expected to be comparable to that of the static bulk contribution (including the material with high values of the dielectric constant) and to scale as the bulk value of the dielectric constant (similar to the bulk contribution). This finding implies that if the experimentally measured flexoelectric coefficient scales as the dielectric constant of the material, this *does not imply* that the measured flexoelectric response is controlled by the static bulk contribution to the flexoelectric effect.'
author:
- 'Alexander K. Tagantsev'
- 'Alexander S. Yurkov'
title: Flexoelectric effect in finite samples
---
Introduction
============
The flexoelectric effect consists of a linear response of the dielectric polarization to a strain gradient. This is a high-order electromechanical which is expected, in general, to be rather weak. However, some of its features make this effect to be of interest from both fundamental and applied points of view. This has stimulated recent intensive experimental [@Ma; @Cross; @Zubko]and theoretical [@Sharma; @Maranganti; @Eliseev]activity in the field.
On the fundamental side, it is of interest that this effect cannot be considered just as a non-local generalization of the piezoelectric effect. In contrast to the later, the flexoelectric effect (response) is controlled by 4 mechanisms of different physical nature, the contributions of which can be comparable[@Tag1985]. On the practical side, of first importance is that this effect, in contrast to the piezoelectric effect, is allowed in centro-symmetric material. It is believed that it is the flexoelectric effect that is responsible for the generation of an electric field in acoustic shock waves propagating in centro-symmetric solids[@Harris]. It was recently shown that this effect plays also an essential role in electromechanical properties of materials with a moderate level of electronic and ionic conductivity[@Moroz2011]. However, the most applied interest is focused on the “piezoelectric metamaterial” – composites made of non-piezoelectric components, which exhibit effective piezoelectric response generated due to the flexoelectric effect. The work in this direction was initiated by pioneering experimental studies by Professor Cross with coworkers [@Ma; @Cross] and was later also supported by theory [@Sharma]. Presently, $\text{(Ba,Sr)TiO}_3$-based composites have been shown to yield effective piezoelectric coefficients comparable to those of commercial piezoelectric ceramics [@Chu]. It was argued, based on the constitutive equation for the flexoelectric response, that a mechanical sensor made of such metamaterials should exhibit a very unusual property. Specifically, in contrast to piezoelectric based devices, it will not behave as an actuator[@Chu; @Cross]. There are several reasons to question such statement. First, already in the 60’s of the past century, the group of Professor Bursian reported experimental data on $\text{BaTiO}_3$ crystals[@Bursian1968] and gave arguments based on equilibrium thermodynamics[@Bursian1974], which contradict this statement. Second, the existence of a linear sensor-not-actuator may come into conflict with the general principles of thermodynamics.
The goal of this paper is to address theoretically this conflict situation to demonstrate that despite an explicit asymmetry of the constitutive equations for the bulk flexoelectric effect, when this effect is characterized in a realistic finite sample the apparent asymmetry of the electromechanical response will vanish. In particular, this implies that the aforementioned piezoelectric metamaterial should exhibit the identical piezoelectric constants when characterized in “direct” and “converse” regimes. We will demonstrate this for two leading contributions to the static flexoelectric response: the contribution of static bulk ferroelectricity and that of surface piezoelectricity.
Static bulk flexoelectricity {#bulk}
============================
In the present paper, being interested in the static or quasi-static situation, we are facing three contributions to the flexoelectric response, which are associated with (i) static bulk ferroelectricity, (ii) surface piezoelectricity, and (iii) surface flexoelectricity [@Tag1985]. In this section we will discuss the problem of the relation between the direct and converse effects for contribution (i), reserving the next section for contribution (ii). In view of vanishing practical importance of the surface flexoelectric effect (it is expected not to be enhanced on high-dielectric-constant materials which are the only ones suitable for applications) we will not be treating it in this paper.
The static bulk flexoelectric effect is customarily described by the following free energy (density) expansion (see e.g.[@Eliseev]): $$\label{F}
F = \frac{\chi^{-1}_{ij}}{2} P_i P_j -
\frac{\\f_{ijkl}}{2} \left(P_k\frac{\partial u_{ij}}{\partial x_l} -
u_{ij}\frac{\partial P_k}{\partial x_l}\right) + \frac{c_{ijkl}}{2} u_{ij}u_{kl}$$ where $P_i$ and $u_{ij}$ are the polarization vector and strain tensor, respectively, and where the Einstein summation convention is adopted. We will consider this thermodynamic potential as having the differential $dF=E_idP_i + \sigma_{ij}du_{ij}$. Then calculating the electric field, $E_i$, and stress tensor, $\sigma_{ij}$, as variational derivatives of the free energy of the sample, given by the integral of $F$ over its volume, one arrives at the following constitutive equations: $$\label{E}
E_k=\chi^{-1}_{kj}P_j - f_{ijkl}\frac{\partial u_{ij}}{\partial x_l}$$ and $$\label{sigma}
\sigma_{ij}= f_{ijkl}\frac{\partial P_k}{\partial x_l}+ c_{ijkl}u_{kl}.$$ The first equation describes a linear polarization response to strain gradient (direct flexoelectric effect). The second one describes the converse flexoelectric effect, implying that to get a “mechanical yield”, spatial inhomogeneity of the polarization is needed. From this, one might infer (as customarily done in relevant papers) that the application of a homogeneous electric field to a sample will not lead to its deformation. Even being nearly evident, in reality, the last statement is not correct.
![Plate of the material exposed to bending and the reference frame used in calculations.[]{data-label="fig1"}](cfig1-1.eps){width="8cm"}
Let us show this for the flexural mode. Consider, a (001) plate of a cubic material of thickness $h$ in the reference frame specified in Fig.\[fig1\], with the $X$ and $Y$ dimensions being $L$ and $b$ respectively. To make the analysis transparent, we allow only a cylindrical bending of the plate about $OX_2$ axis. To simplify the discussion further, we set, for the moment, $c_{1122}=c_{1133}=0$. In such simplified model, the plate bending is associated with ${\partial u_{11}}/{\partial x_3}\neq0$, whereas $u_{22}=u_{33}=0$ so that Eq. suggests the appearance of $P_3$ component of the polarization controlled by $f_{1133}$ component of the flexoelectric tensor. To address the reversibility of this effect, one should check if the application of an electric field normal to the plate will cause its bending. A straightforward way do this is to derive the equation of balance of the bending moment for the plate[@Tim] subjected to a homogeneous electric field $E$ normal to its suface by integrating Eq. across a $YZ$ cross-section of the sample: $$\label{Mom}
b\int_{-h/2}^{h/2}\sigma_{11}zdz =
bf_{1133}\int_{-h/2}^{h/2}\frac{\partial P_3}{\partial z}zdz+ bc_{1111}\int_{-h/2}^{h/2}u_{11}zdz.$$ At mechanical equilibrium, the lhs term must be equal to the minus the component of the mechanical moment of the external forces applied to the lefthand (with respect to the cross-section of the integration) part of the plate,$-M_2$. Without the first rhs term, this equation describes that bending of the sample caused by this moment. To identify the role of this term, we first evaluate it using integration by parts: $$\label{Int}
\int_{-h/2}^{h/2}\frac{\partial P_3}{\partial z}zdz =
-\int_{-h/2}^{h/2} P_3dz = -h\langle P_3 \rangle$$ where $\langle P_3 \rangle$ is the averaged polarization induced by the field $E_3$ in the bulk of the plate. In doing so we assume that the polarization changes continuously from its bulk value to zero on the plate boundary. If, however, one explicitly considers nonzero polarization at the sample surface (as was done previously in Ref. [@Eliseev]), then one should revise the traditional boundary conditions of the elasticity theory[@Yurkov]. We will return to this issue later in the paper. Since the spatial scale of the polarization variation at the interface is much smaller than the thickness of the plate, with a good accuracy $\langle P_3 \rangle \approx P$, where $P$ – polarization in the bulk. Thus, the equation for the moment balance can be rewritten as $$\label{Mom1}
-M_2/b+f_{1133}hP =c_{1111}\int_{-h/2}^{h/2}u_{11}zdz.$$
It is clear from this equation that the application of a homogeneous electric field to the plate is equivalent to that of an external bending moment. Thus, we conclude that a finite, mechanically free ($M_2=0$) sample, placed in a homogeneous electric field, will be bent. This conclusion is closely related to that drawn by Eliseev et al[@Eliseev]. These authors have shown that a ferroelectric plate with the out-of-plane orientation of the spontaneous polarization should exhibit spontaneous bending due to the flexoelectric coupling. It was found that this effect is controlled by a factor $\int_{-h/2}^{h/2}({\partial P_3}/{\partial z})zdz$ which was calculated using a numerical solution for the polarization profiles $P(z)$ in the sample. Here it is also worth mentioning that the bending effect addressed, though being proportional to a component of the *bulk* flexoelectric coefficient and the *bulk* value of the induced polarization, is actually controlled by forces applied to the surface of the plate.
It is instructive to illustrate quantitatively the “symmetry” of the direct and converse flexoelectric effects in a finite sample for a situation which is readily mathematically trackable. We will consider the case of a $(001)$ plate of a cubic material in symmetrical flexural mode with the polarization $P$ normal to the plate and homogenous in its bulk. In the case of symmetric bending, the curvature of the plate in all crossections normal to it, $G$, is the same. Enjoying the results of the theory of thin plates [@Landau] we can express the components of the strain tensor in terms of this curvature: $$\label{strain}
u_{11}=u_{22}=zG;~~~ u_{33}=-z\frac{c_{12}}{c_{11}}G;~~~ u_{12}=u_{23}=u_{13}=0.$$ Integrating the free energy density, Eq., with strain coming from Eq., over the plate thickness, one finds the free energy density per unit area of the plate as a function of $P$ and $G$: $$\label{Fi}
\Phi=\frac{\chi^{-1}_{33}}{2}h P^2 + \frac{D_s}{2}G^2 - 2hPG(f_{1133} - \frac{c_{12}}{c_{11}}f_{1111}).$$ $$\label{Ds}
D_s= \frac{h^3}{6}\frac{c_{11}^2+c_{11}c_{12}-2c_{12}^2}{c_{11}}.$$ where $D_s$ is a coefficient controlling the flexural rigidity of the plate for this kind of bending. In derivation of Eq., we have again used Eq. (this gave a factor of 2 in the coupling term from this equation). Similar expression for the free energy of flexoelectric plate in the cylindrical bending mode was offered by Bursian and Trunov [@Bursian1974], based on purely symmetry arguments. Minimizing Eq. with respect to $P$ and $G$ one finds the equation for the direct and converse effects for the plate in symmetric flexural mode: $$\label{G}
G=\frac{2h}{D_s}\mu_{\text{pl}}E .$$ $$\label{P}
D=P= 2\mu_{\text{pl}}G.$$ where the electric displacement, $D$, and $$\label{mu}
\mu_{\text{pl}}=\chi_{33}\frac{c_{11}f_{1133} - c_{12}f_{1111}}{c_{11}}$$ can be treated as an effective flexoelectric coefficient of the plate. The flexural response, given by Eq., is compatible with the results obtained by Eliseev et al [@Eliseev] for the case of spontaneous bending of thin plates with the blocking boundary condition for the polarization.
Obviously, elements of such plate will work as both actuators and sensors. If round pieces of the plate with central loading and symmetric free-edge side support are used as elements of a piezoelectric metamaterial, electromechanical properties of the latter will be characterized by a single effective piezoelectric coefficient $d_{33}$. Using the relation between the cross-section curvature, $G$, and the maximal deflection, $\xi_{\text{max}}$, for symmetric bending of a circular plate: $$\label{xi}
\xi_{\text{max}}= \frac{GR^2}{2}.$$ (where $R$ is the radius of the plate) and using Eq. one readily finds $$\label{d33}
d_{33}= \frac{\mu_{\text{pl}}R^2}{D_s}.$$
Contribution of surface piezoelectricity {#surface}
========================================
As was recognized at the first thorough treatment of the flexoelectric response [@Tag1985], the polarization response to a strain gradient in a finite sample, generally speaking, may not be fully controlled by the contribution of the bulk static flexoelectricity, even in materials with high values of the dielectric constant (high-$K$ materials). The competing effect that is due to surface piezoelectricity, was not, however, properly addressed theoretically. Thus, it is not clear if it can in fact compete in high-$K$ materials with the static bulk flexoelectricity. In high-$K$ materials, the contribution of the static bulk flexoelectricity is enhanced, since it scales as the dielectric susceptibility (cf. Eq. and Eq.). At the same time, the effect associated with surface piezoelectricity originates from the presence of the interface adjacent layers where the piezoelectricity is induced by the inversion-symmetry-breaking effect of the interface. Since the sign of the effective piezoelectric coefficients of the layers on the opposite sides of a plate should be opposite (as controlled by the orientation of the surface normal), bending of the plate should result in dipole moments in these layers, the sign of which are the same. The dipole moment in a layer is proportional to the strain in it, which, in turn, is proportional to the product of the strain gradient and the plate thickness. Having calculated the resulting change of the average polarization of the whole system, this will give rise to a net polarization proportional to the strain gradient. From this reasoning it is not obvious that such response will be enhanced once that the dielectric constant of the bulk of the material is high. However, such reasoning does not provide a proper vision of the whole effect. In what follows, we will show that such enhancement does take place and the considered bulk and surface contributions to the flexoelectric response can be readily comparable in high-$K$ materials. We will also address the problem of the relation between the direct and converse effect for the mechanism related to the surface piezoelectricity.
![Model for the contribution of surface piezoelectricity to the flexoelectric response of a non-piezoelectric material. The surface layers of thickness $\lambda$ model the surface adjacent (atomically thin) layers of the material where the piezoelectricity is induced by the symmetry breaking impact of the surface.[]{data-label="fig2"}](cfig2-1.eps){width="8cm"}
To be specific we will address these problems for the case of symmetric bending of a thin $(001)$ plate of a cubic material. We will model the effect by considering a system consisting of a plate of an “ideally homogenous” material (i.e. its material parameters are the same throughout the plate) with the bulk flexoelectric effect being neglected and two thin surface piezoelectric layers (Fig.\[fig2\]). The thickness of each layer, $\lambda$, is much smaller that that of the plate, $h$. The top layer is characterized by piezoelectric moduli $h_{333}$ and $h_{311}=h_{322}$, whereas for the bottom layer these moduli have the same the absolute value but are of opposite sign. We also ascribe to these layers an out-of-plane component of the dielectric constant, $\varepsilon_\lambda$.
Let us find the extra free energy associated with the top piezoelectric layer when the plate is symmetrically bent with a cross-sectional curvature, $G$, and when out-of-plain polarization in the layer equals $P_{\lambda}$. We assume no difference in the elastic properties of the piezoelectric layers and the plate.
We start with the free energy density in the layer defined as $$\label{FL}
F_{\lambda} = \frac{\alpha}{2} P_{\lambda}^2 -
P_{\lambda}[h_{333}u_{33}+h_{311}(u_{11}+u_{22})]+
\frac{c_{11}}{2}(u_{11}^2+u_{22}^2+u_{33}^2)+ c_{12}(u_{11}u_{33}+u_{22}u_{11}+u_{22}u_{33})$$ where $c_{11}=c_{1111}$, $c_{12}=c_{1122}$, and $\alpha$ is the inverse dielectric susceptibility of the layer if it were fully mechanically clamped. Because the plate is thin we set in the layer $$\label{strainL}
u_{11}=u_{22}=hG/2.$$ As for $u_{33}$, we find it from the condition that the surface of the film is mechanically free, $\partial F_{\lambda}/\partial u_{33}=0$: $$\label{u33}
u_{33} = -\frac{c_{12}}{c_{11}}hG + \frac{h_{333}}{c_{11}}P_{\lambda}.$$ Inserting Eqs. and ) into Eq. and multiplying the result with $\lambda$, one finds the free energy of top piezoelectric layer. The energy of the bottom piezoelectric layer is the same since this layer differs from the top layer by the sign of piezo-moduli and by that of the strain; in the expression for energy these signs cancel each other out. Thus, for the contribution of the two piezoelectric layers to the free energy of the system, we find: $$\label{FIL}
\Phi_{\lambda} = 2 \lambda \left[\frac{\chi_{\lambda}^{-1}}{2} P_{\lambda}^2 -
(h_{311}-\frac{c_{12}}{c_{11}}h_{333})hGP_{\lambda}\right]+\Phi_1$$ where $\chi_{\lambda} = (\alpha - h_{333}^2/c_{11})^{-1}$is the true (under the mixed mechanical conditions) dielectric susceptibility of the piezoelectric layers and $\Phi_1$ is their mechanical bending energy.
To describe the direct flexoelectric response we use the equation of state for the polarization in the piezoelectric layer, $\partial \Phi_{\lambda}/\partial P_{\lambda}=E_{\lambda}$ ($E_{\lambda}$ is the electric filed in the layer), the condition of continuity of the electric displacement, $D$, in the layer, and the short-circuit condition. This leads to the following set of equations: $$\label{dir1}
P_{\lambda} = \chi_{\lambda}E_{\lambda}+ ehG$$ $$\label{dir2}
D=\varepsilon_f E_f = \varepsilon_0 E_{\lambda}+P_{\lambda}$$ $$\label{dir3}
2\lambda E_{\lambda} + h E_f = 0$$ where $E_f$, $\varepsilon_f$, $\varepsilon_0$ are the electric field in the bulk of the plate, its dielectric constant, and the dielectric constant of the free space, respectively, and $$\label{dir4}
e= \chi_{\lambda}(h_{311}-\frac{c_{12}}{c_{11}}h_{333}).$$ Solving this set of equations we find the relation for the direct flexoelectric response: $$\label{D}
D= 2 \widetilde{e}\lambda G.$$ where $$\label{e}
\widetilde{e}= e\frac{h \varepsilon_{f}}{2\lambda \varepsilon_{f} +h \varepsilon_{\lambda}}$$ and $\varepsilon_{\lambda}=\varepsilon_{0}+\chi_{\lambda}$.
To describe the converse flexoelectric response we present the elastic energy of the system at fixed $P_{\lambda}$ in the form $$\label{Fi1}
\Phi= \frac{D_s}{2}G^2 - 2h \lambda e G E_{\lambda}.$$ When writing this equation we have neglected the elastic energy of the surface layer, $\Phi_1$, compared to that of the plate and have simplified down to $P_{\lambda} \approx \chi_{\lambda}E_{\lambda}$. This approximation means that we neglect feedback effect of $G$ on $P_{\lambda}$. This effect will yield some renormalization of $D_s$, but practically such a renormalization is negligible indeed. Minimizing $\Phi$ with respect to $G$ and applying the electrostatic relation used above we arrive at the set of equations $$\label{con1}
G = \frac {2h \lambda e}{D_s}E_{\lambda}$$ $$\label{con2}
\varepsilon_f E_f = \varepsilon_{\lambda}E_{\lambda}$$ $$\label{con3}
2\lambda E_{\lambda} + h E_f = E(h+2\lambda)$$ where $E$ is the applied (average) electric field. This set leads us to the equation for the converse flexoelectric effect in the system: $$\label{G1}
G=\frac{2h}{D_s}\widetilde{e}\lambda E .$$ When writing this relation, only the leading terms to within a small parameter $\lambda/h$ were kept.
The following remarks are to be made concerning the results obtained. First, the relations obtained for the contribution of the surface piezoelectricity into the flexoelectric response, Eqs. and , are identical to those obtained for the case of static bulk flexoelectricity, Eqs. and , to within the replacement $\mu_{\text{pl}}\Rightarrow \widetilde{e}\lambda$. Thus, all conclusions about the direct-converse-effect symmetry drawn in the previous section for the static bulk flexoelectricity hold for the contribution associated with the surface piezoelectricity. Second, in a high-K material the latter contribution scales as its dielectric constant (similar to the case of the static bulk flexoelectricity). Formally, this follows from the expression for $\widetilde{e}$, Eq. . Taking into account that the thickness of the surface piezoelectric layer is expected to be of the order of the lattice constant, for realistic values of the plate thickness, $\widetilde{e}$ can be evaluated as $$\label{e1}
\widetilde{e}= e\frac{\varepsilon_{f}}{\varepsilon_{\lambda}}.$$ Note that there is no reason to consider $\varepsilon_{\lambda}$ in a high-K material (typically it is a “regular” of incipient ferroelectric in the paraelectric phase) to be high, since the special interplay of the atomic forces responsible for the high value of the bulk permittivity will be inevitably destroyed in the surface layer. Such enhancement of a surface driven effect by a factor of the bulk permittivity looks surprising. However, the physical mechanisms behind this effect can be identified.
For the converse effect it is quite transparent. The bending of the system is controlled by the value of the field in the surface layer (cf. Eq. ). Due to its small thickness, this field is enhanced by a factor of $\varepsilon_{f}/\varepsilon_{\lambda}$ , compared to the applied field.
For the direct effect, the explanation is less straightforward. This time, the bending creates polarization in the surface layer. Because of the inhomogeneity of the system, the short-circuiting does not guarantee the absence of the electric field in it so that the polarization in the surface layer induces a depolarizing field both in itself and in the bulk of the plate. It occurs that if the surface layer is thin enough whereas $\varepsilon_{f}/\varepsilon_{\lambda}$ is large, the polarization response is controlled by the depolarizing field in the bulk of the plate. This way the polarization response of the system becomes sensitive to the bulk value of the dielectric constant.
Another important conclusion is that, taking into account the aforementioned effect of enhancement, one expects both contributions to the flexoelectric response discussed to be of the same order of magnitude even in high-K materials. These contributions would be comparable if $\lambda \widetilde{e}/\varepsilon_{f}$ were about the typical value of the components of the flexoelectric tensor $f_{ijkl}$ , $1-10~\text{V}$ (see e.g.[@Eliseev]). For the “atomic values” of the entering parameters ($ \lambda=0.4 ~\text{nm}$ and $e =1~\text{C}/ \text{m}^2$) and $\varepsilon_{\lambda}/\varepsilon_{0}=10$, we evaluate $\lambda \widetilde{e}/\varepsilon_{f}\simeq 4~\text{V}$ to find that this is, in fact, the case.
Discussion and conclusions
==========================
The analysis presented clearly demonstrates that starting from continuous constitutive electromechanical equations one can derive the relations for the direct and converse flexoelectric effects in a finite sample, which exhibit the symmetry required by thermodynamics. On the practical level such symmetry implies that a piezoelectric meta-material based on the flexoelectric effect will exhibit the same effective piezoelectric coefficient in the testing regimes for the direct and converse piezoelectric effect. This is in conflict to the belief of those dealing with such meta-materials [@Chu; @Cross].
At the same time, it is also clear that the analysis, based on the constitutive electromechanical equations presented above, is valid only for the situation where the polarization at the plate surface can be treated as continuously changing from its bulk value to zero. If it is not the case, formally following this analysis one readily finds that the aforementioned symmetry is violated. For example, for the case of free boundary conditions for polarization (${\partial P}/{\partial z}=0$ at the boundary), corresponding to nonzero polarization at the sample surface, the equation of mechanical equilibrium, Eq. , implies the absence of the converse effect. Such conclusion would be fully consistent with that by Eliseev et al [@Eliseev] who argued that the manifestation of the converse flexoelectric effect in a plate is strongly dependent on the boundary conditions for the polarization. At the same time, there is no reason to expect that the free boundary conditions for the polarization will suppress the direct flexoelectric effect. Thus, if we followed the calculating scheme employed by Eliseev et al [@Eliseev] (and used in Sect.\[bulk\]) we would find, for the free polarization boundary conditions, the absence of the converse flexoelectric effect in the presence of the direct effect. This would make an apparent contradiction between the results obtained from the continuous constitutive equations and those obtained from thermodynamics.
We suggest the following resolution to this contradiction. The point is that incorporating the flexoelectric coupling into the free energy density of a material leads to a modification of the boundary conditions for the bulk constitutive equations. Eliseev et al [@Eliseev] have derived modified boundary conditions for the polarization, however these authors have postulated that the classical mechanical boundary conditions are not affected by such incorporation. However, as was recently shown by one of the authors[@Yurkov], generally, the mechanical boundary conditions should be modified as well. It has been shown that such boundary conditions reduce down to the classical mechanical boundary conditions when the polarization vanishes at the surface. This justifies the calculations based on the classical mechanical boundary conditions, which we have presented in Sect.\[bulk\]. For the general case, the problem of the converse flexoelectric effects should be revisited with the correct mechanical boundary conditions which contain the surface value of the polarization. We expect that such treatment will yield the results consistent with the symmetry between direct and converse flexoelectric effects in a finite sample dictated by thermodynamic arguments.
Another important conclusion follows from the results obtained in Sect.\[surface\]. There, it was shown that the contribution of the surface piezoelectricity to the flexoelectric response is expected to be comparable to that of the static bulk contribution (including the material with high values of the dielectric constant) and to scale with the bulk value of the dielectric constant (similar to the bulk contribution). The latter statement actually implies identical (or at least similar) temperature dependences of these contributions. Note that in earlier publications [@Tag1985] it was hypothesized that these depedences are expected to be different. Based on this hypothesis, the fact that the experimentally measured flexoelectric coefficient scales as the dielectric constant of the material was customarily taken as an indication that the measured flexoelectric response is controlled by the static bulk contribution to the flexoelectric effect. The results from Sect.\[surface\] essentially change the situation. Now one can state that the fact that the experimentally measured flexoelectric coefficient scales as the dielectric constant of the material *does not imply* that the measured flexoelectric response is controlled by the static bulk contribution to the flexoelectric effect.
Acknowledgements {#Acknowledgments}
================
This project was supported by Swiss National Science Foundation. Andrey Zakurdaev is acknowledged for reading the manuscript.
[999]{} W. Ma and L. E. Cross, Applied Physics Letters **78**, 2920 (2001). L. E. Cross, Journal of Materials Science **41**, 53 (2006). P. Zubko, G. Catalan, A. Buckley, P. R. L. Welche, and J. F. Scott, Physical Review Letters **99**, 167601 (2007). N. D. Sharma, R. Maranganti, and P. Sharma, Journal of the Mechanics and Physics of Solids **55**, 2328-2350 (2007). R. Maranganti and P. Sharma, Physical Review B **80**, 054109 (2009) E. A. Eliseev, A. N. Morozovska, M. D. Glinchuk, and R. Blinc, Physical Review B **79**, 165433 (2009) A.K. Tagantsev, Sov. Phys. JETF, **61**, 1246 (1985); Physical Review B, **34**, 5883 (1986); Sov. Phys. Usp., **30**, 588 (1987); Phase Transitions, **35** 119 (1991). P. Harris, Journal of Applied Physics **36**, 739 (1965). A. N. Morozovska, E. A. Eliseev, A. K. Tagantsev, S. L. Bravina, L.-Q. Chen, and S. V. Kalinin, Physical Review B **83**, 195313 (2011). B. Chu, W. Zhu, N. Li, and L. E. Cross, Journal of Applied Physics **106**, 104109-3 (2009) E. V. Bursian and Zaikovskii.O.I, Sov. Phys. Solid State **10**, 1121 (1968). E. V. Bursian and N. N. Trunov, Sov. Phys. Solid State **16**, 760 (1974). A. S. Yurkov Pis’ma v ZhETF, **94**, 490 (2011), in Russian, translation to appear in JETP Letters. S. Timoshenko and S. Woinowsky-Kreiger, Theory of plates and shells (McGraw-Hill, Inc, New York, 1987) L. D. Landau and E. M. Lifshitz, Theory of Elasticity (Pergamon Press, Oxford, 1975)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper studies a simple class of zero-sum games played by two competing quantum players: each player sends a mixed quantum state to a referee, who performs a joint measurement on the two states to determine the players’ payoffs. We prove that an equilibrium point of any such game can be approximated by means of an efficient parallel algorithm, which implies that one-turn quantum refereed games, wherein the referee is specified by a quantum circuit, can be simulated in polynomial space.'
author:
- |
Rahul Jain John Watrous\
[*Institute for Quantum Computing and School of Computer Science*]{}\
[*University of Waterloo, Waterloo, Ontario, Canada*]{}\
[{rjain,watrous}@cs.uwaterloo.ca]{}
date: 'August 20, 2008'
title: |
**Parallel approximation of non-interactive\
zero-sum quantum games**
---
Introduction
============
The theory of games has been studied extensively in mathematics and in several other disciplines for which it has applications. In theoretical computer science, computational aspects of game theory represent an important focus of the field.
There are several settings of interest to quantum computation and quantum cryptography that are naturally modeled by [*quantum games*]{}, which are games involving the exchange and processing of quantum information. For instance, multi-prover quantum interactive proofs [@KobayashiM03; @CleveHTW04; @KempeKMTV08; @KempeKMV08] can be modeled as cooperative quantum games; quantum coin-flipping [@Ambainis01; @Kitaev02; @SpekkensR02; @Mochon04; @Mochon07] is naturally modeled as a game between two players that directly exchange quantum information; and quantum refereed games [@Gutoski05; @GutoskiW05; @GutoskiW07] are competitive games that model quantum interactive proofs with competing provers.
In this paper we consider a simple type of non-interactive, zero-sum quantum game: two competing players (hereafter called [*Alice*]{} and [*Bob*]{}) each send a mixed quantum state to a [*referee*]{}, who performs a joint measurement on the two states to determine the players’ payoffs. For a fixed description of the referee, let $\phi(\rho,\sigma)$ denote Alice’s expected payoff when she sends a mixed state $\rho$ to the referee and Bob sends a mixed state $\sigma$. (For zero-sum games, Bob’s payoff is then given by $-\phi(\rho,\sigma)$.) The theory of quantum information requires the function $\phi(\rho,\sigma)$ to be bilinear, from which it follows that $$\label{eq:min-max-informal}
\max_{\rho} \min_{\sigma} \phi(\rho,\sigma)
= \min_{\sigma} \max_{\rho} \phi(\rho,\sigma)$$ from well-known variants of the Min-Max Theorem. (Indeed, such a fact holds for a much more general class of quantum zero-sum games that can allow for many rounds of interaction among the referee and players [@GutoskiW07].) The value represented by the two sides of the equation is called the [*value*]{} of the game. An [*equilibrium point*]{} of such a game is a pair of quantum states $(\rho,\sigma)$ such that $$\max_{\rho'} \phi(\rho',\sigma)
= \phi(\rho,\sigma)
= \min_{\sigma'} \phi(\rho,\sigma'),$$ the existence of which follows from the equation . In other words, when one player plays one of the states of an equilibrium point, the other has no incentive to play a state different from the other state in the pair. (These notions are, of course, similar to those for classical zero-sum games, but with some technical differences due to the nature of quantum information. In particular, there is a continuum of pure strategies for quantum players, corresponding to what are known as pure quantum states.) An equilibrium point of a zero-sum quantum game, given as and explicit description of the referee’s measurement, can be efficiently computed by means of semidefinite programming.
The main result of this paper is an efficient [*parallel*]{} algorithm to find approximate equilibrium points of non-interactive zero-sum quantum games. For the case where the referee is specified by a quantum circuit rather than in explicit matrix form, this algorithm implies that the value of such a game can be approximated in polynomial space. More succinctly, it implies that the complexity class ${\textup{QRG}}(1)$ of problems having one-turn quantum refereed games is contained in ${\textup{PSPACE}}$.
Our algorithm is an example of the [*multiplicative weights update method*]{}, which is discussed in the papers [@AroraHK05a; @TsudaRW05], for instance, and is explained in detail in the PhD thesis of S. Kale [@Kale07]. This general method captures many previously discovered (and sometimes re-discovered) algorithms, and has origins in learning theory, game theory, and optimization. The specific formulation of our algorithm is a non-commutative extension of an (unpublished) algorithm of Rohit Khandekar and the first author (Rahul Jain) that approximates equilibrium points of classical games.
In the sections that follow, we give relevant definitions from the theory of quantum information, present the algorithm and its analysis, and discuss the containment ${\textup{QRG}}(1)\subseteq{\textup{PSPACE}}$ that follows. We also explain how the problem of finding equilibrium points of quantum games relates to the problem of approximating positive instances of semidefinite programs.
Preliminaries and definitions {#sec:definitions}
=============================
This section gives a brief summary of the quantum information-theoretic concepts that are needed in the paper, and then defines non-interactive zero-sum quantum games. A few additional definitions that will be helpful later in the paper are also discussed.
Basic quantum information-theoretic notions
-------------------------------------------
In this paper we require just a few basic concepts about quantum information; so it is not necessarily required that the reader has any prior familiarity with it.
When we refer to a [*quantum register*]{} we simply mean a discrete quantum system that we wish to consider, such as a collection of qubits representing a message transmitted from one party to another. With any quantum register we associate some vector space $\X = \complex^n$ for a positive integer $n$ that intuitively represents the maximum number of distinct classical states that could be stored in the register without error. A [*state*]{} of such a register is represented by a [*density matrix*]{}, which is a $n\times n$ positive semidefinite matrix having trace equal to 1. Density matrices may reasonably be viewed as the quantum information-theoretic analogue to a vector of probabilities, representing a probability distribution. We will write ${{\mathrm{D}}\left(\X\right)}$ to denote the set of all density matrices associated with a register that is described by $\X$. It is natural to view such density matrices as [*linear operators*]{} acting on $\X$, and for this reason the term [*density operator*]{} is commonly used in place of [*density matrix*]{}.
When two registers having associated spaces $\X = \complex^n$ and $\Y=\complex^m$ are considered as a single compound register, the associated space becomes the tensor product space $\X\otimes\Y = \complex^{nm}$. If the two registers are independently prepared in states described by $\rho$ and $\sigma$, respectively, then the joint state is described by the $nm\times nm$ density matrix $\rho\otimes\sigma$. This matrix may be written in block form as $$\rho\otimes\sigma = \begin{pmatrix}
\rho_{1,1} \sigma & \cdots & \rho_{1,n} \sigma\\
\vdots & \ddots & \vdots \\
\rho_{n,1} \sigma & \cdots & \rho_{n,n} \sigma
\end{pmatrix}.$$
In general, for a vector space $\X = \complex^n$, we write ${{\mathrm{L}}\left(\X\right)}$ to denote the set of [*all*]{} $n\times n$ complex matrices, or linear operators mapping $\X$ to itself, and we write ${{\mathrm{Herm}}\left(\X\right)}$ to refer to the subset of ${{\mathrm{L}}\left(\X\right)}$ given by the [*Hermitian*]{} matrices. These are the matrices $A$ satisfying $A = A^{\ast}$, where $A^{\ast}$ denotes the [*adjoint*]{} or [*conjugate transpose*]{} of $A$. The set ${{\mathrm{Herm}}\left(\X\right)}$ forms a vector space over $\real$, and many optimization methods designed for real-valued symmetric matrices extend to ${{\mathrm{Herm}}\left(\X\right)}$ with little or no special consideration. Finally, we write ${{\mathrm{Pos}}\left(\X\right)}$ to denote the subset of ${{\mathrm{Herm}}\left(\X\right)}$ that consists of all [*positive semidefinite*]{} $n\times n$ matrices (or operators acting on $\X$).
The [*Hilbert-Schmidt inner product*]{} on ${{\mathrm{L}}\left(\X\right)}$ is defined as $${\left\langle A , B\right\rangle} = {\operatorname{Tr}}(A^{\ast} B)$$ for all $A,B\in{{\mathrm{L}}\left(\X\right)}$. It holds that ${\left\langle A , B\right\rangle}$ is a real number for all choices of Hermitian matrices $A$ and $B$, and is a nonnegative real number for all choices of positive semidefinite matrices $A$ and $B$.
A [*measurement*]{} of a register, having an associated vector space $\X = \complex^n$, is described by a collection of $n\times n$ positive semidefinite matrices that sum to the identity. Specifically, a measurement that has some finite, non-empty set $\Sigma$ of possible outcomes is described by a collection $\{P_a\::\:a\in\Sigma\}\subset{{\mathrm{Pos}}\left(\X\right)}$ satisfying $$\sum_{a\in\Sigma} P_a = \I_{\X}.$$ Here, $\I_{\X}$ denotes the $n\times n$ identity matrix, or identity operator on $\X$. (The subscript $\X$ is dropped when it is implicitly clear.) If the register corresponding to $\X$ is in a state described by the density matrix $\rho\in{{\mathrm{D}}\left(\X\right)}$, and this measurement described by $\{P_a\::\:a\in\Sigma\}\subset{{\mathrm{Pos}}\left(\X\right)}$ is performed, each outcome $a\in\Sigma$ will be observed with probability ${\left\langle P_a , \rho\right\rangle}$.
Non-interactive zero-sum quantum games
--------------------------------------
In a non-interactive zero-sum quantum game, Alice and Bob each send a quantum state to a referee, who performs a measurement on these two states to determine their payoffs. Hereafter we will let $\A = \complex^n$ and $\B=\complex^m$ refer to the vector spaces corresponding to the states that Alice and Bob send to the referee.
When the referee performs a measurement to determine Alice and Bob’s payoffs, a [*joint*]{} measurement is used. In other words, Alice’s and Bob’s states are together viewed as a single state of a register. We therefore have that the referee’s measurement is described by a collection $$\{R_a\,:\,a\in\Sigma\} \subset {{\mathrm{Pos}}\left(\A\otimes\B\right)}$$ that satisfies the condition $$\sum_{a\in\Sigma} R_{a} = \I_{\A\otimes\B}.$$ If Alice sends the state $\rho$ and Bob sends the state $\sigma$, then each possible measurement outcome $a\in\Sigma$ appears with probability ${\left\langle R_a , \rho\otimes\sigma\right\rangle}$.
A payoff for each player is associated with each possible measurement outcome $a\in\Sigma$. As we consider only zero-sum games, it is sufficient to describe these payoffs by a function $v:\Sigma\rightarrow\real$; with Alice’s payoff for outcome $a$ being $v(a)$ and Bob’s payoff being $-v(a)$. For a given choice of states $\rho$ and $\sigma$, it holds that Alice’s expected payoff is given by $$\sum_{a\in\Sigma} v(a) {\left\langle R_a , \rho\otimes\sigma\right\rangle}
= {\left\langle R , \rho\otimes\sigma\right\rangle}$$ for $$R = \sum_{a\in\Sigma} v(a) R_a.$$ Bob’s expected payoff is given by $-{\left\langle R , \rho\otimes\sigma\right\rangle}$. When one is interested only in the expected payoff of a given game, it is therefore sufficient to consider that the game is simply determined by $R$. We will refer to $R$ as a [*payoff observable*]{}, given that a matrix that arises in this way from a measurement and a real-valued function on its outcomes is sometimes called an observable.
A necessary and sufficient condition for a matrix $R$ acting on $\A\otimes\B$ to arise from some measurement and real-valued payoff function $v$ as just described is that $R$ is Hermitian. The sort of payoff function $\phi(\rho,\sigma)$ discussed in the introduction therefore takes the form $\phi(\rho,\sigma) = {\left\langle R , \rho\otimes\sigma\right\rangle}$ for $R$ ranging over the set of Hermitian matrices of the appropriate size. As the tensor product is a universal bilinear function, and every real-valued linear function on ${{\mathrm{Herm}}\left(\A\right)}\otimes{{\mathrm{Herm}}\left(\B\right)}$ can be expressed as an inner product with some Hermitian matrix $R$, we have that a necessary and sufficient condition for $\phi(\rho,\sigma)$ to be a physically valid payoff function is that $\phi$ is a real-valued bilinear function.
Now, given that the sets ${{\mathrm{D}}\left(\A\right)}$ and ${{\mathrm{D}}\left(\B\right)}$ are convex and compact, and that Alice’s expected payoff ${\left\langle R , \rho\otimes\sigma\right\rangle}$ is a bilinear function on ${{\mathrm{D}}\left(\A\right)}\times{{\mathrm{D}}\left(\B\right)}$, it follows from well-known extensions of von Neumann’s Min-Max Theorem [@vonNeumann28] that $$\label{eq:min-max-R}
\max_{\rho\in{{\mathrm{D}}\left(\A\right)}} \min_{\sigma\in{{\mathrm{D}}\left(\B\right)}}
{\left\langle R , \rho\otimes\sigma\right\rangle}
=
\min_{\sigma\in{{\mathrm{D}}\left(\B\right)}} \max_{\rho\in{{\mathrm{D}}\left(\A\right)}}
{\left\langle R , \rho\otimes\sigma\right\rangle}.$$ (See, for instance, [@Fan53].) We define $\alpha(R)$ to be the [*value*]{} of the game determined by $R$, which is the quantity represented by the two sides of the above equation . A pair of quantum states $(\rho, \sigma)$ is called an [*equilibrium point*]{} for $R$ if both $\rho$ and $\sigma$ independently achieve the maximum and minimum, respectively, in equation ; or, equivalently, that $$\min_{\sigma' \in {{\mathrm{D}}\left(\B\right)}} {\left\langle R , \rho\otimes\sigma'\right\rangle}
= {\left\langle R , \rho\otimes\sigma\right\rangle}
= \max_{\rho' \in {{\mathrm{D}}\left(\A\right)}} {\left\langle R , \rho'\otimes\sigma\right\rangle}.$$ Again, the existence of an equilibrium point follows easily from equation .
We define that an [*$\epsilon$-approximate equilibrium point*]{} of a game with payoff observable $R$ is a pair of states $(\rho,\sigma)$ such that $$\max_{\rho' \in {{\mathrm{D}}\left(\A\right)}} {\left\langle R , \rho'\otimes\sigma\right\rangle} - \varepsilon
{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}
\leq {\left\langle R , \rho\otimes\sigma\right\rangle}
\leq \min_{\sigma' \in {{\mathrm{D}}\left(\B\right)}} {\left\langle R , \rho\otimes\sigma'\right\rangle} +
\varepsilon {\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}.$$ Note that this is an approximation in an additive sense, and is relative to the maximum absolute value of any payoff (which is reflected by the presence of the factor ${\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}$ in the error).
Additional definitions and notation {#sec:additional-notation}
-----------------------------------
This section summarizes some additional terminology and notation that will be used in the paper. First, a linear mapping of the form $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$ is called a [*super-operator*]{} (as it maps linear operators to linear operators). The [*adjoint super-operator*]{} to $\Phi$ has the form $\Phi^{\ast}: {{\mathrm{L}}\left(\A\right)}\rightarrow{{\mathrm{L}}\left(\B\right)}$, and is uniquely determined by the condition $${\left\langle A , \Phi(B)\right\rangle} = {\left\langle \Phi^{\ast}(A) , B\right\rangle}$$ for all $A\in{{\mathrm{L}}\left(\A\right)}$ and $B\in{{\mathrm{L}}\left(\B\right)}$. A super-operator $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$ is said to be [*positive*]{} if it holds that $\Phi(P)$ is positive semidefinite for every choice of a positive semidefinite operator $P\in{{\mathrm{Pos}}\left(\B\right)}$. It is the case that $\Phi^{\ast}$ is positive if and only if $\Phi$ is positive.
There is a one-to-one and onto linear correspondence between the collection of operators of the form $R\in{{\mathrm{L}}\left(\A\otimes\B\right)}$ and the collection of super-operators of the form $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$, which is sometimes known as the [*Choi-Jamio[ł]{}kowski isomorphism*]{}. Specifically, for every super-operator $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$, one defines an operator $R \in {{\mathrm{L}}\left(\A\otimes\B\right)}$, called the Choi-Jamio[ł]{}kowski representation of $\Phi$, by the equation $$R = \sum_{1\leq i,j \leq m} \Phi(E_{i,j}) \otimes E_{i,j}\;,$$ where $E_{i,j}$ is the matrix with a 1 in entry $(i,j)$ and 0 in every other entry. Conversely, given an operator $R \in {{\mathrm{L}}\left(\A\otimes\B\right)}$, one defines a super-operator $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$ by means of the formula $$\label{eq:CJ}
\Phi(B) = {\operatorname{Tr}}_{\B} \left( R \left(\I_{\A}\otimes B^{{{\scriptscriptstyle\mathsf{T}}}}\right)\right).$$ It follows that $${\operatorname{Tr}}\left(R(A\otimes B)\right) = {\operatorname{Tr}}\left(A\,\Phi(B^{{{\scriptscriptstyle\mathsf{T}}}})\right)$$ for every choice of $A\in{{\mathrm{L}}\left(\A\right)}$ and $B\in{{\mathrm{L}}\left(\B\right)}$. These correspondences are both linear, and are inverse to one another—so one is free to translate back and forth between the two as necessary for a given application. The assumption that $R$ is positive semidefinite implies that the corresponding super-operator $\Phi$ is positive. (In fact, $\Phi$ has the stronger property of being [*completely positive*]{} if and only if $R$ is positive semidefinite.)
For a given quantum game, we may equally well calculate expected payoffs and equilibrium points by using the unique super-operator $\Phi$ determined by rather than the payoff observable $R$. In particular, $(\rho,\sigma^{{{\scriptscriptstyle\mathsf{T}}}})$ is an equilibrium point of the game defining $R$ if and only if $$\label{eq:equilibrium-Phi}
\min_{\sigma' \in {{\mathrm{D}}\left(\B\right)}} {\left\langle \rho , \Phi(\sigma')\right\rangle}
= {\left\langle \rho , \Phi(\sigma)\right\rangle}
= \max_{\rho' \in {{\mathrm{D}}\left(\A\right)}}{\left\langle \rho' , \Phi(\sigma)\right\rangle},$$ and the value of this game is alternately expressed as $$\label{eq:value-Phi}
\alpha(\Phi) {\stackrel{\smash{\text{\tiny def}}}{=}}\max_{\rho\in{{\mathrm{D}}\left(\A\right)}}
\min_{\sigma\in{{\mathrm{D}}\left(\B\right)}}
{\left\langle \rho , \Phi(\sigma)\right\rangle} =
\min_{\sigma\in{{\mathrm{D}}\left(\B\right)}}
\max_{\rho\in{{\mathrm{D}}\left(\A\right)}}
{\left\langle \rho , \Phi(\sigma)\right\rangle}.$$
For a Hermitian $n\times n$ matrix $A$, one denotes the eigenvalues of $A$ by $$\lambda_1(A) \geq \lambda_2(A) \geq \cdots \geq \lambda_n(A),$$ sorted from largest to smallest and including each eigenvalue a number of times equal to its multiplicity. For every $n\times n$ Hermitian matrix $A$, the [*spectral norm*]{} is denoted ${\left\lVert\tinyspaceA{\mspace{1mu}}\right\rVert}$ and satisfies $${\left\lVert\tinyspaceA{\mspace{1mu}}\right\rVert} = \max\{{\left\lvert{\mspace{1mu}}\lambda_1(A) {\mspace{1mu}}\right\rvert},\ldots,{\left\lvert{\mspace{1mu}}\lambda_n(A) {\mspace{1mu}}\right\rvert}\},$$ while the [*trace norm*]{} is denoted ${\left\lVert\tinyspaceA{\mspace{1mu}}\right\rVert}_1$ and satisfies $${\left\lVert\tinyspaceA{\mspace{1mu}}\right\rVert}_1 = {\left\lvert{\mspace{1mu}}\lambda_1(A) {\mspace{1mu}}\right\rvert}+ \cdots + {\left\lvert{\mspace{1mu}}\lambda_n(A) {\mspace{1mu}}\right\rvert}.$$ (Note that both of these formulas assume that $A$ is Hermitian.)
Using the above notation, we may express the equations and in simpler terms: $(\rho,\sigma^{{{\scriptscriptstyle\mathsf{T}}}})$ is an equilibrium point of the game defining $\Phi$ if and only if $$\lambda_1(\Phi(\sigma))
= {\left\langle \rho , \Phi(\sigma)\right\rangle}
= \lambda_m(\Phi^{\ast}(\rho)),$$ while the value of this game satisfies $$\alpha(\Phi) =
\min_{\sigma\in{{\mathrm{D}}\left(\B\right)}}
\lambda_1(\Phi(\sigma))
=
\max_{\rho\in{{\mathrm{D}}\left(\A\right)}}
\lambda_m(\Phi^{\ast}(\rho)).$$
Finally, for future reference we note that if a payoff observable $R$ satisfies $0\leq R\leq \I$, then it holds that $0\leq \Phi(\sigma)\leq
\I$ and $0\leq \Phi^{\ast}(\rho)\leq \I$ for all choices of density matrices $\rho\in{{\mathrm{D}}\left(\A\right)}$ and $\sigma\in{{\mathrm{D}}\left(\B\right)}$. Moreover, for arbitrary Hermitian matrices $A\in{{\mathrm{Herm}}\left(\A\right)}$ and $B\in{{\mathrm{Herm}}\left(\B\right)}$, we have ${\left\lVert{\mspace{1mu}}\Phi(B){\mspace{1mu}}\right\rVert} \leq {\left\lVert\tinyspaceB{\mspace{1mu}}\right\rVert}_1$ and ${\left\lVert{\mspace{1mu}}\Phi^{\ast}(A){\mspace{1mu}}\right\rVert} \leq {\left\lVert\tinyspaceA{\mspace{1mu}}\right\rVert}_1$.
The Main Result {#sec:main}
===============
We now present the main result of the paper, which is a parallel algorithm to approximate the value of a non-interactive zero-sum quantum game. This fact is stated as Theorem \[theorem:approximation\] below, following a few comments on the assumed form of the input.
We suppose that a given non-interactive zero-sum quantum game is described by payoff observable $R\in{{\mathrm{Herm}}\left(\A\otimes\B\right)}$, for $\A = \complex^n$ and $\B = \complex^m$ as discussed in the previous section. More precisely, we assume that $R$ is given as an $nm\times nm$ matrix, along with a specification of the dimensions $n$ and $m$. Each entry of $R$ is a complex number, which we assume has rational real and imaginary parts, each represented as the ratio of two integers expressed in binary notation. We let $k$ be the maximum length of the binary representation over all of these integers, and define ${\operatorname{size}}(R)$ to be $(nm)^2 k$. It is clear that $O({\operatorname{size}}(R))$ bits suffice to encode $R$.
In addition to $R$, $n$ and $m$, an accuracy parameter $\varepsilon > 0$ is also given as input. For technical reasons it is most convenient to assume that $\varepsilon$ is represented in [*unary notation*]{}: each string $1^r$, for a positive integer $r$, denotes the value $\varepsilon=1/r$. This assumption on the input form of $\varepsilon$ reflects the fact that our algorithm does not scale well with respect to accuracy—it forces the length of the input to be proportional to $1/\varepsilon$ rather than $\log(1/\varepsilon)$, and therefore permits our algorithm to be described by circuits with size polynomial in the input length.
The output of the algorithm will be a pair of density matrices $(\rho,\sigma)$ where $\rho\in{{\mathrm{D}}\left(\A\right)}$ and $\sigma\in{{\mathrm{D}}\left(\B\right)}$. They are assumed to be represented in a manner similar to the input matrix $R$.
\[theorem:approximation\] An $\varepsilon$-approximate equilibrium point $(\rho,\sigma)$ for a given payoff observable $R$ can be computed by a logarithmic-space uniform family of Boolean circuits having depth polynomial in $\log({\operatorname{size}}(R))$ and $1/\varepsilon$.
Parallel algorithm for positive games {#sec:algorithm}
-------------------------------------
Our algorithm is most naturally described for the case that the payoff observable $R$ satisfies $0 \leq R \leq \I$. We therefore begin with this case, which will imply Theorem \[theorem:approximation\] by an appropriate translation and rescaling of $R$. The algorithm is described in Figure \[fig:algorithm\].
[**Algorithm**]{}
1. Let $\mu = \varepsilon/8$ and let $N = \left\lceil 64 \ln(nm)/\varepsilon^2\right\rceil$.
2. Initialize: $A_0 = \I_{\A}$, $\rho_0 = A_0/{\operatorname{Tr}}(A_0)$, $B_0 = \I_{\B}$, and $\sigma_0 = B_0/{\operatorname{Tr}}(B_0)$.
3. For each $j$ from 1 to $N$, let $A_j$, $\rho_j$, $B_j$, and $\sigma_j$ be as follows: $$\begin{aligned}
A_j & = \exp\left(\mu \sum_{i = 0}^{j-1} \Phi(\sigma_i)\right),\\
\rho_j & = A_j/{\operatorname{Tr}}(A_j),\\
B_j & = \exp\left(-\mu \sum_{i = 0}^{j-1} \Phi^{\ast}(\rho_i)\right),\\
\sigma_j & = B_j/{\operatorname{Tr}}(B_j).\end{aligned}$$
4. Output the pair $\left(\rho,\sigma^{{{\scriptscriptstyle\mathsf{T}}}}\right)$, where $$\rho = \frac{1}{N} \sum_{j = 0}^{N-1} \rho_j
\quad\quad \text{and} \quad\quad
\sigma = \frac{1}{N} \sum_{j = 0}^{N-1} \sigma_j.$$
### Accuracy of the algorithm
In this section, the accuracy of the algorithm described in Figure \[fig:algorithm\] is analyzed. We note that a similar type of analysis has appeared in previous works on the multiplicative weights update method and its predecessors, and in particular the reader is referred to [@Kale07] for information on the generality of the approach.
At this point in the analysis we are concerned only with the idealized algorithm described in Figure \[fig:algorithm\]—numerical issues concerning the required precision with which the idealized operations are performed are discussed in the next subsection. We begin by noting some facts concerning matrix exponentials. First, the [*Golden-Thompson Inequality*]{} (see Section IX.3 of [@Bhatia97]) states that, for any two Hermitian matrices $X$ and $Y$ of equal dimension, we have $${\operatorname{Tr}}\left(e^{X + Y}\right) \leq {\operatorname{Tr}}\left(e^X e^Y\right).$$ Second is the following simple pair of inequalities concerning the matrix exponential of positive and negative semidefinite matrices.
\[lemma:exp-inequalities\] Let $P$ be an operator satisfying $0\leq P\leq \I$. Then for every real number $\mu > 0$, the following two inequalities hold: $$\begin{aligned}
\exp(\mu P) & \leq \I + \mu \exp(\mu)P,\\
\exp(-\mu P) & \leq \I - \mu \exp(-\mu)P.\end{aligned}$$
It is sufficient to prove the inequalities for $P$ replaced by a scalar $\lambda\in[0,1]$, for then the operator inequalities follow by considering a spectral decomposition of $P$. If $\lambda=0$ both inequalities are immediate, so let us assume $\lambda>0$. By the Mean Value Theorem there exists a value $\lambda_0\in(0,\lambda)$ such that $$\frac{\exp(\mu \lambda) - 1}{\lambda} = \mu \exp(\mu \lambda_0)
\leq \mu \exp(\mu),$$ from which the first inequality follows. Similarly, there exists a value $\lambda_0\in(0,\lambda)$ such that $$\frac{\exp(-\mu\lambda)-1}{\lambda} = -\mu\exp(-\mu\lambda_0)
\leq -\mu \exp(-\mu),$$ which yields the second inequality.
We now proceed to the main part of the accuracy analysis, which comprises two bounds on the eigenvalues of $\Phi^{\ast}(\rho)$ and $\Phi(\sigma)$, where $(\rho,\sigma^{{{\scriptscriptstyle\mathsf{T}}}})$ is the output of the algorithm.
\[lemma:eigenvalue-bounds\] The following inequalities hold: $$\begin{aligned}
\lambda_1(\Phi(\sigma)) & \leq \frac{\exp(\mu)}{N}\sum_{j = 1}^N
{\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle} + \frac{\ln(n)}{\mu N},
\label{eq:A}\\
\lambda_m(\Phi^{\ast}(\rho)) & \geq \frac{\exp(-\mu)}{N}\sum_{j = 1}^N
{\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle} - \frac{\ln(m)}{\mu N}.
\label{eq:B}\end{aligned}$$
Let us begin by noting that each of the operators $A_j$ and $B_j$ (for $j = 0,\ldots,N$) that are obtained during the course of the algorithm are positive definite, and therefore have positive trace. It follows that each of the operators $\rho_j$ and $\sigma_j$ is a well-defined density operator.
To prove the first inequality, observe that $$A_N = \exp\left(\mu\sum_{j = 1}^N \Phi(\sigma_{j-1})\right)
= \exp\left(\mu N \Phi(\sigma)\right).$$ Given that $A_N$ is positive definite, it holds that $${\operatorname{Tr}}(A_N) \geq \lambda_1(A_n) = \exp(\mu N\lambda_1(\Phi(\sigma))),$$ and therefore $$\label{eq:X-inequality-1}
\lambda_1(\Phi(\sigma)) \leq \frac{\ln({\operatorname{Tr}}(A_N))}{\mu N}.$$
The inequality will now follow by bounding $\ln({\operatorname{Tr}}(X_N))$, which can be done as follows. First, note that we may alternately write $$A_j = {\operatorname{exp}}\left(\ln(A_{j-1}) + \mu\,\Phi(\sigma_{j-1})\right)$$ for each $j\geq 1$, and therefore $${\operatorname{Tr}}(A_j) = {\operatorname{Tr}}\left({\operatorname{exp}}\left(\ln(A_{j-1})
+ \mu\,\Phi(\sigma_{j-1})\right)\right)
\leq {\operatorname{Tr}}\left(A_{j-1}
\exp\left(\mu\,\Phi(\sigma_{j-1})\right)\right)$$ by the Golden-Thompson inequality. As $\sigma_{j-1}$ is a density operator, it holds that $\Phi(\sigma_{j-1})\leq \I$, and therefore $$\exp\left(\mu\,\Phi(\sigma_{j-1})\right)
\leq \I + \mu\exp(\mu)\Phi(\sigma_{j-1})$$ by Lemma \[lemma:exp-inequalities\]. Thus, using the fact that ${\operatorname{Tr}}(X Y_1)\leq {\operatorname{Tr}}(X Y_2)$ for all choices of matrices $X$, $Y_1$, and $Y_2$ with $X\geq 0$ and $Y_1\leq Y_2$, we have $${\operatorname{Tr}}(A_j)
\leq
{\operatorname{Tr}}\left(A_{j-1}\left(\I+\mu\exp(\mu)\Phi(\sigma_{j-1})\right)\right)
=
{\operatorname{Tr}}(A_{j-1})\left(1+\mu\exp(\mu){\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle}\right).$$ It now follows from the inequality $1 + \mu \leq \exp(\mu)$ that $${\operatorname{Tr}}(A_j)
\leq {\operatorname{Tr}}(A_{j-1})\exp\left(\mu \exp(\mu)
{\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle}\right).$$ Applying this inequality recursively, and using the fact that ${\operatorname{Tr}}(A_0) = n$, we obtain $$\label{eq:X-inequality-2}
{\operatorname{Tr}}(A_N) \leq
\exp\left(\mu\exp(\mu)
\sum_{j = 1}^N {\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle} + \ln(n)\right).$$ Combining and yields $$\lambda_1(\Phi(\sigma)) \leq
\frac{\exp(\mu)}{N}\sum_{j = 1}^N
{\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle} + \frac{\ln(n)}{\mu N},$$ as required.
The second inequality follows by similar reasoning, except with a few differences that we now highlight. We first observe that $$B_N = \exp\left(-\mu N \Phi^{\ast}(\rho)\right).$$ This time we have $${\operatorname{Tr}}(B_N) \geq \lambda_1(B_N) = \exp(-\mu N \lambda_m(\Phi^{\ast}(\rho))),$$ where the switch from the largest eigenvalue to the smallest is caused by the minus sign in the exponential function. Thus, $$\label{eq:Y-inequality-1}
\lambda_m(\Phi^{\ast}(\rho)) \geq -\frac{\ln({\operatorname{Tr}}(B_N))}{\mu N}.$$ The quantity ${\operatorname{Tr}}(B_N)$ is now bounded in the same way as ${\operatorname{Tr}}(A_N)$, except that we need the second inequality in Lemma \[lemma:exp-inequalities\]. Specifically, the Golden-Thompson inequality implies $${\operatorname{Tr}}(B_j) = {\operatorname{Tr}}(\exp(\ln(B_{j-1}) - \mu\Phi^{\ast}(\rho_{j-1})))
\leq {\operatorname{Tr}}(B_{j-1}\exp(-\mu\Phi^{\ast}(\rho_{j-1}))).$$ As $\Phi^{\ast}(\rho_{j-1})\leq\I$ we have $$\exp(-\mu\Phi^{\ast}(\rho_{j-1}))
\leq \I - \mu\exp(-\mu) \Phi^{\ast}(\rho_{j-1}),$$ and therefore $${\operatorname{Tr}}(B_j) \leq {\operatorname{Tr}}(B_{j-1})\exp(-\mu\exp(-\mu)
{\left\langle \sigma_{j-1} , \Phi^{\ast}(\rho_{j-1})\right\rangle}).$$ It follows that $$\label{eq:Y-inequality-2}
{\operatorname{Tr}}(B_N)\leq \exp\left(-\mu\exp(-\mu)
\sum_{j=1}^N {\left\langle \sigma_{j-1} , \Phi^{\ast}(\rho_{j-1})\right\rangle} + \ln(m)\right).$$ Combining and , along with the fact that ${\left\langle \sigma_j , \Phi^{\ast}(\rho_j)\right\rangle} = {\left\langle \rho_j , \Phi(\sigma_j)\right\rangle}$ for every choice of $j$, yields $$\lambda_m(\Phi^{\ast}(\rho))
\geq \frac{\exp(-\mu)}{N}\sum_{j = 1}^N
{\left\langle \rho_{j-1} , \Phi(\sigma_{j-1})\right\rangle} - \frac{\ln(m)}{\mu N}$$ and completes the proof.
It is now possible to verify that the output $(\rho,\sigma^{{{\scriptscriptstyle\mathsf{T}}}})$ of the algorithm satisfies $$\max_{\rho' \in {{\mathrm{D}}\left(\A\right)}} {\left\langle R , \rho'\otimes\sigma^{{{\scriptscriptstyle\mathsf{T}}}}\right\rangle} - \varepsilon/2
\leq {\left\langle R , \rho\otimes\sigma^{{{\scriptscriptstyle\mathsf{T}}}}\right\rangle}
\leq \min_{\sigma' \in {{\mathrm{D}}\left(\B\right)}} {\left\langle R , \rho\otimes\sigma'\right\rangle}+\varepsilon/2,$$ which is expressed in terms of the mapping $\Phi$ as $$\lambda_1(\Phi(\sigma)) - \varepsilon/2
\leq {\left\langle \rho , \Phi(\sigma)\right\rangle}
\leq \lambda_m(\Phi^{\ast}(\rho)) + \varepsilon/2.$$ It follow from Lemma \[lemma:eigenvalue-bounds\] that $$\lambda_1(\Phi(\sigma)) - \lambda_m(\Phi^{\ast}(\rho)) \leq
\frac{\exp(\mu) - \exp(-\mu)}{N}
\sum_{j = 1}^N{\left\langle \sigma_{j-1} , \Phi(\rho_{j-1})\right\rangle} +
\frac{\ln(nm)}{\mu N},$$ and given that each of the quantities ${\left\langle \sigma_{j-1} , \Phi(\rho_{j-1})\right\rangle}$ is at most 1, we have $$\lambda_1(\Phi(\sigma)) - \lambda_m(\Phi^{\ast}(\rho))
\leq 2\sinh(\mu) + \frac{\ln(nm)}{\mu N}
< 3\mu + \frac{\ln(nm)}{\mu N} \leq \varepsilon/2.$$ Thus, given that $\lambda_m(\Phi^{\ast}(\rho)) \leq {\left\langle \rho , \Phi(\sigma)\right\rangle} \leq
\lambda_1(\Phi(\sigma))$, we have $$\label{eq:final-accuracy}
\lambda_1(\Phi(\sigma)) - \varepsilon/2
\leq \lambda_m(\Phi^{\ast}(\rho))
\leq {\left\langle \rho , \Phi(\sigma)\right\rangle}
\leq \lambda_1(\Phi^{\ast}(\rho))
\leq \lambda_m(\Phi(\sigma)) + \varepsilon/2$$ as claimed.
### Numerical precision and complexity of the algorithm
Let us now consider the complexity of the algorithm described in Figure \[fig:algorithm\]. It is the goal of this section to demonstrate that this algorithm can be implemented, by a logarithmic-space uniform family of Boolean circuits with depth polynomial in $\log({\operatorname{size}}(R))+1/\varepsilon$, with sufficient accuracy to obtain an $\varepsilon$-approximate equilibrium point for the input payoff observable $R$. Throughout the analysis, we (sometimes grossly) overestimate errors for the sake of simpler expressions involving as few variables as possible.
Each iteration performed in step 3 of the algorithm requires the evaluation of $\Phi$ and $\Phi^{\ast}$, two matrix exponential computations, and a constant number of elementary matrix operations (in this case: addition, scalar multiplication, and computation of the trace). Were it not for the matrix exponentials, it would be straightforward to perform all of the required operations within the claimed size and depth bounds using exact computations. Given that the matrix exponentials will generate irrational numbers, however, we must settle for approximations over the course of the algorithm. To guarantee that the algorithm is sufficiently accurate, it will suffice to perform all computations to within an additive error of $(\varepsilon/2)\exp(-8N^2)$, as is shown below. (We could afford to take a much smaller error with respect to ${\operatorname{size}}(R)$, but there is no need to do this.)
Let us begin by making a few simple observations about the matrices computed throughout the course of the algorithm. The matrices $\sigma_0,\ldots,\sigma_{N-1}$ are density matrices, and therefore it holds that ${\left\lVert{\mspace{1mu}}\Phi(\sigma_j){\mspace{1mu}}\right\rVert}\leq 1$ for each choice of $j = 0,\ldots,N-1$. Likewise, ${\left\lVert{\mspace{1mu}}\Phi^{\ast}(\rho_j){\mspace{1mu}}\right\rVert}\leq 1$ for each choice of $j = 0,\ldots,N-1$. Consequently, we have ${\left\lVert\tinyspaceA_j{\mspace{1mu}}\right\rVert}\leq e^{N}$, ${\left\lVert\tinyspaceB_j{\mspace{1mu}}\right\rVert}\leq e^{N}$, $$1\leq {\operatorname{Tr}}(A_j)\leq e^{2N}\quad\quad\text{and}\quad\quad
e^{-N} \leq {\operatorname{Tr}}(B_j) \leq e^{N}$$ for $j = 0,\ldots,N-1$.
Next, let us represent the actual matrices computed during the course of the algorithm by placing a tilde over the variables representing the idealized values that are expressed in Figure \[fig:algorithm\]. It will suffice to prove that ${\left\lVert{\mspace{1mu}}\rho - \widetilde{\rho}{\mspace{1mu}}\right\rVert}_1 \leq \varepsilon/2$ and ${\left\lVert{\mspace{1mu}}\sigma - \widetilde{\sigma}{\mspace{1mu}}\right\rVert}_1 \leq \varepsilon/2$, for then the inequalities $${\left\lvert{\mspace{1mu}}\lambda_1(\Phi(\sigma)) -
\lambda_1\left(\Phi\left(\widetilde{\sigma}\right)\right) {\mspace{1mu}}\right\rvert} \leq
\frac{\varepsilon}{2}
\quad\quad\text{and}\quad\quad
{\left\lvert{\mspace{1mu}}\lambda_m(\Phi^{\ast}(\rho)) -
\lambda_m\left(\Phi^{\ast}\left(\widetilde{\rho}\right)\right) {\mspace{1mu}}\right\rvert}
\leq \frac{\varepsilon}{2}$$ hold. Combined with , we obtain the required accuracy.
Now, each iteration of step 3 of the algorithm will introduce some error into the calculation of the final answer. Let us consider the $j$-th iteration, and assume that a positive real number $\delta_j\in(0,1)$ is given such that ${\left\lVert{\mspace{1mu}}\sigma_i - \widetilde{\sigma}_i{\mspace{1mu}}\right\rVert}_1 \leq \delta_j$ for $i = 0,\ldots,j-1$. Let us define $$X_j = \mu \sum_{i = 0}^{j-1}\Phi(\sigma_i)
\quad\quad\text{and}\quad\quad
\widetilde{X}_j
= \mu \sum_{i = 0}^{j-1}\Phi\left(\widetilde{\sigma}_i\right).$$ Then ${\left\lVert\tinyspaceX_j - \widetilde{X}_j{\mspace{1mu}}\right\rVert} \leq \delta_j N$ and ${\left\lVert\tinyspaceX_j{\mspace{1mu}}\right\rVert}\leq
N$, and therefore $${\left\lVert{\mspace{1mu}}\exp(X_j) - \exp\left(\widetilde{X}_j\right){\mspace{1mu}}\right\rVert} \leq
{\left\lVert\tinyspaceX_j - \widetilde{X}_j{\mspace{1mu}}\right\rVert} e^{{\left\lVert\tinyspaceX_j-\widetilde{X}_j{\mspace{1mu}}\right\rVert}} e^{{\left\lVert\tinyspaceX_j{\mspace{1mu}}\right\rVert}}
< \delta_j e^{3N},$$ where the first inequality follows from Corollary 6.2.32 of [@HornJ91]. By computing the matrix exponential with accuracy $\delta_j$ we therefore have ${\left\lVert\tinyspaceA_j - \widetilde{A}_j{\mspace{1mu}}\right\rVert} \leq \delta_j e^{4N}$, and thus ${\left\lVert\tinyspaceA_j - \widetilde{A}_j{\mspace{1mu}}\right\rVert}_1 \leq \delta_j e^{5N}$. It follows that $${\left\lVert{\mspace{1mu}}\rho_j - \widetilde{\rho}_j{\mspace{1mu}}\right\rVert}_1
\leq
\frac{1}{{\operatorname{Tr}}(A_j)} {\left\lVert\tinyspaceA_j - \widetilde{A}_j{\mspace{1mu}}\right\rVert}_1
+ {\left\lVert{\mspace{1mu}}\widetilde{A}_j{\mspace{1mu}}\right\rVert}_1 {\left\lvert{\mspace{1mu}}\frac{1}{{\operatorname{Tr}}(A_j)} -
\frac{1}{{\operatorname{Tr}}\left(\widetilde{A}_j\right)} {\mspace{1mu}}\right\rvert} \leq \delta_j e^{8N}.$$ By similar reasoning, if it holds that ${\left\lVert{\mspace{1mu}}\rho_i - \widetilde{\rho}_i{\mspace{1mu}}\right\rVert}_1 \leq \delta_j$ for $i = 0,\ldots,j-1$, then $${\left\lVert{\mspace{1mu}}\sigma_j - \widetilde{\sigma}_j{\mspace{1mu}}\right\rVert}_1
\leq
\frac{1}{{\operatorname{Tr}}(B_j)} {\left\lVert\tinyspaceB_j - \widetilde{B}_j{\mspace{1mu}}\right\rVert}_1
+ {\left\lVert{\mspace{1mu}}\widetilde{B}_j{\mspace{1mu}}\right\rVert}_1 {\left\lvert{\mspace{1mu}}\frac{1}{{\operatorname{Tr}}(B_j)} -
\frac{1}{{\operatorname{Tr}}\left(\widetilde{B}_j\right)} {\mspace{1mu}}\right\rvert} \leq \delta_j e^{8N}.$$ We conclude from these bounds that taking $\delta_j =
(\varepsilon/2)e^{-8 N^2}$ guarantees that ${\left\lVert{\mspace{1mu}}\rho - \widetilde{\rho}{\mspace{1mu}}\right\rVert}_1 \leq \varepsilon/2$ and ${\left\lVert{\mspace{1mu}}\sigma - \widetilde{\sigma}{\mspace{1mu}}\right\rVert}_1 \leq \varepsilon/2$.
The required precision for the matrix exponentials is easily obtained by taking sufficiently many terms in the series $e^X = \I + X + X^2/2 + X^3/6 + \cdots$. (This of course is not the most efficient way to compute matrix exponentials, but it suffices to prove the main theorem.) For instance, taking $9N^2$ terms guarantees that the required accuracy $(\varepsilon/2)e^{-8N^2}$ is achieved.
At this point, the parallel complexity of the algorithm is easily bounded. Each of the matrices stored by the algorithm has entries whose real and imaginary parts are represented in binary notation using $O(N^2)$ bits. For each iteration in step 3 of the algorithm, the evaluations of $\Phi$ and $\Phi^{\ast}$, as well as the elementary matrix operations, may therefore be performed by standard parallel algorithms (see, for instance, [@vzGathen93]) by logarithmic-space uniform Boolean circuits (with size that is necessarily polynomial in ${\operatorname{size}}(R)$ and $1/\varepsilon$ given this uniformity constraint), within depth that is polynomial in $\log({\operatorname{size}}(R))$ and $1/\varepsilon$. The number of iterations performed is $N$, which results in total depth polynomial in $\log({\operatorname{size}}(R))$ and $1/\varepsilon$.
Extensions to arbitrary payoff observables
------------------------------------------
For an arbitrary payoff observable $R$, the algorithm from the previous section is not guaranteed to function correctly, as we have used the positivity of the corresponding super-operator $\Phi$ several times during the analysis.
It is straightforward, however, to translate and scale an arbitrary payoff observable in a way that allows the algorithm to be used. For an arbitrary positive semidefinite payoff observable $R$, this is essentially trivial—one simply runs the algorithm on the payoff observable $P = R/{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}$. For a negative semidefinite payoff observable $R$, one simply exchanges the roles of Alice and Bob and considers the payoff observable $-R$ (with the spaces $\A$ and $\B$ swapped).
Let us now consider the general case of a payoff observable $R$ for which $\lambda_1(R) > 0 > \lambda_{nm}(R)$. Define $$P = \frac{R - \lambda_{nm}(R)\I}{\lambda_1(R) - \lambda_{nm}(R)}.$$ Then $0\leq P \leq \I$, and so the algorithm from the previous section may be used to obtain an $\varepsilon$-approximation $(\rho,\sigma)$ for $P$. The point $(\rho,\sigma)$ is easily verified to be a $\delta$-approximate equilibrium point for $R$, where $$\delta = \frac{\lambda_1(R) - \lambda_{nm}(R)}{{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}}\,\varepsilon
\leq 2 \varepsilon.$$
${\textup{QRG(1)}}$ is contained in {#sec:qsp2}
====================================
Quantum interactive proof systems with two competing provers are naturally represented as games between two competing players, moderated by a referee. The two players (Alice and Bob) play the roles of competing provers, while the referee corresponds to the verifier. Quantum refereed games have been studied in [@GutoskiW05; @Gutoski05; @GutoskiW07], and represent a quantum analogue to the classical refereed games model studied in [@FeigeK97].
The simplest form of a refereed quantum game has the general form defined in Section \[sec:definitions\]; meaning that there is no communication from the referee to the players. The players each send a quantum state and the referee measures to determine the winner. With this picture in mind, one defines the complexity class ${\textup{QRG(1)}}$ to be the class consisting of all promise problems $A = (A_{\mathrm{yes}},A_{\mathrm{no}})$ for which there exists a polynomial-time uniform family $Q = \{Q_n\,:\,n\in\natural\}$ of quantum circuits, where each circuit $Q_n$ takes $n + 2 p(n)$ input qubits for some polynomial bounded function $p$, such that the following properties hold:
For every string $x\in A_{\mathrm{yes}}$ it holds that $$\max_\rho \min_\sigma {\operatorname{Pr}}[Q(x,\rho,\sigma) = 1]
\geq \frac{2}{3}.$$
For every string $x\in A_{\mathrm{no}}$ it holds that $$\max_\rho \min_\sigma {\operatorname{Pr}}[Q(x,\rho,\sigma) = 1]
\leq \frac{1}{3}.$$
Here, the maximum and minimum are both over all quantum states on $p({\left\lvert{\mspace{1mu}}x {\mspace{1mu}}\right\rvert})$ qubits, and the notation $Q(x,\rho,\sigma) = 1$ is shorthand for the event that a measurement of some fixed output qubit of the circuit $Q_{{\left\lvert{\mspace{1mu}}x {\mspace{1mu}}\right\rvert}}$ (with respect to the standard basis) yields 1, assuming that the input to the circuit is the state ${\vert x \rangle}{\langle x\vert}\otimes \rho\otimes \sigma$. The name ${\textup{QRG}}(1)$ refers to the fact that these are quantum refereed games with 1 turn, during which the players send quantum states to the referee in parallel.
The class ${\textup{QRG}}(1)$ may be viewed as a simple variant of ${\textup{QMA}}$, where there are two competing provers rather than a single prover. It is obvious that ${\textup{QMA}}\subseteq{\textup{QRG(1)}}$, and that ${\textup{QRG}}(1)$ is closed under complementation (and thus ${\textup{co-QMA}}\subseteq{\textup{QRG}}(1)$).
The class ${\textup{QRG}}(1)$ may roughly be thought of as a quantum analogue to the class $S_2^P$ that was defined by [@Canetti96] and [@RussellS98], and it is easily observed that $S_2^P\subseteq{\textup{QRG}}(1)$. There is one subtlety, however, which is that the definition of ${\textup{QRG}}(1)$ does not allow one prover to see the other’s message (which would not make sense in the quantum setting anyway), whereas the standard definition of $S_2^P$ does.
${\textup{QRG}}(1) \subseteq{\textup{PSPACE}}$.
Suppose that $A$ is a promise problem contained in ${\textup{QRG}}(1)$, and that $\{Q_n\}$ is a polynomial-time uniform family of quantum circuits that witnesses this fact. For each input $x$, let $R_x$ denote the payoff observable that corresponds to the game played by the players Alice and Bob on input $x$, where the payoff for Alice is defined as 1 for acceptance and 0 for rejection. The expected payoff is therefore the probability of acceptance, which Alice tries to maximize and Bob tries to minimize.
We denote by ${\textup{NC}}(\mathit{poly})$ the class of promise problems computed by polynomial-space uniform Boolean circuits with polynomial depth. It holds that ${\textup{NC}}(\mathit{poly}) \subseteq {\textup{PSPACE}}$ [@Borodin77], so it therefore suffices to prove that $A\in{\textup{NC}}(\mathit{poly})$. This is easily accomplished by composing three families of Boolean circuits:
A family of Boolean circuits that outputs a description of the payoff observable $R_x$ associated with the game on input $x$.
The family of Boolean circuits given by Theorem \[theorem:approximation\], that finds an $\varepsilon$-approximate equilibrium point $(\rho,\sigma)$ of the payoff observable $R_x$, for $\varepsilon = 1/8$.
A family of Boolean circuits that computes the expected payoff ${\left\langle R_x , \rho\otimes\sigma\right\rangle}$, and accepts if the value is greater than 1/2 (rejecting otherwise).
The first family is easily derived from the circuits $\{Q_n\}$, by computing the product of a polynomial number of exponential-size matrices that correspond to the quantum gates of the appropriate circuit $Q_n$. This family may be taken to be polynomial-space uniform, with polynomial depth. The second family is, as suggested above, given by Theorem \[theorem:approximation\]. This family is logarithmic-space uniform and has polynomial-size and poly-logarithmic depth with respect to ${\operatorname{size}}(R_x)$. Thus, with respect to the input length ${\left\lvert{\mspace{1mu}}x {\mspace{1mu}}\right\rvert}$, this family is polynomial-space uniform, and has polynomial depth. The last family is required only to perform elementary matrix and arithmetic operations, and can be taken to have similar properties as the first two: polynomial-space uniformity and polynomial depth. Composing these families appropriately demonstrates that $A\in{\textup{NC}}(\mathit{poly})$ as required.
Parallel Approximation of Positive Semidefinite Programs {#sec:positive-SDPS}
========================================================
We now discuss the connection between equilibrium points of non-interactive zero-sum quantum games and semidefinite programs. The main focus of this section will, in particular, be on [*positive*]{} instances of semidefinite programs, and on the question of whether good parallel methods to approximate them exist. We will first discuss the general notion of positive instances of semidefinite programs and then explain how our algorithm may be used in their approximation, albeit with poor accuracy in some cases.
The multiplicative weights update method has been applied to semidefinite programming in [@AroraHK05b; @AroraK07], and the general connection between equilibrium points of different types of games and linear/semidefinite programs is well-known. Once again, the reader is referred to Kale [@Kale07] for further details and historical remarks.
Positive instances of semidefinite programs in super-operator form
------------------------------------------------------------------
Suppose that the following input has been given:
1. a Hermitian matrix $A\in\complex^{n\times n}$,
2. a Hermitian matrix $B\in\complex^{m\times m}$, and
3. a linear mapping $\Phi:\complex^{m\times m}\rightarrow\complex^{n\times n}$ (i.e., a super-operator) that preserves Hermiticity.
To say that $\Phi$ preserves Hermiticity means that $\Phi(Y)$ is Hermitian for every choice of a Hermitian matrix $Y\in\complex^{m\times m}$. This condition is equivalent to the Choi-Jamio[ł]{}kowski representation $R$ of $\Phi$ being a Hermitian matrix.
Given this input, let us consider the following semidefinite programming problem, which we say is in the [*super-operator form*]{}:
$$\begin{aligned}
\text{maximize:}\quad & {\left\langle B , Y\right\rangle}\\
\text{subject to:}\quad & \Phi(Y) \leq A,\\
& Y\geq 0.
\end{aligned}$$
$$\begin{aligned}
\text{minimize:}\quad & {\left\langle A , X\right\rangle}\\
\text{subject to:}\quad & \Phi^{\ast}(X) \geq B,\\
& X\geq 0.
\end{aligned}$$
Here, $X$ and $Y$ range over all (positive semidefinite) matrices in $\complex^{n\times n}$ and $\complex^{m\times m}$, respectively. This form is completely general: it is possible to translate semidefinite programs in so-called [*standard form*]{} to the super-operator form, and vice versa. It can be shown that strong duality holds for semidefinite programs in the super-operator form under conditions that are similar to those for standard form semidefinite programs. In particular, the existence of either of the following implies that strong duality holds:
1. a positive definite matrix $Y$ for which $\Phi(Y) < A$, or
2. a positive definite matrix $X$ for which $\Phi^{\ast}(X) > B$.
We define that such a problem instance is [*positive*]{} if $A$ and $B$ are positive semidefinite matrices and $\Phi$ is a positive super-operator. Let us also define that such a problem is [*strictly positive*]{} if it holds that $A$ and $B$ are positive definite and $\Phi$ is a strictly positive super-operator (which means that $\Phi(\I)$ is positive definite). Strong duality necessarily holds for all strictly positive semidefinite programs in the super-operator form.
Parallel approximation of strictly positive semidefinite programs
-----------------------------------------------------------------
Parallel algorithms for approximately solving positive linear programs have been given by Luby and Nisan [@LubyN93] and Young [@Young01]. To our knowledge, an analogous problem for semidefinite programs has not been considered.
The algorithm from Section \[sec:algorithm\] can be used to approximate, in parallel, strictly positive instances of semidefinite programs as we now explain. First, let us note that an arbitrary strictly positive semidefinite program in the super-operator form can be transformed into one of the following simpler form:
$$\begin{aligned}
\text{maximize:}\quad & {\operatorname{Tr}}(Y)\\
\text{subject to:}\quad & \Phi(Y) \leq \I,\\
& Y\geq 0.
\end{aligned}$$
$$\begin{aligned}
\text{minimize:}\quad & {\operatorname{Tr}}(X)\\
\text{subject to:}\quad & \Phi^{\ast}(X) \geq \I,\\
& X\geq 0.
\end{aligned}$$
This may be done by defining $$\Phi(Y) = A^{-\frac{1}{2}}\,\Psi\left(B^{-\frac{1}{2}} Y
B^{-\frac{1}{2}}\right) A^{-\frac{1}{2}}$$ for a given problem instance defined by $A>0$, $B>0$, and a strictly positive super-operator $\Psi$.
Now, to make the connection with the algorithm from the previous section clear, let us recall that we define $\A=\complex^n$ and $\B=\complex^m$, and suppose that the super-operator $\Phi$ that represents the above semidefinite program takes the form $\Phi:{{\mathrm{L}}\left(\B\right)}\rightarrow{{\mathrm{L}}\left(\A\right)}$. Let us also define ${\operatorname{opt}}(\Phi)$ to be the optimal value of the primal problem (which is the same as the optimal value of the dual problem by strong duality). It is clear that ${\operatorname{opt}}(\Phi) > 0$, for some positive scalar multiple of the identity must be primal feasible. Let us also recall that we have defined $$\alpha(\Phi)
= \max_{\rho\in{{\mathrm{D}}\left(\A\right)}} \min_{\sigma\in{{\mathrm{D}}\left(\B\right)}}
{\left\langle \rho , \Phi(\sigma)\right\rangle}.$$
For all strictly positive super-operators $\Phi$ we have $\alpha(\Phi) = 1/{\operatorname{opt}}(\Phi)$.
Let $(\rho,\sigma)$ be an equilibrium point of $\Phi$, meaning that $$\lambda_1(\Phi(\sigma)) = {\left\langle \rho , \Phi(\sigma)\right\rangle} =
\lambda_m(\Phi^{\ast}(\rho)) = \alpha(\Phi).$$ The assumption that $\Phi$ is strictly positive implies that $\alpha(\Phi)$ is positive.
We now observe that $\sigma/\alpha(\Phi)$ is primal feasible, as it is positive semidefinite and satisfies $$\lambda_1\left(\Phi\left(\sigma/\alpha(\Phi)\right)\right) = 1,$$ which implies $\Phi\left(\sigma/\alpha(\Phi)\right) \leq \I_{\A}$. Likewise, $\rho/\alpha(\Phi)$ is dual feasible as it is positive semidefinite and satisfies $$\lambda_m(\Phi^{\ast}(\rho/\alpha(\Phi))) = 1,$$ which implies $\Phi^{\ast}(\rho/\alpha(\Phi))\geq \I_{\B}$. Both result in the same objective value $1/\alpha(\Phi)$, and so the proposition follows by (weak) duality.
It follows that the algorithm from Section \[sec:algorithm\] may be used to approximate ${\operatorname{opt}}(\Phi)$, albeit with limited accuracy for some choices of $\Phi$, by taking the reciprocal of the value of the game associated with $\Phi$. To be more specific, let $R$ be the Choi-Jamio[ł]{}kowski representation of the super-operator $\Phi$ as discussed in Section \[sec:additional-notation\]. Let us write $\widetilde{\alpha}(\Phi)$ to denote the approximate value of the game described by $R$ that results from the algorithm’s $\varepsilon$-approximate equilibrium point of $R$, and let us also write $\widetilde{{\operatorname{opt}}}(\Phi) = 1/\widetilde{\alpha}(\Phi)$. We then have $$\left(1 - \frac{\varepsilon {\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}}{\alpha(\Phi)}\right) \alpha(\Phi)
\leq
\widetilde{\alpha}(\Phi)
\leq
\left(1 + \frac{\varepsilon{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}}{\alpha(\Phi)}\right) \alpha(\Phi)$$ and therefore $$\left(1 + \frac{\varepsilon{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}}{\alpha(\Phi)}\right)^{-1}
{\operatorname{opt}}(\Phi) \leq \widetilde{{\operatorname{opt}}}(\Phi) \leq
\left(1 - \frac{\varepsilon{\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}}{\alpha(\Phi)}\right)^{-1}
{\operatorname{opt}}(\Phi).$$
For choices of $\Phi$ for which $\alpha(\Phi)$ is large and ${\left\lVert\tinyspaceR{\mspace{1mu}}\right\rVert}$ is small (bounded below and above by constants, say), a reasonable approximation to ${\operatorname{opt}}(\Phi)$ may be obtained. For many choices of $\Phi$, however, our method is clearly not suitable, and we believe it is an interesting problem for future research to find more accurate parallel algorithms for this problem.
Conclusion
==========
In this paper we have shown that equilibrium points of non-interactive zero-sum quantum games can be efficiently computed in parallel, using the multiplicative weights update method. As a consequence, we have that one-turn quantum refereed games can be simulated in polynomial space, or ${\textup{QRG(1)}}\subseteq{\textup{PSPACE}}$. We have also illustrated the connection between values of quantum games and positive instances of semidefinite programming problems.
The main open question that we wish to raise concerns the existence of efficient parallel algorithms for positive instances of semidefinite programming problems. The class of such problems for which our algorithm gives accurate solutions is limited. To what extent can this task be performed for more general classes?
Acknowledgements {#acknowledgements .unnumbered}
----------------
The parallel algorithm presented in this work was inspired by an algorithm appearing in an unpublished joint work, concerning classical games, of Rohit Khandekar and the first author. We therefore thank Rohit Khandekar for his implicit and indirect contribution to this work, and for allowing us to include it in this paper. Rahul Jain’s research is supported by ARO/NSA USA, and John Watrous’s research is supported by Canada’s NSERC and the Canadian Institute for Advanced Research.
[KKM[[$^{+}$]{}]{}08]{}
S. Arora, E. Hazan, and S. Kale. Fast algorithms for approximate semidefinite programming using the multiplicative weights update method. , pages 339–348, 2005.
S. Arora, E. Hazan, and S. Kale. The multiplicative weights update method: a meta algorithm and applications. Manuscript, 2005.
S. Arora and S. Kale. A combinatorial, primal-dual approach to semidefinite programs. , pages 227–236, 2007.
A. Ambainis. A new protocol and lower bounds for quantum coin flipping. In [*Proceedings of the Thirty-Third Annual ACM Symposium on Theory of Computing*]{}, pages 134–142, 2001.
R. Bhatia. . Springer, 1997.
A. Borodin. On relating time and space to size and depth. , 6:733–744, 1977.
R. Canetti. On [BPP]{} and the polynomial-time hierarchy. , 57:237–241, 1996.
R. Cleve, P. H[ø]{}yer, B. Toner, and J. Watrous. Consequences and limits of nonlocal strategies. In [*Proceedings of the 19th Annual IEEE Conference on Computational Complexity*]{}, pages 236–249, 2004.
K. Fan. Minimax theorems. , 39:42–47, 1953.
U. Feige and J. Kilian. Making games short. In [*Proceedings of the Twenty-Ninth Annual ACM Symposium on Theory of Computing*]{}, pages 506–516, 1997.
J. von zur Gathen. Parallel linear algebra. In J. Reif, editor, [*Synthesis of Parallel Algorithms*]{}, chapter 13. Morgan Kaufmann Publishers, Inc., 1993.
G. Gutoski. Upper bounds for quantum interactive proofs with competing provers. In [*Proceedings of the 20th Annual IEEE Conference on Computational Complexity*]{}, pages 334–343, 2005.
G. Gutoski and J. Watrous. Quantum interactive proofs with competing provers. In [*Proceedings of the 22nd Symposium on Theoretical Aspects of Computer Science*]{}, volume 3404 of [*Lecture Notes in Computer Science*]{}, pages 605–616. Springer, 2005.
G. Gutoski and J. Watrous. Toward a general theory of quantum games. In [*Proceedings of the 39th ACM Symposium on Theory of Computing*]{}, pages 565–574, 2007.
R. Horn and C. Johnson. . Cambridge University Press, 1991.
S. Kale. . PhD thesis, Princeton University, 2007.
A. Kitaev. Quantum coin-flipping. Presentation at the 6th Workshop on [*Quantum Information Processing*]{} (QIP 2003), 2002.
J. Kempe, H. Kobayashi, K. Matsumoto, B. Toner, and T. Vidick. Entangled games are hard to approximate. In [*Proceedings of the 49th Annual IEEE Symposium on Foundations of Computer Science*]{}, 2008.
J. Kempe, H. Kobayashi, K. Matsumoto, and T. Vidick. Using entanglement in quantum multi-prover interactive proofs. In [*Proceedings of the 23rd Annual Conference on Computational Complexity*]{}, 2008.
H. Kobayashi and K. Matsumoto. Quantum multi-prover interactive proof systems with limited prior entanglement. , 66(3), 2003.
M. Luby and N. Nisan. A parallel approximation algorithm for positive linear programming. , pages 448–457, 1993.
C. Mochon. Quantum weak coin-flipping with bias of 0.192. In [*45th Annual IEEE Symposium on Foundations of Computer Science*]{}, pages 2–11, 2004.
C. Mochon. Quantum weak coin flipping with arbitrarily small bias. , 2007.
A. Russell and R. Sundaram. Symmetric alternation captures [BPP]{}. , 7:152–162, 1998.
R. Spekkens and T. Rudolph. Quantum protocol for cheat-sensitive weak coin flipping. , 89(22):227901, 2002.
K. Tsuda, G. R[ä]{}tsch, and M. Warmuth. Matrix exponentiated gradient updates for on-line learning and [Bregman]{} projection. , 6:995–1018, 2005.
J. von Neumann. Zur theorie der gesellschaftsspiele. , 100(1928):295–320, 1928.
N. Young. Sequential and parallel algorithms for mixed packing and covering. , pages 538–546, 2001.
| {
"pile_set_name": "ArXiv"
} |
---
title: Performance metrics
---
It can be argued that the quality of science cannot be measured purely by quantitative metrics, and certainly not by a single one that is supposed to fit all. Experimental groups and theoretical groups have different types of output and it is difficult to envisage a way in which the impact or quality of these can be compared either objectively or subjectively. A single instrument built by an experimental group may well lay the foundation for very many scientific results and papers by opening up a new field, and a single theoretical paper may well cause a paradigm shift. In those branches of physics where close links with industrial applications can be forged, yet another measure of succes would need to be found.
Even for those research groups for which the primary scientific output could be considered to be scientific papers, the rate at which papers are published, and therefore also are cited depends on the field. Even within the area of astrophysics, in some subfields papers are published at rather larger rates than in others. No-one could reasonably claim that the former sub-fields have greater scientific significance or importance than the latter. In this sense it seems that the conclusions of Pearce (1994) on ‘good researchers’ should not be used without additional judgment, as indeed pointed out by that author.
To illustrate this point I have compiled some citation statistics by making use of the Smithsonian/NASA Astrophysics Data System (ADS) which also tracks citation statistics (the UK mirror site can be found at http://ukads.nottingham.ac.uk/). The number of bibliographic entries, both refereed and non-refereed, per author in the ADS covers a wide range. Evidently there are many authors with one or a few entries, and at the other end of the scale there is an author for which there are in excess of 11000 entries. A histogram showing the distribution of authors in the ADS is shown : each bin shows the amount of authors, against the number of bibliographic entries per author.
An author with very many entries is likely to be better known, but it does not seem likely that the scientific impact of e.g. each of 11000+ entries is the same. The number of citations a paper receives by itself does not necessarily reflect impact either. A paper that is authored by many can easily get a high citation ranking if each of the authors subsequently writes a further paper in which the original paper is cited. If an average paper in a given field of research has a certain a-priori likelihood of being cited, i.e. the standard impact of a paper in that field, the number of citations it receives must scale with the number of papers published in that field. Also, the rate at which a typical author publishes will reflect the habits of that subfield so that one should expect that for an unexceptional author the number of citations scales with the number of papers that this author publishes to some power that lies between 1 and 2. $$N_{\rm expected\ citations} \propto N_{\rm papers\
published}^\gamma\hskip 1cm 1<\gamma <2$$ From the ADS it is possible to extract the normalised citation rate of papers of any author, which is calculated by taking for each paper the number of citations it has received, divides by the number of authors, and then sums over all entries/papers on which this author appears. This to some extent accounts for differences between many author and single author papers.
I have randomly sampled, from the distribution shown in Fig. 1, authors for whom I have subsequently extracted this normalised citation count for those bibliographic entries which are refereed papers. The sampling was carried out as a constant fraction $0.02$ per bin, rather than for the distribution as a whole in order to reduce the statistical fluctuations and cover the entire range reasonably evenly. Also, authors with less than 50 bibliographic entries were omitted from the analysis in order to reduce statistical fluctuations in citation counts.
The random sampling within each bin sometimes produced a name for which bibliographic entries could not uniquely be assigned. Whenever this occured the name was discarded and a different, randomly selected, name was substituted. The diagram resulting from this is shown on log-log scales in which the solid line is the least squares fit with a $\gamma = 1.52$. The behaviour appears to correspond to the expectation outlined above. In absolute terms authors appearing in the upper right-hand corner of this diagram, such as e.g. Prof. Sir Martin Rees, have great visibility and therefore impact in astrophysics. An equally useful measure of the relative importance of individual authors or groups of authors is the distance above or below the fitted function indicating a relatively high or low number of citations given the number of papers published by that author. An example of this is given by the strongest upward outlier to this distribution that I have found which is Alan Guth, whose seminal work in inflation theory clearly has a very wide impact in current cosmology. His entry is not part of the random sample, and therefore also not present in Fig. 2 but corresponds to a $3-\sigma$ upwards outlier. The bottom panel shows the distribution orthogonal to the fitted function, with positive offset corresponding to authors with normalised citation rates above the fit, and negative offset to those below that fit, with overplotted a Gaussian distribution. The distribution is quite evidently skewed, with a long left-ward tail, and possibly somewhat more peaked than a Gaussian distribution. $$\begin{aligned}
{\rm skewness} &=& -1.09 \pm 0.25 \\
{\rm kurtosis} &=& 1.4 \pm 0.6\end{aligned}$$ Creating a distribution such as this for a department or the national community requires having available a complete citation and publication record. Even if this were to exist, as argued above, it does not measure reliably the performance of those groups for which scientific papers is not the only or primary scientific output. Any attempt at measuring performance solely in this manner is therefore in practice impossible and in principle flawed.
Moreover, a measure such as this covers the entire career of individuals. It is known that exceptional papers distinguish themselves primarily through being cited consistently over decades, rather than passing through a brief peak of citations and then disappearing from citation records. Even if the above were a reliable measure it is not amenable to adaptation for the very short-term measures sought for RAE assessment exercises and the like. To the authors’ knowledge there is no metric that reliably measures performance over periods as short as 5 years, and none that is an indicator of future performance. Specifically it seems inappropriate to use numbers of citations, even after ‘normalisation’ in the sense of the ADS, as a reliable direct metric of impact.
This research has made use of NASA’s Astrophysics Data System Bibliographic Services.
Pearce, F., (1994), Astronomy & Geophysics, [**45**]{}, 2.15
| {
"pile_set_name": "ArXiv"
} |
2.3truecm
[**LAPLACE-BELTRAMI EQUATION ON HYPERSURFACES AND $\Gamma$-CONVERGENCE**]{} 0.7truecm [Tengiz BUCHUKURI, Roland DUDUCHAVA & George TEPHNADZE ]{}[^1]
0.7truecm
0.25in
> [**Abstract.**]{} We investigate a mixed boundary value problem for the stationary heat transfer equation in a thin layer with a mid hypersurface $\cC$ in $\bR^3$ with the boundary. The main object is to trace what happens in $\Gamma$-limit when the thickness of the layer converges to zero. The limit Dirichlet BVP for the Laplace-Beltrami equation on the surface is described explicitly and we show how the Neumann boundary conditions in the initial BVP transform in the $\Gamma$-limit. For this we apply the variational formulation and the calculus of Günter’s tangential differential operators on a hypersurface and layers, which allow global representation of basic differential operators and of corresponding boundary value problems in terms of the standard Euclidean coordinates of the ambient space $\mathbb{R}^n$.
Introduction {#sec0 .unnumbered}
============
The main aim of this paper is to demonstrate what happens with a boundary value problem for the Helmholtz equation in a thin layer $\Omega^\varepsilon$ with a mid hypersurface $\cC$ in $\bR^3$ when the thickness of the layer $2\varepsilon$ diminishes to zero $\varepsilon\to0$. We impose the Neumann boundary conditions on the upper and lower faces of the layer $\mathcal{C}\times\{\pm\varepsilon\}$ and the Dirichlet boundary conditions on the lateral surface $\partial\mathcal{C}\times(-\varepsilon,\varepsilon)$.
The convergence is understood in the $\Gamma$-convergence sense. Equation in the layer is represented in terms of the extended Gunter’s derivatives-the system of tangential Gunter’s derivatives on the surface. the column of surface gradient $$\label{eq1.1}
\mathcal{D}:=(\mathcal{D}_1,\mathcal{D}_2,\mathcal{D}_3)^\top$$ (cf. [@Gu1], [@KGBB1], [@Du1]). Here $\cD_j:=\pa_j-\nu_j\pa_\nub$ is the Günter’s tangential derivative on the mid surface $\cC$ and $\nub=(\nu_1,\ nu_2,\nu_3)^\top$ is the unit normal vector field on $\cC$. The first-order differential operator $\mathcal{D}_j$ is the directional derivative along $\pi\,e_j$, where $\pi:\mathbb{R}^n\to T\mathcal{C}$ is the orthogonal projection onto the tangent plane to $\mathcal{C}$ and, as usual, $e_j=(\delta_{jk})_{1\leq k\leq n}\in\mathbb{R}^n$, with $\delta_{jk}$ denoting the Kronecker symbol.
Calculus of Gunter’s derivatives on a hypersurface allows representation of the most basic partial differential operators (PDO’s), as well as their associated boundary value problems, on a hypersurface $\mathcal{C}$ in $\mathbb{R}^n$, in global form, in terms of the standard spatial coordinates in $\mathbb{R}^n$. Such BVPs arise in a variety of situations and have many practical applications. See, for example, [@Ha1 §[72]{}] for the heat conduction by surfaces, [@Ar1 §[10]{}] for the equations of surface flow, [@Ci1], [[@AC1] for the vacuum Einstein equations describing gravitational fields]{}, [@TZ1] for the Navier-Stokes equations on spherical domains, as well as the references therein.
A hypersurface $\mathcal{C}$ in $\mathbb{R}^3$ has the natural structure of a $2$-dimensional Riemannian manifold and the aforementioned PDE’s are not the immediate analogues of the ones corresponding to the flat, Euclidean case, since they have to take into consideration geometric characteristics of $\mathcal{C}$ such as curvature. Inherently, these PDE’s are originally written in local coordinates, intrinsic to the manifold structure of $\mathcal{C}$.
The operator $\mathcal{D}$ is globally defined on $\mathcal{C}$, and has a relatively simple structure. In terms of (\[eq1.1\]), the Laplace-Beltrami operator on $\mathcal{C}$ simply becomes [(see [@MM1 pp. 2ff and p. 8.])]{} $$\label{eq1.2}
\Delta_{\mathcal{C}}=\mathcal{D}^*\mathcal{D}
\quad\mbox{ on }\quad\mathcal{C}.$$
Alternatively, this is the natural operator associated with the Euler-Lagrange equations for the variational integral
$$\label{eq1.3}
\mathcal{E}[u]=-\frac12\int_{\mathcal{C}}\|{\mathcal{D}}u\|^2\,dS.$$
A similar approach, based on the principle that, at equilibrium, the displacement minimizes the potential energy, leads to the derivation of the equation for the elastic hypersurface (cf. [@DMM1; @Du3] for the isotropic case).
These results are useful in numerical and engineering applications (cf. [@AN1], [@Be1], [@Ce1], [@Co1], [@DL1], [[@BGS1], [@Sm1]]{}) and we plan to treat a number of special surfaces in greater detail in a subsequent publication.
We consider heat conduction by an “isotropic” media, governed by the Laplace equations and with the classical Dirichlet-Neumann mixed boundary conditions on the boundary in the layer domain $\Omega^\ve:=\cC \times(-\ve,\ve)$ of thickness $2\ve$: Let us consider the mixed BVP with zero Dirichlet but non-zero Neumann data: $$\begin{aligned}
\label{e7.37}
\begin{array}{lll}
&\Delta_{\Omega^\ve} T(\cx,t)=f(\cx,t), &(\cx,t)\in\cC\times(-\ve,\ve),\\[2mm]
& T^+(\cx,t)=0, &(\cx,t)\in \pa\cC\times(-\ve,\ve),\\[2mm]
&\pm(\pa_t T)^+(\cx,\pm\ve)=q(\cx,\pm\ve),\qquad &\cx\in\cC,
\end{array}\end{aligned}$$ where $\pm\pa_t=\pa_\nub$ represents the normal derivative on the surfaces $\cC\times{\pm\ve}$. Here $\cC\subset\cS$ is a smooth subsurface of a closed hypersurface $\cS$ with smooth nonempty boundary $\pa\cC$. In the investigation we apply that the Laplace operator $\Delta_{\Omega^\ve}=\pa^2_1+\pa^2_2+\pa^2_3$ is represented as the sum of the Laplace-Beltrami operator on the mid-surface and the square of the transversal derivative: $$\Delta_{\Omega^\ve}T=\sum\limits_{j=1}^{4}\cD_j^2 T=\Delta_{\cC}T+\pa_t^2 T.$$
The BVP can be reformulated as the minimization problem for the functional $$\begin{aligned}
\label{e7.38}
E(T_\ve):&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T)(\cx,\tau)\right|^2+\left|(\pa_\tau T)(\cx,\tau)\right|^2\right]+f(\cx,\tau)T(\cx,\tau)\Big]d\sigma d\tau\nonumber\\
&&+\int\limits_{\cC}\left[q(\cx,+\ve)T^+(\cx,+\ve)-q(\cx,-\ve)T^+(\cx,-\ve)\right]d\sigma.
\end{aligned}$$
After scaling (stretching the variable $t=\ve \tau$ and dividing the entire functional by $\ve$) has the following form $$\begin{aligned}
\label{e7.38a}
E_{\ve}(T_\ve):&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T_\ve)(\cx,\tau)\right|^2+\frac1{\ve^2}\left|(\pa_\tau T_\ve)(\cx,\tau)\right|^2\right]+f_\ve(\cx,\tau)T_\ve(\cx,\tau)\Big]d\sigma d\tau\nonumber\\
&&+\frac1\ve\int\limits_{\cC}\left[q^+(\cx,+\ve)T^+_\ve(\cx,+1)
-q(\cx,-\ve)T^+_\ve(\cx,-1)\right]d\sigma,\\
&&T_\ve(\cx,\tau):=T\left(\cx,\ve\tau\right)\in \mathbb{H}^1(\Omega^1), \qquad f_\ve(\cx,\tau):=f\left(\cx,\ve\tau\right)\in \mathbb{H}^{-1}(\Omega^1).\nonumber
\end{aligned}$$
The main result of the present investigation is the following Theorem \[t0.1\].
\[t0.1\] Let $q(\cx,\pm\ve)\in\bH^{1/2}(\cC)$ and are uniformly bounded in $\bL_2(\cC)$, $f_{\varepsilon}(\cx,t)\to f^0(\cx)$ in $\mathbb{H}^{-1}(\Omega^1)$ and there exists a function $q^0\in\bH^{-1/2}(\cC)$ such that $$\begin{aligned}
\label{e7.39}
\lim_{\ve\to0}\frac1{2\ve}\scal\varphi(\cdot),q(\cdot,\ve)-q(\cdot,-\ve)\scar_\cC
=\scal\varphi,q^0\scar_\cC,\qquad \forall\,\vf\in\bH^{1/2}(\cC).\end{aligned}$$ Then the functional in $\Gamma$-converges to the functional $$\begin{aligned}
\label{e7.40}
E^{(0)}(T)=\int\limits_{-1}^1\int\limits_{\cC}
\Big[\frac12\left|(\cD_{\cC}T)(\cx,0)\right|^2+[f^0(\cx)+q^0(\cx)]
T(\cx,0)\Big]d\sigma dt\nonumber\\
=2\int\limits_{\cC}\left[\frac12\left\langle\cD_{\cC} T(\cx),\cD_{\cC}
T(\cx)\right\rangle + [f^0(\cx)+q^0(\cx)]T(\cx)\right]d\sigma.\end{aligned}$$
The following Dirichlet boundary value problem for Laplace-Beltrami equation on the mid surface $\cC$ $$\begin{aligned}
\label{e7.42}
\begin{array}{l}
\Delta_\cC T(\cx)=f^0(\cx)+q^0(\cx) \quad \cx \in \cC,\\[3mm]
T^+(\cx)=0, \qquad \cx \in \pa\cC.
\end{array}
\end{aligned}$$ is an equivalent reformulation of the minimization problem with the energy functional and, therefore, can be considered as the limit of the initial BVP .
\[r0.2\] It is remarkable to note that the weak derivative $q^0$ of the Neumann condition from the initial BVP (see the comment below) migrated into the right hand side of the limit equation.
Let us comment on the condition . We remind that $q(\cdot,\pm\ve)\in \bH^{1/2}(\cC)$. If $q_1\in\bH^1(\cC \times(-\ve,\ve))$ is an extension of these functions inside the domain $\lim\limits_{t\to\pm\ve}q_1(\cx,t) =q(\cx,\pm\ve)$, then $q^0(\cx)=\dst\frac12(\pa_tq)(\cx,0)\in\bL_2(\cC)$ represents the derivative in the weak sense $$\lim_{\ve\to0}\frac1{2\ve}\scal\vf(\cdot),q(\cdot,\ve)-q(\cdot,-\ve)\scar_\cC
=\lim_{\ve\to0}\frac1{2\ve}\int_{-\ve}^\ve\scal\vf(\cdot),\pa_\tau q_1)(\cdot,\tau)\scar_\cC\,d\tau=\scal\vf,q^0\scar_\cC$$ for all $\vf\in\bL_2(\cC)$ (see Corollary \[c7.8\] of the Lebesgue Differentiation Theorem below).
The layout of the paper is as follows. In §1-§2 we review some basic differential-geometric concepts which are relevant for the work at hand (e.g., hypersurfaces and different methods of their identification). In §3 we identify the most important partial differential operators on hypersurfaces, such as gradient, divergence, Laplace-Beltrami operator. In §4 we consider the energy functional and the associated Euler-Lagrange equation. In sections §5, §6 we apply the aforementioned approach and prove main theorems of the present paper, including Theorem \[t0.1\].
Brief review of the classical theory of hypersurfaces {#sec2}
=====================================================
We commence with the definition of a hypersurface and give two equivalent definitions. Both definitions are important for our purposes.
\[d2.3\] A Subset ${\mathcal S}\subset\mathbb{R}^n$ of the Euclidean space is called a [**hypersurface**]{} if it has a covering ${\mathcal
S}=\bigcup_{j=1}^M{\mathcal S}_j$ and coordinate mappings $$\label{e1.1}
\Theta_j\;:\;\omega_j\rightarrow{\mathcal S}_j:=\Theta_j(\omega_j)
\subset\mathbb{R}^n, \qquad\omega_j\subset\mathbb{R}^{n-1},
\quad j=1,\ldots,M,$$ such that the corresponding differentials $$\label{eA2.1.15}
\hskip-1.8mm\begin{array}{c} D\Theta_j(p)
:=\matr\,[\pa_1\Theta_j(p),\ldots,\pa_{n-1}\Theta_j(p)]\, ,
\end{array}$$ have the full rank $${\rm rank}\,D\Theta_j(p)=n-1\, ,\qquad \forall p\in Y_j\, ,\quad
k=1,\ldots,n\, ,\quad j=1,\ldots,M\, ,$$ i.e. , all points of $\omega_j$ are regular for $\Theta_j$ for all $j=1,\ldots,M$.
Such mapping is called an [**immersion**]{} as well.
The hypersurface is called [**smooth**]{} if the corresponding coordinate diffeomorphisms $\Theta_j$ in are smooth ($ C^\infty$-smooth). Similarly is defined a [**$\mu$-smooth**]{} hypersurface.
The derivatives $${\boldsymbol{g}}_{k}({\scriptstyle{\mathcal{X}}})=\partial _{k}\Theta
_{j}(\Theta _{j}^{-1}({\scriptstyle{\mathcal{X}}})),\qquad {\scriptstyle{%
\mathcal{X}}}\in \mathcal{C},\quad k=1,\ldots ,n-1 \label{e1.84}$$are then tangential vector fields on ${\mathcal{C}}$ and moreover, compose a basis in the space of tangential vector fields $\mathcal{W}({\mathcal{C}})$.
The most important role in the calculus of tangential differential operators we are going to apply belongs to the unit normal vector field $\nub(y)$, $t\in\mathcal{C}$. The [**unit normal vector field**]{} to the surface ${\mathcal{C}}$, also known as the [**Gauß mapping**]{}, is defined by the vector product of the covariant basis $$\label{e1.88}
\nub(\cx):=\pm \frac{\boldsymbol{g}_1(\cx)\wedge\ldots\wedge\boldsymbol{g}_{n-1}(\cx)}{
|\boldsymbol{g}_1(\cx)\wedge\ldots\wedge\boldsymbol{g}_{n-1}(\cx)|},\qquad\cx\in\mathcal{C}.$$
The system of tangential vectors $\left\{ {\boldsymbol{g}}_{k}\right\}_{k=1}^{n-1}$ to ${\mathcal{C}}$ (cf. ) is, by the definition, linearly independent and is known as the **covariant basis**. There exists the unique system $\left\{ {\boldsymbol{g}}^{k}\right\} _{k=1}^{n-1}$ biorthogonal to it-the **contravariant basis**: $$\langle {\boldsymbol{g}}_{j},{\boldsymbol{g}}^{k}\rangle =\delta _{jk}\qquad
j,k=1,\ldots ,n-1.$$ The contravariant basis is defined by the formula: $$\label{e2.5}
{\boldsymbol{g}}^k=\frac{1}{\det \,G_\mathcal{S}}{\boldsymbol{g}}_1\wedge
\cdots \wedge{\boldsymbol{g}}_{k-1}\wedge{\boldsymbol{\nu}} \wedge{%
\boldsymbol{g}}_{k+1}\wedge \cdots \wedge{\boldsymbol{g}}_{n-1},\quad
k=1,\ldots ,n-1,$$ where $$G_{\mathcal{S}}({\scriptstyle{\mathcal{X}}}):=[\langle {\boldsymbol{g}}_{k}({%
\scriptstyle{\mathcal{X}}}),{\boldsymbol{g}}_{m}({\scriptstyle{\mathcal{X}}}%
)\rangle ]_{n-1\times n-1},\qquad p\in \mathcal{S}$$is the [**Gram matrix**]{}.
Next we expose yet another definition of a hypersurface-an [**implicit**]{} one.
\[d2.2\] Let $k\geq1$ and $\omega\subset\bR^n$ be a compact domain. An implicit $ C^k$-smooth (an implicit Lipschitz) hypersurface in $\bR^n$ is defined as the set $$\label{e2.2}
\cS=\Big\{\cx\in \omega\;:\; \Psi_\cS(\cx)=0\Big\}\, ,$$ where $\Psi_\cS\,:\,\omega\rightarrow \bR$ is a $ C^k$-mapping (or is a Lipschitz mapping) which is regular $\nabla\,\Psi(\cx)\not=0$.
(300,140) (-00,40)
-15mm
Note, that Definition \[d2.3\] and Definition \[d2.2\] of a hypersurface $\cS$ are equivalent and by taking a single function $\Psi_\cS$ for the implicit definition of a hypersurface $\cS$ we does not restrict the generality (see e.g., [@Du4]).
It is well known that using implicit surface functions gradient (see ) we can write an alternative definition of the unit normal vector field on the surface (see ): $$\label{e2.4}
\nub(y):=\lim_{x\to t}\frac{(\nabla\Psi _{\mathcal{S}})({x})}{|(\nabla
\Psi_{\mathcal{S}})(x)|},\qquad t\in\mathcal{S}.$$
In applications it is necessary to extend the vector field $\nub(t)$ in a neighborhood of $\cS$, preserving some important features. Here is the precise definition of extension.
\[d2.4.i\] Let ${\cS}$ be a surface in $\bR^n$ with unit normal $\nub$. A vector filed $\cN\in C^1(\Omega^\ve)$ in a neighborhood $\Omega^\ve$ of ${\cS}$, will be referred to as a [**proper extension**]{} if ${\cN}\Big|_\cS={\nub}$, if it is unitary $|{\cN}|=1$ in $\Omega^\ve$ and ${\cN}$ satisfies the following condition in the neighborhood $$\label{e2.2.19x}
\pa_j\cN_k(x)=\pa_k\cN_j(x)\qquad \mbox{for all}\quad x\in\Omega^\ve,\quad j,k=1,\ldots,n.$$
Such extension is needed, for example, to define correctly the normal derivative (the derivative along normal vector fields, outer or inner). It turned out that the “naive” extension (cf. ) $$\label{e2.4x}
\nub(t):=\frac{(\nabla\Psi _{\mathcal{S}})({x})}{|(\nabla
\Psi_{\mathcal{S}})(x)|},\qquad x\in\Omega^\ve$$ is not proper (see [@DST1] for a counterexample).
For the proof of the next Proposition \[p2.4\] and Corollary \[c2.5\] on extension of the normal vector field we refer to [@DST1].
\[p2.4\] Let ${\mathcal{S}}\subset \mathbb{R}^{n}$ be a hypersurface given by an implicit function $$\mathcal{S}=\left\{\cx\in\mathbb{R}^n\;:\;\Phi_\mathcal{S}(\cx)=0\right\}$$ for some $\Phi_\mathcal{S}\in C^1(\Omega^\ve)$. Then the gradient $\nabla\Phi_\mathcal{S}(x)$ of the function $$\label{e4.3}
\Phi_\mathcal{S}(\cx+t\nub(\cx)):=t, \qquad\cx+t\nub(\cx)\in\Omega^\ve,$$ defined in the parameterized neighborhood $$\Omega^\ve:=\left\{x=\cx+t\nub(\cx)\;:\;\cx\in \mathcal{S},\quad
-\varepsilon<t<\varepsilon\right\}$$ for sufficiently small $\varepsilon$, represents a unique proper extension of the unit normal vector field on the surface $$\nub(\cx)=\lim_{x\to\cx}\nabla\Phi_\mathcal{S}(x),\qquad \cx\in\mathcal{S}.$$
\[c2.5\] For any proper extension $\cN(x)$, $x\in\Omega^\ve\subset\bR^n$ of the unit normal vector field $\nub$ to the surface $\cS\subset\Omega^\ve$ the equality $$\label{eq:N}
\pa_\cN\cN(x)=0 \qquad\mbox{holds \ for \ all}\quad x\in\Omega^\ve.$$
In particular, for the derivatives $$\label{e2.7}
\cD_k=\pa_k-\cN_k\pa_\cN, \qquad k=1,\ldots,n\, ,$$ which are extension into the domain $\Omega^\ve$ of Günter’s derivatives $\cD_k=\pa_k-\nu_k\pa_{\nub}$ on the surface $\cS$, we have the equality: $$\begin{aligned}
\label{eq:D}
\cD_k\cN_j=\pa_k\cN_j-\cN_k\pa_\cN=\pa_k\cN_j, \qquad \cD_k\cN_j=\cD_j\cN_k,\\
\text{for all}\quad j,k=1,\ldots,n.\nonumber
\end{aligned}$$
In the sequel we will dwell on a proper extension and apply the above properties of $\cN$.
Important role in surface geometry goes to the [**Weingarten matrix**]{} $$\begin{aligned}
\label{e2.12}
\cW_\cS(\cx):=\left[\cD_j\nu_k(\cx)\right]_{n\times n},\qquad \cx\in\cS,
\end{aligned}$$ which is, due to the second equality in , a symmetric matrix. The mean trace of the Weingarten matrix is a [**mean curvature of the surface**]{}: $$\begin{aligned}
\label{e2.13}
\begin{array}{l}
\cH(\cx):=\dst\frac1{n-1}\text{Tr}\,\cW_\cS(\cx)=\dst\frac1{n-1}
\dst\sum\limits_{k=1}^n\cD_k\nu_k(\cx)=\dst\sum\limits_{k=1}^n
\lambda_k(\cx),\qquad \cx\in\cS,
\end{array}
\end{aligned}$$ where $\lambda_1(\cx),\ldots,\lambda_n(\cx)$ are the eigenvalues of $\cW_\cS(\cx)$. The Weingarten matrix is degenerated $$\det\,\cW_\cS(\cx)\equiv0\qquad \text{for all}\quad \cx\in\cS$$ because $\cW_\cS(\cx)\nub(\cx)\equiv0$. Therefore one of the eigenvalues is zero, say $\lambda_n(\cx)\equiv0$ for all $\cx\in\cS$. The [**Gauss curvature**]{} of the surface coincides with the product of non-degenerated eigenvalues of the Weingarten matrix: $$\begin{aligned}
\label{e2.14}
\cG_\cS(\cx):=\lambda_1(\cx)\cdots\lambda_{n-1}(\cx),\qquad \cx\in\cS
\end{aligned}$$ (cf. [@Du4; @Du5; @DK1] for details).
Calculus of tangential differential operators {#sec3}
=============================================
The content of the present section partly follows [@DMM1 § 4] and [@Du4 §§ 4,5].
In the present section we consider a hypersurface $\cS$, which is the boundary of some domain $\Omega\subset\bR^n$. The boundary hypersurface $\cS$ is given by an immersion . For the sake of simplicity we drop the indices $\Theta_1,\ldots,\Theta_m$, $\omega_1,\ldots,\omega_m$, but will resume indexing if necessary. $\nub(t)=(\nu_1(t),\ldots, \nu_n(t))^\top$ is the outer (with respect to $\Omega$) unit normal vector field to $\cS$ (cf. and ) and $\cN(x)$ is the proper extention of $\nub$ in a neighborhood $\omega_\cS$ of $\cS$ (cf. Definition \[d2.4.i\]).
A hypersurface $\cC$ is a part of $\cS$ and has a smooth boundary $\Gamma=\pa\cC$, given by another immersion $$\begin{aligned}
\label{e2.5ay}
\Theta_\Gamma\,:\,\partial\omega\rightarrow\Gamma:=\pa\cC\, ,
\qquad\partial\omega\subset\bR^{n-2}.
\end{aligned}$$ $\nub_\Gamma(t)$ is the outer normal vector field to the boundary $\Gamma$, which is tangential to $\cC$ (and to $\cS$).
By $\cV(\cS)$ we denote the set of all smooth vector fields, tangential to the hypersurface $\cS$: $$\label{eA1.1.1}
\U\;:\;\omega\rightarrow\bR^n\, ,\qquad \U(x)=\sum\limits_{j=1}^n
U_j(x)\e^j\, ,\qquad \langle\U(\cx),\nub(\cx)\rangle\equiv0$$ where $U^j\in C^\infty_0(\cS)$ and $\{\e^j\}_{j=1}^n$ is the natural Cartesian basis in $\bR^n$ $$\label{e1.1.x}
\e^1:=(1,0,\ldots,0),\ldots, \e^n:=(0,\ldots,0,1),$$ while $\langle x,y\rangle$ denotes the scalar product in $\mathbb{R}^n$: $$\langle x,y \rangle:=\sum_{j=1}^nx_jy_j,\qquad x,y\in\mathbb{R}^n.$$
A [**curve**]{} on a smooth surface $\cS$ is a mapping $$\label{eA2.1.1}
\gm\;:\;\cI\mapsto\cS\, ,\qquad \cI:=[0,1]\subset\bR\, ,$$ of a line interval $\cI$ to $\cS$.
Let $\U\in\cV(\cS)$ and consider the corresponding ordinary differential equations (ODE): $$\label{eA1.1.3}
y'=\U(y)\, ,\qquad y(0)=\cx\, ,\qquad \cx\in\cS\, .$$ A solution $y(t)$ of is called an [**integral curve**]{} (or [**orbit**]{}) of the vector field $\U$ and represents a subset of the surface $\cS$. The mapping $$\label{eA1.1.4}
y=y(t,\cx)=\cF^t_{\U}(\cx)\;:\;\cI\times\cS\rightarrow\cS\subset\bR^n,\qquad
\cI:=[0,1],$$ is called the [**flow**]{} generated by the vector field $\U$ at the point $\cx$.
A vector field $\U\in\cV(\Omega)$ defines the [**first order differential operator**]{} $$\label{eA1.1.5}
\U f(\cx)=\partial_{\U} f(\cx):=\lim\limits_{h\to0}\frac{f\left(
\cF^h_{\U}(\cx)\right)-f(\cx)}{h}=\frac d{dt}\left.f\left(
\cF^t_{\U}(\cx)\right)\right|_{t=0}$$ for a function defined on the surface $\cS$, which is called the derivative along $\U$. If $f(x)$ is defined in the neighbourhood of the surface $\cS$, by applying the chain rule to we get $$\begin{aligned}
\label{eA1.1.7}
\partial_{\U} f(x)=\langle \U(x),\nabla f(x)\rangle
=\sum\limits_{j=1}^nU_j(x)\frac{\pa f}{\pa x_j}.
\end{aligned}$$
In particular, the Günter’s derivatives $$\cD_j:=\pa_j-\nu_j\partial_{\nub}=\pa_j-\nu_j\sum_{k=1}^n\nu_k\pa_k\, ,
\qquad j=1,\ldots,n,$$ introduced in , are tangential. Another set of tangential derivatives on the surface $\cS$ is the Stokes’ derivatives $$\label{e3.14}
\cM_{jk}=\nu_j\pa_k-\nu_k\pa_j,\qquad j,k=1,\ldots,n.$$ Gunter’s and Stockes derivatives differentiate functions along the following tangent vector fields $$\begin{aligned}
\label{e2.x3}
\begin{array}{c}
\cD_j:=\partial_{\d^j}=\d^j\cdot\nabla\, ,\qquad
\cM_{jk}:=\partial_{\fm_{jk}}=\fm_{jk}\cdot\nabla\, ,\\[2mm]
\d^j:=\pi_\cS \e^j=\e^j-\nu_j\nub,\quad \fm_{jk}:=\nu_j\e_k-\nu_k\e_j\, ,\\[2mm]
\langle \d^j,\nub\rangle=0\, ,\qquad \langle\fm_{jk},\nub\rangle=0\, ,\quad j,k=1,\ldots,n\, .
\end{array}
\end{aligned}$$ The following reciprocal representations are easy to verify: $$\begin{aligned}
\label{e3.16}
\cD_j:=\sum \nu_k\cM_{kj},\qquad\cM_{jk}=\nu_j\cD_k-\nu_k\cD_j,\qquad j,k=1,\ldots,n.
\end{aligned}$$
The generating vector fields $\big\{\d^j\big\}_{j=1}^n$ and $\big\{\fm_{jk}\big\}_{j,k=1}^n$ are not bases in the space of tangential vectors to $\cS$, since they are linearly dependent $$\begin{aligned}
\label{e2.x4}
\sum_{j=1}^n\nu_j(\cx)\d^j(\cx)\equiv0, \qquad \fm_{jj}=0, \qquad \fm_{jk}=-\fm_{kj},
\end{aligned}$$ but both systems $\big\{\d^j\big\}_{j=1}^n$ and $\big\{\fm_{jk}\big\}_{0\leqslant j<k\leqslant n}$ are full and any tangential vector field $\U\in\cV(\cS)$ is represented as follows $$\begin{aligned}
\label{e2.x5}
\U(\cx)=\sum_{j=1}^nU^j(\cx)\d^j(\cx)=\sum_{0\leqslant j<k\leqslant n}^nc_{jk}
(\cx)\fm_{jk}(\cx)\, .
\end{aligned}$$
For a properly extended normal vector field $\cN$ (cf. Definition \[d2.4.i\]). we can extend the operators $\cD_j$ and $\cM_{jk}$ (cf. ) $$\begin{aligned}
\label{e3.19}
\cD_j=\pa_j-\cN_j\partial_\cN\, , \qquad \cM_{jk} :=\cN_j\pa_k
-\cN_k\pa_j\, ,\qquad 1\leqslant j,k\leqslant n
\end{aligned}$$ In the sequel, we shall make no distinction between the operator $\cD_j$ or $\cM_{jk}$ on $\cS$ and the extended one in $\bR^n$ given by .
Throughout the paper we use the following notation for the scalar products $$\label{e4}
\scal u,v\scar_\cS:=\oint\limits_\cS u^\top(t)\ov{v(t)}d\sigma\, ,
\qquad \scal\vf,\ v\scar_\Gm:=\oint\limits_\Gm\vf^\top(s)\ov{\ v(s)}d\fs\, .$$
For a tangential differential operator $P$ on a closed hypersurface $\cS$ let $P_\cS^*$ denote the “surface” adjoint: $$\begin{aligned}
\label{e10.6.0}
\scal P\vf,\psi\scar_\cS:=\oint\limits_{\cS}\langle P\vf,
\psi\rangle\,d\sigma=\oint\limits_{\cS}\langle\vf,P^*_\cS
\psi\rangle\,d\sigma=\scal\vf,P^*_\cS \psi\scar_\cS\\
\forall\,\vf,\,\psi\in C^1(\cS). \nonumber\end{aligned}$$
In [@DMM1] is shown that for a tangential differential operator $P\vf=\sum_{j=1}^na_j\pa_j \vf + b\vf$ the surface-adjoint and the formally adjoint operators coincide, i.e., $$\label{e2.a}
P^*_\cS\vf=P^*\vf=-\sum_{j=1}^n\pa_ja^\top_j\vf +b^\top\vf\, .$$
In particular, the Stokes’ derivatives are skew-symmetric $$\begin{aligned}
\label{e2.b}
\big(\cM^*_{jk}\big)_\cS=\cM^*_{jk}=-\cM_{jk}=\cM_{kj}\qquad
\forall\,j,k=1,\ldots,n\, ,\end{aligned}$$ while the adjoint operator to the operator $\cD_j$ is given by formula $$\label{e2.4.13}
\big(\cD_j\big)^*_\cS\vf=\cD_j^*\vf=-{\cD}_j\vf-\nu_j\cH^0_\cS\vf\, ,
\qquad \vf\in C^1(\cS),$$ where $\cH^0_\cS(\cx)=(n-1)\cH_\cS(\cx)$ is proportional to the mean curvature (see ).
\[t4.1\] The surface gradient and the surface divergence represented in Gunter’s derivatives have the following form $$\begin{aligned}
\label{e4.1}
\nabla_\cS\vf =\Bigl\{{\cD}_1\vf,{\cD}_2\vf,...,
{\cD}_n\vf\Bigr\}^\top,\\
\label{e4.2}
\Div_{\cS}\,\V=-\nabla^*_\cS \V:=\sum\limits_{j=1}^n\cD_jV^j,
\end{aligned}$$ where $\vf\in C^1(\cS)$ is a scalar function and $\V=\sum_{j=1}^nV^je_j\in\cV(\cS)$ is a 1-smooth tangential vector field. The Laplace-Beltrami operator $\Dlb_{\cS}$ on $\cS$ has the form $$\begin{aligned}
\label{e2.5.23}
\Dlb_\cS\,\psi=\sum\limits_{j=1}^n\cD_j^2\psi=\sum_{j<k}\cM_{jk}^2\psi
=\frac12\sum_{j,k=1}^n\cM_{jk}^2\psi\qquad\forall\,\psi\in C^2(\cS)\, .
\end{aligned}$$
The following Proposition \[p3.2\] is important while considering boundary value problems for Laplace-Beltrami equation (cf. [@Du3] for a proof).
\[p3.2\] For $\vf\in C^1(\cS)$ the surface gradient vanishes $\nabla_\cS\vf\equiv0$ if and only if $\vf(\cx)\equiv\const$.
Let $1<p<\infty$, $s\in\bR$. For the definition of Bessel potential $\bH^s_p(\cS)$ and Sobolev-Slobodeckii $\bW{}^s_p(\cS)$ spaces for a closed smooth manifold $\cS$ we refer to [@Tr1] (also see [@Du2; @Hr1] etc.). For $p=2$ the Sobolev–Slobodetski $\bW^s_2(\cS)$ and Bessel potential $\bH^s_2(\cS$ spaces coincide (i.e., the norms are equivalent). For an integer $m=1,2,\ldots$ the spaces $\bW{}^m_p(\cS)$ and $\bH^s_p(\cS)$ coincide with the Sobolev space and an equivalent norm in the Sobolev space is defined with the help of Gunter’s derivatives (the derivatives are understood in distributional sense) $$\|\vf\,\big|\,\bW_p^\ell(\cS)\,\|:=\left[\sum\limits_{|\al|\leq\ell}
\|\cD_\al\vf\,\big|\,\bL_p(\cS)\|\right]^{1/p}.$$
By $\bX_p^s(\cS)$ denote one of the following: Bessel potential $\bH^s_p(\cS)$ or Sobolev-Slobodeckii $\bW{}^s_p(\cS)$ space. Consider the space $$\begin{aligned}
\label{e1.21b}
\bX^s_{p,\#}(\cS):=\left\{\vf\in\bX^s_2(\cS)\;:\;\scal\vf,1\scar_\cS=0\right\}.
\end{aligned}$$ It is obvious, that $\bX^s_{p,\#}(\cS)$ does not contain constants: if $c_0={\rm const}\in\bX^s_{p,\#}(\cS)$ than $$0=\scal c_0,1\scar_\cS=c_0\scal1,1\scar_\cS=c_0\mes\,\cS$$ and $c_0=0$. Moreover, $\bX^s_p(\cS)$ decomposes into the direct sum $$\begin{aligned}
\label{e1.21c}
\bX^s_p(\cS)=\bX^s_{p,\#}(\cS)+\{{\rm const}\}
\end{aligned}$$ and the dual (adjoint) space is (see [@DTT1] for details) $$\begin{aligned}
\label{e1.21d}
(\bX^s_{p,\#}(\cS))^*=\bX^{-s}_{p',\#}(\cS), \qquad p':=\frac p{p-1}.
\end{aligned}$$
\[t3.3\] Let $\cS$ be an $\ell$-smooth closed hypersurface, $\ell=1,2,\ldots$, $1<p<\infty$ and $|s|\leqslant\ell$. Let $\bX^s_p(\cS)$ be the same as in - .
Let $\cA$ be a positive definite matrix-function $$\label{e3.26}
\langle \mathcal{A}(\cx)\xi,\xi\rangle\geqslant C\|\xi\|^2, \qquad \xi\in\bR^n$$ for all $\cx\in\cC$. Then the “anisotropic” Laplace-Beltrami operator between the spaces with detached constants (see ) $$\begin{aligned}
\Div_\cS(\mathcal{A}\,\nabla_\cS)\;:\;\bX^{s+1}_{p,\#}(\cS)\to\bX^{s-1}_{p,\#}(\cS).
\end{aligned}$$ is invertible. Moreover, in the setting $$\begin{aligned}
-\Div_\cS(\mathcal{A}\,\nabla_\cS)\;:\;\bW^1_{2,\#}(\cS)\to\bW^{-1}_{2,\#}(\cS)
\end{aligned}$$ the operator is self adjoint and positive definite: $$\begin{aligned}
\label{e3.28}
&&\hskip-15mm\scal-\Div_\cS(\mathcal{A}\nabla_\cS\vf),\psi\scar_\cS
=\scal\vf,-\Div_\cS(\mathcal{A}\nabla_\cS\psi)\scar_\cS,\\[2mm]
\label{e3.29}
&&\hskip-15mm\scal-\Div_\cS(\mathcal{A}\nabla_\cS\vf),\vf\scar_\cS\geq
C\|\vf\big|\bW^1_{2,\#}(\cS)\|^2 \qquad \text{for all} \quad\vf,\psi\in\bW^1_{2,\#}(\cS).
\end{aligned}$$
[**Proof:**]{} For the proof see [@DTT1 Theorem 1.10]).
Now let $\cC\subset\cS$ be a smooth subsurface of a closed hypersurface $\cS$ and $\gm=\pa\cC\not=\emptyset$ be its smooth boundary $\pa\cC=\Gm$ (see Fig. 2).
(300,140) (-00,40)
-15mm
The space $\wt {\bH}_p^s(\cC)$ is defined as a subspace of those functions $\vf\in \bH_p^s(\cS)$, which are supported in the subsurface, $\supp\,\vf\subset\ov{\cC}$, whereas $\bH_p^s(\cC)$ denotes the quotient space $\bH_p^s(\cC)=\bH_p^s(\cS)\Big/\wt{\bH}_p^s(\cC^c)$, where $\cC^c:=\cS\setminus\ov{\cC}$ is the complementary surface to $\cC$. The space $\bH_p^s(\cC)$ can be identified with the space of distributions $\vf$ on $\cC$ which have an extension to a distribution $\ell\vf\in\bH_p^s(\cS)$. Therefore $r_\cC\bH_p^s(\cS)=\bH_p^s(\cC)$, where $r_\cC$ denotes the restriction operator of functions (distributions) from the surface $\cS$ to the subsurface $\cC$.
The spaces $\wt {\bW}_p^s(\cC)$ and $\bW_p^s(\cC)$ are defined similarly (see [@Tr1] and also [@Du2; @Hr1] etc.).
The subspaces $\wt{\bH}_{p,\#}^s(\cC)$, $\wt{\bW}_{p,\#}^s(\cC)$, $\bH_{p,\#}^s(\cC)$ and $ \bW_{p,\#}^s(\cC)$ are defined similarly as in : they consist of functions from the corresponding spaces which have mean value zero $\scal\vf,1\scar_\cC=0$.
Let us consider the following boundary value problems for the ”anisotropic” Laplace equation with mixed boundary conditions $$\label{e3.32}
\left\{\begin{array}{ll}
\Div_\cC(\mathcal{A}\,\nabla_\cC u)(t)=f(t),\qquad & t\in\cC, \\[0.2cm]
u^+(s)=g(s), \qquad & {\rm on} \quad \Gm_D, \\[0.2cm]
\langle\nub_\Gm(s),(\mathcal{A}\,\nabla_\cC u)^+(s)\rangle=h(s),\qquad &{\rm on}\quad \Gm_N,
\end{array}\right.$$ where $\pa\cC=\Gm=\Gm_D\cup\Gm_N$ is a decomposition of the boundary in two connected parts and $\mathcal{A}=\{a_{ij}\}$ is $n\times n$ strictly positive definite matrix-function (see .
The BVP we consider in the following weak classical setting $$\label{e3.33}
f\in\wt\bH^{-1}(\cC),\qquad g\in\bH^{1/2}(\Gamma_D),\qquad
h\in\bH^{-1/2}(\Gamma_N).$$
\[t3.4\] The mixed boundary value problem in the weak classical setting has a unique solution in the space $\bW^1(\cC)$.
[**Proof:**]{} For the proof we quote [@DTT1 Theorem 2.2].
In conclusion of the present section let us recall the definition of surface $\delta$-function (see, e.g., [@Du2 (4.30) in § 4]).
Let $k=1,2,\ldots$ and $\cC$ be a $C^k$-smooth hypersurface in $\bR^n$, open or closed. The surface $\dl$-function is defined by the equality $$\label{e3.32x}
(g\otimes\delta_\cC,v)_{\cC}:=(g,\gm_\cS v)_{\cC}=\int\limits_\cC g(\tau)\gm_\cS
v(\tau)d\sigma,\qquad g\in C^k(\cC),\quad v\in C_0^k(\bR^n),$$ where $\gm_\cS v(\tau)$ denotes the trace on the boundary surface. Obviously, $\supp(g\otimes\delta_\cC)=\supp\,g\subset\cC$.
In the next Lemma \[l3.5\] the definition is extended to less regular functions.
\[l3.5\] Let $1<p<\infty$ $(1\leq q\leq\infty),$ $s<0$, $g\in{\bH}^s_p(\cC)$ (or $g\in {\bB}^s_{p,q}(\cC)$). Then $$g\otimes\delta_\cC\in\bH_p^{s-\frac1{p'}}(\bR^n\setminus\cC)\, , \quad
\left(g\otimes\delta_\cS\in\bB_{p,q}^{s-\frac1{p'}}(\bR^n\setminus\cC)\right),$$ where $p'=p/(p-1)$. In particular, if $g\in{\bH}^{-1/2}(\cC)$, than $g\otimes\delta_\cC\in\bH^{-1}(\bR^n\setminus\cC)$.
We conclude the section with some auxiliary results on Lebesgue points of integrable functions, which we apply in proofs later in § \[sec7\].
Let $B(x)$ be a ball in the Euclidean space $B\subset\mathbb{R}^n$ centered at $x$. [*The derivative of the integral at $x$*]{} is defined to be $$\begin{aligned}
\label{e3.32a}
\lim_{B(x) \rightarrow x} \frac{1}{|B(x)|} \int_{B(x)}f(y) \, \mathrm{d}y,\end{aligned}$$ where $|B(x)|$ denotes the volume (i.e., the Lebesgue measure) of $B(x)$, and $B(x)\rightarrow x$ means that the diameter of $B(x)$ tends to $0$. Note that $$\begin{aligned}
\label{e3.32b}
\left|\frac{1}{|B(x)|} \int_{B(x)}f(y)\,dy - f(x)\right| = \left|\frac{1}{|B(x)|} \int_{B(x)}[f(y) -f(x)]\,dy\right|\nonumber\\
\leqslant \frac{1}{|B(x)|} \int_{B(x)}|f(y) -f(x)|\,dy.\end{aligned}$$ The points $x$ for which the right hand side tends to zero are called the [*Lebesgue points of $f$.*]{}
\[t7.7\] For an integrable function $f\in\bL_1(\Omega)$ the derivative of the integral exists and is equal to $f(x)$ at almost every point $x\in\Omega$.
Moreover, almost every point $x\in\Omega$ is a Lebesgue point of $f$ (see ).
\[c7.8\] If $g\in\bL_2(\Omega)$, $f\in\bL_2(\Omega\times(-1,1))$, then $$\begin{aligned}
\label{e7.45}
\lim_{\ve\rightarrow0}\frac{1}{2\ve}\int_{t-\ve}^{t+\ve}\scal g(\cdot),f(\cdot,\tau))\scar_{\Omega}d\tau=\scal g(\cdot),f(\cdot,t))\scar_{\Omega}
\end{aligned}$$ for almost all $t\in(-1,1)$.
[**Proof:**]{} It is clear, that $g\cdot f\in\bL_1(\Omega\times(-1,1))$ and for the function $h(t):=\scal g(\cdot),f(\cdot,t) \scar_{\Omega}$ the inclusion $h\in\bL_1((-1,1))$ is true. Thence we can apply Theorem \[t7.7\] to the function $h(t)$ and get .
Laplace operator in a layer domain {#sec4}
==================================
We will keep the notation of §2: $\Theta$, $\Omega^\ve$, $\omega$, $\cS$ and $\cC$. We consider a [**layer domain**]{} $$\begin{aligned}
\label{e5.1}
\Omega^\ve:=\Big\{\cx_t\in\bR^n\,:\,\cx_t=\cx+t\nub(\cx)=\Theta(x)
+t\nub\big(\Theta(x)\big)\, ,\quad x\in\omega\, ,\;-\ve<t<\ve\Big\}\nonumber\\[2mm]
=\mathcal{C}\times(-\varepsilon,\varepsilon),
\end{aligned}$$ where $\nub(\cx)=\nub(\Theta(x))$ for $\cx=\Theta(x)\in\cS$, is the outer unit normal vector field (see and ). The surface $\cC$ is a mid-surface for the layer domain.
We will also use the notation $\nub(x):=\nub(\Theta(x))$ for brevity unless this does not leads to a confusion. The coordinate $t$ will be referred to as the [**transverse variable**]{}.
Without going into detail let us remark only that if the hypersurface $\cS$ is $C^1$-smooth and $1/\ve$ is more than the maximum of modules of all principal curvatures of the surface $\cS$ (i.e., of all eigenvalues $|\lambda_1(\cx)|,\ldots,|\lambda_{n-1}(\cx)|,|\lambda_n(\cx)|$ of the Weingarten matrix $\cW_\cS(\cx)$, $\cx\in\cS$), then the mapping $$\begin{aligned}
\label{e5.2}
\begin{array}{c}
\Theta^\ve\,:\,\omega^\ve:=\omega\times(-\ve,\ve)\rightarrow \Omega^\ve\, ,
\quad\omega^\ve\subset\bR^n\, ,\\[2mm]
\Theta^\ve(y,t):=\Theta(y)+t\nub(y)\, ,\qquad (y,t)\in\omega^\ve
\end{array}
\end{aligned}$$ is a diffeomorphism.
We will also suppose that $\cN$ is a proper extension of the outer unit normal vector field $\nub(t)$ into the layer neighborhood $\Omega^\ve$ (cf. Definition \[d2.4.i\]).
The n-tuple $\g_1:=\pa_1\Theta,\ldots,\g_{n-1}:=\pa_{n-1}\Theta,\g_n:=\cN$, where $\cN$ is the proper extension of $\nub$ in the neighborhood $\Omega^\ve$, is a basis in $\Omega^\ve$ and arbitrary vector field $\U=\sum_{j=1}^nU^0_j\e^j\in\cV(\Omega^\ve)$ is represented with this basis in “curvilinear coordinates".
Let us consider the system of $(n+1)$-vectors $$\begin{aligned}
\label{e5.15}
\d\,^j:=\e^j-\cN_j\cN\, ,\qquad j=1,\ldots,n \quad {\rm and}\quad
\d\,^{n+1}:=\cN,
\end{aligned}$$ where $\e^1,\ldots,\e^n$ is the Cartesian basis in $\bR^n$ (cf. ); the first $n$ vectors $\d\,^1,\ldots,\d\,^n$ are tangential to the surface $\cC$, while the last one $\d\,^{n+1}=\cN$ is orthogonal to all $\d\,^1,\ldots,\d\,^n$. This system is, obviously, linearly dependent, but full and any vector field $\U\in\cV(\Omega^\ve)$ is written in the following form: $$\begin{aligned}
\label{e5.16}
\U=\dst\sum_{j=1}^nU_j\e^j=\dst\sum_{j=1}^{n+1}U^0_j\d\,^j.
\end{aligned}$$
Since the system $\big\{\d\,^j\big\}_{j=1}^{n+1}$ is linearly dependent $$\begin{aligned}
\label{e5.17}
\dst\sum_{j=1}^n\cN_j\d\,^j=0,\qquad \langle\cN_j,\d^j\rangle=0,\quad j=1,\ldots,n,
\end{aligned}$$ the representation is not unique. To fix the unique representation in we will keep the following convention: $$\begin{aligned}
\label{e5.19}
U^0_j:=U_j-\langle\cN,\U\rangle\cN_j,\quad j=1,\ldots,n,\quad
U^0_{n+1}=\langle\cN,\U\rangle=\dst\sum_{j=1}^nU_j\cN_j.
\end{aligned}$$ The convention is natural because if the vector $\U(\cx)$ is tangent to $\cC$ for $\cx\in\cC$, than $U^0_j(\cx):=U_j(\cx)$ for $j=1,\ldots,n$ and $U^0_{n+1}(\cx)=0$.
Moreover, if the scalar product of vectors $$\begin{aligned}
\label{e5.32}
\U:=\dst\sum_{j=1}^nU_j\d\,^j=\dst\sum_{j=1}^{n+1}U^0_j\d\,^j,\quad
\V:=\dst\sum_{j=1}^nV_j\d\,^j=\dst\sum_{j=1}^{n+1}V^0_j\d\,^j
\end{aligned}$$ is defined by the equality $$\langle\U^0,\V^0\rangle:=\sum_{j=1}^{n+1}U^0_j\overline{V^0_j},$$ than the “new” and the “old” scalar products coincide: $$\begin{aligned}
\label{e5.33}
\langle\U,\V\rangle^0&\hskip-3mm=&\hskip-3mm\sum_{j=1}^{n+1}U^0_j\overline{V^0_j}
=\sum_{j=1}^n(U_j-\cN_j\langle\cN,\U\rangle)(\overline{V_j-\cN_j\langle
\cN,\V\rangle})+\langle\cN,\U\rangle\langle\cN,\V\rangle\nonumber\\
&\hskip-3mm=&\hskip-3mm\sum_{j=1}^nU_j\overline{V_j}=\langle\U,\V\rangle.
\end{aligned}$$ In particular, $$\label{e5.34}
\|\U\|^0:=\sum_{j=1}^{n+1}|U^0_j|^2=\sum_{j=1}^n|U_j|^2=\|\U\|.$$
Note for a later use, that due to the equalities and the convention we get $$\begin{aligned}
\label{e5.22}
\pa_{\U}&\hskip-3mm=&\hskip-3mm\dst\sum_{j=1}^nU_j\pa_j=\sum_{j=1}^n[U^0_j\pa_j
+\langle\cN,\U\rangle\cN_j\pa_j=\sum_{j=1}^nU^0_j(\pa_j-\cN_j\pa_\cN)
+\langle\cN,\U\rangle\pa_\cN\nonumber\\
&\hskip-3mm=&\hskip-3mm\sum_{j=1}^nU^0_j\cD_j+U_{n+1}\cD_{n+1}
=\sum_{j=1}^{n+1}U^0_j\cD_j=:\cD_\U.
\end{aligned}$$
\[d5.6\] For a function $\vf\in\bW^1(\Omega^{\ve})$ the extended gradient is $$\label{e5.23}
\cD_{\Omega^\ve}\,\vf=\Bigl\{\cD_1\vf,...,\cD_n\vf,
\cD_{n+1}\vf\Bigr\}^\top
=\sum_{j=1}^{n+1}(\cD_j\vf)\d^j,\qquad \cD_{n+1}\vf:=\pa_\cN\vf$$ and for a smooth vector field $\U=\dst\sum_{j=1}^{n+1}U^0_j\d\,^j\in\cV( \Omega^\ve)$ (see , ) the extended divergence is $$\begin{aligned}
\label{e5.24}
\Div_{\Omega^\ve}\,\U:=\dst\sum\limits_{j=1}^{n+1}\cD_jU^0_j
+\cH^0_\cC\langle\cN,\U\rangle=-\nabla^*_{\Omega^\ve}\U,
\end{aligned}$$ since $$\begin{aligned}
\label{e5.25}
\cH^0_{\Omega^\ve}(x):=\sum_{j=1}^n\pa_j\cN_j(x)=\sum_{j=1}^{n+1}\cD_j\cN_j(x)
=\sum_{j=1}^n\cD_j\nu_j(t)=\cH^0_\cC(t),\\
x\in\Omega^\ve,\qquad t=\pi_\cS x\nonumber
\end{aligned}$$ and $\cH^0_\cC(t)$ differs from the mean curvature $\cH_{\cC}(t)$ (see ) by the constant multiplier $\cH^0_\cC(t)=(n-1)\cH_\cC(t)$.
\[l5.7\] The classical gradient $\nabla\vf:=\Bigl\{\pa_1 \vf,...,\pa_n\vf\Bigr\}^\top$, written in the full system of vectors $\big\{\d\,^j\big\}_{j=1}^{n+1}$ in coincides with the extended gradient $\cD_{\Omega^\ve}\,\vf$ in .
Similarly: the classical divergence $\Div\,\U:=\dst\sum\limits_{j=1}^n \pa_jU_j$ of a vector field $\U:=\dst\sum\limits_{j=1}^nU_j\e^j$, written in the full system , coincides with the extended divergence $\Div\,\U=\Div_{\Omega^\ve}\,\U$ in .
The extended gradient and the negative extended divergence are dual $\nabla^*_{\Omega^\ve}=-\Div_{\Omega^\ve}$ and $\Div^*_{\Omega^\ve}=-\cD_{\Omega^\ve}$.
The Laplace-Beltrami operator $\Delta_{\Omega^\ve}:=\Div_{\Omega^\ve} \cD_{\Omega^\ve}\,\vf=-\nabla^*_{\Omega^\ve}\,\Big(\cD_{\Omega^\ve}\vf\Big)$ on $\Omega^\ve$, written in the full system , acquires the following form $$\begin{aligned}
\label{e5.26}
\Delta_{\Omega^\ve}\vf=\sum\limits_{j=1}^{n+1}\cD_j^2\vf\, ,\quad
\vf\in\bW^2(\Omega^{\ve})\, .
\end{aligned}$$
[*Proof:*]{} A similar lemma is proved in [@Du5 Lemma 4.3], but definition of the divergence $\Div_{\Omega^\ve}$ is different there. Therefore we expose the full proof below.
That the gradients coincide follows from the choice of the full system : $$\begin{aligned}
\label{e5.27}
\nabla\vf&\hskip-3mm:=&\hskip-3mm\Bigl\{\pa_1\vf,...,\pa_n\vf
\Bigr\}^\top=\sum_{j=1}^n(\pa_j\vf)\e^j=\sum_{j=1}^n
(\cD_j\vf+\cN_j\cD_{n+1}\vf)\e^j\nonumber\\
&\hskip-3mm=&\hskip-3mm\sum_{j=1}^n(\cD_j\vf)\d\,^j+(\cD_{n+1}\vf)\cN
=\sum_{j=1}^{n+1}(\cD_j\vf)\d\,^j=\cD_{\Omega^\ve}\vf
\end{aligned}$$ since $$\begin{aligned}
\label{e5.28}
\begin{array}{c}
\e^j=\d^j+\cN_j\cN,\quad \pa_j=\cD_j+\cN_j\cN,\\
\dst\sum_{j=1}^n\cN_j\cD_j=0,\quad
\dst\sum_{j=1}^n(\cD_j\vf)\e^j=\dst\sum_{j=1}^n(\cD_j\vf)\d\,^j.
\end{array}
\end{aligned}$$ By applying and we proceed as follows: $$\begin{aligned}
\label{e5.29}
\Div\,\U&\hskip-3mm=&\hskip-3mm\sum_{j=1}^n\pa_jU_j=\sum_{j=1}^n\cD_jU_j
+\sum_{j=1}^n\cN_j\pa_\cN U_j=\sum_{j=1}^n\cD_j\left[U^0_j+
\cN_j\langle\cN,\U\rangle\right]\nonumber\\
&&+\sum_{j=1}^n\pa_\cN\big(\cN_jU_j\big)=\sum_{j=1}^n\cD_jU^0_j
+\sum_{j=1}^n(\cD_j\cN_j)\langle\cN,\U\rangle
+\cD_{n+1} U^0_{n+1}\nonumber\\
&\hskip-3mm=&\hskip-3mm\sum_{j=1}^{n+1}\cD_jU^0_j+\cH^0_\cC\langle\cN,\U\rangle
=\Div_{\Omega^\ve}\U.
\end{aligned}$$
The proved equality and the classical equality $\nabla^*=-\Div$, ensure the both claimed equalities $\nabla^*_{\Omega^\ve}=-\Div_{\Omega^\ve}$ and $\Div^*_{\Omega^\ve}=-\cD_{\Omega^\ve}$: $$\scal\cD_{\Omega^\ve}\vf,\U\scar=\scal\nabla\,\vf,\U\scar
=-\scal\vf,\Div\,\U\scar=-\scal\vf,\Div_{\Omega^\ve}\U\scar.$$
Formula for the Laplace-Beltrami operator is a direct consequence of equalities , and definitions $$\Dlb\vf=\Div\,\nabla\vf=\Div_{\Omega^\ve}\cD_{\Omega^\ve}\vf=\sum_{j=1}^{n+1}
\cD^2_j\vf+\langle\cN,\cD_{\Omega^\ve}\vf\rangle
=\sum_{j=1}^{n+1}\cD^2_j\vf,$$ because (see the third formula in ) $\langle\cN,\cD_{\Omega^\ve}\vf\rangle=\dst\sum_{j=1}^n\cN_j\cD_j\vf=0$.
Let us check the following equalities for a later use: $$\begin{aligned}
\label{e5.30}
\cD_{\Omega^\ve}\U=\big[\cD_jU^0_k\big]_{n+1\times n+1}+\langle\cN,\U\rangle\cW_{\Omega^\ve},
\end{aligned}$$ where $$\begin{aligned}
\U:=\dst\sum_{m=1}^{n+1}U^0_m\d\,^m=\dst\sum_{m=1}^nU_m\e^m,\quad
U^0_{n+1}=\dst\sum_{m=1}^n\cN_mU_m,\quad \cD_{n+1}:=\pa_\cN,\quad \d^{n+1}:=\cN.
\end{aligned}$$ $\cW_{\Omega^\ve}$ is the extended Weingarten matrix (cf. ) $$\begin{aligned}
\label{e5.31}
\cW_{\Omega^\ve}:=\big[\cD_j\cN_k\big]_{n+1\times n+1}
\end{aligned}$$ and its last column and last row are $0$, because $\cD_j\cN_{n+1}=\cD_{n+1}\cN_j=\cD_{n+1}\cN_{n+1}=0$ for $j=1,\ldots,n$.
In fact (see fore some further details of calculation): $$\begin{aligned}
\cD_{\Omega^\ve}\U&\hskip-3mm:=&\hskip-3mm
\big[\pa_jU_k\big]_{n\times n}=\sum_{j,ky=1}^n\pa_jU_k\e^j\otimes\e^k\\
&\hskip-3mm:=&\hskip-3mm\sum_{j,k=1}^n[\cD_j+\cN_j\pa_\cN][U^0_k+\cN_k\langle
\cN,\U\rangle][\d^j+\cN_j\cN]\otimes[\d^k+\cN_k\cN]\\
&\hskip-3mm=&\hskip-3mm\sum_{j,k=1}^n(\cD_jU^0_k)\d^j\otimes[\d^k+\cN_k\cN]
+\sum_{j,k=1}^n\cD_j[\cN_k\langle\cN,\U\rangle]\d^j\otimes[\d^k+\cN_k\cN]\\
&&+\sum_{j,k=1}^n\cN^2_j(\pa_\cN U^0_k)\cN\otimes[\d^k+\cN_k\cN]
+\sum_{j,k=1}^n\cN^2_j\cN^2_k\pa_\cN\langle\cN,\U\rangle\cN\otimes\cN
\end{aligned}$$ -10mm $$\begin{aligned}
&\hskip-3mm=&\hskip-3mm\sum_{j,k=1}^n(\cD_jU^0_k)\d^j\otimes\d^k
+\sum_{j,k=1}^n\cN_k(\cD_jU^0_k)\d^j\otimes\d^{n+1}\\
&&+\sum_{j,k=1}^n\langle\cN,\U\rangle(\cD_j\cN_k)\d^j\otimes[\d^k+\cN_k\cN]
+\sum_{j,k=1}^n\cN_k^2\cD_j\langle\cN,\U\rangle\d^j\otimes\d^{n+1}\\
&&+\sum_{k=1}^n(\cD_{n+1}U^0_k)\d^{n+1}\otimes\d^k
+\sum_{k=1}^n\left[\cN_k\cD_{n+1}U^0_k
+\cD_{n+1}U^0_{n+1}\right]\d^{n+1}\otimes\d^{n+1}\\
&\hskip-3mm=&\hskip-3mm\sum_{j,k=1}^n(\cD_jU^0_k)\d^j\otimes\d^k
+\sum_{j,k=1}^n\left[\cD_j(\cN_kU^0_k)-U^0_k\cD_j\cN_k
\right]\d^j\otimes\d^{n+1}\\
&&+\langle\cN,\U\rangle\sum_{j,k=1}^n(\cD_j\cN_k)\d^j\otimes\d^k
+\sum_{j=1}^n\cD_j\langle\cN,\U\rangle\d^j\otimes\d^{n+1}\\
&&+\sum_{k=1}^n(\cD_{n+1}U^0_k)\d^{n+1}\otimes\d^k
+(\cD_{n+1}U^0_{n+1})\d^{n+1}\otimes\d^{n+1}\\
&\hskip-3mm=&\hskip-3mm\sum_{j,k=1}^{n+1}(\cD_jU^0_k)\d^j\otimes\d^k
-\sum_{j,k=1}^nU^0_k(\cD_j\cN_k)\d^j\otimes\d^{n+1}
+\langle\cN,\U\rangle\sum_{j,k=1}^n(\cD_j\cN_k)\d^j\otimes\d^k\\
&\hskip-3mm=&\hskip-3mm\big[\cD_jU_k\big]_{(n+1)\times(n+1)}
+\langle\cN,\U\rangle\cW_{\Omega^\varepsilon}
-\sum_{j,k=1}^nU^0_k(\cD_j\cN_k)\d^j\otimes\d^{n+1}
\end{aligned}$$ $$\begin{aligned}
&\hskip-3mm=&\hskip-3mm\big[\cD_jU_k\big]_{(n+1)\times(n+1)}
+\langle\cN,\U\rangle\cW_{\Omega^\varepsilon}
-\big[(\cW_{\Omega^\varepsilon}\U^0)_j\delta_{j,n+1}\big]_{(n+1)\times(n+1)},
\end{aligned}$$ since $$\begin{aligned}
\pa_\cN\cN_j=0,\quad\sum_{j,k=1}^n\cN^2_j=1,\quad \sum_{j=1}^n\cN_j\cD_j=0,
\quad \sum_{j=1}^n\cN_j\d^j=0,\\
\sum_{k=1}^n\cN_kU^0_k=0,\quad\sum_{k=1}^n\cN_k\cD_j\cN_k=\frac12\cD_j\sum_{k=1}^n\cN^2_k=\frac12\cD_j1=0,
\quad j=1,2\ldots,n+1.
\end{aligned}$$
Convex energies {#sec5}
===============
Let again $\Omega^\varepsilon$ be a layer domain of width $2\varepsilon$ in the direction transversal to the mid-surface $\cC$ (see § \[sec4\]).
Any minimizer $u$ of the energy functional $$\label{e6.1}
\cE^\varepsilon(u):=\int_{\Omega^\varepsilon}\langle\nabla\,u,\nabla\,u
\rangle\,dy, \qquad u\in C^\infty(\Omega^\varepsilon)$$ should satisfy $$\begin{aligned}
\label{e6.2}
0&\hskip-3mm=&\hskip-3mm\frac{d}{dt}\cE^\varepsilon(u+tv)\Bigl|_{t=0}
=\int_{\Omega^\varepsilon}\left[\langle\nabla\,u,\nabla\,v\rangle
+\langle\nabla\,v,\nabla\,u\rangle\right]\,dy \nonumber\\
&\hskip-3mm=&\hskip-3mm2{\rm Re}\int_{\Omega^\varepsilon}\langle\nabla\,u,
\nabla\,v\rangle\,dy=-2{\rm Re}\int_{\Omega^\varepsilon}\langle\Div\nabla\,
u,v\rangle\,dy=-2{\rm Re}\int_{\Omega^\varepsilon}\langle\Delta\,
u,v\rangle\,dy
\end{aligned}$$ for arbitrary $u\in C^\infty(\Omega^\varepsilon)$ and $v\in C^\infty_0(\Omega^\varepsilon)$, which implies $$\label{e6.3}
\Delta\,u=0\qquad\mbox{ on }\quad \Omega^\varepsilon.$$
In other words, is the Euler-Lagrange equation associated with the energy functional .
Similarly, minimizers of the energy functional $$\label{e6.4}
\cE_0(u):=\int_{\cC}\langle\nabla_\cC u,\nabla_\cC u\rangle\,d\sigma, \qquad
u\in C^\infty(\cC)$$ on the hypersurface $\cC$ should satisfy the following Laplace-Beltrami equation $$\label{e6.5}
\Delta_\cC u:=\Div_\cC\nabla_\cC u=0\qquad\text{on}\quad \cC.$$ To treat the dimension reduction problem for the Laplace equation (see [@Br1] for a similar consideration in case of a flat 3D body), we assume, without restricting generality, that $\Omega^1$ (i.e., for $\varepsilon=1$) is still a layer domain. Otherwise we can first change the variable $\cx_n=\ve_0\bar\cx_n$, $0<\bar\cx_n<1$, where $0<\ve_0<1$ is such that $\Omega^{\varepsilon_0}$ is still a layer domain.
Next we introduce a new coordinate system (cf. ) $$\begin{aligned}
\label{e6.6}
\begin{array}{c}
x:=\dst\sum_{m=1}^nx_m\e^m=\dst\sum_{m=1}^n\cx_m\d\,^m+t\d\,^{n+1},\\
\cx_k:=x_k-\cN_k\langle\cN,x\rangle,\quad k=1,\ldots,n,\quad t=\cx_{n+1}
:=\langle x,\cN\rangle=\dst\sum_{m=1}^nx_m\cN_m
\end{array}
\end{aligned}$$ and the scalar product of elements $$\cx:=\dst\sum_{m=1}^{n+1}\cx_m\d\,^m,\quad \cy:=\dst\sum_{m=1}^{n+1}\cy_m\d\,^m$$ define by the equality (cf. similar in ) $$\langle\cx,\cy\rangle:=\sum_{j=1}^{n+1}\cx_j\overline{\cy_j}.$$ Then (cf. -) $$\begin{aligned}
\label{e6.7}
\langle\cx,\cy\rangle=\sum_{j=1}^{n+1}\cx_j\overline{\cy_j}
=\sum_{j=1}^n(x_j-\cN_j\langle\cN,x\rangle)(\overline{(y_j-\cN_j\langle
\cN,y\rangle)})+\langle\cN,x\rangle\langle\cN,y\rangle\nonumber\\
=\sum_{j=1}^nx_j\overline{y_j}=\langle x,y\rangle.
\end{aligned}$$ In particular, $$\|\cx\|:=\sum_{j=1}^{n+1}|\cx_j|^2=\sum_{j=1}^n|x_j|^2=\|x\|.$$
Due to Lemma \[l5.7\] the classical gradient in the energy functional can be replaced by the extended gradient $$\begin{aligned}
\label{e6.9}
\cE^\varepsilon(u)&\hskip-3mm:=&\hskip-3mm\int_{\Omega^\varepsilon}\langle
\cD_{\Omega^\varepsilon}u(y),\cD_{\Omega^\varepsilon} u(y)\rangle\,dy\nonumber\\
&\hskip-3mm=&\hskip-3mm\int_{-\varepsilon}^\varepsilon
\int_\cC\left[\langle\cD_\cC u(\cx,t),\cD_\cC u(\cx,t)\rangle
+|\pa_tu(\cx,t)|^2\right]\,d\sigma\,dt,\\
&&\hskip50mm\cD_\cC:=\left(\cD_1,\ldots,\cD_n\right)^\top\nonumber
\end{aligned}$$ for arbitrary $u\in\bW^1(\Omega^\varepsilon)$, because $\cD_{n+1}=\pa_\cN =\pa_t$. Here $\cC$ is the mid surface of the layer domain $\Omega^\varepsilon=\cC\times(-\varepsilon,\varepsilon)$ and $d\sigma$ is the surface measure on $\cC$.
Due to the representation and the new coordinate system we can apply the scaling with respect to the variable $t$ and study the scaled energy. The approach is based on $\Gamma$-convergence (see [@Br1; @FJM1]) and can be applied to a general energy functional which is convex and has square growth. The problem we have in mind is the following: *Do these energies defined on thin n-dimensional domains $\Omega^\varepsilon$ converge (and in which sense) to an energy defined on the $n-1$ dimensional Hypersurface $\cC$ (the mid-surface of $\Omega^\varepsilon$) when the domain $\Omega^\varepsilon$ is “squeezed” infinitely in the transversal direction to $\cC$?*
In the next two sections we apply the results developed in the present paper to boundary value problems for the heat conduction by a hypersurface. In particular we shall show, that when the thickness of the layer domain $\Omega^\varepsilon$, with the mid-surface $\cC$, tends to zero, a solution to the linear heat conduction equation Gamma-converges to a solution to the certain boundary value problem Laplace-Beltrami equation on the mid-surface $\cC$ written explicitly (see \[sec4\]).
Variational reformulation of a heat transfer problems {#sec6}
=====================================================
Let $\Omega$ be a bounded Lipschitz domain in $\mathbb{R}^3$ with piecewise smooth boundary $\partial \Omega = \overline{\cC}_D \cup \overline{\cC}_N$, where $\cC_D$ and $\cC_N$ are open non-intersecting surfaces $\cC_D\cap\cC_N= \varnothing$ and their common boundary is a smooth arc. Denote by $\nub=(\nu_1,\nu_2,\nu_3)^\top$ the unit normal on $\cC$, external with respect to $\Omega$.
We consider the general steady-state, linear heat transfer problem for a medium occupying domain $\Omega$. We assume that on the $\cC_D$ part of the boundary $\partial \Omega$ temperature $g$ is prescribed, while on the $\cC_N$ part of $\partial \Omega$ is prescribed heat flux $q$.
We look for a temperature distribution $T(x)$ in $\Omega$, which satisfies the linear heat conduction equation $$\label{ht_1}
\Div(\cA(x)\nabla T)(x)=f(x), \qquad x\in \Omega$$ and boundary conditions $$\begin{aligned}
\label{ht_2}
& T^+(y)=g(y)\qquad {\rm on} \;\; \cC_D,\\
\label{ht_3}
&-\langle\nub(y),\cA^+(y)(\nabla T)^+(y)\rangle=q(y)\qquad{\rm on}\;\; \cC_N,\end{aligned}$$ where $\cA$ is the thermal conductivity, $f$ is the heat source, $g$ is the distribution of temperature and $q$ is the heat flux, which are supposed known.
We will suppose, that $\cA(x)$ is a continuous $3\times3$ matrix-function, positive definite in the following sense (see ), which implies the inequality $$\begin{aligned}
\label{ht_02}
\scal\cA\U,\U\scar\geqslant C\|\U|\bL_2(\Omega)\|^2
\end{aligned}$$ valid for all 3-vectors $\U=(U_1,U_2,U_3)^\top\in\bL_2(\Omega)$. The conditions on $\cA$ imply that the traces $\cA^+(y)$ at the boundary $\cC$ exist and $\cA^+$ has the same properties, namely, is a continuous positive definite matrix function.
We impose the following natural constraints on the solution $T$ and functions $f$, $g$ and $q$, which are prescribed: $$\begin{aligned}
\label{ht_0}
T\in\bH^1(\Omega),\quad f\in\wt\bH^{-1}(\Omega), \quad
g\in\bH^{1/2}(\cC_D),\quad q\in\bH^{-1/2}(\cC_N).
\end{aligned}$$ The existence of the traces $\langle\nub(y),\cA^+(y)(\nabla T)^+\rangle \in\bH^{-1/2} (\cC_3)$, which is not ensured by the trace theorem, follows from the Green formula $$\begin{aligned}
\label{ht_03}
\int_\Omega(\Div\,\cA(x)\nabla T)(x)\psi(x)dx=\int_\cC\langle\nub(y),\cA^+(y)
(\nabla T)^+(y)\rangle\psi^+(y)\,d\sigma\nonumber\\
-\int_\Omega\langle \cA(x)\nabla T(x),\nabla\psi(x)\rangle\,dx
\end{aligned}$$ by the duality between the spaces $\bH^{1/2}(\cC)$ and $\bH^{-1/2}(\cC)$ due to the fact that $T$ is a solution to the equation . For this we rewrite in the form $$\int_\cC\langle\nub(y),\cA^+(y)(\nabla T)^+(y)\rangle\psi^+(y)\,d\sigma
=\int_\Omega f(x)\psi(x)dx+
\int_\Omega\langle \cA(x)\nabla T(x),\nabla\psi(x)\rangle\,dx,$$ and note that $\psi\in\bH^1(\Omega)$ is arbitrary and, therefore, $\psi^+\in\bH^{1/2}(\cC)$ is arbitrary.
Let $\Omega\subset\bR^n$ be a domain with a Lipshitz boundary $\cM:=\partial\Omega$ and $\cM_0\subset\partial\Omega$-be a subsurface of the boundary surface which has the non-zero measure. By $\wt\bH^1(\Omega,\cM_0)$ we denote a subspace of $\wt\bH^1(\Omega)$ of those functions which have vanishing traces on the part of the boundary $$\begin{aligned}
\label{ht_06}
\wt\bH^1(\Omega,\cM_0):=\left\{\vf\in\bH^1(\Omega)\;:\;
\vf^+(y)=0\quad\forall\, y\in\cM_0\right\}.
\end{aligned}$$ This space inherits the standard norm from $\bH^1(\Omega)$: $$\begin{aligned}
\|\vf\,\big|\,\bH^1(\Omega)\,\|:&=\left[\|\varphi\,\big|\,\bL(\Omega)\,\|^2
+\sum\limits_{j=1}^n\|\pa_j\vf\,\big|\,\bL_2(\Omega)\|^2\right]^{1/2}.\end{aligned}$$
Consider the functional $$\begin{aligned}
\label{ht_5}
\Phi (T)=\int\limits_{\Omega}\left[\frac12\langle\cA(x)\nabla T(x),\nabla T(x)\rangle
+f(x)T(x)\right]dx+\int\limits_{\cC_N}q(y) T^+(y)d\sigma
\end{aligned}$$ where $f$ and $q$ satisfy conditions and $T\in\bH^1(\Omega)$ has vanishing traces on $\cC_D$, i.e., $T\in\wt\bH^1(\Omega,\cC_D)$ (see ).
The second summand in the in integral on $\Omega$ is understood in the sense of duality between the spaces $\wt\bH^{-1}(\Omega)$ and $\bH^1(\Omega)$. Concerning the integral on $\cC_N$: it is understood in the sense of duality between the spaces $\wt\bH^{1/2}(\cC_N)$ and $\bH^{-1/2}(\cC_N)$ because $q\in\bH^{-1/2}(\cC_N)$ and, due to the condition inclusion $T\in\wt\bH^1(\Omega,\cC_D)$, $\supp\,T^+\subset\cC_N$ which implies $T^+\in\wt\bH^{1/2}(\cC_N)$.
\[t6.1\] The problem - with vanishing Dirichlet condition $T^+(y)=g(y)=0$ for all $y\in\cC_D$ is reformulated into the following equivalent variational problem: Let $f$ and $q$ satisfy conditions and look for a temperature distribution $T\in\wt\bH^1(\Omega,\cC_D)$ (see ) which is a stationary point of the functional .
[**Proof:**]{} Let $T(x)$ be a stationary point of the functional , where $\Phi(T)$ attains a local infimum. Consider the variation $$\begin{aligned}
\label{e6.8}
\delta\Phi=\frac{d}{d\varepsilon}\Phi(T+\varepsilon\V)|_{\varepsilon=0}
=\int\limits_{\Omega}\big[\langle \cA(x)\nabla T(x),\nabla\V(x)\rangle+f(x)\V(x)\big] dx\nonumber\\
+\int\limits_{\cC_N}q(y)\V^+(y).
\end{aligned}$$ The trial function $\V\in\bH^1(\Omega)$ is such that $T+\varepsilon\V$ satisfies the boundary conditions. Then from the equalities $T^+(y)+\V^+(y)=0=T^+(y)$ on $\cC_D$ follows that $T^+(y)=\V^+(y)=0$ on $\cC_D$, i.e., $T$ and $\V$ have the traces zero on the part $\cC_D$ of the boundary.
It is clear, that for those $\V$ for which the functional $\Phi(T+\ve\V)$ attains infimum, we have $\delta\Phi=0$. By applying the Gauß theorem to the first summand under the integral on $\Omega$ in , we obtain the associated Euler-Lagrange equation $$\begin{aligned}
\label{ht_7}
\int\limits_{\Omega}\big[-\Div \cA(x)\nabla T(x) + f(x)\big]
\V(x)\,dx + \int\limits_{\cC_D}\langle \nub(y),\cA^+(y)(\nabla T)^+(y)\rangle\V^+(y) d\sigma\nonumber\\
+\int\limits_{\cC_N}\Big[q(y)+\langle\nub(y),\cA^+(y)(\nabla T)^+(y)
\rangle\Big]\V^+(y)d\sigma=0.
\end{aligned}$$
Since the trial function $\V$ vanishes on $\cC_D$ (see ), the integral on $\cC_D$ in vanishes. Now taking arbitrary function $\V\in C_0^{\infty}(\Omega)$ (vanishing in the vicinity of the boundary $\cC$), all summands in except the first one vanish and we obtain $$\label{ht_8}
\int\limits_\Omega \big[-\Div \cA(x)\nabla T(x) + f(x)\big]\V(x)\,dx=0,$$ which is equivalent to the basic differential equation in .
Therefore from follows that $$\begin{aligned}
\label{ht_9}
&&\int\limits_{\cC_N}\Big[q(y)+\langle\nub(y),\cA^+(y)(\nabla T)^+(y)\rangle
\Big]\V^+(y)\,d\sigma=0.\end{aligned}$$ The trace $\V^+$ of a trial function in is arbitrary, we derive, the boundary condition .
Vice versa: Let $T$ be a solution to the mixed problem - with vanishing Dirichlet traces $T^+(y)=g(y)=0$ on $\cC$, by taking the scalar product of the basic equation in with the solution $T$, by applying the Green formulae and the boundary conditions with $g=0$, we get the following equality: $$\begin{aligned}
0&\hskip-3mm=&\hskip-3mm\int\limits_{\Omega}\big[-\Div \cA(x)\nabla T(x)
+ f(x)\big] T(x)\,dx=\int\limits_{\Omega}\big[\cA(x)\nabla T(x) + f(x)\big]\nabla T(x)\,dx\nonumber\\
&&+\int\limits_{\cC_D\cup\cC_N}\langle \nub(y),\cA^+(y)(\nabla T)^+(y)\rangle T^+(y)
d\sigma\nonumber\\
&\hskip-3mm=&\hskip-3mm\int\limits_{\Omega}\big[\cA(x)\nabla T(x) + f(x)\big]\nabla
T(x)\,dx\int\limits_{\cC_N}q(y)T^+(y)d\sigma.
\end{aligned}$$ Therefore, $T$ is a stationary point of the functional $\Phi$ in .
\[c.cor\] The minimization problem for the functional is an equivalent reformulation of the BVP .
If $\cC_D=\cC, \; \cC_N=\emptyset$, the problem - reduces to the problem with a Dirichlet boundary condition $$\label{ht_10}
T^+(y)=0 \qquad {\rm on}\quad \cC$$ and the corresponding functional $\Phi$ in variational formulation (see ) takes the form $$\label{ht_11}
\Phi_D(T)=\frac12\int\limits_{\Omega}\Big[\langle \cA(x)\nabla T(x),\nabla T(x)\rangle+ f(x)T(x)\Big]\,dx.$$ If $\cC_D=\emptyset, \; \cC_N=\cC$, from - we get the problem with Neumann boundary condition $$\label{ht_12}
-\langle \cA^+(y)\nub(y),(\nabla T)^+(y)=q(y) \qquad {\rm on}\quad \cC$$ and the corresponding functional in variational formulation (see ) takes the form $$\label{ht_13}
\Phi_N(T)=\frac12\int\limits_{\Omega}\Big[\langle \cA(x)\nabla T(x),\nabla T(x)
\rangle + f(x)T(x)\Big]\,dx + \int\limits_\cC q(y)T^+(y)d\sigma.$$
Heat transfer in thin Layers {#sec7}
============================
Let $\cC$ be a $C^2$ smooth orientable surface in $\mathbb{R}^3$ given by a single chart (immersion) $$\theta\;:\;\omega \rightarrow \cC, \qquad \omega \subset \mathbb{R}^2$$ and let $\nu(\cx),\; \cx \in \cC$ be the unit normal on $\cC$ with the chosen orientation. Chart is supposed to be single just for convenience and multi-chart case can be considered similarly. Denote by $\Omega^\varepsilon $ the layer domain i.e. the set of all points in $\mathbb{R}^3$ in the distance less then $\varepsilon$ from $\cC$. Then for sufficiently small $\varepsilon$ the map $\Theta\;:\; \cC \times (-\varepsilon, \varepsilon)\rightarrow \Omega^\varepsilon $ $$\label{ht_16}
\Theta(\cx,t)=\cx+t\nu(\cx)=\theta(x)+t\nu(\theta(x)), \qquad x\in \omega$$ is $C^1$ homeomorphism and $\Theta(\cC \times \{0\})=\cC$.
As noted above we can properly extend normal field on the entire $\Omega^\varepsilon $ assuming $$\label{ht_16_1}
\nub(\cx+t\nub(\cx))=\nub(\cx),\qquad \cx\in\cC, \qquad -\varepsilon<t<\varepsilon.$$ If $\varepsilon$ is sufficiently small, the boundary $\cM^\varepsilon:=\partial \Omega^\varepsilon $ is represented as the union of three $C^1$-smooth surfaces $\cM^\ve=\cM_{\ve,D}\cup\cM_{\ve,N}^{-}\cup\cM_{\ve,N}^{+}$, where $\cM_{\ve,D}=\partial \cC\times[-\varepsilon,\varepsilon]$ is the lateral surface, $\cM^+_{\ve,N}=\cC\times \{+\varepsilon\}$ is the upper surface and $\cM^-_{\ve,N}=\cC\times\{-\varepsilon\}$ is the lower surface of the of the boundary $\cM^\ve$ of layer domain $\Omega^\ve$.
In the present section we will consider heat conduction by an “isotropic” media, governed by the Laplace equations (the case $\cA(x)\equiv1$ in -). The case of an “anisotropic” media will be treated in a forthcoming publication in a thin layer domain $\Omega^\ve:=\cC\times(-\ve,\ve)=\Theta^{-1}(\Omega^\ve)$: $$\begin{aligned}
\label{ht_17}
\begin{array}{ll}
\Delta_{\Omega^\ve} T(\cx,t)=f(\cx,t), \qquad &(\cx,t)\in
\cC\times(-\ve,\ve),\\[3mm]
T^+(\cx,t)= 0, & (\cx,t)\in \pa\cC\times(-\ve,\ve),\\[3mm]
\pm(\pa_t T)^+(\cx,\pm\ve)=q^\pm(\cx), & \cx\in\cC,
\end{array}\end{aligned}$$ where (see , ) $$\Delta_{\Omega^\ve}T=\sum\limits_{j=1}^{4}\cD_j^2 T=\Delta_{\cC}T+\pa_t^2 T.$$ The different signs $\pm(\pa_t T)^+(\cx,\pm\ve)$ in the third equality in is due to the different orientation of the outer unit normal vector $\nub(\cx)$ at the upper and lower surfaces $\cC\times\{\pm\ve\}$.
We impose the following constraints $$\begin{aligned}
\label{ht_171}
\begin{array}{c}
T\in\bH^{1}(\Omega^{\varepsilon}),\qquad f\in\bL_2(\Omega^1),\\
0 \;\; \text{is the Lebesgue point for the function}\;\; F(t):=\dst\int_\cC |f(\cx,t)|^2d\sigma
\end{array}
\end{aligned}$$ (see and note that $\|F\big|\bL_1(-1,1)\|\leqslant\|f\big| \bL_2(\Omega^1)\|^2$). The latter constraint implies that $F(0)$ exists and, due to Theorem \[t7.7\], $$\begin{aligned}
\label{e7.7}
\frac1\ve \int_{-\ve}^\ve F(t)dt=\frac1\ve\int_{-\ve}^\ve\int_\cC|f(\cx,t)|^2d\sigma dt
\leqslant 2F(0)<\infty
\end{aligned}$$ for all $0<\ve<\ve_0$ and some small $\ve_0>0$ (cf. the definition of a Lebesgue point ).
Conditions are slightly more restrictive on $f$ than is necessary for the solvability (see ) and is needed for the $\Gamma$-convergence.
The next example demonstrates that not all functions in $\bL_2(\Omega^1)$ have the property . Let $$f\left( \cx,t\right)=\begin{cases}
\sqrt{\left( -\dst\frac1{\ln t}\right)^\prime}=\dst\frac1{t^{1/2}\log t},\qquad
&\text{for}\quad t\in\left(0,\dst\frac12\right),\\
0, &\text{for}\quad t\notin \left( 0,\dst\frac12\right).
\end{cases}$$
It is easy to show that$$\begin{aligned}
\left\Vert f\left( \cx,t\right) \big|\,\bL_{2}(\Omega ^{1})\right\Vert
&=&\int_{-1}^{1}\int_{\cC}|f(\cx,t)|^{2}d\sigma dt=\int_{\cC}d\sigma
\int_{0}^{1/2}\frac{dt}{t\ln^2t} \\
&=&-C\int_{0}^{1/2}\left(\frac1{\ln t}\right)^\prime dt=-\left.\frac C{\ln t}\right\vert_0^{1/2}=\frac C{\ln 2}<\infty.\end{aligned}$$
On the other hand, if $F(t)$ is defined in , $$\begin{aligned}
\frac{1}{\ve}\int_{-\ve}^{\ve}F(t)dt &=&\frac{1}{\ve}\int_{-\ve}^{\ve}\int_{\cC}
|f(\cx,t)|^{2}d\sigma dt=\int_{\cC}d\sigma \frac{1}{\ve}\int_{0}^{\ve}
\frac{dt}{t\ln^2t} \\
&=&-\frac{C}{\ve}\int_{0}^{\ve}\left( \frac1{\ln t}\right)^\prime dt=-\left.\frac C
{\ve\ln t}\right\vert_0^\ve=-\frac C{\ve\ln\ve}
\rightarrow\infty,\quad \text{as}\quad\ve\rightarrow 0\end{aligned}$$ and $0$ is not the Lebesgue point for the function $F(t)$.
\[r7.5\] Note, taking the Dirichlet and the Neumann traces zero $T^+(\cx,t)=0$ on $\pa\cC\times(-\ve,\ve)$ and on $\cC\times\{\pm\ve\}$, (see ) we need to prove the $\Gamma$-convergence (see the Remark \[r0.2\] above;
On the other hand, a BVP $$\begin{aligned}
\label{ht_17_0}
\Delta_{\Omega^\ve}T_0(\cx,t)&\hskip-3mm=&\hskip-3mmf(\cx,t), \hskip18mm
(\cx,t)\in\cC\times(-\ve,\ve),\\
\label{ht_18_0}
T_0^+(\cx,t)&\hskip-3mm=&\hskip-3mm g(\cx,t), \hskip18mm (\cx,t)\in \pa\cC\times(-\ve,\ve),\\
\label{ht_19_0}
\pm(\pa_t T_0)^+(\cx,\pm\ve)&\hskip-3mm=&\hskip-3mm q^\pm(\cx), \hskip25mm \cx \in \cC\times\{\pm\ve\}\end{aligned}$$ with the non-zero Dirichlet and Neumann traces on the boundary and the standard constraints reduces to the equivalent BVP .
Indeed, let $G\in\bH^1(\Omega^\ve)$ be a solution to the Mixed boundary value problem $$\label{cc1}
\begin{array}{ll}
\Delta_{\Omega^\ve}G(x)=0, &x\in\Omega^\ve,\\[2mm]
G^+(\cx)=g(\cx), &(\cx,t)\in \pa\cC\times(-\ve,\ve),\\[2mm]
\pm(\pa_t G)^\pm(\cx,\ve)=q^\pm(\cx), \qquad &\cx \in \cC\times\{\pm\ve\}.
\end{array}$$ The unique solvability of the problem is a classical result and follows, for example, from the Lax-Milgram Lemma.
Then the difference $T:=T_0-G$ solves the BVP .
The formulated BVP governs a heat transfer in the body $\Omega^\varepsilon$ when there are thermal sources or sinks in $\Omega^\varepsilon $. The temperature on the lateral surface $\pa\cC\times(-\ve,\ve)$ is zero and heat fluxes are equal and fixed on the upper and lover $\cC^\pm:=\cC\times\{\pm\ve)$ surfaces. It is well known, that the boundary value problem as well as it’s equivalent problem - have the unique solution $T\in \mathbb{H}^1(\Omega^\varepsilon)$ (respectively, $T_0\in \mathbb{H}^1(\Omega^\varepsilon)$; see, e.g., [@DTT1]).
The energy functional associated with the problem - reads (cf. Theorem \[t6.1\]) $$\begin{aligned}
\label{ht_5x}
E(T)&\hskip-3mm=&\hskip-3mm\int\limits_{\Omega^\ve}\Big[\frac12\langle(\cD_{\Omega^\ve}T)(x),
(\cD_{\Omega^\ve}T)(x)\rangle+f(x)T(x)\Big]dx.
\end{aligned}$$ To justify the equality , we remind that expressing the Cartesian derivatives by means of Günter’s derivatives, according to we get $$\begin{aligned}
\langle\nabla T,\nabla T\rangle=\sum\limits_{j=1}^3|\partial_j T|^2
=\sum\limits_{j=1}^4|\cD_j T|^2=\sum\limits_{j=1}^3 |\cD_j T|^2+|\partial_\nub T|^2=\langle\cD_{\Omega^\ve}T,\cD_{\Omega^\ve}T\rangle.\end{aligned}$$ More generally, we consider the non-linear functional $$\label{ht_21}
E(T)=\int\limits_{\Omega^\varepsilon}\cK((\cD_{\Omega^\ve}T)(x),\,T(x))\,dx,$$ in the case of the functional we have $$\begin{aligned}
\label{ht_22}
&&\hskip-22mm\cK(\cD_{\Omega^\ve}T,\,T)=
\frac12\langle\cD_{\Omega^\ve}T(\cx,t),\cD_{\Omega^\ve}T(\cx,t)\rangle+f(\cx,t)\,T(\cx,t).\end{aligned}$$
\[l7.1\] Let $\Omega$ be a domain in $\mathbb{R}^n$ with the Lipshitz boundary $\cM:=\partial\Omega$ and $\cM_0\subset\cM$ be a subsurface of non-zero measure. Then the inequality $$\label{Poinc}
\|\varphi\,\big|\,\bL_2(\Omega)\|\leqslant C\|\nabla\varphi\,\big|\,\bL_2(\Omega)\|
=C\left[\sum\limits_{j=1}^n\|\pa_j\varphi\,\big|\,\bL_2(\Omega)\|^2\right]^{1/2}$$ holds for all functions $\varphi\in \wt{\bH}{}^1(\Omega,\cM_0)$ and the constant $C$ is independent of $\varphi$.
Moreover, Let $\cC\subset\mathbb{R}^n$ be a smooth hypersurface with the Lipschitz boundary $\Gamma=\partial\cC$, $\Omega=\cC\times[a,b]$ is a cylinder with the base $\cC$ and $\cM_0:=\Gamma_0\times[a,b]$, $\Gamma_0\subset\Gamma$. Then for all functions $\varphi\in \wt{\bH}{}^1(\Omega,\cM_0)$ the inequality $$\label{Poinc_2}
\|\varphi\,\big|\,\bL_2(\Omega)\|\leqslant C'\|\nabla_\cC\varphi\,\big|
\,\bL_2(\Omega)\|=C'\left[\sum\limits_{j=1}^n\|\cD_j\varphi\,\big|\, \bL_2(\Omega)\|^2\right]^{1/2}$$ holds with only surface gradient $\nabla_\cC:=(\cD_1,\ldots,\cD_n)^\top$ and the constant $C'$ is independent of $\varphi$.
[**Proof:**]{} The formula $$\begin{aligned}
\label{ht_7.124}
\|\varphi\,\big|\,\wt{\bH}^1(\Omega,\cM_0)\,\|
:=\|\nabla\varphi\,\big|\,\bL_2(\Omega)\|\end{aligned}$$ defines an equivalent norm in the space $\wt{\bH}{}^1(\Omega,\cM_0)$. Indeed, the inequality $\|\vf\,\big|\,\bH^1(\Omega,\cM_0)\,\|$ $\leqslant\|\vf\, \big|\,\bH^1(\Omega)\|$ with the standard norm $\|\vf\,\big|\,\bH^1(\Omega)\|$ on $\wt{\bH}{}^1(\Omega,\cM_0)$ is trivial. On the other hand, $\|\varphi\,\big|\,\wt{\bH}^1(\Omega,\cM_0)\,\|$ has all properties of a norm. Since other properties are trivial to check, we will only check that $\|\varphi\,\big|\,\wt{\bH}^1(\Omega,\cM_0)\,\|=\|\nabla\varphi\,\big|\,\bL_2(\Omega)\,\|=0$ implies $\varphi=0$. Indeed, the trivial norm implies that the gradient vanishes $\nabla\varphi=0$, which means that the corresponding function is constant $\varphi={\rm const}$; since $\varphi=0$ on $\cM_0$, it follows $\varphi\equiv0$.
If we apply the open mapping theorem of Banach (see [@Ru73 Theorem 2.11, Corollary 2.12.b], we conclude that the inverse inequality $$\|\vf\,\big|\,\bH^1(\Omega)\|\leqslant C_1\|\vf\,\big|\,\bH^1(\Omega,\cM_0)\,\|
=C_1\|\nabla\varphi\,\big|\,\bL_2(\Omega)\,\|$$ holds with some constant $C_1<\infty$. Since $$\|\vf\,\big|\,\bH^1(\Omega)\|^2=\|\varphi\,\big|\,\bL_2(\Omega)\,\|^2
+\|\nabla\varphi\,\big|\,\bL_2(\Omega)\,\|^2\leqslant C^2_1\|\nabla\varphi\,\big|\,\bL_2(\Omega)\,\|^2,$$ The claimed inequality follows with the constant $C:=\sqrt{C_1^2-1}$.
Now if $\cC$ is a hypersurface, $\Omega=\cC\times[a,b]$ is a cylinder with the base $\cC$ and $\cM_0:=\Gamma_0\times[a,b]$, $\Gamma_0\subset\Gamma :=\partial\cC$, in the space $\wt{\bH}{}^1(\Omega,\cM_0)$ we consider the semi-norm $$\label{Poinc_3}
\|\varphi\,\big|\,\bH^1(\Omega,\cM_0)\|:=\|\nabla_\cC\varphi\,\big|\,\bL_2(\Omega)\|
=\left[\sum\limits_{j=1}^n\|\cD_j\varphi\,\big|\,\bL_2(\Omega)\|^2
\right]^{1/2},$$ which turns out to be a norm. Indeed, from $\|\varphi\,\big|\,\bH^1(\Omega,\cM_0)\|=\|\nabla_\cC\varphi\,\big|\,\bL_2
(\Omega)\|=0$ follows that $\varphi(\cx,t)=\varphi(t)$ is independent of the variable $\cx\in\cC$. Since $\varphi(t)=\varphi(\cx,t)=0$ for all $\cx\in\Gamma_0$ and all $t\in(a,b)$, it follows $\varphi\equiv0$. The proof is accomplished as in the foregoing case.
\[l7.2\] If $f\in\mathbb{L}_2(\Omega^1)$, $q^\pm_0\in\bH^{-1/2}(\cC)$, then the energy functional $E(T)$ in - is correctly defined on the space $\wt{\bH}^1(\Omega^{\ve},\pa\cC\times(-\ve,\ve))$, is strictly convex and has the following quadratic estimate $$\begin{aligned}
\label{ht_121}
\begin{array}{l}
E(t T_1+(1-t)T_2)\leqslant t E(T_1)+(1-t)E(T_2),\\[3mm]
C_1\dst\int\limits_{\Omega^\ve}|(\cD_{\Omega^\ve}T)(x)|^2dx-C_2\leqslant E(T)\leqslant C_3\left[1 + \dst\int\limits_{\Omega^\ve}|(\cD_{\Omega^\ve}T)(x)|^2dx\right],\\
\hskip20mm\forall\,T_1,\ T_2\in\mathbb{H}^1(\Omega^\ve),\qquad \forall\,T \in\wt\bH^1(\Omega^{\ve},\pa\cC\times(-\ve,\ve))
\end{array}
\end{aligned}$$ for some positive constants $C_1,C_2$ and $C_3$.
[*Proof:*]{} Let us decompose the functional $E(T)$ in into the sum of bilinear and linear parts $$\begin{aligned}
\label{e7.13b}
\begin{array}{c}
E(T)=E^{(1)}(T)+E^{(2)}(T)\\[3mm]
E^{(1)}(T):=\dst\frac12\dst\int_{\Omega^\ve}\langle(\cD_{\Omega^\ve}T)(x),
(\cD_{\Omega^\ve}T)(x)\rangle dx,\\[3mm]
E^{(2)}(T):=\dst\int_{\Omega^\ve}f(x)T(x)dx.
\end{array}\end{aligned}$$
The quadratic function $F(x)=x^2$ is strictly convex $[tx_1+(1-t)x_2]^2<tx^2_1+(1-t)x^2_2$ for all $x_1,x_2\in\bR$, $x_1 \neq x_2$, $0<t<1$ and, therefore, the functional $E^{(1)}(T)$ is strictly convex. Since $E^{(2)}(T)$ is linear, the sum $E(T)=E^{(1)}(T)+E^{(2)}(T))$ is, obviously, strictly convex (see the first inequality in ).
Next let us prove the second two-sided estimate in . To this end note, that the first functional $$\label{e7.22}
E^{(1)}(T)=\dst\frac12\dst\int_{\Omega^\ve}|(\cD_{\Omega^\ve} T)(x)|^2dx$$ is quadratic itself. We will prove the following estimate for the second functional $$\label{ht_7.13c}
|E^{(2)}(T)|\leqslant M +M \left(\dst\int_{\Omega^\ve}\left|(\cD_{\Omega^\ve}T)(x)\right|^2
dx\right)^{1/2}.$$ Since $f\in\mathbb{L}_2(\Omega^1)$ and $T\in\wt\bH^1(\Omega^{\ve},\pa\cC\times(-\ve,\ve)) \subset\mathbb{L}_2(\Omega^\ve)$, due to Lemma \[l7.1\] we can write $$\int_{\Omega^\ve}f(x)T(x)dx\leqslant \|f|\mathbb{L}_2(\Omega^\ve)\|\
\|T|\mathbb{L}_2(\Omega^\ve)\|\leqslant M
\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|\leqslant M\Big(1+
\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|^2\Big).$$ The proved inequalities justify the estimate .
The right inequality in the second line of is a direct consequence of , and .
Let us prove the left inequality in the second line of . We have $$\begin{aligned}
|E^{(2)}(T)|\leqslant \|f|\mathbb{L}_2(\Omega^\ve)\|\|T|\mathbb{L}_2(\Omega^\ve)\|
\leqslant\|f|\mathbb{L}_2(\Omega^\ve)\|\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|\\[2mm]
\leqslant\frac1{2\eta}\|f|\mathbb{L}_2(\Omega^\ve)\|^2
+\frac\eta2\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|^2\leqslant\frac1{2\eta}\|f|\mathbb{L}_2(\Omega^1)\|^2
+\frac\eta2\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|^2\end{aligned}$$ for any $\eta>0$. Choosing $\eta<1$ and by taking $C_1=\displaystyle\frac{1-\eta}2$, $$C_2\geqslant\frac1{2\eta}\|f|\mathbb{L}_2(\Omega^1)\|^2$$ we get $$E(T)\geqslant E^{(1)}(T)-|E^{(2)}(T)|\geqslant C_1\|\nabla T|\mathbb{L}_2(\Omega^\ve)\|^2 - C_2.$$ -9mm
Now we perform the scaling of the variable $t=\varepsilon \tau$, $-1<\tau<1$ and study the functionals in the fixed domain $\Omega^1=\cC \times (-1,1)$ $$\begin{aligned}
\label{ht_23}
E_{\varepsilon}\left(T_{\ve}\right)&=&\int_{\Omega^1}\cK\left( \cD_{1} T_{\ve},\,\cD_{2}T_{\ve},\,\cD_{3}T_{\ve},\,\frac{1}{\varepsilon}
\cD_{4}T_{\ve},T_{\ve}\right) dx \nonumber\\
&=& \int_{-1}^{1}\int_{\cC}\cK\left(\cD_{\cC} T_{\ve},\,\frac{1}{\varepsilon}\pa_t T_{\ve},T_{\ve}\right)d\sigma dt,\end{aligned}$$ where $\cD_{\cC}=( \cD_{1},\,\cD_{2},\,\cD_{3}),\;\cD_{4}=\pa_{t}$. The functionals $E_{\varepsilon}\left(T_{\ve}\right)$ are related to the original functional $E\left(T\right)$ by the equality $$\label{ht_24}
E_{\varepsilon }\left( T_{\ve}\right) =\frac{1}{\varepsilon }E
\left(T\right),\text{ \ \ \ where \ \ \ }T_{\ve}(x,t)
=T\left(\cx_1,\cx_2,\cx_3,\varepsilon t\right).$$ Assume, that $T_j\in\mathbb{H}^1\left(\Omega^1\right), \; j\in\mathbb{N},$ are the scaled solutions to the problem , with $\ve=\ve_j, \; f_j(\cx,t)=f(\cx,\ve_j t),\; 0<\ve_j\leq 1, \; \lim\limits_{j \to \infty} \ve_j =0$; then from the Euler-Lagrange equation, associated with the functional (see ), follows that $$\begin{aligned}
\label{ht_25}
E_{\ve_j}(T_j):&\hskip-3mm=&\hskip-3mm\int_{\Omega^1}\cK\left( \cD_{\cC} T_j,\,\frac{1}{\varepsilon_j}\pa_tT_j,T_j\right) dx \\
&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T_j)(\cx,t)\right|^2+\frac1{\varepsilon^2_j}\left|(\pa_tT_j)(\cx,t)
\right|^2\right]+f_j(\cx,t)T_j(\cx,t)\Big]d\sigma dt=0. \nonumber
\end{aligned}$$ From , , and Lemma \[l7.2\] follows $$\begin{aligned}
\label{ht_26}
C_0\|T_j|\mathbb{H}^1(\Omega^1)\|^2&\hskip-3mm\leqslant&\hskip-3mm
\|\nabla_{\Omega^1} T_j|\mathbb{L}_2(\Omega^1)\|^2=\int_{\Omega^1}
\left[\frac{1}{2}\left\langle\cD_{\cC} T_j,\cD_{\cC} T_j\right\rangle+
\frac1{2\ve^2_j}|\pa_t T_j|^2\right]dx\nonumber\\
&\hskip-3mm=&\hskip-3mm\left|\int\limits_{-1}^1\int\limits_{\cC}f_j(\cx,t)T_j(\cx,t)
d\sigma\,dt\right|\leqslant\|f_j|\mathbb{L}_2(\Omega^1)\|\|T_j|\mathbb{L}_2(\Omega^1)\|
\nonumber\\[2mm]
&\hskip-3mm\leqslant&\hskip-3mm\|f_j|\mathbb{L}_2(\Omega^1)\|\|T_j|\mathbb{H}^1(\Omega^1)\|\end{aligned}$$ and, consequently, $$\begin{aligned}
\label{e7.25}
\left(\int_{\Omega^1}\left(\frac12\left\langle
\cD_{\cC} T_j,\cD_{\cC} T_j\right\rangle+\frac1{2\ve^2_j}|\pa_t T_j|^2\right)
dx\right)^{1/2}\leqslant\|T_j|\mathbb{H}^1(\Omega^1)\|\nonumber\\
\leqslant\frac1{C_0}\|f_j|\mathbb{L}_2(\Omega^1)\|\leqslant\frac2{C_0}
\|f|\mathbb{L}_2(\Omega^1)\|,\end{aligned}$$ for all $\ve_j<\ve_0$ and some $\ve_0>0$, because (cf. ) $$\begin{aligned}
\|f_j|\mathbb{L}_2(\Omega^1)\|=\left[\int_{\cC}\int_{-1}^1|f(\cx,\ve_jt)|^2d\sigma dt
\right]^{1/2}=\left[\frac1{\varepsilon_j}\int_{-\ve_j}^{\ve_j}\int_{\cC}|f(\cx,\tau)|^2
d\sigma d\tau\right]^{1/2}\\
\leqslant 2\|f(\cdot,0)|\mathbb{L}_2(\cC)\|,\qquad \ve_j<\ve_0.\end{aligned}$$
From follows $$\label{ht_26.1}
\sup_{j}\int_{\Omega^1 }\left\vert T_{j}\right\vert
^{2}dx<\infty, \quad \sup_{j}\int_{\Omega^1 }\left\vert \cD_{\cC} T_{j}\right\vert
^{2}dx<\infty ,\quad \sup_{j}\text{\ }\frac{1}{
\varepsilon _{j}^{2}}\int_{\Omega^1 }\left\vert \pa_t T_{j}\right\vert ^{2}dx < \infty.$$
Due to and Lemma \[l7.1\] the sequence $\left\{T_j\right\}_{j=1}^\infty$ is uniformly bounded in $\mathbb{H}^1\left(\Omega^1\right)$ and a weakly converging subsequence (say $\left\{T_j\right\}_{j=1}^\infty$ itself) to a function $T$ in $\mathbb{H}^1\left(\Omega^1\right)$ can be extracted.
The functional $$E_3(T)=\int\limits_{\Omega^1}\left\vert \pa_t T\right\vert ^{2}dx$$ is convex and continuous in $\mathbb{H}^1\left(\Omega^1 \right)$; then it is weakly lower semi-continuous and $\pa_t T=0$ a.e., because $$\label{e7.28}
\int_{\Omega^1 }\left\vert \pa_t T\right\vert ^{2}dx =E_3(T)\leq\lim_j \inf E_3(T_j)
=\lim_j\inf\int_{\Omega^1}\left|\pa_tT_{j}\right|^2dx=0$$ (see the last inequality in ). Hence $T(\cx,t)$ is independent of $t$, i.e. $$\label{ht_26a}
T(\cx,t)=T(\cx), \qquad \cx \in \cC, \quad -1\leq t \leq 1.$$ Let $f_j(\cx,t):=f(\cx,\ve_jt)\to f(\cx,0)$ in $\mathbb{L}^2(\Omega^1)$. Set $$\begin{aligned}
\label{ht_26b}
\begin{array}{c}
E^{(0)}(T)=E^{(1)}(T)+E^{(2)}(T)\\[3mm]
E^{(1)}(T):=\dst\frac12\dst\int_{\Omega^\ve}\langle(\cD_{\Omega^\ve}T)(\cx),
(\cD_{\Omega^\ve}T)(\cx)\rangle d\cx\,dt\\
\hskip30mm=\dst\int_\cC\langle(\cD_{\Omega^\ve}T)(\cx),
(\cD_{\Omega^\ve}T)(\cx)\rangle d\cx,\\[3mm]
E^{(2)}(T):=\dst\int_{\Omega^\ve}f(\cx,0)T(\cx)d\cx\,dt
=2\dst\int_{\cC}f(\cx,0)T(\cx)d\cx.
\end{array}\end{aligned}$$
Let us check that the $E_{j}$ sequence $\Gamma$-convergs to $E^{(0)}$ in $\mathbb{H}^{1}\left( \Omega^1 \right)$. Indeed, if $T_{j}\rightharpoonup T$ in $\mathbb{H}^1\left(\Omega^1\right) $, We have $$E_j(T)=E_j^{(1)}(T)+E_j^{(2)}(T),$$ where $$E_j^{(1)}(T)=\int\limits_{\Omega^1 }\left( \frac{1}{2}\left\langle\cD_{\cC} T,\cD_{\cC} T\right\rangle +\frac{1}{\ve_j^2}|\pa_t T|^2 \right)dx,\quad E_j^{(2)}(T)=\int\limits_{\Omega^1 } f_jT dx.$$ The functional $E^{(1)}(T)$ is convex and continuous and so it is weakly lower semicontinuous in $\mathbb{H}^{1}\left( \Omega^1 \right)$, therefore $$\lim \inf_j E_j^{(1)}(T_j)\geqslant \lim \inf_j E^{(1)}(T_j) \geqslant E^{(1)}(T).$$ Sequence $E_j^{(2)}(T_j)$ converges to $E^{(2)}(T)$, because $f_j(\cx,t) \rightarrow f(\cx,0)$ and $T_j\rightharpoonup T$ in $\mathbb{L}^2(\Omega^1)$. Consequently $$\lim \inf_j E_j (T_j)\geqslant E^{(0)}(T).$$ This proves $\lim \inf$ inequality for the sequence $E_j$.
Note, that $$\label{ht_27}
E^{(2)}(T)=\int\limits_{\cC}\int\limits_{-1}^1 f(\cx,0) T(\cx) dt\,
d\sigma=2\int\limits_{\cC}f(\cx,0)T(\cx)d\sigma.$$ To show that the lower bound is reached i.e. to build a recovery sequence $T_j$ we fix $T\in\mathbb{H}^{1}\left( \cC \right) $ and set $T\left( \cx,t\right) =T \left(\cx\right), \; \cx\in \cC,\; t \in (-1,1) $. Define recovery sequence as $T_j(x,t)=T(x,t)=T(x)$ Then $\partial_t T_j=\partial_t T=0$ and $$\lim_{j\to\infty}E_j(T_j)=\lim_{j\to\infty}E_j^{(1)}(T)
+\lim_{j\to\infty}E_j^{(2)}(T)=E^{(1)}(T)+E^{(2)}(T)=E^{(0)}(T).$$ We have proved the following result.
\[t7.3\] If $\varepsilon \rightarrow 0$ and $f_{\varepsilon}(\cx,t):=f(\cx,\varepsilon t)\to f(\cx,0)$ in $\mathbb{H}^{-1}(\Omega^1)$, then the functional in $\Gamma$-converges to the functional $$\begin{aligned}
\label{ht_28}
E^{(0)}(T)=\int\limits_{-1}^1\int\limits_{\cC}\left[ \frac{1}{2}\left\langle\cD_{\cC}
T(\cx),\cD_{\cC} T(\cx)\right\rangle +f(\cx,0)T(\cx)\right]
d\sigma\,dt\nonumber\\
=2\int\limits_{\cC}\left[\frac{1}{2}\left\langle\cD_{\cC} T(\cx),\cD_{\cC}
T(\cx)\right\rangle + f(\cx,0)T(\cx)\right]d\sigma.\end{aligned}$$
The following Dirichlet boundary value problem for Laplace-Beltrami equation on the mid surface $\cC$ $$\begin{aligned}
\label{e7.cc}
\begin{array}{l}
\Delta_\cC T(\cx)=f(\cx,0) \quad \cx \in \cC,\\[3mm]
T^+(\cx)=0, \qquad \cx \in \pa\cC
\end{array}
\end{aligned}$$ is an equivalent reformulation of the minimization problem with the energy functional (see Theorem \[t6.1\]) and, therefore, can be considered as the $\Gamma$-limit of the initial BVP
Now we ar able to prove the main Theorem \[t0.1\] formulated in the introduction.
[**Proof of Theorem \[t0.1\]**]{}: Due to Corollary \[c.cor\] the minimization problem for the functional is an equivalent reformulation of the BVP . Let us rewrite the scaled energy functional as follows: $$\begin{aligned}
E_{\ve_j}(T_j)&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T_j)(\cx,t)\right|^2+\frac1{\ve^2_j}\left|(\pa_tT_j)(\cx,t)
\right|^2\right]+f_j(\cx,t)T_j(\cx,t)\Big]d\sigma dt\nonumber\\
&&+\frac1{\ve_j}\int\limits_{\cC}q(\cx,\ve_j)\left[T_j(\cx,\ve_j)-T_j(\cx,-\ve_j)\right]
d\sigma\nonumber\\
&&+\frac1{\ve_j}\int\limits_{\cC}T_j(\cx,-\ve_j)\left[q(\cx,\ve_j) -
q(\cx,-\ve_j)\right]d\sigma\nonumber\\
\end{aligned}$$ $$\begin{aligned}
\label{e7.43}
\hskip15mm&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T_j)(\cx,t)\right|^2+\frac1{\ve^2_j}\left|(\pa_tT_j)(\cx,t)
\right|^2\right]+f_j(\cx,t)T_j(\cx,t)\Big]d\sigma dt\nonumber\\
&&+\frac1{\ve_j}\int\limits_{\cC}\int\limits_{-\ve_j}^{\ve_j}q(\cx,\ve_j)(\pa_tT_j)(\cx,t)
d\sigma\,dt\nonumber\\
&&+\frac1{\ve_j}\int\limits_{\cC}T_j(\cx,-\ve_j)\left[q(\cx,\ve_j) -
q(\cx,-\ve_j)\right]d\sigma.
\end{aligned}$$ Since $(\pa_tT_j)(\cx,t)$ converges weakly to $\pa_tT(\cx)\equiv0$ as $j\to\infty$, $\dst\frac1{\ve_j}\pa_tT_j$ is uniformly bounded in $\bL_2(\cC,\times(-1,1))$ (see ), $q(\cdot,\ve_j)$ is uniformly bounded in $\bL_2(\cC)$, (see ), Corollary \[c7.8\] of the Lebesgue Differentiation Theorem applies and we get: $$\begin{aligned}
&&\frac1{\ve_j}\int\limits_{\cC}\int\limits_{-\ve_j}^{\ve_j}q(\cx,\ve_j)(\pa_tT_j)(\cx,t)
d\sigma\,dt\\
&&=\frac1{\ve_j}\int\limits_{\cC}\int\limits_{-\ve_j}^{\ve_j}q(\cx,\ve_j)
\left[(\pa_tT_j)(\cx,t)-(\pa_tT)(\cx)\right]d\sigma\,dt=\co(\ve_j).
\end{aligned}$$ Now we can continue as follows: $$\begin{aligned}
\label{e7.44}
E_{\ve_j}(T_j):&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}\Big[\frac12
\left[\left|(\cD_{\cC}T_j)(\cx,t)\right|^2+\frac1{\ve^2_j}\left|(\pa_tT_j)(\cx,t)
\right|^2\right]+f_j(\cx,t)T_j(\cx,t)\Big]d\sigma dt\nonumber\\
&&+2\int\limits_{\cC}T_j(\cx)q^0(\cx)d\sigma + \co(\ve_j).
\end{aligned}$$ From and we get finally $$\begin{aligned}
\lim_{\ve_j\to0}E_{\ve_j}(T_j)&\hskip-3mm=&\hskip-3mm\int\limits_{-1}^1\int\limits_{\cC}
\Big[\frac12\left|(\cD_{\cC}T)(\cx)\right|^2+\left[f(\cx,0)+q^0(\cx)\right]
T(\cx)\Big]d\sigma dt\\
&\hskip-3mm=&\hskip-3mm2\int\limits_{\cC}
\frac12\Big[\left|(\cD_{\cC}T)(\cx)\right|^2+\left[f^0(\cx)+q^0(\cx)\right]
T(\cx)\Big]d\sigma dt\end{aligned}$$ and is proved.
The concluding assertion, that the BVP is an equivalent reformulation of the minimization problem with the energy functional , is explained in Theorem \[t6.1\]).
\[r7.6\] If we take non-zero Dirichlet data in the BVP , we can not reformulate the BVP into the variational form (see Theorem \[t6.1\]), because $T^+\not\in\wt\bH^{1/2}(\cC_N)$ while $q\in\bH^{-1/2}(\cC_N)$ and the existence of the integral in the last summand of the functional $\Phi(T)$ can be ensured only for $q\in\wt\bH^{-1/2}(\cC_N)$. Moreover, in the functional $\Phi (T)$ will emerge a new summand $$\int\limits_{\cC_D}(\pa_tT)(y)g(y)d\sigma$$ and to ensure its existence we have to impose even more constraint on the data $q\in\wt\bH^{1/2}(\cC_N)$.
[0]{} H.Ammari and J.C.Nédélec, [*Generalized impedance boundary conditions for the Maxwell equations as singular perturbations problems*]{}, Commun. PDE [**24**]{} 1999, 821–849. [ L.Andersson, P.T.Chrusciel, Cauchy data for vacuum Einstein equations and obstructions to smoothness of null infinity. [*Phys. Rev. Lett.*]{} [**70**]{}, 1993, 2829–2832.]{} R.Aris, [*Vectors, Tensors, and the Basic Equations of Fluid Mechanics*]{}, Prentice-Hall, Englewood Cliffs, N.J., 1962. A.Bendali, [*Numerical analysis of the exterior boundary value problem for the time-harmonic Maxwell equations by a boundary finite element method. Part I: The continuous problem*]{}, Mathematics of Computation, Vol.43 (1984), 29–46. [ Bonner B. D., Graham I. G. and Smyshlyaev V.P., [*The computation of conical diffraction coefficients in high-frequency acoustic wave scattering*]{}, Preprint N103047, Isaac Newton Institute, Cambridge, July 2003.]{} A. Braides, $\Gamma $-convergence for beginners, Oxpord lecture series in mathematics and its applications, Oxford university press, 2007. M.Cessenat, [*Mathematical methods in electromagnetism. Linear theory and applications*]{}, Series on Advances in Mathematics for Applied Sciences, Vol.41, World Scientific Publishing Co., Inc., River Edge, NJ, 1996. P.G.Ciarlet, [*Introduction to Linear Shell Theory*]{}, Series in Applied Mathematics, Vol.1, Gauthier-Villars, Éditions Scientifiques et Médicales Elsevier, Paris, North-Holland, Amsterdam, 1998. P.G. Ciarlet, [*Mathematical Elasticity, Vol.III: Theory of Shells*]{}, Studies in Mathematics and Applications, 29, Elsevier, North-Holland, Amsterdam, 2000. M.Costabel, [*A coercive bilinear form for Maxwell’s equations*]{}, J. Math. Anal. Appl., Vol.157 (1991), 527–541. M. Costabel, E. Stephan, An improved boundary element Galerkin method for three-dimensional crack problems, [*Integral Equations and Operator Theory*]{} [**10**]{}, 1987, 467-504. R.Dautray and J.-L.Lions, [*Mathematical Analysis and Numerical Methods for Science and Technology*]{}, Springer-Verlag, Berlin, 1990. R. Duduchava, On multidimensional singular integral operators. I-II: The case of compact manifolds, [*J. Oper. Theory*]{}, [**11**]{}, 41-76, 199-214, 1984. R. Duduchava, The Green formula and layer potentials, [*Integral Equations and Operator Theory*]{} [**41**]{}, 2, 2001, 127–178. R. Duduchava, Lions’s lemma, Korn’s inequalities and Lam’e operator on hypersurfaces, pp. 1-35. Acceoted in: Proceedings of the Conference in Honour of Professor Nikolay Vasilevski at SINVESTAV, Mexico City, Mexico 2008. R. Duduchava, Partial differential equations on hypersurfaces, [*Memoirs on Differential Equations and Mathematical Physics*]{} [**48**]{}, 2009, 19-74. R. Duduchava, A revised asymptotic model of a shell. [*Memoirs on Differential Equations and Mathematical Physics*]{} [**52**]{}, 2011, 65-108. R. Duduchava, Continuation of functions from hypersurfaces. [*Complex Analysis and Differential Equations*]{} [**57**]{}, Issue 6 (2012), 625-651. R. Duduchava, D. Kapanadze, Extended Normal Vector Field and the Weingarten Map on Hypersurfaces, [*Georgian Mathematical Journal*]{} [**15**]{} (2008), No. 3, 485-500. R.Duduchava, D.Mitrea, M.Mitrea, Differential operators and boundary value problems on surfaces. [*Mathematische Nachrichten*]{} [**9-10**]{}, 2006, 996-1023. R. Duduchava, D. Natroshvili, E. Shargorodsky, Basic boundary value problems of thermoelasticity for anisotropic bodies with cuts I-II, [*Georgian Mathematical journal*]{}, [**2**]{}, 123–140, 259–276, 1995. R. Duduchava, E. Shargorodsky, G. Tephnadze, Extension of the unit normal vector field from a hypersurface and Eikonal equation. Manuscript. 9999 R. Duduchava, M. Tsaava, T. Tsutsunava, Mixed boundary value problem on hypersurfaces. Accepted in [*International Journal of Differential Equations*]{}. R. Duduchava, W. Wendland, The Wiener-Hopf method for systems of pseudodifferential equations and its application to crack problems, [*Integral Equations and Operator Theory*]{} [**23**]{}, 294-335, 1995. G. Fichera, Existence theorems in lineSar and semi-linear elasticity, [*Z. Agnew. Math. Mech*]{} [**54**]{}, Sonderheft, T24-T36 (1974). G. Friesecke, R.D. James, S. Müller, A theorem on geometric rigidity and the derivation of nonlinear plate theory from three dimensional elasticity, [*Communications on Pure and Applied Mathematics*]{} [bf 55]{}, 11, 2002, 1461–1506. N.Günter, [*Potential Theory and its Application to the Basic Problems of Mathematical Physics*]{}, Fizmatgiz, Moscow 1953 (Russian. Translation in French: Gauthier–Villars, Paris 1994). W.Haack, [*Elementare Differentialgeometrie*]{} (German), Lehrbücher und Monographien aus dem Gebiete der exakten Wissenschaften, 20, Basel-Stuttgart: Birkhäuser Verlag, VIII, 1955. L. Hörmander, [*The Analysis of Linear Partial Differential Operators*]{} [**I-IV**]{}, Springer-Verlag, Heidelberg 1983. R. Kapanadze, On some properties of singular integral operators in spaces with norms, [*Proceedings of the Tbilisi State University*]{} [ **129**]{}, 1968, 263-277. V.Kupradze, T.Gegelia, M.Basheleishvili and T.Burchuladze, [*Three-Dimensional Problems of the Mathematical Theory of Elasticity and Thermoelasticity*]{}, North-Holland, Amsterdam 1979 (Russian edition: Nauka, Moscow 1976). R. Leis, [*Initial Boundary Value Problems in Mathematical Physics*]{}, Teubner, Stuttgart 1986. J.L. Lions, E. Magenes, [*Non-homogeneous Boundary Value Problems and Applications I*]{}, Springer-Verlag, Heidelberg 1972. Massari U., Miranda M., [*Minimal Surfaces of Codimension One*]{}, North-Holland Mathematics Studies, 91. Notas de Matemtica, 95. Amsterdam - New York - Oxford: North-Holland, 1984. W. McLean, [*Strongly elliptic systems and boundary integral equations*]{}. Cambridge: Cambridge University Press [**XIV**]{}, 2000. D.Mitrea, M.Mitrea and M.Taylor, [*Layer potentials, the Hodge Laplacian, and global boundary problems in nonsmooth Riemannian manifolds*]{}, Memoirs Am. Math. Soc., Vol.713, 2001. W. Rudin, [em Functional Analysis]{}, McGraw-Hill Company. New York 1973. [ Smyshlyaev V.P., Diffraction by conical surfaces at high frequencies, [*Wave Motion*]{} Vol. 12, No.4, 329-339 (1990).]{} [ R.Temam and M.Ziane, [*Navier-Stokes equations in thin spherical domains*]{}, Optimization Methods in Partial Differential Equations, Contemporaty Math., AMS, Vol.209, pp.281–314.]{} H. Triebel, [*Interpolation Theory, Function Spaces, Differential Operators*]{}, 2nd edition, Johann Ambrosius Barth Verlag, Heidelberg–Leipzig 1995.
[**T. Buchukuri**]{}, [*A.Razmadze Mathematical Institute, Tbilisi State University, Tamarashvili str. 6, Tbilisi 0177, Georgia*]{}\
[email: t\[email protected]]{}\
\
[**R. Duduchava**]{}, [*A.Razmadze Mathematical Institute, Tbilisi State University, Tamarashvili str. 6, Tbilisi 0177, Georgia*]{}\
[email: [email protected]]{}\
\
[**G. Tephnadze**]{}, [*Department of Mathematics, Faculty of Exact and Natural Sciences, Tbilisi State University, Chavchavadze str. 1, Tbilisi 0128, Georgia*]{}\
[email: [email protected]]{}\
[^1]: The investigation is supported by the grant of the Shota Rustaveli Georgian National Science Foundation GNSF/DI/10/5-101/12
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In these lectures I discuss peculiarities of the critical behaviour of “non-ideal” systems as it is explained by the renormalization group approach. Examples considered here include account of the single-ion anisotropy, structural disorder, frustrations. I introduce main ideas of renormalization and show how it serves the explanation of typical features of criticality in the above systems: softening of the phase transition, changes in the universality class, complicated effective critical behaviour. critical behaviour, renormalization, field-theoretical renormalization group 05.50.+q, 05.70.Jk, 64.60.Ak'
author:
- 'Yu.Holovatch'
---
Introduction {#I}
============
There exist many ways one can choose to approach the subject of these lectures. Taken that the presentation is limited in time (spent by the students and lecturers at the Caribbean seashore during the Mochima school in theoretical physics) and in space (given by the Editors of this volume) this choice becomes a difficult one. The way of presentation I decided to follow was chosen for several reasons. Extremely high theoretical level of certain modern renormalization group (RG) studies of criticality in different systems (i.e. the “language” of these studies) sometimes does not allow uninitiated reader to follow the derivations and even to understand the problem statement and/or physical consequences of the results. Therefore the goal would be to provide a minimal vocabulary, explaining main notions as simple as possible. On the other hand, it is tempting to use such a simple vocabulary to make a short review of state-of-the art RG studies in a certain domain. RG explanation of criticality in “non-ideal” 3d systems might be a good candidate for such a domain: it is a subject of ongoing activity where important results have recently been obtained and still a lot is to be done.
The following account will serve this purpose: after mentioning several examples of criticality and scaling in condensed matter physics (section \[II\]) and introducing model Hamiltonians of the “non-ideal” systems we shall be interested in (section \[III\]) I shall give the main ideas and notions of the renormalization taking as an example a simple 1d Ising model (section \[IV\]). Once the reader is acquainted with the RG transformation, its flow and fixed points, stability, universality and scaling, I shall pass to the “non-ideal” systems showing how to obtain their effective Hamiltonians and to the reviewing of recent results in this domain (sections \[V\] and \[VI\]). Some conclusions and outlook are given in section \[VII\].
Criticality and scaling {#II}
=======================
It is generally recognized that the term “critical point” was introduced in 1869 by Thomas Andrews who studied a special point for carbon dioxide at about 31$^{\circ}$C and 73 atmospheres pressure where the properties of liquid and of gas become indistinguishable [@history]. Approaching the critical point $T_{\mathrm{c}}$ from below, the liquid-gas density difference obeys a power law [*scaling*]{} governed by the [*critical exponent*]{} $\beta$: $$\label{2.1}
\rho_{\rm L} - \rho_{\rm G} \propto (T_{\mathrm{c}} - T)^{\beta},
\qquad T \rightarrow T_{\mathrm{c}}^-,$$ as shown in the figure \[fig1\]a.
(80,170) (20,15) (10,115) (170,95) (180,55) (90,105) (90,25) (100,00) (220,15) (200,115) (370,45) (380,05) (305,00)
Similar power laws were found for some other physical quantities, describing their approach to zero or singularity at $T_{\mathrm{c}}$. Less known is the fact, that the mean field values of critical exponents (which are integers or their simple relations, e.g. $\beta=1/2$ for equation (\[2.1\])) were questioned already over a century ago [@Verschaffelt00][^1]. Therefore much time passed before there appeared the theory that both explained scaling behaviour (\[2.1\]) found in condensed matter physics, and, on a larger scale, not only in physics, as well as offered quantitatively accurate predictions for the exponents. As we now know this theory is based on the renormalization group formalism [@Wilson] and has its origin in quantum field problems [@RGbooks].
Next experimental observations of critical behaviour arrived in 1895, when Pierre Curie showed that the ferromagnet iron also displayed a special point. This point is the highest temperature at which iron can be permanently magnetized at zero external magnetic field which is now called the Curie point. Curie himself noticed the parallelism between density-temperature curve at constant pressure for carbon dioxide and magnetization-temperature curves at constant magnetic field for iron. Behaviour of the [*order parameter*]{}, spontaneous magnetization $M$, in the vicinity of Curie point (we denote it $T_{\mathrm{c}}$ as well) is governed by the familiar scaling law, cf. equation (\[2.1\]): $$\label{2.2}
M \propto (T_{\mathrm{c}} - T)^{\beta}, \qquad T \rightarrow
T_{\mathrm{c}}^-.$$ It appears that different magnetic systems possess different sets of critical exponents but there exist wide classes of magnets characterized by the same values of exponents. For example, uniaxial magnets (like ${\rm FeF_2}$, ${\rm MnF_2}$, ${\rm
K_2CoF_4}$, ${\rm Rb_2CoF_4}$) possess the same critical exponents. Moreover, they coincide with those of simple fluids (${\rm Xe}$, ${\rm SF_6}$, ${\rm CO_2}$ …) and some other systems. In such a case it is said that these systems belong to the same [*universality class*]{}. The reasons for this as well as the main characteristics of a system which define the universality classes will be cleared up in the subsequent sections.
As far as the second derivatives of Gibbs free energy (isothermal magnetic susceptibility for magnets and isothermal compressibility for fluids) diverge at $T_{\mathrm{c}}$ such [phase transitions]{} are referred to as the second order phase transitions. This classification is due to Paul Ehrenfest. The order parameter does not possess a discontinuity at $T_{\mathrm{c}}$: it is an example of a continuous phase transition. Phenomena which occur in the vicinity of the 2nd order phase transition point are called critical phenomena. Changes of system structure at $T_{\mathrm{c}}$ are reflected in the behaviour of the pair correlation function $G_2(r)$ (density-density function for fluids and magnetization-magnetization one for magnets) and described by two universal exponents $\eta$ and $\nu$. The exponential decay of $G_2(r)$ with distance at $T>T_{\mathrm{c}}$ transforms to the power law one: $$\label{2.3}
G_2(r) \sim \re^{-r/\xi}\,, \qquad T>T_{\mathrm{c}}\,; \qquad
G_2(r)\sim \frac{1}{r^{d-2+\eta}}\,, \qquad T=T_{\mathrm{c}}\,,$$ where $\xi$ is a correlation length which diverges at $T_{\mathrm{c}}$: $\xi\sim |T-T_{\mathrm{c}}|^{-\nu}$, $T \rightarrow T_{\mathrm{c}}$.
The XX century witnessed numerous experiments performed after pioneering works of Thomas Andrews and Pierre Curie where it was shown that criticality and scaling accompany not only the second order thermodynamic phase transitions. They are found in quantum phase transitions, percolation, non-equilibrium dissipative phase transitions. Properties of long flexible polymer chains in good solvents are described in terms of critical phenomena as well. This list can be continued. However already the above mentioned field of phenomena is very wide, especially if one takes into account that any of them occurs in objects differing by their microscopic nature. Thermodynamic 2nd order phase transitions occur in magnets, ferroelectrics, quantum liquids; percolation phenomena occurs in resistor networks, random magnets, gas filters. Nevertheless there are at least two essential features common for the above mentioned phenomena. These are: [*the singular character of change of properties*]{} in certain critical points and [*universal behaviour*]{} in the vicinity of these points. As it became clear now the reason for both is anomalous growth of fluctuations and their correlation at very high distances in the vicinity of critical points. Correlation range becomes the only characteristic scale of the system and this causes the insensitivity of its behaviour to the so-called microscopic parameters. In particular this results in the [*scale invariance*]{} of the system where critical phenomena occur.
Since the pioneering work of Ising [@Ising25], theoretical description of many-particle systems is often based on the so-called classical spin models. They appeared to be of primary importance to reveal the main features of criticality. In the next section \[III\] we shall discuss several models currently used in describing the critical phenomena, the analysis of which will be a subject of subsequent sections.
Non-ideal spin Hamiltonians: single ion anisotropy, structural disorder, frustrations {#III}
=====================================================================================
In classical spin models, each particle is imitated as a vector located on a lattice site and “interacting” with other vectors. The interaction is chosen to mimic the interparticle one whereas the dimensionality of the vector is equal to the number of internal degrees of freedom of the particle. Below we start with what we call an ideal system, which is a $m$-vector model and show how to include different types of non-idealities within a spin model formalism.
An ideal system: $m$-vector model {#IIIa}
---------------------------------
The model describes a system of $m$-dimensional classical vectors (“spins”) located in the sites of $d$-dimensional hypercubic lattice. The Hamiltonian of the $m$-vector model in the absence of an external magnetic field reads: $$\label{3.1}
{\cal H} = - \frac{1}{2}\sum_{\bf R, R'} J\left(|{\bf R} - {\bf
R'}|\right) \vec{S}_{{\bf R}} \vec{S}_{{\bf R'}}\,,$$ where $J(|{\bf R} - {\bf R'}|)$ is an interaction between spins $\vec{S}_{{\bf R}}$ and $\vec{S}_{{\bf R'}}$ located in sites ${\bf R}$ and ${\bf R'}$ (we shall consider it to be the short-range one and of ferromagnetic origin: $J(R)>0$) and $\vec{S}_{{\bf R}} \vec{S}_{{\bf R'}}$ means a scalar product. Such a model appears as a natural generalization of the Ising model for the case of $m$-component spin [@Stanley68] and serves as a basic model in describing phase transitions in systems with multicomponent order parameter. Below, we shall be mainly interested in the 3d systems. However, for the sake of completeness let us recall that ferromagnetic ordering does not occur in this model at $d=1$ [@Stanley68], it does not occur for $d=2$ and $m>1$ either [@continuous]. Whereas the celebrated Onsager solution [@Onsager44] of the 2d Ising model ($d=2$, $m=1$) brings about the second order phase transition with non-trivial values of the critical exponents: $\beta=1/8$, $\nu=1$, $\eta=1/4$, $\gamma=7/4$, $\alpha=0$ (exponents $\gamma$ and $\alpha$ govern power law scaling of the magnetic susceptibility and specific heat, the latter being logarithmically divergent for the 2d Ising model). At $d=3$, ferromagnetic ordering occurs at any $m$ (see table \[tab1\] for typical values of exponents), and $d=4$ is the [*upper critical dimension*]{} of the problem: exponents attain their mean field values for $d\geqslant 4$.
$m$ $\gamma$ $\nu$ $\eta$ $\beta$ $\alpha$
----- -------------- -------------- --------------- -------------- --------------
1 $1.2396(13)$ $0.6304(13)$ $0.0335(250)$ $0.3258(14)$ $0.109(4)$
2 $1.3169(20)$ $0.6703(15)$ $0.0354(250)$ $0.3470(16)$ $-0.011(4)$
3 $1.3895(50)$ $0.7073(35)$ $0.0355(250)$ $0.3662(25)$ $-0.122(10)$
: The standard values of the critical exponents of the 3d $m$-vector model (obtained in [@Guida98] from the high-order renormalization group expansions). \[tab1\]
We refer to the above model as to the ideal one: it describes the lattice system without any defects of structure or complexities of interaction which are often encountered in real systems. Let us show how those can be considered within the same formalism.
Single-ion anisotropy {#IIIb}
---------------------
Real substances often are anisotropic. For instance, in cubic crystals one expects the spin interaction to react on the lattice structure (crystalline anisotropy) suggesting additional terms in the Hamiltonian, invariant under the cubic group [@Aharony76]. Such a single-ion anisotropy breaks the rotational symmetry of the $m$-vector magnet (\[3.1\]) and the Hamiltonian reads: $$\label{3.2}
{\cal H} = - \frac{1}{2}\sum_{\bf R, R'} J\left(|{\bf R} - {\bf
R'}|\right) \vec{S}_{{\bf R}} \vec{S}_{{\bf R'}} + V \sum_{\bf R}
\sum_{i=1}^m \left(S^i_{\bf R}\right)^4,$$ where $V$ defines anisotropy strength and makes the order parameter to point either along the edges ($V>0$) or along diagonals ($V<0$) of a $m$-dimensional hypercube. Therefore a model with the spin Hamiltonian (\[3.2\]) is often called a [*cubic model*]{}.
An interesting phenomenon is observed in 3d cubic magnets: for low values of $m$ their critical exponents coincide with those of the “ideal” $m$-vector model (one speaks about isotropization of critical fluctuations), whereas for large $m$ they belong to the new universality class. The value $m_{\mathrm{c}}^{\rm cub}$ at which one regime is changed to the other one is called the marginal dimension. Moreover, transition to the low-temperature phase may also occur via the first-order scenario.
Structural disorder {#IIIc}
-------------------
To treat structural disorder within the lattice model one usually introduces random variables into the Hamiltonian (\[3.1\]) and couples them to spin degrees of freedom. The new model can either mimic a quenched system (new variables are randomly distributed and fixed in a certain configuration) or an annealed, equilibrium one [@Brout59]. Here, we shall be interested in changes of critical behaviour caused by quenched disorder. Furthermore, we shall consider two different examples, showing how to introduce disorder via dilution and random anisotropy[^2].
### Random-site dilution {#IIIc1}
To describe dilution, one may introduce the random-site $m$-vector model, considering situation, when spins $\vec{S}_{{\bf R}}$ in (\[3.1\]) occupy only a part of the lattice sites, $N_1$, the rest $N-N_1$ sites being empty (or occupied by non-magnetic atoms). Magnetic and non-magnetic sites are randomly distributed and fixed in a certain configuration. The model Hamiltonian reads: $$\label{3.3}
{\cal H} = - \frac{1}{2}\sum_{\bf R, R'} J\left(|{\bf R} - {\bf
R'}|\right) c_{{\bf R}}c_{{\bf R'}}\vec{S}_{{\bf R}} \vec{S}_{{\bf
R'}}\,,$$ where $c_{{\bf R}}$ are the occupation numbers: $$\label{3.4}
c_{{\bf R}} = \left \{ \begin{array}{ll} 1, & \quad \mbox{site
${\bf R}$ is occupied},
\\
0, & \quad \mbox{site ${\bf R}$ is empty}.
\end{array}
\right.$$ Here we do not touch upon the phenomena occuring near the percolation threshold $c_{\rm perc}$ and consider the so-called weak dilution, $c=N_1/N\gg c_{\rm perc}$. To complete the model one should choose a certain distribution function for the occupation numbers $c_{{\bf R}}$. Let us consider the case when the site is occupied with the probability $c$ and is empty with the probability $(1-c)$ and this probability does not depend on the occupation numbers on the neighbouring sites. Such a situation corresponds to the following occupation probability ${\cal P}
(\{c_{{\bf R}} \})$: $$\label{3.5}
{\cal P} (\{c_{{\bf R}} \})=\prod_{\bf R} p(c_{{\bf R}}),
\hspace{1cm} p(c_{{\bf R}})= c \delta (c_{{\bf R}}-1) + (1-c)
\delta (c_{{\bf R}}).$$ For different $m$, the above model describes magnetic phase transitions in crystalline alloys of uniaxial magnets and their non-magnetic isomorphs ${\rm Fe_xZn_{1-x}F_2}$, ${\rm
Mn_xZn_{1-x}F_2}$ [@Folk03], diluted Heisenberg-like magnets [@Pelissetto02; @Dudka03], superfluid phase transition in He$^4$ in porous medium [@Yoon97]. Similar to the cubic anisotropy systems, there exists a marginal dimensionality $m_{\mathrm{c}}^{\rm dil}$ separating two different scenarios: for $m<m_{\mathrm{c}}^{\rm dil}$ dilution causes a change in the universality class, whereas the asymptotic critical exponents of $m>m_{\mathrm{c}}^{\rm dil}$-component systems remain unchanged under dilution. The so-called Harris criterion allows us to determine this marginal dimension from the pure system heat capacity behaviour [@Harris74]. It states that critical exponents are not altered by dilution, if the heat capacity of the pure system does not diverge ($\alpha<0$). From Table \[tab1\] one concludes that in three dimensions it is the Ising model which changes the universality class.
### Random anisotropy
Another way of introducing randomness to the model (\[3.1\]) is to consider the case, when each spin is subjected to a local anisotropy of random orientation. The resulting Hamiltonian reads [@Harris73]: $$\label{3.6}
{\cal H} = - \frac{1}{2}\sum_{\bf R, R'} J(|{\bf R} - {\bf R'}|)
\vec{S}_{{\bf R}} \vec{S}_{{\bf R'}}
-D\sum_{{\bf R}} (\hat {x}_{\bf
R}\vec{S}_{\bf R})^{2}.$$ Here, $D>0$ is the strength of the anisotropy and $\hat {x}_{\bf
R}$ is a random unit vector pointing in the direction of the local anisotropy axis. The random anisotropy model (\[3.6\]) is relevant to the description of a wide class of disordered magnets. It was first introduced to describe magnetic properties of amorphous alloys of rare-earth compounds with aspherical electron distributions and transition metals [@Harris73]. Today the majority of the amorphous alloys containing rare-earth elements are recognized to be random anisotropy magnets [@ramreviews].
As in the random-site case, the model should be completed by choosing a certain distribution $p(\hat x_{\bf R})$ for the random variables $\hat x_{\bf R}$. Most often, two different distributions are considered [@Aharony75]. The first is an isotropic one, where the random vector $\hat x_{\bf R}$ points with equal probability in any direction in the $m$-dimensional hyperspace: $$\label{3.7}
p(\hat x_{\bf R}) \equiv\left(\int \rd^m \hat x_{\bf
R}\right)^{-1}= \frac{\Gamma(m/2)}{2\pi^{m/2}}\,.$$ Here $\Gamma(x)$ is Euler gamma-function, and the right-hand side presents the volume of the $m$-dimensional hypersphere of unit radius. This distribution mimics an amorphous system without any preferred direction. The second distribution restricts the vector $\hat x_{\bf R}$ to point with equal probability along one of the $2m$ directions of the axes $\hat k_i$ of a hypercubic lattice: $$\label{3.8}
p(\hat x_{\bf R}) = \frac{1}{2m}\sum_{i=1}^m\left[\delta^{(m)}
(\hat x_{\bf R}-\hat k_i)+\delta^{(m)}(\hat x_{\bf R}+\hat k_i)\right],$$ where $\delta(y)$ are Dirac $\delta$-functions. This distribution (sometimes called a cubic one) corresponds to a situation when an amorphous magnet still remembers the initial cubic lattice structure.
It is generally believed that ferromagnetism does not exist for the 3d model (\[3.6\]) with the isotropic random axis distribution (\[3.7\]) [@ramreviews]. However, anisotropic distribution (\[3.8\]) leads to a magnetically ordered low temperature phase and the transition belongs to the random site Ising model universality class [@Dudka05].
Frustrations {#IIId}
------------
An archetype of a model describing the effect of frustrations is the model of stacked triangular antiferromagnet [@Kawamura88]. In this model, the antiferromagnetically interacting spins are placed on the sites of 2d triangular lattices stacked in register along the orthogonal direction. The sign of the interlayer interaction is unimportant, because there are no frustrations in orthogonal direction. The Hamiltonian reads: $$\label{3.9}
{\cal H} = - \frac{1}{2}\sum_{\langle {\bf R,R'}\rangle} J \,
\vec{S}_{\bf R} \vec{S}_{\bf R'},\qquad
J = \left \{ \begin{array}{ll} J_1 < 0, & \mbox{inside a
plane,}
\\
J_2\,, & \mbox{between planes.}
\end{array}
\right.$$ Sum in (\[3.9\]) runs over the nearest neighbours of the above described lattice. Systems which are characterized by the Hamiltonian (\[3.9\]) exhibit noncollinear spin ordering. An example is given by the famous 120$^\circ$ structure: each spin in a layer forms 120$^\circ$ angles with the neighbouring spins. Although model (\[3.9\]) is formulated for general $m$, of most interest are values $m=2$ and $m=3$. Namely for these values of $m$ the model has experimental realizations [@Pelissetto02; @Delamotte04] and describes noncollinear ordering of stacked triangular antiferromagnets as ${\rm VCl_2}$, ${\rm VBr_2}$, ${\rm CsMnBr_3}$, and helical magnets as ${\rm
Ho}$, ${\rm Dy}$, $\beta-{\rm MnO_2}$. In the latter substances, frustration is induced by the competition of ferromagnetic nearest-neighbour and antiferromagnetic next-nearest-neighbour interactions, which acts only along one lattice axis: $J_1>0$ and $J_2<0$, correspondingly. When the ratio $J_1/J_2$ exceeds a critical value, in the low-temperature phase the spins align ferromagnetically in a plane and form a spiral along the orthogonal axis. Moreover, at $m=3$ Hamiltonian (\[3.9\]) also describes A/B transition in He$^3$.
Numerous experimental and MC studies performed so far have not lead to the definitive conclusion about the order of transition into non-collinear state. There is no unique answer from the theoretical viewpoint either [@Pelissetto02; @Delamotte04].
Renormalization {#IV}
===============
In order to explain the main ideas of the renormalization group (RG) theory and to show how this method works in practice we shall study the critical behaviour of the 1d Ising model by means of the RG approach. The explanations given below are due to Michael Fisher [@Fisher], who compared this study with the Bohr-Sommerfeld picture in quantum mechanics. Passing from the classical mechanics to the full account of quantum mechanics, Bohr-Sommerfeld’s picture represents only a crude approximation. Nevertheless, it introduces some important ideas, just like RG study of the 1d Ising model enables one to introduce ideas of renormalization and scaling in the critical region.
RG transformation {#IVa}
-----------------
We start from the Hamiltonian of the 1d Ising model in the presence of an external magnetic field $H$, which being divided by $k_{\mathrm{B}}T$ is written in the following form: $$-{\cal H}_{\rm eff} = K \sum_{j=1}^{N} S_j S_{j+1} + h
\sum_{j=1}^{N} S_j + C \sum_{j=1}^{N} 1, \label{12.23}$$ where $K=J/(k_{\mathrm{B}}T)$, $h=H/(k_{\mathrm{B}}T)$, and the last term $C \sum_{j=1}^{N} 1=NC$ is added for the convenience of forthcoming calculations. We call this temperature-dependent Hamiltonian an effective one, ${\cal H}_{\rm eff}$. Note, however, that contrary to the effective Hamiltonians considered in the forthcoming section \[V\], it contains full information about the spin Hamiltonian. Given $\{K,h,C\}$ one completely specifies ${\cal H}_{\rm eff}$: thus it can be regarded as a point in a space of $\, {\rm 3} \,$ parameters $\{K,h,C\}$. With change of $\{K,h,C\}$, this point moves.
One of the first approaches to the RG is to regard it as a specific way of calculating the partition function $Z_N\{{\cal
H}\}$: $$Z_N\{{\cal H}_{\rm eff}\} = \frac{1}{2^N} \sum_{S_1=\pm 1} \dots
\sum_{S_N=\pm 1} \re^{-{\cal H}_{\rm eff}}, \label{12.24}$$ where we have normalized the expression for the partition function by the partition function of the ideal model ($Z^{{\rm
ideal}}=2^N$) to obtain $Z_N\{{\cal H}_{\rm eff}\}
\rightarrow_{T\rightarrow \infty} 1$. The RG method of calculating $Z_N$ is to step-by-step perform the summation in (\[12.24\]) in such a way as to try as much as possible to preserve the system as it used to be prior the summation.
Let us do this by means of [*decimation*]{} procedure: performing summation over every second spin variable (see figure \[fig2\]). To show the changes which are introduced by summation over certain spin variable $S_0$ let us represent the total Boltzmann weight in the factorized form: $$\re^{-{\cal H}_{\rm eff}} = \cdots \re^{KS_-S_0 +
\frac{1}{2}h(S_-+S_0)+C} \re^{KS_0S_+ + \frac{1}{2}h(S_0+S_+)+C}
\cdots= \cdots P(S_-,S_0) P(S_0,S_+) \cdots \label{12.26}$$
\[ht\]
![[*Decimation*]{} procedure followed by the [*spatial rescaling*]{}. Summation over every second spin variable is performed and then all lengths are rescaled in such a way that the new lattice spacing $a^{\prime}$ is equal to the old one $a$. Now any distance $R$ in the original lattice becomes $R^{\prime}= {1}/{2} R$ in the new one measured in the units of the lattice spacing. \[fig2\] ](Fig2.eps){width="8cm"}
Summation over $S_0$ which enters only $P(S_-,S_0)$ and $P(S_0,S_+)$ will lead to a new function $P^{\prime}(S_-,S_+)$, defined by: $$P^{\prime}(S_-,S_+) = \frac{1}{2} \sum_{S_0=\pm 1}
P(S_-,S_0) P(S_0,S_+).
\label{12.27}$$ Factor ${1}/{2}$ is included in (\[12.27\]) because with each spin eliminated one must remove factor $\frac{1}{2}$ from the normalizing factor in (\[12.24\]). Now the RG idea is to express the new factor $P^{\prime}$ in the same form as the initial one: $$P^{\prime}(S_-,S_+) = \re^{K^{\prime}S_-S_+ +
\frac{1}{2}h^{\prime}(S_-+S_+)+C^{\prime}}. \label{12.28}$$ The new parameters define the [*renormalized Hamiltonian*]{}: $${\cal H_{\rm eff}^{\prime}} = {\cal H_{\rm eff}^{\prime}}
\{K^{\prime}, h^{\prime}, C^{\prime} \}, \label{12.29}$$ and ${\cal H_{\rm eff}^{\prime}}$ has a half of initial spins. The result can be formally written as : $${\cal H_{\rm eff}^{\prime}} = {\cal R}_b [{\cal H}_{\rm eff} ],
\label{12.30}$$ with a spatial rescaling factor $b$ (in our case $b=2$, see figure \[fig2\]). The resulting number of spins $N^{\prime}$ is connected with the initial one by $N^{\prime} = N/b$, and in the case of $d$-dimensional system this is generalized to: $$N^{\prime} = N/b^d.
\label{12.31}$$
Performing summation (\[12.27\]) and expressing the result in the renormalized form (\[12.28\]) one can get expressions for the renormalized variables $K^{\prime}, h^{\prime}, C^{\prime}$ (it is proposed to the interested reader to do this).
Now the renormalized model is characterized by the Hamiltonian ${\cal H_{\rm eff}^{\prime}}( K^{\prime}, h^{\prime}, C^{\prime})$ having the form similar to the initial one, though the lattice constant after taking the sum over every second spin is equal to $2a$ (see figure \[fig2\]). In order to have the renormalized model looking as the initial one we [*rescale*]{} all lengths in such a way that the new lattice spacing $a^{\prime}$ is equal to the old one. Now any distance $R$ in the original lattice becomes $R^{\prime}= {1}/{2} R$ in the new one measured in the units of the lattice spacing. For the arbitrary $b$ we have the following mapping: $$R \Rightarrow R^{\prime} = R/b.
\label{12.35}$$ Now let us have a look at the behaviour of the spin-spin correlation function $\langle S_0 S_R \rangle$. First let us renumber the remaining spins to have the labels arranged in the subsequent order (see figure \[fig2\]): $ S_2 \Rightarrow
S^{\prime}_1, \quad S_4 \Rightarrow S^{\prime}_2, \quad \dots,
\quad S_{2k} \Rightarrow S^{\prime}_k, \dots $ Or, taking into account (\[12.35\]), $ S_{2R^{\prime}} =
S_{R^{\prime}}^{\prime}$. Since the remaining spins $S^{\prime}$ have not changed under the renormalization procedure, the renormalized correlation function is equal to the original one: $$\langle S_0 S_{2R^{\prime}} \rangle =
\langle S_0^{\prime} S_{R^{\prime}}^{\prime} \rangle.
\label{12.38}$$ It follows that if the original correlation length is $\xi=\xi({\cal H}_{\rm eff})$ then the renormalized correlation length is two times smaller. Or, for a general $b$: $$\xi({\cal H}_{\rm eff}) = b \xi({\cal H}_{\rm eff}^{\prime}).
\label{12.39}$$ The RG procedure has the effect of shrinking the correlation length. Recalling that $\xi$ becomes infinite at $T=T_{\mathrm{c}}$ one can state, that the RG procedure is driving a system away from criticality (if it was not critical). Already here it is seen that the RG transformation we are considering has a deep connection with the critical properties of a system.
Fixed points and RG flows {#IVb}
-------------------------
To proceed in the determination of these properties let us have a closer look at the partition function and at the free energy of the system. First, after taking a partial trace of the partition function $Z_N\{{\cal H_{\rm eff}}\}$ we get the Boltzmann distribution in $N^{\prime}$ spin variables $S^{\prime}$: $$\re^{-{\cal H}_{\rm eff}^{\prime}(S^{\prime})} = Sp_{N^{\prime
\prime}}^{S^{\prime \prime}} \re^{-{\cal H}_{\rm eff}(S)},
\label{12.40}$$ where $Sp_{N^{\prime \prime}}^{S^{\prime \prime}}$ stands for the trace over $N^{\prime \prime}= N - N^{\prime}$ spin variables $S^{\prime \prime}$. Now taking trace over the remaining $N^{\prime}$ spin variables one gets: $$Z_{N^{\prime}}\{{\cal H}_{\rm eff}^{\prime}\} =
Sp_{N^{\prime}}^{S^{\prime}} \re^{-{\cal H}_{\rm
eff}^{\prime}(S^{\prime})}. \label{12.41}$$ Substituting into the right hand side of (\[12.41\]) its explicit form given by (\[12.40\]) one gets: $$\nonumber Z_{N^{\prime}}\{{\cal H}_{\rm eff}^{\prime}\} =
Sp_{N^{\prime}}^{S^{\prime}} Sp_{N^{\prime \prime}}^{S^{\prime
\prime}} \re^{-{\cal H}_{\rm eff}(S)}= Z_{N}\{{\cal H}_{\rm eff}
\}.$$ This relation can be rewritten in terms of the [*flow equations*]{} which describe a motion of a point describing an effective Hamiltonian: $$\begin{aligned}
K^{\prime} &=& {\cal R}_K (K,h),
\label{12.43} \\
h^{\prime} &=& {\cal R}_h (K,h),
\label{12.44} \\
C^{\prime} &=& b^d C + {\cal R}_0 (K,h),
\label{12.45}\end{aligned}$$ for the [*couplings*]{} defining ${\cal H}_{\rm eff}$. As far as the temperature $T$ enters parameter $K$, the above relations imply the flow equations for $T$ as well. To study this let us consider $H=0$ and write the flow equation for the temperature as: $$T \Rightarrow T^{\prime}= {\cal R}(T),
\label{12.46}$$ with ${\cal R}(T)$ being appropriate function of $T$. Suppose that ${\cal R}(T)$ has a form given in the figure \[fig3\]. The important feature there is that ${\cal R}(T)$ crosses the line $T^{\prime}=T$ at some point $T=T^*$. This point is called the [*fixed point*]{} (FP). It is clear that when $T$ is smaller or greater than $T^*$ the successive application of the renormalization procedure drives the system away from the fixed point. On the other hand at $T=T^*$ the system remains at the fixed point under the application of the RG procedure. Recalling the flow equation for the correlation length (\[12.39\]) which can be written as: $$\xi(T) = b \xi(T^{\prime}),
\label{12.47}$$ one has that in the FP this reads: $$\xi(T^*) = b \xi(T^*).
\label{12.48}$$
\[ht\]
![Temperature renormalization function ${\cal R}(T)$. ${\cal R}(T)$ crosses the line $T^{\prime}=T$ at the fixed point $T=T^*$. When $T$ is smaller ($T=T_0$) or greater ($T=T_1$) than $T^*$, the successive application of the renormalization procedure drives the system away from the fixed point. On the other hand at $T=T^*$ the system remains at the fixed point under the application of the RG procedure. \[fig3\] ](flow.eps){width="7cm"}
Since $b>1$ it is possible only when
- $\xi(T^*) = \infty$ and
- $\xi(T^*) = 0$.
The first case characterizes the critical point $T^*=T_{\mathrm{c}}$, whereas the second corresponds to the vanishing of the correlation length at zero or infinite temperatures (when the spins are frozen at a ground state or totally uncoupled). Note, that this corresponds to two FPs more ($T^*=0$, $T^*=\infty$) in the figure \[fig3\].
Having identified the fixed point $T^*$ of the RG transformation as the critical point of the system let us study what knowledge about the critical exponents can be obtained based on the properties of the RG transformation. First let us linearize it in the vicinity of the FP introducing the variable: $$\tau=\frac{T-T_{\mathrm{c}}}{T_{\mathrm{c}}} = \frac{T-T^*}{T^*}
\label{12.49}$$ and replacing the plot of ${\cal R}(T)$ near $T^*$ by its tangent at $T^*$.
Then after the renormalization the temperature deviation will be: $$\tau^{\prime} \equiv \tau^{(1)} \simeq \Lambda_1(b) \tau
\label{12.50}$$ for small enough $\tau$, where $\Lambda_1(b)$ is the slope of the tangent. To find its dependence on $b$ let us apply the renormalization procedure twice: $$\tau^{\prime \prime} \equiv \tau^{(2)} \simeq \Lambda_1(b)
\Lambda_1(b) \tau .
\label{12.51}$$ Such transformation should be equivalent to transforming with a spatial rescaling factor $b^2$: $$\tau^{(2)} \simeq \Lambda_1(b^2) \tau ,
\label{12.52}$$ or $$\Lambda_1(b) \Lambda_1(b) = \Lambda_1(b^2).
\label{12.53}$$ This leads to the conclusion that $\Lambda_1(b)$ should have the following form: $$\Lambda_1(b) = b^{\lambda_1},
\label{12.54}$$ with $\lambda_1$ being constant independent of $b$. Now for the correlation length after $l$ transformations one gets: $$\xi(\tau) = b^l\xi([\Lambda_1(b)]^l \tau) =
b^l\xi(b^{l\lambda_1}\tau).
\label{12.55}$$ As far as (\[12.55\]) holds for any $l$ (and for small $\tau$(!)) let us choose it to satisfy: $$b^{l\lambda_1}\tau =1,$$ or $$b^l = (1/\tau)^{1/\lambda_1}.
\label{12.56}$$ Then from (\[12.55\]) one gets: $$\nonumber
\xi(\tau) = (1/\tau)^{1/\lambda_1} \xi(1)$$ or $$\xi(\tau) \sim \tau^{-1/\lambda_1}. \label{12.57}$$ Comparing (\[12.57\]) with the definition of the critical exponent $\nu$: $\xi(\tau) \sim \tau^{-\nu}$ we get: $$\nu =1/\lambda_1.
\label{12.58}$$ So as we have seen, the knowledge of the linearized RG transformation enables one to determine the critical exponent!
Note, that having carried out similar analysis for the free energy per spin, $f={-}{N^{-1}}\ln Z_N$, one arrives at the expression: $
f(\tau) =b^{-d l}f(b^{\lambda_1 l}\tau)$. Again, choosing $b$ from (\[12.56\]) one gets $f(\tau)
=\tau^{d/\lambda_1}f(1)=\tau^{d \nu}f(1)$. Comparing the latter with the scaling behaviour of the free energy in the vicinity of the critical point $f(\tau)\sim \tau^{2-\alpha}$ one proves the [*hyperscaling relation*]{}: $$2-\alpha= d\nu.
\label{12.61}$$
Similarly, considering the case of non-zero magnetic field one can write the recursion relations for the renormalized temperature and field. Their linearization in the vicinity of the fixed point leads to the [*matrix of the linear RG operator*]{} with two eigenvalues, defining two different critical exponents.
To complete the RG study of the 1d Ising model it remains to perform the above described procedure explicitly and to find the values of the critical exponents (at zero-temperature FP $T=H=0$, [@Fisher]). The above described transformations concisely give the main features of the RG transformation, allowing us, in particular, to define the critical exponents of the system.
From the spin Hamiltonians to the effective ones {#V}
================================================
Very often a starting point for the RG study of critical behaviour of a many-particle system is its [*effective Hamiltonian*]{}. Taken, that in the model description the system is identified with the [*spin Hamiltonian*]{} (see section \[III\]), the effective Hamiltonian arises as a certain “metamodel”. It shares global features of different spin Hamiltonians: their dimensionalities, symmetries, type of interparticle interaction and as a result, brings about common features of their critical behaviour. Below, we shall show how to obtain the effective Hamiltonians for the spin models of section \[III\].
$m$-vector model {#Va}
----------------
To get an effective Hamiltonian we proceed as follows. Let us define the free energy ${\cal F}$ and the partition function ${\cal Z}$ of a spin model (\[3.1\]) as: $$\label{5.2}
{\cal F} = -\beta^{-1} \ln {\cal Z}, \qquad {\cal Z}={\rm Sp}
\re^{-\beta {\cal H}},$$ where $\beta=(k_{\mathrm{B}}T)^{-1}$ and ${\rm Sp}(\dots) $ as usually means the sum over all possible states. In our case it corresponds to the integration over the surface of $m$-dimensional hypersphere (we take it to be of unit radius): $$\label{5.3}
{\rm Sp} (\dots) = \prod_{\bf R} \int {\rm d} \vec{S}_{\bf R}
\delta (|\vec{S}_{\bf R}| - 1) (\dots).$$ Let us introduce the Fourier-transforms of the variables $\vec{S}_{{\bf R}}$ by: $$\label{5.4}
\vec{S}_{{\bf R}} = \frac{1}{\sqrt N} \sum_{{\bf k}} \re^{\ri{\bf
k}{\bf R}} \vec{S}_{{\bf k}}\,, \hspace{1cm} \vec{S}_{{\bf k}} =
\frac{1}{\sqrt N} \sum_{{\bf R}} \re^{-\ri{\bf k}{\bf R}}
\vec{S}_{{\bf R}}\,.$$ Here and below, when it will not be mentioned explicitly, vector ${\bf R}$ spans all sites of the lattice whereas ${\bf k}$ changes in the first Brillouin zone. Now the Hamiltonian (\[3.1\]) can be rewritten as : $$\label{5.5}
{\cal H} = -\frac{1}{2} \sum_{{\bf k}} \nu(k) \vec{S}_{{\bf k}}
\vec{S}_{-{\bf k}}\,,$$ where we have introduced Fourier transform of the potential of interaction by: $$\label{5.6}
J(R) = \frac{1}{ N} \sum_{{\bf k}} \re^{\ri{\bf k}{\bf R}}
\nu(k), \hspace{1cm} \nu(k) = \sum_{{\bf R}} \re^{-\ri{\bf k}{\bf
R}} J(R).$$ Now the partition function (\[5.2\]) reads: $$\label{5.7}
{\cal Z} ={\rm Sp} \re^{\frac{\beta }{2} \sum_{{\bf k}} \nu(k)
\vec{S}_{{\bf k}} \vec{S}_{-{\bf k}}}.$$ In order to take the trace in (\[5.7\]) let us transform it to the expression with linear dependence on $\vec{S}_{{\bf k}}$. To this end, one makes use of the Stratonovich-Hubbard transformation introducing the field variable $\vec{\phi}_{{\bf k}}$ which is conjugated to the spin variable $\vec{S}_{{\bf k}}$ by the identity: $$\begin{aligned}
\nonumber
\re^{-\beta {\cal H}} &=& \prod_{\vec{k}} \re^{\frac{\beta
\nu(k)}{2}\vec{S}_{{\bf k}} \vec{S}_{-{\bf k}}} = \prod_{\vec{k}}
(\frac{1}{2\pi \beta \nu(k)})^{m/2} \int {\rm d} \vec{\phi}_{{\bf
k}} \re^{\frac{-1}{2\beta \nu(k)} \vec{\phi}_{{\bf k}}
\vec{\phi}_{{\bf -k}} + \vec{S}_{{\bf k}} \vec{\phi}_{{\bf -k}}}
\\
&&{}\sim{} \int ({\rm d} \vec{\phi}) \re^{\sum_{\bf k}(
\frac{-1}{2\beta \nu(k)} |\vec{\phi}_{{\bf k}}|^2 + \vec{S}_{{\bf
k}} \vec{\phi}_{{\bf -k}})}. \label{5.8}\end{aligned}$$ As far as the Stratonovich-Hubbard transformation was performed for each function $\vec{S}_{{\bf k}}$ we arrived at the functional integral $\int ({\rm d} \vec{\phi}) = \prod_{{\bf R}}
\prod_{i=1}^m \int_{-\infty}^{\infty} {\rm d} \phi^{(i)}_{{\bf
R}}$ over the field variables $$\label{5.10}
\vec{\phi}_{{\bf R}} = \frac{1}{\sqrt N} \sum_{{\bf k}}
\re^{\ri{\bf k}{\bf R}} \vec{\phi}_{{\bf k}}\,, \hspace{1cm}
\vec{\phi}_{{\bf k}} = \frac{1}{\sqrt N} \sum_{{\bf R}}
\re^{-\ri{\bf k}{\bf R}} \vec{\phi}_{{\bf R}}\,.$$ From now on we omit the coefficients in front of the functional integral.
We have reached our goal: now the trace in the expression for the partition function concerns only the last term of the exponents under integration in (\[5.7\]) which is a linear function of $\vec{S}$: $$\label{5.11}
{\cal Z} = {\rm Sp} \re^{\frac{\beta}{2} \sum_{{\bf k}} \nu(k)
\vec{S}_{{\bf k}} \vec{S}_{-{\bf k}}} \sim \int ({\rm d}
\vec{\phi}) \re^{\frac{-1}{2} \sum_{{\bf k}} \frac{1}{\beta \nu
(k)} |\vec{\phi}_{{\bf k}}|^2} {\rm Sp} \, \re^{\sum_{{\bf
R}}\vec{S}_{{\bf R}} \vec{\phi}_{{\bf R}}}.$$ The last step is to take integral over $\vec{S}_{{\bf R}}$ (recall that the trace is defined by equation (\[5.3\])). This can be achieved by passing from the $m$-dimensional Cartesian coordinates $S^{1}, S^{2}, \dots, S^{m}$ to the $m$-dimensional polar ones $S,
\theta_1, \dots ,\theta_{m-1}$. Performing this integration and representing the result as a series in $\vec{\phi}$ one gets for the partition function: $$\label{5.12}
{\cal Z} \sim \int ({\rm d} \vec{\phi})\re^{\frac{-1}{2}
\sum_{{\bf k}} (\frac{1}{\beta \nu(k)} - u_2) |\vec{\phi}_{{\bf
k}}|^2 - \sum_{{\bf R}} \sum_{l=2}^{\infty}
\frac{u_{2l}}{(2l)!}|\vec{\phi}_{{\bf R}}|^{2l}},$$ where the coefficients $u_{2l}$ readily follow: $u_2 = -1/m$, $u_4=6/m^2(m+2)$, $\dots$ Expression (\[5.12\]) gives the functional representation of the partition function of the $m$-vector model. So far we have not gained a lot by the above described transformations: the difficulty of taking ${\rm Sp}$ of the initial expression (\[5.5\]) now is transformed into the difficulty of calculating a functional integral (\[5.12\]). However, in order to study critical behaviour of the model (\[3.1\]), expression (\[5.12\]) can be further approximated. It appears that all of $\phi$ higher than the fourth powers do not effect the asymptotic critical behaviour at $d=3$: they do not change the value of the fixed point of RG transformation (cf. section \[IV\]) and are [*irrelevant*]{} in the RG sense [@Wilson; @RGbooks]. Therefore one can be restricted to the $\phi^4$ model. Being interested in the long-range correlations arising in the system in the vicinity of a critical point, one substitutes the Fourier image of the interaction potential by its expansion for small wave vector values $k$: $\nu(k)\simeq\nu(0)-{1}/{2}|\nu''(0)|k^2$ and writes the contribution $\sum_{\bf k} k^2 \vec \phi_{\bf k} \vec \phi_{-\bf
k}$ in the form $\sum_{\bf R} (\nabla \vec \phi_{\bf R})^2$ re-scaling variables $\phi$ to get the gradient term without any coefficient. One further passes to the continuous limit $\sum_{\bf
R} \rightarrow \int {\rm d} {\bf R}$ and gets for the free energy: $$\label{5.13}
{\cal F} \sim \ln \int({\rm d}\vec{\phi}) \re^{-{\cal H}_{\rm
eff}}$$ with the [*effective Hamiltonian*]{}: $$\label{5.14}
{\cal H}_{\rm eff} = \int {\rm d}^d R \left\{ \frac{1}{2}
\left((\nabla \phi)^2 +\mu_0^2 \phi^2\right) + \frac{u_0}{4!}
\phi^4\right\},$$ where $\phi\equiv\vec{\phi}_{\bf R}$, $\phi^2\equiv|\vec{\phi}_{\bf R}|^2$, variables $\mu_0$ and $u_0$ are called bare mass and coupling. From the above derivation we know that $u_0$ is positive ($u_0\sim u_4$ from equation (\[5.12\])), which ensures the existence of the integral (\[5.13\]). Effective Hamiltonian (\[5.14\]) shares common global properties with the spin Hamiltonian (\[3.1\]): the dimension of space (i.e. dimension of vectors ${\bf k}$, ${\bf
R}$), dimension of the order parameter ($m$) and its symmetry: the functional representation we obtained is symmetric under group of rotations in the $m$-dimensional space $O(m)$: it depends only on $|\vec{\phi}|$. Let us now see what differences will appear in the functional representation of the non-ideal models.
Cubic model {#Vb}
-----------
It is intuitively clear that an effective Hamiltonian of the cubic model (\[3.2\]) will differ from that of the $m$-vector model (\[5.14\]) by the presence of terms with new, cubic symmetry. Indeed this is the case that one can easily check following the next steps: starting from the Hamiltonian (\[3.2\]) via the Stratonovich-Hubbard transformation one obtains the functional representation for the interaction part (which coincides with equation (\[5.8\])). An additional term $\re^{-\beta V \sum_i
S^i}$ can be represented in a form of the functional derivative $\re^{-\beta V \sum_i \partial^4 / \partial (\phi^i)^4}$, resulting in the following expression for the partition function of the cubic model (c.f. equation (\[5.11\])): $$\label{5.14a}
{\cal Z} \sim \int ({\rm d} \vec{\phi}) \re^{\frac{-1}{2}
\sum_{{\bf k}} \frac{1}{\beta \nu (k)} |\vec{\phi}_{{\bf k}}|^2}
\re^{-\beta V \sum_{\bf R} \sum_i \partial^4 / \partial
(\phi^i_{{\bf R}})^4} {\rm Sp} \, \re^{\sum_{{\bf
R}}\vec{S}_{{\bf R}} \vec{\phi}_{{\bf R}}}.$$ Taking trace in equation (\[5.14a\]) leads to the familiar expression (\[5.12\]). However, the derivative $\partial^4 /
\partial (\phi^i_{\bf R})^4$ gives rise to the contributions of cubic symmetry, the lowest order contribution being proportional to $v_0 \sum_i (\phi^i_{\bf R})^4$. The resulting effective Hamiltonian reads: $$\label{5.14b}
{\cal H}_{\rm eff} = \int {\rm d}^d R \left\{ \frac{1}{2}
\left((\nabla \phi)^2 + \mu_0^2 \phi^2\right) + \frac{u_0}{4!}
\phi^4 + \frac{v_0}{4!} \sum_{i=1}^m (\phi^i)^4 \right\},$$ and contains two bare couplings $u_0$ and $v_0$, corresponding to the $\phi^4$ terms of different symmetries. Coupling $u_0$ is positive, whereas the sign of the coupling $v_0$ coincides with that of $V$ in equation (\[3.2\]). At $v_0=0$ one recovers an effective Hamiltonian of the $m$-vector model (\[5.14\]).
Weakly diluted quenched $m$-vector model {#Vc}
----------------------------------------
The peculiarities of the free energy calculation for the model (\[3.3\]) consist in averaging over quenched disorder. Indeed, for each configuration of empty and occupied sites in (\[3.3\]) one can write a corresponding configuration-dependent partition function $Z_{{\rm conf}}$: $$\label{5.15}
{\cal Z}_{{\rm conf}} = {\rm Sp}_{\vec{S}} \re^{-\beta \cal{H}},$$ where ${\rm Sp}_{\vec{S}}$ concerns spin degrees of freedom and is defined by (\[5.3\]). The free energy is obtained as the configurational average: $$\label{5.16}
{\cal F}= -\beta^{-1} \langle \ln {\cal Z}_{{\rm
conf}}\rangle_{{\rm conf}}.$$ One of the ways to proceed is to make use of the replica trick [@replicas], which allows us to avoid integration of the logarithm in (\[5.16\]) substituting it by a power function: $$\label{5.17}
\ln {\cal Z} = \lim_{n \rightarrow 0} \frac{{\cal Z}^n -1}{n}\,.$$ Then $Z_{{\rm conf}}^n$ can be written as: $$\label{5.18}
{\cal Z}_{{\rm conf}}^n = {\rm Sp} \re^{\frac{\beta}{2}\sum_{\bf
R,R'} J(|{\bf R} - {\bf R'}|) \sum_{\alpha=1}^n
\vec{\sigma}^{\alpha}_{{\bf R}} \vec{\sigma}^{\alpha}_{{\bf R'}}}.$$ with obvious notations ${\rm Sp} (\dots) = \prod_{\alpha=1}^n {\rm
Sp}_{\vec{S}^{\alpha}} (\dots)$ and $\vec{\sigma}^{\alpha}_{{\bf
R}} \equiv c_{{\bf R}} \vec{S}^{\alpha}_{{\bf R}}$. Upcoming calculations closely follow lines of the subsection \[Va\]. Introducing by the Stratonovich-Hubbard transformation field variables $\vec{\phi}^{\alpha}_{{\bf R}}$, conjugated to $\vec{\sigma}^{\alpha}_{{\bf R}}$ one is able to take the trace over spin subsystem and is left with the configuration-dependent partition function: $$\label{5.19}
{\cal Z}_{\rm conf}^n \sim \int ({\rm d} \vec{\phi})
\re^{\frac{-1}{2} \sum_{{\bf k}} \frac{1}{\beta \nu
(k)}\sum_{\alpha=1}^n \vec{\phi}_{{\bf k}}^{\alpha}
\vec{\phi}_{-{\bf k}}^{\alpha} - \sum_{{\bf R}}
\sum_{l=1}^{\infty} \frac{u_{2l}}{(2l)!} \sum_{\alpha=1}^n
|\vec{\phi}^{\alpha}_{{\bf R}}|^{2l} c_{{\bf R}}}.$$ Note that $c_{{\bf R}}$ appears in (\[5.19\]) in the first power because from $c_{{\bf R}}=\{0,1\}$ it follows that $(c_{{\bf
R}})^{l} \equiv c_{{\bf R}}$. The last step is to perform configurational averaging of (\[5.19\]) with the distribution function (\[3.5\]). Let us represent the result of averaging in the exponential form $$\label{5.20}
\left\langle \re^{-\sum_{{\bf R}} \sum_{l=1}^{\infty}
\frac{u_{2l}}{(2l)!} \sum_{\alpha=1}^n |\vec{\phi}^{\alpha}_{{\bf
R}}|^{2l} c_{{\bf R}}} \right\rangle_{\rm conf} = \prod_{{\bf R}}
\re^{\sum_{p \geqslant 1} \frac{(-1)^p}{p!} \kappa_p(c) \Big (
\sum_{l=1}^{\infty} \frac{u_{2l}}{(2l)!} \sum_{\alpha=1}^n
|\vec{\phi}^{\alpha}_{{\bf R}}|^{2l} \Big )^p}\,,$$ where $\kappa_p(c)$ are cumulants of random variables $c_{{\bf
R}}$ and can be easily calculated for the random variable distribution (\[3.5\]): $$\begin{aligned}
\nonumber \kappa_1 &=& \sum_{c_{{\bf R}}=\{0,1\}} c_{{\bf R}}
p(c_{{\bf R}})=c,
\\ \label{5.21}
\kappa_2 &=& \sum_{c_{{\bf R}}=\{0,1\}} (c_{{\bf R}})^2 p(c_{{\bf
R}}) - \left(\sum_{c_{{\bf R}}=\{0,1\}} c_{{\bf R}} p(c_{{\bf
R}})\right)^2 =c(1-c), \dots .\end{aligned}$$ For the free energy one gets: $${\cal F}= -\beta^{-1} \langle \ln Z_{{\rm conf}}\rangle_{{\rm
conf}}= -\beta^{-1} \lim_{n \rightarrow 0} \left\{ \frac{c_1}{n}
\int({\rm d} \vec{\phi})^n \re^{-F[\phi]} -1/n \right\},
\label{5.22}$$ where the ($n$-dependent) coefficient $c_1$ can be recast explicitly following all the steps of calculations described above and the free energy functional $F[ \phi ]$ is given by: $$F[ \phi ] = \frac{1}{2}\sum_{{\bf k}} \sum_{\alpha=1}^n
\frac{1}{\beta \nu(k)} |\vec{\phi}^{\alpha}_{{\bf k}}|^2 - \sum_{p
\geqslant 1} \frac{(-1)^p}{p!} \kappa_p(c) \sum_{{\bf R}} \left(
\sum_{l=1}^{\infty} \frac{u_{2l}}{(2l)!} \sum_{\alpha=1}^n
|\vec{\phi}^{\alpha}_{{\bf R}}|^{2l} \right)^p. \label{5.23}$$ Again, as in the previous subsection, restricting ourselves to the $\phi^4$ terms, expanding the short-range interaction potential, re-scaling the fields and passing to the continuous limit one gets for the free energy: $$\label{5.24}
{\cal F} \sim \int({\rm d}\vec{\phi}) \re^{-{\cal H}_{\rm eff}},$$ now the proportionality sign also hides, besides the coefficient, the replica limit, as written explicitly in (\[5.22\]). The effective Hamiltonian reads: $$\label{5.25}
{\cal H}_{\rm eff} = \int {\rm d}^d R \left\{ \frac{1}{2}
\sum_{\alpha=1}^n \left[(\nabla \phi^\alpha)^2+\mu_0^2
(\phi^\alpha)^2\right] + \frac{u_0}{4!} \sum_{\alpha=1}^n
(\phi^\alpha)^4 + \frac{v_0}{4!} \sum_{\alpha,\beta=1}^n
(\phi^\alpha)^2 (\phi^\beta)^2 \right\}.$$ Here, the coupling $u_0$ is positive (being proportional to $cu_4$, (\[5.12\])) whereas the coupling $v_0$ is proportional to $c(c-1)u_2^2$ and is negative. The last term in (\[5.25\]) is present only for non-zero dilution: it is directly responsible for the effective interaction between replicas due to the presence of impurities.
Random anisotropy model {#Vd}
-----------------------
To treat the random anisotropy in the spin Hamiltonian (\[3.6\]), one first writes the configuration-dependent partition function for a fixed local anisotropy axes configuration $\{\hat x\}$: $$\label{5.26}
{\cal Z}_{\rm conf}(\{\hat x\})={\rm Sp}_{\vec{S}}\re^{-\beta
{\cal H}}.$$ As in equation (\[5.15\]), trace in equation (\[5.26\]) concerns only the spin degrees of freedom. Applying the Stratonovich-Hubbard transformation to the interaction part of the Hamiltonian one presents (\[5.26\]) in the form of the functional integral and is able to take the trace: $$\label{5.27}
{\cal Z}_{\rm conf}\left(\{\hat x\}\right)\sim\int(\rd\vec{\phi})
\re^{-{\cal H}(\hat x)},$$ with $$\label{5.28}
{\cal H}({\hat x}) = \int{\rm d}^d R \Big \{
\frac{1}{2}\left[(\nabla \phi)^2 + \mu_1^2 \phi^2\right]-D_1 (\phi
\hat x)^2+v_1\phi^4+ z_1 \phi^2(\phi \hat x)^2+ \cdots \Big \} .$$ Again, the expansion for small $k$ was performed and the continuous limit has been taken. Explicit values for the coefficients $\mu_1$, $D_1$, $v_1$, $z_1$ are given in [@Dudka05].
The rest of calculations follow the steps outlined in the section \[Vc\]: for the quenched disorder the free energy is defined by (\[5.16\]), where averaging is performed over the random axis distribution (given by (\[3.7\]) or (\[3.8\])). Subsequently, one substitutes $\ln {\cal Z}$ by a power function via the replica trick (\[5.17\]). For the isotropic random axis distribution (\[3.7\]) one gets the effective Hamiltonian [@Aharony75] containing three $\phi^4$ couplings of different symmetry: $$\label{5.29}
{\cal H}_{\rm eff} = \int {\rm d}^d R \left\{ \frac{1}{2}\left[
(\nabla \varphi)^2 + \mu_0^2 \varphi^2 \right]+u_0\varphi^4+
v_0\sum_{\alpha=1}^n (\phi^{\alpha})^4+
w_0\sum_{\alpha,\beta=1}^n\sum_{i,j=1}^m
\phi_i^{\alpha}\phi_j^{\alpha}\phi_i^{\beta}\phi_j^{\beta} \right
\}.$$ Here and below $\varphi^2=\sum_{\alpha=1}^n(\phi^\alpha)^2$. One can check the signs of the couplings: $u_0\sim \frac{D^2}{m^2(m+2)}>0$, $v_0\sim D>0$, $w_0\sim \frac{-D^2}{m(m+2)}<0$. Moreover, from the explicit expressions for the couplings $w_0$ and $u_0$ one gets for their ratio $w_0/u_0=-m$. The latter relation determines a region of physically allowed initial values in the $(u-v-w)$-space of couplings.
For the cubic distribution (\[3.8\]) the average over the random variables $\{\hat x\}$ leads to the effective Hamiltonian [@Aharony75] with four couplings: $$\begin{aligned}
\nonumber
{\cal H}_{\rm eff}&=& \int{\rm d}^d R \left\{
\vphantom{ w_0\sum_{i=1}^m\sum_{\alpha,\beta{=}1}^n
\left(\phi_i^{\alpha}\right)^2\left(\phi_i^{\beta}\right)^2}
\frac{1}{2}\left[(\nabla\varphi)^2 + \mu_0^2\varphi^2
\right]+u_0\varphi^4+v_0\sum_{\alpha=1}^n(\phi^{\alpha})^4\right.
\\ &&{}\left. \label{5.30}+
w_0\sum_{i=1}^m\sum_{\alpha,\beta{=}1}^n
\left(\phi_i^{\alpha}\right)^2\left(\phi_i^{\beta}\right)^2+
y_0\sum_{i=1}^m\sum_{\alpha=1}^n\left(\phi_i^{\alpha}\right)^4
\right\}.\end{aligned}$$ The signs of the couplings are as follows: $u_0 \sim
{D^2}/{(2m^2)}>0$, $v_0\sim D>0$, $w_0\sim {-D^2}/{(2m)}<0$. The last term in (\[5.30\]) is of cubic symmetry. It has to be included since it is generated if one further applies the RG transformation. Therefore $y_0$ can be of either sign. The symmetries of $w_0$ terms in (\[5.29\]) and (\[5.30\]) differ. However the ratio $w_0/u_0=-m$ holds for the effective Hamiltonian (\[5.30\]) as well.
Stacked triangular antiferromagnet {#Ve}
----------------------------------
It is an interaction potential (\[3.9\]) that makes a difference in a derivation of the effective Hamiltonian for the frustrated magnet and the regular one. One can find the details of the procedure to derive an effective Hamiltonian via the Stratonovich-Hubbard transformation in [@Kawamura88]. The main difference from the above described calculations for the $m$-vector magnet (subsection \[Va\]) arises when one proceeds with the Gaussian term in equation (\[5.12\]): $$\label{5.31}
\sum_{{\bf k}} \left(\frac{1}{\beta \nu(k)} - u_2\right)
|\vec{\phi}_{{\bf k}}|^2 \equiv \sum_{{\bf k}} d(k)
|\vec{\phi}_{{\bf k}}|^2.$$ The sum over ${\bf k}$ in (\[5.31\]) spans the 1st Brillouin zone. For the $m$-vector model, $d(k)$ has a minimum at $k=0$ and an expansion in $k$ around this minimum results in the effective Hamiltonian (\[5.14\]). Note, that one minimum in $d(k)$ leads to one field variable $\phi_{\bf k}$ in the effective Hamiltonian. For the frustrated model, the $d(k)$ has a maximum at $k=0$ and two independent minima at $k\neq 0$ in the 1st Brillouin zone. Subsequently, the zone can be rearranged into separate subzones and the field $\phi_{\bf k}$ over the original Brillouin zone can be decomposed into two fields, $\phi_{1,{\bf k}}$ and $\phi_{2,{\bf k}}$, with ${\bf k}$ confined to the new subzone. Now, the minima of $d(k)$ occur at the origin and pertain to the fields $\phi_{1}$ and $\phi_{2}$. The above decomposition results in the appearance of the terms of new symmetry in the effective Hamiltonian, which now reads: $$\label{5.32}
{\cal H}_{\rm eff} \!=\! \int{\rm d}^d R \left\{
\frac{1}{2}\left[\mu_0^2 (\phi_1^2+\phi_2^2) +(\nabla\phi_1)^2+
(\nabla\phi_2)^2 \right] +
\frac{u_0}{4!}\left[\phi_1^2+\phi_2^2\right]^2 +
\frac{v_0}{4!}\left[(\phi_1 \cdot \phi_2)^2- \phi_1^2\phi_2^2
\right] \right \},$$ with a scalar product of $m$-component fields $\phi_1 \cdot \phi_2
= \sum_{i=1}^m\phi_1^i \phi_2^i$. The coupling $u_0$ in (\[5.32\]) is positive, whereas the sign of the coupling $v_0$ determines a type of spin ordering: a non-collinear order occurs for $v_0>0$. For $v_0<0$ the fields $\phi_1$ and $\phi_2$ tend to allign parallel or antiparallel, this corresponds to the sinusoidal ordering or the linearly-polarized spin-density wave.
RG explanation of criticality in non-ideal systems {#VI}
==================================================
Now with the effective Hamiltonians (section \[V\]) and the RG method (section \[IV\]) at hand one can proceed further, applying the method in order to study the critical behaviour of the models. However, to be honest with the reader, he should be warned that the whole story is not that simple as shown in the section \[IV\]. It is not only because the very models we are interested in are much more complicated as compared to the 1d Ising model: in what follows below we will be interested in the criticality of 3d systems. The [*real space renormalization*]{} described in the section \[IV\] works for the low-dimensional systems, whereas for realistic 3d systems one rather applies the RG equations in the momentum space. Moreover, the very approach to RG as to the way of calculating the partition function of the system [@Wilson] has its alternative known as the field-theoretical RG [@RGbooks]. The latter is a tool to cope with the divergencies of correlation functions (vertex functions). Here, the RG procedure consists in the controlled rearrangement of the perturbation theory expansions giving rise to the RG equations. However, the underlying notions of both procedures are the same: given the effective Hamiltonian one applies the RG transformation and studies the flow equations for the couplings of the Hamiltonian. If the flow equations possess the fixed point (FP), which is stable and reachable from the initial conditions, it corresponds to the critical point of the system. Scaling arises in the vicinity of this FP and the universal values of the exponents governing scaling of different physical quantities may be found. Having this preamble in mind, we can make a brief review of what results for the RG flows, FPs and exponents of the non-ideal models have been found so far and how they differ from those of the “ideal” $m$-vector model.
$m$-vector magnet {#VIa}
-----------------
Put in a formal way, the question: does the 3d $m$-vector magnet exhibit a critical point (a Curie point for a ferromagnet or a Néel point for an antiferromagnet), transforms into the question: is there a stable and reachable FP in the flow equations (cf. (\[12.43\])–(\[12.45\])) for the couplings of its effective Hamiltonian? The flow equation for the coupling $u$ of the effective Hamiltonian (\[5.14\]) can be written in a general form of an ordinary first order differential equation: $$\label{6.1}
\frac{\rd u}{\rd\ln \ell}=\beta(u),$$ with the RG [*flow parameter*]{} $\ell$ and [*$\beta$-function*]{} $\beta(u)$. The parameter $\ell$ may serve to measure a distance to the critical point: $\ell\rightarrow 0$ corresponds to $T\rightarrow T_{\mathrm{c}}$, and a specific form of the $\beta$-function depends on a choice of the RG procedure. The data discussed in this chapter have been obtained within the field-theoretic RG [@RGbooks] approach at $d=3$. Giving numbers, we shall refer to the source, where they were obtained. However we shall try to be not too specific, mentioning several review papers for the interested reader.
(80,170) (20,20) (5,140) (170,140) (128,140) (95,140) (97,35) (140,35) (180,80) (85,-5) (220,20) (210,140) (330,140) (362,140) (373,140) (390,140) (390,103) (400,80) (325,-5)
To give an idea about the RG expansions and their treatment, we write down several first terms of the $\beta$-function (\[6.1\]) [@Kleinert91]: $$\label{6.2}
\beta(u)= -u(\varepsilon -u + 3u^2(3m+14)/(m+8)+ \cdots),$$ with $\varepsilon=4-d$. For a quantitative analysis, one can develop an [*$\varepsilon$-expansion*]{} looking for solutions of the FP equation $$\label{6.2a}
\beta(u^*)=0$$ in a form of a series $u^*=\sum_iu^{(i)}\varepsilon^i$. Alternatively, one can solve the non-linear equation (\[6.2a\]) directly at $d=3$ ($\varepsilon=1$). However, as it is well known by now, the perturbative RG expansions have zero radius of convergence and are asymptotic at best [@RGbooks] (cf. behaviour of the function (\[6.2\]), figure \[fig4\][**a**]{}). Special procedures of [*resummation*]{} have been elaborated to deal with them. We give an example of how one of these procedures works transforming a divergent series, figure \[fig4\][**a**]{}, into a convergent one, figure \[fig4\][**b**]{}. In the last figure, the function (\[6.2\]) has been resummed by the Padé-Borel resummation. The procedure consists of several steps. First, assuming the factorial growth of the coefficients $c_i$ in the expansion $\beta(u)=\sum_i c_iu^i$ (\[6.2\]), one constructs the Borel trransform of the initial function $\beta(u)$ via: $$\label{6.2b}
\beta^{\rm B}(u)= \sum_{i} \frac{c_iu^i}{i!}\,.$$ Then, the Borel transform is extrapolated by a Padé approximant $[K/L](u)$. The last is the ratio of two polynomials of order $K$ and $L$ such that its truncated Taylor expansion is equal to $\beta^{\rm B}(u)$. The resummed function is then calculated by an inverse Borel transform of this approximant: $$\label{6.2c}
\beta(u)=\int_0^{\infty}{\rm d}t \exp(-t)[K/L](ut).$$ Similar techniques are currently widely used in analyzing the RG expansions [@RGbooks; @Holovatch02]. In particular, the numerical estimates of different physical quantities given in this chapter (and in Table \[tab1\] as well) have been obtained using resummation techniques.
One sees the presence of two FPs in figure \[fig4\][**b**]{}: an unstable Gaussian FP [**G**]{} $u^*=0$ (the slope of the $\beta$-function is negative, $\partial \beta(u)/\partial u
|_{u^*} < 0$) and a stable one, $u^*\neq 0$. Here, the slope of the $\beta$-function is positive: starting form any initial conditions with $u>0$ the solution of the differential equation (\[6.1\]) reaches its FP value. This FP corresponds to the critical point $T_{\mathrm{c}}$ of the 3d Ising model. Similar behaviour of the $\beta$-function is found for other values of $m$, therefore the FP with $u\neq 0$ for general $m$ is called a $O(m)$-symmetric or Heisenberg FP [**H**]{}. We do not show the procedure of calculating the critical exponents: the other RG functions being evaluated at this FP bring about the asymptotic values of the critical exponents (in particular, those given in Table \[tab1\]). These exponents govern criticality of systems of different microscopic nature, which can be described by the effective Hamiltonian (\[5.14\]). It is said that these systems belong to the $O(m)$ [*universality class*]{}. In the RG picture, different microscopic origin is reflected in different initial conditions for the flow. However, the FP location and stability is defined solely by the global features: dimensionality, symmetry, interaction type.[^3] Systems which share the global features belong to the same universality class.
Cubic model {#VIb}
-----------
The main difference of the effective Hamiltonian of the cubic model (\[5.14b\]) as compared to that of the $m$-vector model (\[5.14\]) is that it contains one more coupling $v$ of different symmetry. Therefore, two $\beta$-functions describe the RG flow: $$\label{6.3}
\frac{\rd u}{\rd\ln \ell}=\beta_u(u,v), \hspace{3em}
\frac{\rd v}{\rd\ln \ell}=\beta_v(u,v).$$ Stability of a FP is now defined by the stability matrix composed of the $\beta$-function derivatives $\partial \beta_{u,v}/\partial
\{u,v\}$. The FP picture, which arises form an analysis of the $\beta$-functions at $d=3$ [@Aharony76] is schematically shown in figure \[fig5\]. Four FPs are obtained: unstable Gaussian [**G**]{} ($u=v=0$) and Ising [**I**]{} ($u=0, v\neq 0$) as well as Heisenberg [**H**]{} ($u\neq 0, v= 0$) and mixed [**M**]{} ($u\neq0,
v\neq0$). The stability of the two last FPs depends on the value of $m$.
(80,170) (20,10) (95,170) (92,142) (90,97) (162,97) (152,62) (180,85) (100,-5) (220,10) (295,170) (292,142) (290,97) (362,97) (359,117) (380,85) (305,-5)
Two different regimes for the RG flows are observed. For small $m<m_{\mathrm{c}}^{\rm cub}$ the FP [**H**]{} is stable. At this FP the system does not feel a presence of the cubic coupling, $v=0$. Therefore, for $m<m_{\mathrm{c}}^{\rm cub}$ the cubic model belongs to the $O(m)$ [*universality class*]{}: its exponents coincide with the exponents of the $m$-vector model, table \[tab1\]. However, with an increase of $m$ the FP [**M**]{} approaches [**H**]{} and at $m=m_{\mathrm{c}}^{\rm cub}$ both FPs coincide: a [*crossover*]{} to the new regime occurs. For $m>m_{\mathrm{c}}^{\rm cub}$ the FP [**M**]{} becomes stable and governs the critical properties of the cubic model in the new universality class. The [*marginal dimension*]{} value is slightly less than three: $m_{\mathrm{c}}^{\rm cub}=2.862(5)$ [@Folk00]. From this estimate it follows in particular, that a cubic Heisenberg ($m=3$) magnet does not belong to the $O(3)$ universality class. Its critical exponent being estimated in the fixed point [**M**]{} read [@Carmona00]: $$\label{6.4}
\gamma= 1.390(12), \quad
\nu= 0.706(6), \quad
\eta= 0.0333(26) ,\quad
\beta= 0.364(15), \quad
\alpha= -0.118(18).$$ Numerically, these values are close to their counterparts for the $m$-vector model (cf. Table \[tab1\]). However the principal difference arises from the above analysis: as one can easily check solving the system of differential equations (\[6.3\]) for $m>m_{\mathrm{c}}^{\rm cub}$ the RG flows with $v<0$ cannot reach the stable FP (an abscissa $v=0$ serves as a separatrix for the flows). As it follows from sections \[IIIb\], \[Vb\] negative $v$ corresponds to ordering along diagonals of $m$-dimensional hypercube. Therefore, the RG analysis results in a statement that such ordering cannot occur via a 2nd order phase transition: ferromagnetic crystals with three easy axes should undergo a 1st order phase transition. It is worth noting here, that whereas the presence of stable and reachable FP brings about the 2nd order phase transition, its absence signalling only that the 2nd order phase transition does not occur. The nature of the low-temperature phase and the scenario of how it is attained remains to be checked by other methods.
Random-site dilution {#VIc}
--------------------
Formally, an analysis of the weakly diluted quenched $m$-vector model resembles those we discussed in the former subsection \[VIb\]. Indeed, both effective Hamiltonians (\[5.14b\]), (\[5.25\]) contain two couplings of different symmetry and give rise to the already familiar FP picture. Moreover, the effective Hamiltonian (\[5.25\]) at $m=1$ coincides with (\[5.14b\]).[^4] However, our goal is to analyze it in the replica limit $n=0$. Note, that now the physically meaningful values of couplings are $u>0$, $v<0$, see section \[Vc\]. The FP picture and the RG flows at $d=3$ are shown in the figure \[fig6\] [@Folk03]. Besides the familiar FPs [**G**]{}, [**H**]{}, and [**M**]{} a polymer FP [**P**]{} is present. It is stable and corresponds to the $O(m=0)$ universality class, however it is never reached from the initial conditions $u>0$, $v<0$.
(80,170) (20,10) (95,170) (92,145) (90,97) (162,97) (152,62) (180,85) (100,-5) (220,10) (295,170) (292,145) (290,97) (362,97) (356,123) (329,88) (340,85) (359,67) (380,85) (305,-5)
Again, the new marginal dimension $m_{\mathrm{c}}^{\rm dil}$ governs the crossover between the new and the $O(m)$ universality classes: FP [**M**]{} is stable for $m<m_{\mathrm{c}}^{\rm dil}$. A search for the value of $m_{\mathrm{c}}^{\rm dil}$ shows, that at $m=m_{\mathrm{c}}^{\rm dil}$ the heat capacity critical exponent of an undiluted system changes its sign: $\alpha(m_{\mathrm{c}}^{\rm dil})=0$. In this way one recovers the Harris criterion [@Harris74], section \[IIIc1\], translated into the RG “language”. The numerical value of $m_{\mathrm{c}}^{\rm dil}$ being slightly less than two, $m_{\mathrm{c}}^{\rm dil}=1.912(4)$ [@Holovatch01], only the Ising model ($m=1$) changes its exponents upon dilution. Indeed, the numerical values of the exponents read [@Pelissetto00]: $$\label{6.5}
\gamma= 1.330(17), \quad
\nu= 0.678(10), \quad
\eta= 0.030(3),\quad
\beta= 0.349(5),\quad
\alpha= -0.034(30)$$ and differ essentially from those of the 3d Ising model (table \[tab1\]).
However, both in the experiments and in the MC simulations one deals with the system not yet in an asymptotic region, where exponents do not attain their FP values and the [*effective exponents*]{} are found. Being non-universal, they can be calculated in the RG treatment as functions of the flow-dependent couplings. In figure \[fig7\], we show an effective critical exponent $\gamma_{\rm eff}$ measured recently for the ac susceptibility of the a-${\rm Fe_{86}Mn_4Zr_{10}}$ amorphous alloy [@Perumal03] and compare them with the theoretical RG calculations of the effective exponents of the weakly diluted quenched $m=3$ model. The exponent was calculated along different RG flows labeled by numbers in figure \[fig6\][**b**]{}. This calculation serves an example how the non-asymptotic effects may be taken into account in the RG analysis. Although direct correspondence between the temperature distance to the critical point $\tau$ and the RG flow parameter $\ell$ is problematic, the RG serves as a useful tool of accompanying the studies of effective critical behaviour.
(80,170) (20,10) (25,170) (195,17) (100,-5) (220,4) (205,170) (390,17) (329,91) (329,115) (329,140) (305,-5)
Random anisotropy {#VId}
-----------------
On this example we shall show how the RG predicts two different phenomena occuring in the random anisotropy magnets. It appears that the type of local random axis distribution crucially effects an origin of the low-temperature phase in random anisotropy systems [@ramreviews]. First, we consider the results obtained for an isotropic distribution (\[3.7\]), which leads to the effective Hamiltonian (\[5.29\]) [@Aharony75]. It contains three couplings, $u,v,w$, hence three $\beta$-functions define the RG flows. Solving the FP equations one arrives at the FP picture shown in figure \[fig8\][**a**]{} [@Aharony75; @Dudka05].
![Fixed points of the random anisotropy model. The filled boxes show the stable fixed points, the cross denotes typical initial values of couplings. (a): isotropic local random axis distribution. The stable FP [**P**]{} cannot be reached by the RG flow which starts from the region shown by a cross in the figure. (b): cubic local random axis distribution. The stable FP [**P**]{} cannot be reached, but the random Ising FP [**M**]{} is stable and reachable for the RG flow.[]{data-label="fig8"}](isotr1.eps "fig:"){width="48.00000%"}![Fixed points of the random anisotropy model. The filled boxes show the stable fixed points, the cross denotes typical initial values of couplings. (a): isotropic local random axis distribution. The stable FP [**P**]{} cannot be reached by the RG flow which starts from the region shown by a cross in the figure. (b): cubic local random axis distribution. The stable FP [**P**]{} cannot be reached, but the random Ising FP [**M**]{} is stable and reachable for the RG flow.[]{data-label="fig8"}](cubic.eps "fig:"){width="48.00000%"}\
Let us recall (cf. section \[Vd\]), that physically meaningful values of couplings are $u>0,\linebreak v>0,w<0$. Therefore only the FPs located in this region are shown in the figure. However, there is another condition found for the ratio of couplings: $w_0/u_0=-m$. The region of typical initial conditions to study the RG flow is shown in figure \[fig8\] by a cross. The only stable FP found, a polymer FP [**P**]{} is not reachable from the initial conditions. The [*run-away*]{} solutions of the RG equations bring about an absence of a 2nd order phase transition.
A different picture is obtained for a cubic random axis distribution (\[3.8\]) [@Dudka05]. Here, the effective Hamiltonian (\[5.30\]) contains four couplings of different symmetry, $u$, $v$, $w$, $y$. The physical initial values for the couplings lay in the region (section \[Vd\]): $u>0,v>0,w<0,w/u=-m$. A typical FP picture is shown in figure \[fig8\][**b**]{} for $v=0$. Similar to the former case of isotropic random axis distribution, the stable FP [**P**]{} cannot be reached. However, one more stable FP [**M**]{} is present. It is reachable for the RG flow that starts from the initial conditions marked by a cross in the figure. This FP is a FP of the random-site Ising model (subsection \[VIc\]) for any value of $m$. It means that the ferromagnetic 2nd order phase transition in the $m$-vector magnet with the cubic random axis distribution belongs to the universality class of the random-site Ising model and is governed in asymptotics by the exponents (\[6.5\]).
Stacked triangular antiferromagnet {#VIe}
----------------------------------
Again, as in the former subsections \[VIa\]–\[VId\], the RG answer about a possibility of a 2nd order phase transition in the 3d stacked triangular antiferromagnet would be a presence of a stable accessible FP for the couplings $u,v$ of the effective Hamiltonian (\[5.32\]). The model possesses a rather complicated FP structure, sketched in figure \[fig9\].
(80,390) (20,215) (85,200) (95,375) (180,290) (90,302) (162,302) (130,355) (167,330) (220,215) (295,375) (380,290) (290,302) (362,302) (260,200) (20,15) (95,175) (90,102) (162,102) (152,67) (130,45) (180,90) (62,0) (220,15) (295,175) (290,102) (362,102) (356,128) (330,45) (380,90) (285,0)
FP picture changes with $m$ and one finds three marginal dimensions $m_i^{\rm chir}$ that govern its topology. For large $m>m_3^{\rm chir}$ the stable chiral FP [**C**]{}$_+$ is present and it can be reached from the initial values of the couplings $u,v\geqslant 0$. Therefore, the phase transition to the non-collinear chiral state is of the second order. At $m=m_3^{\rm
chir}$, [**C**]{}$_+$ merges with the anti-chiral FP [**C**]{}$_-$ and disappears, only one unstable Heisenberg FP [**H**]{} is found for $m_3^{\rm chir}>m>m_2^{\rm chir}$. With further decrease of $m$, two more FPs appear at $m=m_2^{\rm chir}$, [**S**]{}$_-$ and [**S**]{}$_+$. The last one is stable, however it describes the sinusoidal phase, which occurs for $v<0$. Finally, at $m=m_1^{\rm
chir}$ this FP merges with [**H**]{} and looses its stability with further increase of $m$. It follows from this analysis, that the 2nd order phase transition into non-collinear phase can occur only in the magnets with $m>m_3^{\rm chir}$.
In principle, the above described FP picture has been known since the first RG studies of the problem [@Kawamura88]. However, reliable numerical estimates for marginal dimensions $m_i^{\rm
chir}$ have been obtained only recently. The question of interest is, what is the order of the phase transition at $m=2;3$ when the model has its physical realization? The estimates: $m_3^{\rm
chir}=6.23(21)$, $m_2^{\rm chir}=1.99(4)$, $m_1^{\rm
chir}=1.43(2)$ [@Holovatch04] clearly rule out the possibility of a 2nd order phase transition for $m=2;3$: the FP picture is shown in the second figure of figure \[fig9\] and no accessible FP is found.
Note, however, certain controversy in the RG studies performed so far: whereas the above FP picture is confirmed by the perturbative RG expansions and the non-perturbative RG approach [@Delamotte04; @Holovatch04], an analysis of the FP equations by direct solution of the non-linear FP equations for the resummed $\beta$-functions brings about a presence of the stable reachable FP for $m=2;3$ [@Pelissetto01]. The last is associated with the critical point of the 2nd order phase transition.
Conclusions and outlook {#VII}
=======================
How do the changes in structure effect the critical behaviour of the matter? We tried to give an answer to this question taking a 3d $m$-vector model as an ideal reference system and showing what will happen to it under an effect of different non-idealities. The examples considered include anisotropy, structural disorder, frustrations: the features one often encounters dealing with realistic condensed matter objects. The response of a system to such non-idealities appears to be very different, ranging from an insensitivity (the cubic $m=2$ magnet remains in the $O(2)$ universality class), through softening (the heat capacity of the random site $m=1$ magnet does not diverge) to disappearance of a 2nd order phase transition (isotropically distributed local random axis destroys long-range order). These various asymptotic features are accompanied by a complicated non-asymptotic effective critical behaviour.
It is astonishing that all this bunch of phenomena can be explained and accurately described within one theoretical framework, the RG approach. An application of basic RG notions of flows, fixed points and their accessibility, marginal dimensions and crossovers supported by an elaborate machinery to perform and analyze the RG transformation resulted in a coherent picture of phenomena in the vicinity of a critical point. A lot remains to be cleared up in this picture. Maybe a participant of the school or a reader of these lectures will decide to make his or her contribution? Good luck!
Acknowledgements {#A .unnumbered}
================
I am grateful to Bertrand Berche, Arnaldo Donoso, and Ricardo Paredes for the invitation to lecture at the Spring school on Foundations of statistical and mesoscopic physics (Mochima, Venezuela, June 20th–24th 2005) and to all participants of the school for the wonderful atmosphere created there. Bertrand Berche is further acknowledged for his encouragement, advice during preparation of the manuscript and, last but not least, for his stories about the Cagniard de la Tour state, the critical state discovered as early as in 1822 [@history]! I thank my colleagues Viktoria Blavats’ka, Bertrand Delamotte, Maxym Dudka, Christian von Ferber, Reinhard Folk, Dmytro Ivaneiko, Taras Yavors’kii – some of the results mentioned in the last part of these lectures are due to our common work.
This work was supported by Austrian Fonds zur Förderung der wissenschaftlichen Forschung under Project No. P16574.
[99]{}
A historical overview of the experimental and theoretical studies of critical phenomena may be found in J.S. Rowlinson introductary essay to the English edition of J.D. van der Waals Leiden thesis of 1873: J.D. van der Waals: On the Continuity of the Gaseous and Liquid States. Edited with an Introductory Essay by J.S. Rowlinson. North-Holland Physics Publishing, Amsterdam, 1988; and in: C. Domb. The Critical Point. Taylor & Francis, London, 1996.
J.E. Verschaffelt. Koninklijke Akademie van Wetenschappen, Proc. Sect. Scienc., 1990, [**II**]{}, 588. English version: http://www.knaw.nl/waals/pdf/Verschaffelt01.pdf
Wilson K.G., Kogut J., Phys. Rep., 1974, [**12**]{}, 75.
The field theoretical RG in its applications to the phase transition theory is explained in: E. Brézin, J. C. Le Guillou, and J. Zinn-Justin, in: Phase Transitions and Critical Phenomena, edited by C. Domb and M. S. Green, Vol. 6, Academic Press, London, 1976; Amit D.J. Field Theory, the Renormalization Group, and Critical Phenomena. World Scientific, Singapore, 1989; Zinn-Justin J. Quantum Field Theory and Critical Phenomena. Oxford University Press, Oxford, 1996.
Ising E., Z. Phys., 1925, [**31**]{}, 255.
Stanley H.E., Phys. Rev. Lett., 1968, [**20**]{}, 589.
Mermin N.D., Wagner H., Phys. Rev. Lett., 1966, [**17**]{}, 1133; Erratum: [*ibid.*]{}, 1966, [**17**]{}, 1307; Hohenberg P.C., Phys. Rev., 1967, [**158**]{}, 383.
Onsager L., Phys. Rev., 1944, [**65**]{}, 117.
Guida R., Zinn-Justin J., J. Phys. A, 1998, [**31**]{}, 8103.
Aharony A. Phase Transitions and Critical Phenomena, edited by C. Domb and M. S. Green, Vol.6. Academic Press, London, 1976.
Brout R., Phys. Rev., 1959, [**115**]{}, 824.
Belanger D.P., Young A.P., J. Magn. Magn. Mater., 1991, [**100**]{}, 272.
A review of the random 3d Ising model critical behaviour may be found in: Folk R., Holovatch Yu., Yavors’kii T., Physics – Uspekhi, 2003, [**46**]{}, 169 \[Uspekhi Fizicheskikh Nauk, [**173**]{}, 175\] and in [@Pelissetto02].
A review on the critical behaviour of systems with complex effective Hamiltonians: Pelissetto A., Vicari E., Phys. Rep., 2002, [**368**]{}, 549.
Dudka M., Folk R., Holovatch Yu., Ivaneiko D., J. Magn. Magn. Mater., 2003, [**256**]{}, 243.
Yoon J., Chan M.H.W., Phys. Rev. Lett., 1997, [**78**]{}, 4801.
Harris A.B., J. Phys. C, 1974, [**7**]{}, 1671.
Harris R., Plischke M., Zuckermann M.J., Phys. Rev. Lett., 1973, [**31**]{}, 160.
A review of early work on random anisotropy magnets may be found in: Cochrane R.W., Harris R., Zuckermann M.J., Phys. Rep., 1978, [**48**]{}, 1. Recent experimental, numerical, and theoretical studies are reviewed in [@Dudka05].
Dudka M., Folk R., Holovatch Yu., Journ. Mag. Mag. Mat., 2005, [**294**]{}, 305.
Aharony A., Phys. Rev. B, 1975, [**12**]{} 1038.
Kawamura H., Phys. Rev. B, 1988, [**38**]{} 4916.
A review on frustrated magnets with an emphasis on the results obtained within the non-perturbative RG approach: Delamotte B., Mouhanna D., Tissier M., Phys. Rev. B, 2004, [**69**]{}, 134413.
Fisher M.E., in: Critical Phenomena. Proceedings of the Summer School, Stellenbosch, South Africa. Ed. by F. J. W. Hahne. Lecture Notes in Physics, [**186**]{}, Springer Verlag, Berlin, 1983; D. R. Nelson, Fisher M.E., Ann. Phys. (N.Y.), 1975, [**91**]{}, 226.
Emery V.J., Phys. Rev. B, 1975, [**11**]{}, 239; Edwards S.F., Anderson P.W., J. Phys. F, 1975, [**5**]{}, 965.
Kleinert H., Neu J., Schulte-Frohlinde V., Chetyrkin K.G., Larin S.A., Phys. Lett. B, 1991, [**272**]{}, 39; Erratum: [*ibid.*]{}, 1993, [**319**]{} 545.
A review on application of resummation technique to divergent RG expansions of disordered models: Holovatch Yu., Blavats’ka V., Dudka M., von Ferber C., Folk R., Yavors’kii T., Int. J. Mod. Phys. B, 2002, [**16**]{}, 4027.
Folk R., Holovatch Yu., Yavors’kii T., Phys. Rev. B, 2000, [**62**]{}, 12195.
Carmona J.M., Pelissetto A., Vicari E., Phys. Rev. B, 2000, [**61**]{}, 15136.
Holovatch Yu., Dudka M., Yavors’kii T., J. Phys. Stud., 2001, [**5**]{}, 233.
Pelissetto A., Vicari E., Phys. Rev. B, 2000, [**62**]{}, 6393.
Perumal A., Srivinas V., Rao V.V., Dunlap R.A., Phys. Rev. Lett., 2003, [**91**]{}, 137202.
Holovatch Yu., Ivaneiko D., Delamotte B., J. Phys. A, 2004, [**37**]{}, 3569.
Pelissetto A., Rossi P., Vicari E., Phys. Rev. B, 2001, [**63**]{}, 140414.
Questions and answers {#questions-and-answers .unnumbered}
=====================
- [*(Alexander López)*]{}: How do the signs of the couplings $u$ and $v$ in the effective Hamiltonian (\[5.14b\]) determine a type of the low-temperature ordering?
- Neglecting fluctuations (taking a function $\phi(r)$ to be just a variable $\phi$) you can think about the effective Hamiltonian as of the Landau free energy. Now, let the reference system display a 2nd order phase transition. This means that $u>0$. Minimizing Landau free energy and looking for the spontaneous magnetization one finds at $T<T_{\mathrm{c}}$ two different non-trivial solutions: $\vec{\phi}=(\phi/\sqrt{m},
\dots, \phi/\sqrt{m})$, it exists at $v<0$, and $\vec{\phi}=(\phi,
0, \dots, 0)$ at $v>0$. They correspond to two types of ordering: along the diagonals or along the edges of a $m$-dimensional hypercube. For $m=3$, these are directions \[111\] and \[100\] correspondingly.
- [*(Bertrand Berche)*]{}: What are experimental realizations of the cubic model?
- I have already mentioned ferromagnetic crystals. Besides, at $m=3$ the model describes a ferroelectric phase transition which occurs in ${\rm SrTiO_3}$ at 105 K ( Cowley R.A., Bruce A.D., J. Phys. C, 1973, [**6**]{}, L191). Moreover, since the model provides an example of a system with an arbitrary weak first-order phase transition, it is also used as a testing ground to describe an elecroweak transition in the early Universe (P. Arnold, S Sharpe, L. Yaffe, Y. Zhang, Phys. Rev. Lett., 1997, [**78**]{}, 2062).
- [*(Carlos Vásquez)*]{}: In 1983, Weinrib and Halperin proposed a model to describe an effect of extended (correlated) disorder on magnetic 2nd order phase transition. There, the impurity-impurity correlation function decays for large separations as $g(r)\sim r^{-a}$. Currently, there exist two different predictions for the critical exponents of such a model. What are the methods used to obtain them?
- Indeed, the original result of Weinrib and Halperin ( A. Weinrib, Halperin B.I., Phys. Rev. B, 1983, [**27**]{}, 417) was obtained in the first order of the expansion in $\varepsilon=4-d$, $\delta=4-a$. Recent two-loop estimates (V. V. Prudnikov, Prudnikov P.V., Fedorenko A.A., J. Phys. A, 1999, [**32**]{}, L399) are due to the fixed $d$, $a$ RG technique. Qualitative answer of both approaches is that for $a<d$ the disorder is relevant if the correlation length critical exponent of the system without defects obeys $\nu<2/a$.
- [*(Bertrand Berche)*]{}: Could you comment on logarithmic corrections to the scaling laws? Is the critical behaviour at marginal dimensions you were speaking about (e.g. at $m_{\mathrm{c}}^{\rm dil}\simeq1.91$ for the 3d diluted $m$-vector magnet) governed by such corrections?
- Logarithmic corrections arise at upper critical dimension $d^{\rm up}$: that is, at the [*space*]{} dimension, above which the mean-field theory holds[^5]. For the models I was speaking about, the upper critical dimension is four. Indeed, the divergence of an isothermal susceptibility of the $m$-vector model at $d=4$ is governed by the mean field exponent $\gamma=1$. However, the power law singularity is accompanied by a logarithmic one: $$\label{7.1}
\chi^{-1} \sim|\tau|^{-1}{\ln |\tau|}^{-(m+2)/(m+8)},$$ as first derived in: Larkin A.I., Khmelnitskii D.E., JETP, 1969, [**29**]{}, 1123. In the RG scheme, one can see the origin of such corrections solving the flow equation (\[6.1\]) at $d=4$ ($\varepsilon=0$): $$\label{7.2}
\frac{\rd
u}{\rd\ln \ell}=-u^2,$$ here the right-hand side is the leading term of the $\beta$-function (\[6.2\]). The solution of equation (\[7.2\]) $$\label{7.3}
u=\frac{1}{|\ln \ell|} \hspace{1em} + \hspace{1em} {\rm const},
\hspace{2em} \ell \rightarrow 0$$ being substituted into an appropriate expression for the susceptibility leads to the above dependence (\[7.1\]). Now, let us return to the marginal dimensions $m_{\mathrm{c}}$ we were discussing in these lectures. They are [*field*]{} dimensions and we estimated them at [*space*]{} dimension $d=3$. As far as the space dimension was lower than $d^{\rm up}=4$, the logarithmic corrections do not appear. Returning back to the example given by formulas (\[7.2\]), (\[7.3\]): even at $m=m_{\mathrm{c}}$, the first power of couplings is present in the right-hand side of functions (\[7.2\]) for $d=3$ and the solutions will rather behave as $u \sim \ell^{\rm const}$.
- [*(Dragi Karevski)*]{}: Do you know other citeria, similar to the Harris one, which predict changes in the critical behaviour caused by different types of disorder? For example, what happens when disorder is coupled to the order parameter?
- Indeed, Harris criterion concerns the systems, where disorder is coupled to the energy density (look for example at the Hamiltonian (\[3.3\]), where random variables $c_{\bf R}$ are coupled to the product of spins). It states that critical exponents of a disordered system do not change, if the heat capacity of the pure system does not diverge. Later, a statement that the correlation length critical exponent of $d$-dimensional systems with such type of disorder should obey an inequality $\nu>2/d$ was proven ( Chayes J.T., L. Chayes, Fisher D.S., T. Spenser, Phys. Rev. Lett., 1986, [**57**]{}, 2999). For the extended, long-range correlated disorder the generalized Harris criterion holds (see the above mentioned paper of Weinrib and Halperin as well as D. Boyanovsky, Cardy J.L., Phys. Rev. B, 1983, [**27**]{}, 6971).
Y. Imry and S.-k Ma have shown (Phys. Rev. Lett., 1975, [**35**]{}, 1399) that even arbitrary weak disorder coupled to the order parameter of continuous symmetry (i.e. for $m\geqslant 2$ vector model) destroys ferromagnetism at $d<4$. For the random-field Ising model ($m=1$) the lower critical dimension is $d=2$. As I already have mentioned in the lectures, the isotropically distributed random axis destroys ferromagnetism at $d<4$. One of the ways of showing this is to exploit the arguments similar to those of Imry and Ma for the random-field systems (Pelcovits R.A., Pytte E., Rudnick J., Phys. Rev. Lett., 1978, [**40**]{}, 476).
Last but not least, let me mention the Luck criterion for the connectivity disorder, as explained in the lecture by Wolfhard Janke.
\[last@page\]
[^1]: I thank Reinhard Folk for pointing me the work of Jules-Émile Verschaffelt.
[^2]: Another option would be to consider a random field disorder [@Belanger91]. However, we do not introduce it since an appropriate model differs from those considered here by its upper critical dimension and this will make our account even broader. The same concerns strong dilution at the percolation threshold.
[^3]: In particular, here we consider systems with the short-range interaction.
[^4]: One can check it by further substitution $n\rightarrow m$, $\{u_0,v_0\} \rightarrow \{v_0,u_0\}$.
[^5]: Note added in proof: We do not discuss here the logarithmic corrections appearing at low dimensions as those in 2d diluted Ising model or 2d $q=4$ Potts model. For a list of systems where logarithmic corrections appear and for the scaling relations between them see: Kenna R., Johnston D.A., Janke W., Phys. Rev. Lett., 2006, **96**, 115701.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Upper limits for the mass-radius ratio and total charge are derived for stable charged general relativistic matter distributions. For charged compact objects the mass-radius ratio exceeds the value 4/9 corresponding to neutral stars. General restrictions for the red shift and total energy (including the gravitational contribution) are also obtained.'
address:
- |
Department of Physics, The University of Science and Technology,\
Clear Water Bay, Hong Kong, P. R. China
- 'Department of Physics, The University of Hong Kong, Pokfulam Road, Hong Kong, P. R. China'
author:
- 'M.K. Mak[^1] and Peter N. Dobson, Jr.[^2]'
- 'T. Harko[^3]'
title: 'Maximum Mass-Radius Ratios for Charged Compact General Relativistic Objects'
---
Pacs Numbers: 04.20.-q, 97.60.-s
Introduction
============
It is generally accepted today that black holes are uniquely characterized by their total mass-energy $E$, charge $Q$ and angular momentum $J$ [@Ch99]. Most of the investigations of the astrophysical objects have been done under the assumption of the electric charge neutrality of the stellar matter. However, as a result of accretion neutron stars can acquire a net charge, if accretion produces luminosity close to the Eddington limit $L_{E}$ [@Sh71]. Let us consider a star of mass $M$ that undergoes spherical accretion and assume, for simplicity, that the accreting material is ionized hydrogen. If the accreting luminosity is $L$, the infalling electrons experience a radiative force $F_{R}=\frac{\sigma _{T}L}{4\pi cr^{2}}$, where $\sigma _{T}$ is the Thomson cross section. Since the radiation drag acting on the protons is a factor $\left( m_{e}/m_{p}\right) ^{2}$ smaller, electrons and protons are subject to different accelerations, and the star acquires a net positive charge $Q=\frac{GMm_{p}}{e}\frac{L}{L_{E}}$ [@Tu97], where $%
L_{E}=4\pi GMm_{p}c/\sigma _{T}$ is the Eddington luminosity. The astrophysical conditions under which this phenomenon can take place are rather extreme but in principle they could lead to a charged astrophysical configuration. This mechanism has been recently proposed, via vacuum breakdown near a charged black hole, as a source of $\gamma $-ray bursts [@Pr98]. A phase transition of neutron matter to quark matter at zero temperature or temperatures small compared to degeneracy temperature allows the existence of hybrid stars, i.e. stars having a quark core and a crust of neutron matter [@Wi84]. In fact, quark matter with electrically charged constituents rather than neutron matter could hold the large magnetic field of the pulsars [@Kh95] and hence it is possible that for strange-matter made stars the effects of the non-zero electrical charge be important.
By using the static spherically symmetric gravitational field equations Buchdahl [@Bu59] has obtained an absolute constraint of the maximally allowable mass $M$- radius $R$ for isotropic fluid spheres of the form $\frac{2M}{R}<\frac{8}{9}$ (where natural units $c=G=1$ have been used).
It is the purpose of the present Letter to obtain the maximum allowable mass -radius ratio in the case of stable charged compact general relativistic objects. This is achieved by generalizing to the charged case the method described for neutral stars in Buchdahl [@Bu59] and Straumann [@St84].
Maximum mass-radius ratio for charged general relativistic compact objects
==========================================================================
For a static general relativistic spherically symmetric configuration the interior line element is given by $ds^{2}=e^{\nu }dt^{2}-e^{\lambda }dr^{2}-r^{2}\left( d\theta ^{2}+\sin
^{2}\theta d\varphi ^{2}\right) $.
The properties of the charged compact object can be completely described by the structure equations, which are given by [@Be71]: $$\label{1}
\frac{dm}{dr}=4\pi \rho r^{2}+\frac{Q}{r}\frac{dQ}{dr},$$ $$\label{2}
\frac{dp}{dr}=-\frac{\left( \rho +p\right) \left( m+4\pi r^{3}p-\frac{%
Q^{2}}{r}\right) }{r^{2}\left( 1-\frac{2m}{r}+\frac{Q^{2}}{r^{2}}\right) }+%
\frac{Q}{4\pi r^{4}}\frac{dQ}{dr},$$ $$\label{3}
\frac{d\nu }{dr}=\frac{2\left( m+4\pi r^{3}p-\frac{Q^{2}}{r}\right) }{%
r^{2}\left( 1-\frac{2m}{r}+\frac{Q^{2}}{r^{2}}\right) },$$ where $\rho $ is the energy density of the matter, $p$ is the thermodynamic pressure, $m(r)$ is the mass and $Q(r)=4\pi \int_{0}^{r}e^{\frac{\nu +\lambda }{2}%
}r^{2}j^{0}dr$ is the charge inside radius $r$, respectively. The electric current inside the star is given by $j^{i}=\left(
j^{0},0,0,0\right) $ .
The structure equations (\[1\])-(\[3\]) must be considered together with the boundary conditions $p(R)=0$, $p(0)=p_{c}$ and $\rho (0)=\rho _{c}$, where $\rho _{c}$, $p_{c}$ are the central density and pressure, respectively.
With the use of Eqs. (\[1\])-(\[3\]) it is easy to show that the function $\zeta =e^{\frac{\nu }{2}}>0,\forall r\in \lbrack 0,R\rbrack $ obeys the equation $$\label{4}
\sqrt{1-\frac{2m}{r}+\frac{Q^{2}}{r^{2}}}\frac{1}{r}\frac{d}{dr}\left[
\sqrt{1-\frac{2m}{r}+\frac{Q^{2}}{r^{2}}}\frac{1}{r}\frac{d\zeta }{dr}%
\right] =\frac{\zeta }{r}\left[ \frac{d}{dr}\frac{m}{r^{3}}+\frac{Q^{2}}{%
r^{5}}\right].$$
For $Q=0$ we obtain the equation considered in [@St84]. Since the density $\rho $ does not increase with increasing $r$, the mean density of the matter $<\rho >=\frac{3m}{4\pi r^{3}}$ inside radius $r$ does not increase either. Therefore we assume that inside a compact general relativistic object the condition $\frac{d}{dr}\frac{m}{r^{3}}<0$ holds independently of the equation of state of dense matter. By defining a new function $$\label{5}
\eta (r)=\int_{0}^{r}\frac{r^{\prime }}{\sqrt{1-\frac{2m\left( r^{\prime
}\right) }{r^{\prime }}+\frac{Q^{2}\left( r^{\prime }\right) }{r^{\prime 2}}}%
}\left[ \int_{0}^{r^{\prime }}\frac{Q^{2}\left( r^{\prime \prime }\right)
\zeta \left( r^{\prime \prime }\right) }{r^{\prime \prime 5}\sqrt{1-\frac{%
2m\left( r^{\prime \prime }\right) }{r^{\prime \prime }}+\frac{Q^{2}\left( r^{\prime
\prime }\right) }{r^{\prime \prime 2}}}}dr^{\prime \prime }\right]
dr^{\prime },$$ denoting $\Psi=\zeta -\eta $, and introducing a new independent variable $\xi =\int_{0}^{r}r^{\prime
}\left( 1-\frac{2m(r^{\prime })}{r^{\prime }}+\frac{Q^{2}\left( r^{\prime
}\right) }{r^{\prime 2}}\right) ^{-\frac{1}{2}}dr^{\prime }$ [@St84], from Eq.(\[5\]) we obtain the basic result that inside all stable stellar type charged general relativistic matter distributions the condition $\frac{d^{2}\Psi }{d\xi ^{2}}<0$ must hold for all $r\in \left[ 0,R\right] $. Using the mean value theorem we conclude $\frac{d\Psi }{d\xi }\leq \frac{%
\Psi \left( \xi \right) -\Psi (0)}{\xi }$, or, taking into account that $%
\Psi (0)>0$ it follows that, $$\label{6}
\Psi ^{-1}\frac{d\Psi }{d\xi }\leq \frac{1}{\xi }.$$
In the initial variables the inequality (\[6\]) takes the form $$\begin{aligned}
\label{7}
&\frac{1}{r}\left( 1-\frac{2m(r)}{r}+\frac{Q^{2}(r)}{r^{2}}\right) ^{\frac{1%
}{2}}\left[ \frac{1}{2}\frac{d\nu }{dr}e^{\frac{\nu (r)}{2}}-\frac{r}{\sqrt{%
1-\frac{2m}{r}+\frac{Q^{2}}{r^{2}}}}\int_{0}^{r}\frac{Q^{2}\left( r^{\prime
}\right) e^{\frac{\nu \left( r^{\prime }\right) }{2}}}{r^{\prime 5}\sqrt{1-%
\frac{2m\left( r^{\prime }\right) }{r^{\prime }}+\frac{Q^{2}\left( r^{\prime
}\right) }{r^{\prime 2}}}}dr^{\prime }\right] \leq \nonumber\\
&\frac{e^{\frac{\nu (r)}{2}}-\int_{0}^{r}r^{\prime }\left( 1-\frac{2m\left(
r^{\prime }\right) }{r^{\prime }}+\frac{Q^{2}\left( r^{\prime }\right) }{%
r^{\prime 2}}\right) ^{-\frac{1}{2}}\left[ \int_{0}^{r^{\prime }}\left( 1-%
\frac{2m\left( r^{\prime \prime }\right) }{r^{\prime \prime }}+\frac{%
Q^{2}\left( r^{\prime \prime }\right) }{r^{\prime \prime 2}}\right) ^{-\frac{%
1}{2}}\frac{Q^{2}\left( r^{\prime \prime }\right) e^{\frac{\nu \left(
r^{\prime \prime }\right) }{2}}}{r^{\prime \prime 5}}dr^{\prime \prime
}\right] dr^{\prime }}{\int_{0}^{r}r^{\prime }\left( 1-\frac{2m(r^{\prime })%
}{r^{\prime }}+\frac{Q^{2}\left( r^{\prime }\right) }{r^{\prime 2}}\right)
^{-\frac{1}{2}}dr^{\prime }}.\nonumber\\\end{aligned}$$
In the following we denote $\alpha (r)=1-\frac{Q^{2}(r)}{2m(r)r}$. For stable stellar type compact objects $\frac{m}{r^{3}}$ does not increase outwards. We suppose that for all $r^{\prime }\leq r$ we have $\frac{\alpha
\left( r^{\prime }\right) m(r^{\prime })}{r^{\prime }}\geq \frac{\alpha
\left( r\right) m(r)}{r}\left( \frac{r^{\prime }}{r}\right) ^{2}$ or, equivalently, $\frac{2m\left( r^{\prime }\right) }{r^{\prime }}-\frac{2m(r)}{%
r}\left( \frac{r^{\prime }}{r}\right) ^{2}\geq \frac{Q^{2}(r^{\prime })}{%
r^{\prime 2}}-\frac{Q^{2}(r)}{r^{2}}\left( \frac{r^{\prime }}{r}\right) ^{2}$ We assume that inside the compact stellar object the charge $Q(r)$ satisfies the general condition $$\label{8}
\frac{Q(r^{\prime \prime })e^{\frac{\nu \left( r^{\prime \prime }\right) }{2}%
}}{r^{\prime \prime 5}}\geq \frac{Q(r^{\prime })e^{\frac{\nu \left(
r^{\prime }\right) }{2}}}{r^{\prime 5}}\geq \frac{Q(r)e^{\frac{\nu (r)}{2}}}{%
r^{5}},r^{\prime \prime }\leq r^{\prime }\leq r.$$
Therefore we can evaluate the terms in equation (\[7\]) as follows. For the term in the denominator of the RHS of Eq.(\[7\]) we obtain: $$\label{9}
\left[ \int_{0}^{r}r^{\prime }\left( 1-\frac{2m\left( r^{\prime }\right) }{%
r^{\prime }}+\frac{Q^{2}\left( r^{\prime }\right) }{r^{\prime 2}}\right) ^{-%
\frac{1}{2}}dr^{\prime }\right] ^{-1}\leq \frac{2\alpha (r)m(r)}{r^{3}}%
\left[ 1-\left( 1-\frac{2\alpha (r)m(r)}{r}\right) ^{\frac{1}{2}}\right]
^{-1}.$$
For the second term in the bracket of the LHS of Eq.(\[7\]) we have $$\begin{aligned}
\label{10}
&\int_{0}^{r}\left( 1-\frac{2m(r^{\prime })}{r^{\prime }}+\frac{Q^{2}\left(
r^{\prime }\right) }{r^{\prime 2}}\right) ^{-\frac{1}{2}}\frac{Q^{2}\left(
r^{\prime }\right) e^{\frac{\nu \left( r^{\prime }\right) }{2}}}{r^{\prime 5}%
}dr^{\prime }\geq \nonumber \\
&\frac{Q^{2}(r)e^{\frac{\nu (r)}{2}}}{r^{5}}\int_{0}^{r}\left( 1-\frac{2m(r)%
}{r^{3}}r^{\prime 2}+\frac{Q^{2}(r)}{r^{6}}r^{\prime 2}\right) ^{-\frac{1}{2}%
}dr^{\prime }= \nonumber\\
&\frac{Q^{2}(r)e^{\frac{\nu (r)}{2}}}{r^{5}}\left( \frac{2\alpha (r)m(r)}{%
r^{3}}\right) ^{-\frac{1}{2}}\arcsin \left( \sqrt{\frac{2\alpha (r)m(r)}{r}}%
\right).\nonumber \\\end{aligned}$$
The second term in the bracket of the RHS of Eq. (\[7\]) can be evaluated as $$\begin{aligned}
\label{11}
&\int_{0}^{r}r^{\prime }\left( 1-\frac{2m\left( r^{\prime }\right) }{%
r^{\prime }}+\frac{Q^{2}\left( r^{\prime }\right) }{r^{\prime 2}}\right) ^{-%
\frac{1}{2}}\left[ \int_{0}^{r^{\prime }}\left( 1-\frac{2m\left( r^{\prime
\prime }\right) }{r^{\prime \prime }}+\frac{Q^{2}\left( r^{\prime \prime
}\right) }{r^{\prime \prime 2}}\right) ^{-\frac{1}{2}}\frac{Q^{2}\left(
r^{\prime \prime }\right) e^{\frac{\nu \left( r^{\prime \prime }\right) }{2}}%
}{r^{\prime \prime 5}}dr^{\prime \prime }\right] dr^{\prime }\geq \nonumber \\
&\int_{0}^{r}r^{\prime }\left( 1-\frac{2m(r^{\prime })}{r^{\prime }}+\frac{%
Q^{2}\left( r^{\prime }\right) }{r^{\prime 2}}\right) ^{-\frac{1}{2}}\frac{%
Q^{2}\left( r^{\prime }\right) e^{\frac{\nu \left( r^{\prime }\right) }{2}}}{%
r^{\prime 4}}\left( \frac{2\alpha (r^{\prime })m(r^{\prime })}{r^{\prime }}%
\right) ^{-\frac{1}{2}}\arcsin \left( \sqrt{\frac{2\alpha (r^{\prime
})m(r^{\prime })}{r^{\prime }}}\right) dr^{\prime }\geq \nonumber \\
&\frac{Q^{2}(r)e^{\frac{\nu (r)}{2}}}{r^{5}}\int_{0}^{r}r^{\prime }\left( 1-%
\frac{2\alpha (r)m(r)}{r}r^{\prime 2}\right) ^{-\frac{1}{2}}\left( \frac{%
2\alpha (r)m(r)}{r^{3}}r^{\prime 2}\right) ^{-\frac{1}{2}}\arcsin \left(
\sqrt{\frac{2\alpha (r)m(r)}{r^{3}}}r^{\prime }\right) dr^{\prime }= \nonumber \\
&\frac{Q^{2}(r)e^{\frac{\nu (r)}{2}}}{r^{5}}r^{2}\left( \frac{2\alpha
(r)m(r)}{r}\right) ^{-\frac{3}{2}}\left[ \sqrt{\frac{2(r)m(r)}{r}}-\sqrt{1-%
\frac{2\alpha (r)m(r)}{r}}\arcsin \left( \sqrt{\frac{2\alpha (r)m(r)}{r}}%
\right) \right].\nonumber \\\end{aligned}$$
In order to obtain the inequality (\[11\]) we have also used the property of monotonic increase in the interval $x\in \left[ 0,1\right] $ of the function $\frac{\arcsin x}{x}$ .
Using Eqs.(\[9\])-(\[11\]), Eq.(\[7\]) becomes: $$\label{12}
\left[ 1-\left( 1-\frac{2\alpha (r)m(r)}{r}\right) ^{\frac{1}{2}}\right]
\frac{m(r)+4\pi r^{3}p-\frac{Q^{2}}{r}}{r^{3}\sqrt{1-\frac{2\alpha (r)m(r)}{r%
}}} \leq \frac{2m(r)}{r^{3}}+\frac{Q^{2}}{r^{4}}\left[ \frac{\arcsin \left(
\sqrt{\frac{2\alpha (r)m(r)}{r}}\right) }{\sqrt{\frac{2\alpha (r)m(r)}{r}}}%
-1\right].$$
Eq. (\[12\]) is valid for all $r$ inside the star. Consider first the neutral case $Q=0$. By evaluating (\[12\]) for $r=R$ we obtain $\frac{1}{%
\sqrt{1-\frac{2M}{R}}}\leq 2\left[ 1-\left( 1-\frac{2M}{R}\right) ^{\frac{1}{%
2}}\right] ^{-1},$ leading to the well-known result $\frac{2M}{R}\leq \frac{8%
}{9}$ [@St84].
Next consider the case $Q\neq 0$. We denote $$\label{13}
f\left( M,R,Q\right) =\frac{Q^{2}(R)}{R^{4}}\left( \frac{\alpha (R)M}{R^{3}}%
\right) ^{-1}\sqrt{1-\frac{2\alpha (R)M}{R}}\left[ \frac{\arcsin \left(
\sqrt{\frac{2\alpha (R)M}{R}}\right) }{\sqrt{\frac{2\alpha (R)M}{R}}}%
-1\right].$$
The function $f\left( M,R,Q\right) \geq 0,\forall M,R,Q$. Then (\[12\]) leads to the following restriction on the mass-radius ratio for compact charged general relativistic objects: $$\label{14}
\frac{2M}{R}\leq \frac{8}{9}+\frac{2f\left( M,R\right) }{9}-\frac{%
f^{2}\left( M,R\right) }{9}.$$
The variation of the maximum mass-radius ratio $u=\frac{M}{R}$ for a charged compact object as a function of the charge-mass ratio $q=\frac{Q}{M}$ is represented in Fig.1.
Due to the presence of the charge the maximum mass-radius ratio is only slightly modified as compared to the non-charged case. In the uncharged case the bound $2\frac{M}{R}<\frac{8}{9}$ is very close to the limit $2\frac{M}{R}<1$ arising from black hole considerations. But for a charged compact general relativistic object the bound $\frac{M}{R}<1$ obtained from horizon considerations is much larger than the limit following from Eq.(\[14\]).
In order to find a general restriction for the total charge $Q$ a compact stable object can acquire we shall consider the behavior of the Ricci invariants $r_{0}=R_{i}^{i}=R$, $r_{1}=R_{ij}R^{ij}$ and $%
r_{2}=R_{ijkl}R^{ijkl}$. If the general static line element is regular, satisfying the conditions $e^{\nu (0)}=const.\neq 0$ and $e^{\lambda (0)}=1$ , then the Ricci invariants are also non-singular functions throughout the star. In particular for a regular space-time the invariants are non-vanishing at the origin $r=0$. For the invariant $r_{2}$ we find $$\begin{aligned}
\label{15}
&r_{2}=\left[ 8\pi \left( \rho +p\right) -\frac{4m}{r^{3}}+\frac{6Q^{2}}{%
r^{4}}\right] ^{2}+2\left( 8\pi p+\frac{2m}{r^{3}}-\frac{2Q^{2}}{r^{4}}%
\right) ^{2}+ \nonumber \\
&2\left( 8\pi \rho -\frac{2m}{r^{3}}+\frac{2Q^{2}}{r^{4}}\right)
^{2}+4\left( \frac{2m}{r^{3}}-\frac{Q^{2}}{r^{4}}\right) ^{2}. \nonumber \\\end{aligned}$$
For a monotonically decreasing interior electric field $\frac{Q^{2}}{8\pi
r^{4}}$, the function $r_{2}$ is regular and motonically decreasing throughout the star. Therefore it satisfies the condition $r_{2}(R)<r_{2}(0)$, leading to the following general constraint on the value of the electric field at the surface of the compact object: $$\label{16}
6\frac{M^{2}}{R^{6}}<\frac{12M}{R^{3}}\frac{Q^{2}}{R^{4}}-7\left( \frac{Q^{2}%
}{R^{4}}\right) ^{2}+4\pi ^{2}\left( 6\rho _{c}^{2}+4\rho
_{c}p_{c}+6p_{c}^{2}\right),$$ where we assumed that at the surface of the star the matter density vanishes, $\rho (R)=0$.
Another condition on $Q(R)$ can be obtained from the study of the scalar $$\label{17}
r_{1}=\left( 8\pi \rho +\frac{Q^{2}}{r^{4}}\right) ^{2}+3\left( 8\pi p-\frac{%
Q^{2}}{r^{4}}\right) ^{2}+\frac{64\pi pQ^{2}}{r^{4}}.$$
Under the same assumptions of regularity and monotonicity for the function $%
r_{1}$ and considering that the surface density is vanishing we obtain for the surface value of the monotonically decreasing electric field the upper bound $$\label{18}
\frac{Q^{2}}{R^{4}}<4\pi \rho _{c}\sqrt{1+3\left( \frac{p_{c}}{\rho _{c}}%
\right) ^{2}}.$$
The invariant $r_{0}$ leads to the trace condition $\rho _{c}>3p_{c}$ of the energy-momentum tensor that holds at the center of the fluid spheres.
Discussions and final remarks
=============================
The existence of a limiting value of the mass-radius ratio leads to limiting values for other physical quantities of observational interest. One of these quantities is the surface red shift $z$ of the compact object, defined according to $z=\left( 1-\frac{2M}{R}\right) ^{-\frac{1}{2}}-1$. For an electrically neutral star Eq.(\[7\]) leads to the well-known constraint $z\leq 2$. For the charged star the surface red shift must obey the more general restriction $$\label{20}
z\leq 2+\frac{Q^{2}}{R^{4}}\left( \frac{\alpha (R)M}{R^{3}}\right)
^{-1}\left[ \frac{\arcsin \left( \sqrt{\frac{2\alpha (R)M}{R}}\right) }{%
\sqrt{\frac{2\alpha (R)M}{R}}}-1\right].$$
The variation as a function of the charge-mass ratio of the maximum red shift for charged compact objects for a given mass-radius ratio is presented in Fig.2.
Therefore higher surface red shifts than 2 could be observational criteria indicating the presence of electrically charged ultra compact matter distributions.
As another application of obtained limiting mass -radius ratios for charged stars we shall derive an explicit limit for the total energy of compact general relativistic objects. The total energy (including the gravitational field contribution) inside an equipotential surface $S$ can be defined to be [@Ka88] $$\label{21}
E=E_{M}+E_{F}=\frac{1}{8\pi }\xi _{s}\int_{S}\left[ K\right] dS,$$ where $\xi ^{i}$ is a Killing field of time translation, $\xi _{s}$ its value at $S$ and $\left[ K\right] $ is the jump across the shell of the trace of the extrinsic curvature of $S$, considered as embedded in the 2-space $t=const.$.$E_{M}=\int_{S}T_{i}^{k}\xi ^{i}\sqrt{-g}dS_{k}$ and $%
E_{F}$ are the energy of the matter and of the gravitational field, respectively. This definition is manifestly coordinate invariant. In the case of static spherically symmetric matter distribution we obtain for the total energy (also including the gravitational contribution) the exact expression $E=-re^{\frac{\nu -\lambda }{2}}$ [@Ka88]. Hence the total energy (including the gravitational contribution) of a charged compact general relativistic object is $E=-R\left[ 1-\frac{2M}{R}+\frac{Q^{2}(R)}{R^{2}}\right] $
For a neutral matter distribution $Q=0$ and for the total energy of the star we find the upper limit $E\leq -\frac{R}{9}$. In the charged case we obtain $$\label{22}
E\leq -\frac{R}{9}+\frac{2f}{9}R-\frac{f^{2}}{9}R,$$ with the function $f$ defined in Eq. (\[13\]).
In the present Letter we have considered the mass-radius ratio limit for charged stable compact general relativistic objects. Also in this case it is possible to obtain explicit inequalities involving $\frac{2M}{R}$ as an explicit function of the charge $Q$. The surface red shift and the total energy (including the gravitational one) are modified due to the presence of a strong electric field inside the compact object. The mass-radius ratio depends on the value of the total charge of the star, with the increases in mass, red shift or total energy proportional to the charge parameter.
Chandrasekhar S., [*The Mathematical Theory of Black Holes*]{}, (1992) Oxford, Oxford University Press.
Shvartsman V.F., [*Soviet Physics-JETP*]{}, [**33**]{} 1971 475.
Turolla R., Zane S., Treves A. and Illarionov A., [*Astrophys. J.*]{}, [**482**]{} (1997) 377; Michel F.C., [*Astrophys. Space Science*]{}, [**15**]{} (1972) 153; Anile A.M. and Treves A., [*Astrophys. Space Science*]{}, [**19**]{} (1972) 411; Maraschi I., Reina C. and Treves A., [*Astron. Astrophys.*]{}, [**35**]{} (1974) 389; Meszaros P., [*Astron. Astrophys.*]{}, [**44**]{} (1975) 59; Maraschi I., Reina C. and Treves A., [*Astron. Astrophys.*]{}, [**66**]{} (1978) 99; Treves A. and Turolla R., preprint astro-ph/9812383 (1998).
Preparata G., Ruffini R. and Xue S.-S., preprint astro-ph/9810182 (1998); Ruffini R., preprint astro-ph/9811232 (1998); Ruffini R., Salmonson J.D., Wilson J.R.and Xue S.-S., preprint astro-ph/0004257 (2000); Ruffini R., preprint astro-ph/0001425 (2000).
Witten E., [*Phys. Rev. D*]{}, [**30**]{} (1984) 272.
Khadrikar S.B., Mishra A. and Mishra H., [*Mod. Phys. Lett. A*]{}, [**10**]{} (1995) 2651.
Buchdahl H.A., [*Phys. Rev.*]{}, [**116**]{} (1959) 1027.
Straumann N., [*General Relativity and Relativistic Astrophysics*]{} (1984), Springer Verlag, Berlin.
Beckenstein J.D., [*Phys. Rev. D*]{}, [**4**]{} (1971) 2185.
Katz J., Lynden-Bell D. and Israel W., [*Class. Quantum. Grav.*]{}, [**5**]{} (1988) 971; Gron O. and Johannesen S., [*Astrophys. Space Science*]{}, [**19**]{} 1992 411.
[^1]: E-mail:[email protected]
[^2]: E-mail:[email protected]
[^3]: E-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We use an estimate on the Thurston–Bennequin invariant of a Legendrian link in terms of its Kauffman–polynomial to show that links of topological unknots, e.g. the Borromean rings or the Whithead link, may not be represented by Legendrian links of Legendrian unknots.'
author:
- 'Klaus Mohnke[^1]'
title: Legendrian links of topological unknots
---
In [@Eliashberg:Legendre] Eliashberg classified all Legendre knots representing the unknot in terms of their Thurston–Bennequin number, $tb$, and their rotation, $r$. Bennequin’s inequality in these cases reads as $$tb + |r|\le -1.$$ Thus the Legendre knot given by the wavefront of the ’eye’ which has $tb=-1, r=0$ is referred to as the [*trivial Legendrian knot*]{}.
Back then there was no other obvious obstruction for Legendrian [*links*]{} consisting of (topological) unknots and Eliashberg asked the question: ”Given a link of topological unknots, can it be realized as a link of \[…\] Legendrian unknots?”
The answer is negative in general and the new obstructions are given by a sharper inequality on the Thurston–Bennequin number governed by the Kauffman polynomial $K(x,t)$ which was found by Lee Rudolph in [@Rudolph] (for further reference see e.g. [@Chmutov/Goryunov] and [@Fuks/Tabachnikov]). It simply states that the Thurston–Bennequin number is not bigger than the the minimal degree in the variable $x$ of the Kauffman polynomial: $$tb\le -\max\mbox{-}\deg_x K.$$
The contribution of the author is to apply this to links of topological unknots. We had to be careful because the two groups of authors [@Chmutov/Goryunov; @Fuks/Tabachnikov] used different Kauffman polynomials and thus obtain slightly different inequalities: here we work with the Dubrovnik–version Chmutov and Goryunov used. Let us first recall the definition of the Thurston–Bennequin number:
Let $L$ be an oriented Legendrian link given by a wave front projection. Then the [*Thurston–Bennequin number*]{} of $L$, $tb(L)$, is the number of sideward crossings minus the number of up– or downward crossings minus half the number of cusps.
From that the following observation is immediate
Let $L=\coprod_iL_i$ be a Legendrian link with pairwise unlinked components $L_i$. Then the Thurston–Bennequin number of that link is simply given by the sum of those of the components $$tb(L) = \sum_i tb(L_i).$$ In particular, it does not depend on the orientation of the components.
To investigate Eliashberg’s question we took the most simplest examples we knew: The Borromean rings $B$ and the Whitehead link $W$. The Kauffman polynomials are given by $$\split
K_W(x,y) &=
yx^5-2x^4-(2y^3+6y)x^3+(-y^4-y^2+6+y^{-2})x^2+(3y^3+9y+2y^{-1})x\\
&\qquad+(y^4+y^2-5-2y^{-2})-(y^3+4y-2y^{-1})x^{-1}+(2+y^{-2})x^{-2}
\endsplit$$ and $$\split
K_B(x,t) & =y^2x^4+(-4y+y^{-3})x^3+(-3y^4-10y^2+3y^{-2})x^2\\
&\qquad+(-2y^5-2y^3+14y+3y^{-1}-3y^{-3})x + (6y^4+18y^2+1-6y^{-2})\\
&\qquad-(-2y^5-2y^3+14y+3y^{-1}-3y^{-3})x^{-1}+(-3y^4-10y^2+3y^{-2})x^{-2}\\
&\qquad-(-4y+y^{-3})x^{-3}+y^2x^4.
\endsplit$$ From that we easily deduce the main result of that note
\(1) For any Legendrian representaion of the Borromean rings we have $$tb\le -4.$$ (2) For any Legendrian representation of the Whitehaed link we have $$tb\le -5.$$ Thus for both not all components may be Legendrian unknots.
\(1) In the cases of the Borromean rings and the Whitehead link the inequality is sharp as the following pictures show.\
(2) In the case of the Whitehead link the Kauffman polynomial gives no further obstruction for its mirror. Indeed it can be represented as a Legendrian link of Legendrian unknots.
On the other hand it is possible to give two different Legendrian Whithead links with $tb=-5$. One consists of components with Thurston–Bennequin numbers $-4$ and $-1$ the other with $-3$ and $-2$.
\(1) Are there sharper bounds on the Thurston–Bennequin number than that given in [@Chmutov/Goryunov; @Fuks/Tabachnikov]?\
(2) Are there sharper bounds if one imposes additional conditions? E.g. consider Brunnian links, i.e. links of unknots which fall appart if one removes one component.\
(3) Is it at least true that a link of the type described in (2) may be represented as a Legendrian link consisting of Legendrian unknots iff the maximal degree in $x$ of its Kauffman polynomial is equal to the number of components of the link (note that this degree is never less than the number of components)?\
(4) Are there further restrictions for the distribution of the Thurston–Bennequin number on the components of a Legendrian link?
[*Acknowledgements.*]{} I would like to thank Uwe Kaiser and Alexander Pilz for patiently checking my computations. I apologize to Lee Rudolph for wrongly stating the reference for the inequality used in this note in the first version and thank him for correcting me.
[1]{}
S. Chmutov and V. Goryunov. Polynomial invariants of [L]{}egendrian links and their fronts. In Shin’chin Suzuki, editor, [*Proceedings of [K]{}nots 96*]{}, pages 239–256. World Scintific Publishing Co., 1997.
Yakov Eliashberg. Legendrian and transversal knots in tight contact $3$-manifolds. In [*Topological methods in modern mathematics*]{}. Publish or Perish, 1993.
Dmitry Fuks and Serge Tabachnikov. Invariants of [L]{}egendrian and transverse knots in the standard contact space. , 36(5):1025–1053, 1997.
Lee Rudolph. A congruence between link polynomials. , 107:319–327, 1990.
[^1]: Universität-Gesamthochschule Siegen, e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
By requiring co-ordination to take place using explicit message passing instead of relying on shared memory, actor-based programming languages have been shown to be effective tools for building reliable and fault-tolerant distributed systems. Although naturally communication-centric, communication patterns in actor-based applications remain informally specified, meaning that errors in communication are detected late, if at all.
Multiparty session types are a formalism to describe, at a global level, the interactions between multiple communicating entities. This article describes the implementation of a prototype framework for monitoring Erlang/OTP `gen_server` applications against multiparty session types, showing how previous work on multiparty session actors can be adapted to a purely actor-based language, and how monitor violations and termination of session participants can be reported in line with the Erlang mantra of ‘let it fail’. Finally, the framework is used to implement two case studies: an adaptation of a freely-available DNS server, and a chat server.
author:
- Simon Fowler
bibliography:
- 'refs.bib'
title: An Erlang Implementation of Multiparty Session Actors
---
#### Acknowledgements
Thanks to Sam Lindley, Garrett Morris, and Philip Wadler for useful discussions, and to the anonymous reviewers for detailed and insightful comments. This work was supported by EPSRC grant EP/L01503X/1 (University of Edinburgh CDT in Pervasive Parallelism).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A new moderately $r$-process-enhanced metal-poor star, [RAVE J093730.5$-$062655]{}, has been identified in the Milky Way halo as part of an ongoing survey by the $R$-Process Alliance. The temperature and surface gravity indicate that [J0937$-$0626 ]{}is likely a horizontal branch star. At $[\rm{Fe/H}] = -1.86$, [J0937$-$0626 ]{}is found to have subsolar \[X/Fe\] ratios for nearly every light, $\alpha$, and Fe-peak element. The low $[\alpha/\rm{Fe}]$ ratios can be explained by an $\sim0.6$ dex excess of Fe; [J0937$-$0626 ]{}is therefore similar to the subclass of “iron-enhanced” metal-poor stars. A comparison with Milky Way field stars at $[\rm{Fe/H}] = -2.5$ suggests that [J0937$-$0626 ]{}was enriched in material from an event, possibly a Type Ia supernova, that created a significant amount of Cr, Mn, Fe, and Ni and smaller amounts of Ca, Sc, Ti, and Zn. The $r$-process enhancement of [J0937$-$0626 ]{}is likely due to a separate event, which suggests that its birth environment was highly enriched in $r$-process elements. The kinematics of [J0937$-$0626]{}, based on [*Gaia*]{} DR2 data, indicate a retrograde orbit in the Milky Way halo; [J0937$-$0626 ]{}was therefore likely accreted from a dwarf galaxy that had significant $r$-process enrichment.'
author:
- 'Charli M. Sakari'
- 'Ian U. Roederer'
- 'Vinicius M. Placco'
- 'Timothy C. Beers'
- Rana Ezzeddine
- Anna Frebel
- Terese Hansen
- Christopher Sneden
- 'John J. Cowan'
- George Wallerstein
- 'Elizabeth M. Farrell'
- 'Kim A. Venn'
- Gal Matijevič
- 'Rosemary F.G. Wyse'
- 'Joss Bland-Hawthorn'
- Cristina Chiappini
- 'Kenneth C. Freeman'
- 'Brad K. Gibson'
- 'Eva K. Grebel'
- Amina Helmi
- Georges Kordopatis
- Andrea Kunder
- Julio Navarro
- Warren Reid
- George Seabroke
- Matthias Steinmetz
- Fred Watson
title: 'The $R$-Process Alliance: Discovery of a low-$\alpha$, $r$-Process-Enhanced Metal-Poor Star in the Galactic Halo'
---
Introduction {#sec:Intro}
============
The advent of large surveys has provided insight into the formation and evolution of the Milky Way (MW) and its satellites, particularly the nucleosynthesis of the elements and chemical evolution in galaxies of different masses. Many open questions remain, however, including the astrophysical site for the creation of the heaviest elements in the Universe. These elements are created by the rapid ($r$-) neutron capture process; suggestions that $r$-process nucleosynthesis could occur during a neutron star merger [@LattimerSchramm1974; @Rosswog2014; @Lippuner2017] have now been confirmed through observations of GW 170817 [@Abbott2017; @Chornock2017; @Drout2017; @Shappee2017]. However, core-collapse supernovae from strongly magnetic stars (the so-called “jet-supernovae”) may also be a viable site of the $r$-process (e.g., @Winteler2012, @Cescutti2015, @Cote2018). One of the most useful sites for probing the environments, yields, and occurrence rates for $r$-process nucleosynthesis are the $r$-process-enhanced metal-poor stars, which retain a relatively pure $r$-process signature and whose spectra are not overly contaminated from metal lines.
A new collaboration, the $R$-Process Alliance (RPA), has begun a campaign to identify more of these $r$-process-enhanced metal-poor stars (with $[\rm{Ba/Eu}]<0$), with the ultimate goal of constraining the site(s) of the $r$-process across cosmic time. Initial results from the Northern and Southern hemisphere surveys (@Sakari2018b [@Hansen2018], plus additional papers from @Placco2017, @Sakari2018a, @Cain2018, @Gull2018, @Holmbeck2018, and @Roederer2018b) have identified many more of these stars, including 18 new highly-enhanced $r$-II stars (with $[\rm{Eu/Fe}]>+1.0$) and 101 new moderately-enhanced $r$-I stars (with $+0.3\le[\rm{Eu/Fe}]\le+1.0$), according to the classifications from @BeersChristlieb2005. These new discoveries enable the $r$-process-enhanced metal-poor stars to be studied as stellar populations, so that their chemical and kinematic properties can be assessed as a whole.
Though they serve as useful laboratories for studying the $r$-process, it is still not known how or where $r$-process-enhanced stars form, including how they have retained such a strong $r$-process signal without being significantly diluted by the nucleosynthetic products of other stars (e.g., core collapse supernovae). One theory is that the $r$-process-enhanced stars form in the lower mass ultra-faint dwarfs which are later accreted into the MW halo. This framework is supported by both observations and simulations: $r$-process-enhanced stars have been found in ultra faint dwarfs, notably Reticulum II [@Ji2016; @Roederer2016], while simulations suggest that low mass dwarf galaxies are capable of retaining the ejecta from an $r$-process nucleosynthetic event (e.g., @BlandHawthorn2015, @Beniamini2018). In addition, the $r$-process-enhanced stars are also predominantly old (e.g., @Placco2017 [@Holmbeck2018; @Sakari2018b; @Valentini2018]), and simulations indicate that many of the oldest stars in the MW halo may have been accreted (e.g., @SteinmetzMuller1994 [@Brook2007; @Brook2012; @ElBadry2018]).
Another convincing piece of evidence for an extragalactic origin for the $r$-process-enhanced stars comes from kinematics. Several $r$-II stars have orbits consistent with accretion from a satellite [@Roederer2018], while many of the highly enhanced $r$-II and $r$-I stars have retrograde orbits in the MW halo which indicate an extragalactic origin (e.g., @Sakari2018b). An increased number of $r$-I and $r$-II stars, combined with increasingly better data from [*Gaia*]{} [@GaiaREF], will enable detailed orbits and more subgroups to be identified, as was done in @Koppelman2018 and @Roederer2018.
Some dwarf galaxy stars can also be identified chemically, as a result of differing chemical evolution in massive and low-mass galaxies (see, e.g., @Tolstoy2009). This is generally only possible for intermediate-metallicity stars that have formed after several previous generations of stars (i.e., after enough time has passed to allow chemical evolution to proceed differently in the low-mass environment). This also requires that the accreted dwarf galaxy experienced extended epochs of star formation, rather than a single burst (see @Webster2015 for evidence that this is possible, even in the lowest mass ultra faint dwarfs). The majority of metal-poor stars are unlikely to show the chemical signatures of more metal-rich dwarf galaxy stars. A few exceptions have been identified in the MW halo, notably the class of “Fe-enhanced” metal-poor stars which show low \[X/Fe\] ratios at $[\rm{Fe/H}]<-1$ (e.g., @Yong2013); generally, however, these stars are fairly rare. Until now, none of these stars in the MW halo have been $r$-process-enhanced.
This paper reports the discovery of an $r$-process-enhanced, metal-poor star that exhibits the typical chemical signatures of dwarf galaxy stars (notably low \[$\alpha$/Fe\] ratios). Section \[sec:Observations\] describes the observations, data reduction, and atmospheric parameters of this star, while Section \[sec:Abunds\] presents the abundances. The implications of these abundances, the kinematics, and comparisons with other MW halo stars and dwarf galaxy stars are discussed in Section \[sec:Discussion\].
Observations, Data Reduction, and Atmospheric Parameters {#sec:Observations}
========================================================
[J0937$-$0626 ]{}was identified as a metal-poor star in Data Release 4 of the RAdial Velocity Experiment (RAVE; @RAVEref [@RAVEDR4ref]) and the subsequent re-analysis by @Matijevic2017. It was then targeted for a medium-resolution, optical analysis by @Placco2018. [J0937$-$0626 ]{}was then observed at high spectral resolution in 2016 and 2017 using the Astrophysical Research Consortium (ARC) 3.5-m telescope at Apache Point Observatory, as part of the Northern Hemisphere survey of the RPA [@Sakari2018b]. The ARC Echelle Spectrograph was used in its default mode, leading to a spectral resolution of $R\sim31,500$ and coverage of nearly the full optical range, from $3800$ to $10400$ Å. The exposure times were selected to ensure high S/N ratios in the red and the blue, as shown in Table \[table:Targets\]. The data were reduced in the Image Reduction and Analysis Facility program (IRAF)[^1] using standard techniques, as described in @Sakari2018b. The heliocentric radial velocity was found by cross-correlating the spectrum with a high-resolution, high-S/N spectrum of Arcturus [@Hinkle2003]. The radial velocity is in excellent agreement with the value from RAVE DR5 (see Table \[table:Targets\]).
Equivalent widths (EWs) of and lines from @Fulbright2006, @Venn2012, and @McWilliam2013 were found using the automated program [DAOSPEC]{} [@DAOSPECref]. Fe abundances were then determined using the 2017 version of [ MOOG]{} [@Sneden], with an appropriate treatment of scattering [@Sobeck2011].[^2] The $<$3D$>$, non-Local Thermodynamic Equilibrium (NLTE) corrections from @Amarsi2016 were applied to each line, as discussed in @Sakari2018b. The temperature and microturbulent velocity of [J0937$-$0626 ]{}were determined by removing trends in the NLTE abundances with wavelength, reduced EW, and excitation potential (see Figure \[fig:Trends\]); the surface gravity was found by forcing agreement between NLTE and abundances. The final adopted parameters are listed in Table \[table:Targets\], along with the LTE parameters and the parameters derived with the 1D NLTE corrections of @Ezzeddine2017 [also see @Ezzeddine2016 and @Sakari2018b for more details]. The Ezzeddine et al. corrections lead to similar parameters as the Amarsi et al. $<$3D$>$ NLTE corrections. The largest effect of the NLTE corrections to the lines is to raise the surface gravity and the \[Fe/H\] over the LTE values.
@Schuster2004 and @Beers2007 obtained photometry of [J0937$-$0626]{}, finding colors that are consistent with the spectroscopic parameters derived here. @Schuster2004 classified [J0937$-$0626 ]{}as a “red-horizontal-branch-asymptotic-giant-branch transition” star, while @Beers2007 found that it was displaced from the metal-poor main sequence, potentially as a result of its lower surface gravity. Indeed, the spectroscopic parameters for [J0937$-$0626 ]{}place it in the expected region for old, moderately metal-poor HB stars. The spectroscopic temperature is also in agreement with the photometric analysis by @Munari2014 and the spectroscopic RAVE DR5 value [@RAVEDR5ref], while the temperature, surface gravity, and metallicity are in agreement with the medium-resolution analysis of @Placco2018. [*Gaia*]{} has provided a parallax for [J0937$-$0626 ]{}in Data Release 2 [@GaiaREF; @GaiaDR2REF], which gives a distance; @BailerJones2018 also provide a statistically-determined distance (see Table \[table:Targets\]). These distances, combined with the $E(B-V)$ from the @SchlaflyFinkbeiner2011 reddening maps, indicate that [J0937$-$0626 ]{}likely has an absolute magnitude of $M_{\rm{V}}~=~-0.120\pm0.18$ (with the inverse parallax distance) or $M_{\rm{V}}~=~0.067\pm0.18$ (with the @BailerJones2018 distance). These magnitudes are both consistent with [J0937$-$0626 ]{}being a red horizontal branch star.
Very few red horizontal branch stars have been observed by the RPA; the targets are mainly red giant branch stars. However, though [J0937$-$0626 ]{}may be a horizontal branch star, its abundances should reflect the composition of typical MW stars (with the exception of C; see the discussion in Section \[subsec:Scenarios\] and Figure 2 in @Roederer2018b). Also note that its atmospheric parameters place [J0937$-$0626 ]{}within the extent of the NLTE grid from @Amarsi2016 [see their Table 2]. To further confirm that the NLTE corrections are appropriate for a red horizontal branch star, the more metal-rich $r$-II star HD 222925 from @Roederer2018b was re-analyzed. Its parameters with the $<$3D$>$ NLTE Amarsi et al. corrections ($T_{\rm{eff}}~=~5625$ K, $\log~g~=~2.3$, $\xi~=~1.75$ km s$^{-1}$, and \[/H\] $=~-1.44$) agree with the photometric parameters from Roederer et al. ($T_{\rm{eff}}~=~5636$ K, $\log~g~=~2.54$) and are higher than the spectroscopic parameters ($\xi~=~2.20$ km s$^{-1}$, and \[/H\] $=~-~1.58$; though note than Roederer et al. find \[/H\] $=~-1.47$). The offsets in the metallicity and microturbulent velocity are consistent with the general trends found in LTE vs. NLTE comparison (e.g., @Amarsi2016). However, in HD 222925 these atmospheric parameter offsets only lead to small differences in the \[X/Fe\] ratios ($\la 0.1$ dex); this indicates that the $<$3D$>$ NLTE corrections produce reasonable results for red horizontal branch stars.
Carbon abundances were found by synthesizing the CH [*G*]{}-band region at 4312 Å. [J0937$-$0626 ]{}is found to have a subsolar $[\rm{C/Fe}]~=~-0.55\pm0.40$, a reasonably low value given its advanced evolutionary state. Taking the evolutionary corrections of @Placco2014c into account, the “natal” carbon abundance was likely higher, at $[\rm{C/Fe}] \sim 0.1$. [J0937$-$0626 ]{}is not (and never was) a CEMP star.
[@lcl]{} ID & [RAVE J093730.5$-$062655]{}& Other IDs: TYC 4900-1967-1, BS 17576-0027, 2MASS J09373053-0626551\
RA (J2000) & 09:37:30.54 &\
Dec (J2000) & -06:26:55.0 &\
$V$ & 11.81 &\
$K$ & 10.13 &\
$E(B-V)$ & 0.0266 & Average value from @SchlaflyFinkbeiner2011 maps\
$d$ (kpc) & $2.342^{+0.230}_{-0.192}$ & Inverse parallax distance\
& $2.149^{+0.190}_{-0.163}$ & @BailerJones2018\
$M_{\rm{V}}$ & $-0.120\pm0.18$ & Calculated from the inverse parallax distance\
& $0.067\pm0.18$ & Calculated from the @BailerJones2018 distance\
Observation Dates & 28 Jan, 11 Feb 2016, 2 Mar 2017 & Seeing = 0.9, 1.09, 1.12\
Exposure Time (s) & 3240 &\
S/N, 4400 Å& 100 & Per pixel; there are 2.5 pixels per resolution element\
S/N, 6500 Å& 170 & Per pixel; there are 2.5 pixels per resolution element\
& &\
$v_{\rm{helio}}$ (km s$^{-1}$) & $268.8\pm1.0$ & This work\
& $268.436\pm1.297$ & RAVE DR5\
& &\
$T_{\rm{eff}}$ (K) & $\boldsymbol{5875\pm55}^{a}$ & Spectroscopic, with $<$3D$>$ NLTE correction; this work\
& $5875\pm55$ & Spectroscopic, LTE; this work\
& $5850\pm50$ & Spectroscopic, with 1D NLTE @Ezzeddine2017 corrections; this work\
& $6091$ & Spectroscopic, @Placco2018\
& $5667.31\pm214$ & Spectroscopic, RAVE DR5\
& 5606 & Photometric, @RamirezMelendez2005 calibration\
& 5752 & Photometric, @Casagrande2010 calibration\
$\log g$ & $\boldsymbol{2.61\pm0.16}^{a}$ & Spectroscopic, this work\
& $2.31\pm0.16$ & Spectroscopic, LTE; this work\
& $2.70\pm0.20$ & Spectroscopic, with 1D NLTE @Ezzeddine2017 corrections; this work\
& $2.52$ & Spectroscopic, @Placco2018\
& $2.81\pm0.48$ & Spectroscopic, RAVE DR5\
$\xi$ (km/s) & $\boldsymbol{2.09\pm0.25}^{a}$ & This work\
& $2.14\pm0.25$ & Spectroscopic, LTE; this work\
& $2.20\pm0.20$ & Spectroscopic, with 1D NLTE @Ezzeddine2017 corrections; this work\
$[$M/H$]$ & -2.04 & RAVE DR5\
$[$Fe/H$]$ & $\boldsymbol{-1.86\pm0.02}^{a}$ & This work\
& $-2.03\pm0.02$ & Spectroscopic, LTE; this work\
& $-1.89\pm0.15$ & Spectroscopic, with 1D NLTE @Ezzeddine2017 corrections; this work\
& $-1.70$ & @Placco2018\
$[$C/Fe$]$ & $\boldsymbol{-0.55\pm0.40}^{a}$ & Measured value, this work\
& $\sim+0.1$ & “Natal” value, calculated with the evolutionary corrections of @Placco2014c\
& $+0.38$ & Measured value, @Placco2018\
Detailed Abundances {#sec:Abunds}
===================
Abundances of Fe, Ca, Sc, Ti, Cr, Mn, Co, and Ni were determined from EWs; all other elements were determined from spectrum syntheses. The lines for the EW analyses are from the line lists of @Fulbright2006 [@Fulbright2007] and @McWilliam2013. Corrections for hyperfine structure and (if necessary) isotopic splitting were included for Sc, V, Mn, and Co, using the data from the Kurucz database[^3] and @McWilliam2013. The spectrum synthesis line lists were generated with the [linemake]{} code.[^4] Hyperfine structure, isotopic splitting, and molecular lines from CH, C$_{2}$, and CN were included in the synthetic spectrum line lists. All \[X/H\] ratios were calculated line-by-line, where the solar abundance has been determined from the Kurucz solar spectrum[^5] if the lines are sufficiently weak and unblended, using the same atomic data; otherwise, the @Asplund2009 solar values are adopted. Note that unlike @Sakari2018b, a differential analysis has not been utilized, because there is not a suitable standard star in this metallicity range.
Table \[table:LineAbunds\] shows the line-by-line EWs or, for lines whose abundances were derived from spectrum syntheses, abundances. Table \[table:Abunds\] shows the final mean abundances. For the EW-based abundances, the random errors represent the line-to-line dispersion, with a minimum error in a single line of $0.05-0.1$ dex, depending on the strength of the line and S/N; for abundances that were determined via spectrum syntheses, the random errors are based on the quality of the syntheses. Table \[table:Abunds\] also shows the total error, which is a quadrature sum of the random error and the systematic error due to uncertainties in the atmospheric parameters. The systematic errors were determined from the variances and covariances of the atmospheric parameters, according to the techniques outlined in @McWilliam2013 and @Sakari2018b. Table \[table:Abunds\] also provides the abundance offsets that occur if LTE parameters are adopted. The offsets are all $\la 0.2$ dex; all \[X/Fe\] ratios relative to are negligible. These offsets reflect the abundance sensitivities to differences in $\log
g$, microturbulent velocity, and \[/H\]. NLTE corrections were not applied to elements other than ; significant NLTE sensitivities are generally noted below.
[@ccccccc]{} & 5895.92 & 0.00 & -0.180 & & 4.13 & SYN\
& 4167.28 & 4.34 & -0.745 & & 5.56 & SYN\
& 4703.00 & 4.34 & -0.670 & & 5.46 & SYN\
& 5528.41 & 4.34 & -0.480 & & 5.44 & SYN\
& 3944.00 & 0.00 & -0.640 & & 3.17 & SYN\
& 3961.52 & 0.01 & -0.340 & & 3.24 & SYN\
& 3905.52 & 1.91 & -1.090 & & 5.20 & SYN\
& 4283.01 & 1.89 & -0.220 & 35.7 & & EW\
& 4289.37 & 1.88 & -0.300 & 29.7 & & EW\
& 4302.54 & 1.90 & 0.275 & 63.1 & & EW\
[@lcDccDcD]{} & 84 & 5.64 & 0.01 & 0.05 & -1.86 & 0.05 & -0.17\
& 20 & 5.64 & 0.03 & 0.11 & -1.86 & 0.11 & -0.11\
& 1 & 4.13 & 0.10 & 0.15 & -0.25 & 0.13 & 0.14\
& 3 & 5.49 & 0.03 & 0.05 & -0.25 & 0.03 & 0.15\
& 2 & 3.19 & 0.06 & 0.09 & -1.40 & 0.07 & 0.10\
& 1 & 5.30 & 0.20 & 0.23 & -0.35 & 0.21 & 0.14\
& 17 & 4.30 & 0.02 & 0.04 & -0.18 & 0.02 & 0.17\
& 7 & 0.86 & 0.02 & 0.09 & -0.43 & 0.05 & 0.01\
& 6 & 3.13 & 0.02 & 0.05 & 0.04 & 0.03 & 0.16\
& 25 & 3.05 & 0.02 & 0.09 & -0.04 & 0.05 & 0.01\
& 7 & 3.83 & 0.04 & 0.08 & 0.05 & 0.05 & 0.16\
& 4 & 3.90 & 0.04 & 0.11 & 0.12 & 0.05 & 0.01\
& 1 & 3.12 & 0.10 & 0.11 & -0.45 & 0.10 & 0.17\
& 1 & 2.65 & 0.10 & 0.11 & -0.48 & 0.11 & 0.17\
& 2 & 4.18 & 0.06 & 0.07 & -0.16 & 0.06 & 0.14\
& 2 & 2.40 & 0.06 & 0.07 & -0.30 & 0.06 & 0.15\
& 1 & 0.96 & 0.20 & 0.32 & -0.05 & 0.24 & -0.01\
& 2 & 0.05 & 0.08 & 0.10 & -0.30 & 0.10 & 0.01\
& 2 & 0.62 & 0.07 & 0.10 & -0.10 & 0.10 & 0.01\
& 3 & 0.40 & 0.05 & 0.16 & 0.08 & 0.08 & 0.02\
& 3 & -0.35 & 0.03 & 0.08 & 0.41 & 0.08 & 0.01\
& 1 & 0.07 & 0.05 & 0.09 & 0.35 & 0.08 & 0.02\
& 1 & -0.44 & 0.10 & 0.12 & 0.70 & 0.12 & 0.02\
& 2 & 0.14 & 0.09 & 0.11 & 0.58 & 0.11 & 0.02\
& 3 & -0.49 & 0.06 & 0.12 & 0.85 & 0.08 & 0.01\
& 1 & -0.04 & 0.05 & 0.11 & 0.70 & 0.08 & 0.01\
& 1 & 0.19 & 0.10 & 0.12 & 0.95 & 0.12 & 0.01\
& 1 & <-0.79 & & & <1.05 & &\
Light Elements: Na and Al {#subsec:Light}
-------------------------
The Na abundance was determined from a synthesis of the 5895 Å$\;$ line (the 5889 Å$\;$ line is too strong), indicating a slightly subsolar \[Na/Fe\] ratio. The lines at 3944 and 3961 Å$\;$ yield a significantly subsolar $[\rm{Al/Fe}]~=~-1.40\pm0.06$. Table \[table:Abunds\] shows the LTE abundances, but both the Na and Al lines likely suffer from NLTE effects. The [INSPECT]{} database[^6] [@Lind2011] indicates that the 5895 Å$\;$ line should have a NLTE correction of -0.41 dex, which would make the Na abundance significantly subsolar. NLTE corrections to the Al lines may be as large as $+0.6$ to $0.8$ dex in this temperature, surface gravity, and metallicity range [@NordlanderLind2017].
$\alpha$ Elements {#subsec:Alphas}
-----------------
The $\alpha$-elements with detectable lines in [J0937$-$0626 ]{}include Mg, Si, Ca, and Ti; the O lines are too weak. The three lines at 4167, 4703, and 5528 Å$\;$ and the line at 3905 Å$\;$ were synthesized (see Figure \[fig:AlphaSynths\]), while EWs were measured for 17 lines, 6 lines, and 25 lines. @Mashonkina2016 show that NLTE corrections $\sim 0.1$ dex may be required for Ca, with larger and smaller corrections for and , respectively (though note that the highest temperature they consider is 5000 K). The \[X/Fe\] ratios for Mg, Si, and Ca are subsolar, while \[Ti/Fe\] is roughly solar; these ratios are subsolar even when LTE parameters are used. @Placco2018 also found a low $[\alpha/\rm{Fe}]~=~-0.09$ based on a medium-resolution spectrum.
![Syntheses of Mg, Al, and Si lines. The solid line shows the best-fit, while the dashed lines show the $1\sigma$ uncertainties. \[fig:AlphaSynths\]](AlphaSynths.eps)
Iron-Peak Elements and Zinc {#subsec:FePeak}
---------------------------
EWs were measured for 7 , 7 , 4 , 1 , 1 , and 3 lines. The \[/Fe\] ratio is expected to suffer from small NLTE effects; [@BergemannCescutti2010] find corrections $<0.2$ dex. The and lines also require NLTE corrections on the order of $+0.4$ [@BergemannGehren2008] and $+0.6$ dex [@Bergemann2010], respectively, according the MPIA NLTE correction database[^7] (though note that none of the models extend to [J0937$-$0626]{}’s surface gravity). None of these NLTE corrections were applied. The lines at 4722 and 4810 Å$\;$ were synthesized. The LTE \[Sc/Fe\], \[Mn/Fe\], \[Co/Fe\], and \[Ni/Fe\] ratios are all subsolar, \[Cr/Fe\] is slightly enhanced, and \[Zn/Fe\] is slightly subsolar.
![Syntheses of Ba, Pr, and Eu lines. \[fig:NCSynths\]](NCSynths.eps)
Neutron-Capture Elements {#subsec:NeutronCaptures}
------------------------
The Sr abundance in [J0937$-$0626 ]{}was derived from the relatively strong 4215 Å$\;$ line, Y was derived from the weak 4883 and 4900 Å$\;$ lines, and Zr was derived from the 4161 and 4208 Å$\;$ lines. Unlike the lighter elements, Sr and Zr yield approximately solar \[X/Fe\] ratios; Y is slightly subsolar.
Barium and europium are the elements used for classification of r-I and r-II stars. The 5853, 6141, and 6496 Å lines were used (the 4554 Å$\;$ line is too strong), while the 3819, 4129, and 4205 Å lines were used. [J0937$-$0626 ]{}has a roughly solar $[\rm{Ba/Fe}] = 0.08\pm0.05$ but an enhanced $[\rm{Eu/Fe}]~=~0.85\pm 0.06$, making it an $r$-I star. Its low $[\rm{Ba/Eu}]~=~-0.77\pm0.07$ indicates that it has received minimal contamination from the main $s$-process. Lines of , , , , and are detectable and also indicate enhancement. Figure \[fig:NCSynths\] shows syntheses of Ba, Pr, and Eu lines. An upper limit is derived from the line at 4019 Å.
Discussion {#sec:Discussion}
==========
Relative to Fe, [J0937$-$0626 ]{}has subsolar \[X/Fe\] ratios for nearly all elements (with the exception of Cr and possibly Ti) other than the neutron-capture elements. It has approximately solar \[X/Fe\] ratios for Sr, Y, Zr, and Ba, yet is enhanced in La, Ce, Pr, Nd, Eu, Gd, and Dy. Below, [J0937$-$0626]{}’s abundance patterns are compared to other MW halo stars and dwarf galaxy stars (Section \[subsec:MWComp\]), including the subset of “Fe-enhanced” metal-poor stars. Scenarios to explain the light, $\alpha$, and Fe-peak elements are explored in Section \[subsec:Scenarios\]. The $r$-process enhancement and patterns in [J0937$-$0626 ]{}are discussed in Section \[subsec:Patterns\], while the kinematics are discussed in Section \[subsec:Kinematics\].
Comparisons with Milky Way and Dwarf Galaxy Stars {#subsec:MWComp}
-------------------------------------------------
Figure \[fig:alphaFe\] demonstrates that [J0937$-$0626 ]{}has low \[Mg/Fe\] and \[Ca/Fe\] relative to MW field stars, along with \[Si/Fe\] and \[Ti/Fe\]. [J0937$-$0626 ]{}therefore has a deficiency of $\alpha$-elements, relative to Fe, compared to MW field stars. This is a phenomenon usually seen in dwarf galaxy stars (e.g., @Shetrone2003 [@Tolstoy2009]), where lower $[\alpha/\rm{Fe}]$ at a given $[\rm{Fe/H}]$ is usually interpreted as a sign of enrichment from Type Ia supernovae which produce lots of Fe, but few $\alpha$-elements (e.g., @Lanfranchi2008).[^8] However, with the exception of the neutron-capture elements (which will be discussed in Section \[subsec:Patterns\]), most of [J0937$-$0626]{}’s \[X/Fe\] ratios are lower than MW field stars at the same metallicity, including Sc and Ni (Figure \[fig:FePeakFe\]; though note that Mn and Cr agree with MW stars). In this respect, [J0937$-$0626 ]{}resembles the class of MW and dwarf galaxy stars that have been called “Fe-enhanced” metal-poor stars [@Yong2013], stars which show subsolar \[X/Fe\] ratios in nearly every element.[^9] Also shown in Figures \[fig:alphaFe\] and \[fig:FePeakFe\] are a selection of low-$\alpha$ or “Fe-enhanced” metal-poor stars from the MW halo, HE 1207$-$3108 [@Yong2013], HE 0533$-$5340 [@Cohen2013], SDSS J001820.5-093939.2 [@Aoki2014], and BD$+80^{\circ}$ 245 [@Carney1997; @Ivans2003; @Roederer2014b], along with low-$\alpha$ stars from four dwarf galaxies, Ursa Minor (UMi COS 171; @CohenHuang2010), Carina (Car 612; @Venn2012), Horologium I (Hor I, three stars; @Nagasawa2018), and Ret II (DES J033548$-$540349, hereafter [DES J0335$-$5403]{}; @Ji2016). These comparison stars indeed show the characteristic subsolar \[X/Fe\] ratios found in [J0937$-$0626]{}, albeit with slightly different values. [J0937$-$0626 ]{}can therefore be considered to be another one of these “Fe-enhanced” metal-poor stars.
A more detailed element-by-element comparison is shown in Figure \[fig:DelXFe\], which plots \[X/Fe\] ratios for each element. Following @Yong2013, only stars with similar metallicities are shown—here, [J0937$-$0626 ]{}($[\rm{Fe/H}]=-1.86$) is shown along with [DES J0335$-$5403 ]{}($[\rm{Fe/H}]=-2.19$; @Ji2016) and BD$+80^{\circ}$ 245 ($[\rm{Fe/H}]=-2.04$; @Ivans2003 [@Roederer2014b]). For the elements through Zn, [J0937$-$0626]{}’s abundance ratios are similar to BD$+80^{\circ}$ 245. [DES J0335$-$5403 ]{}shows a similar abundance pattern, with the exception of Ca, Ti, Co, and Zn, which all have higher \[X/Fe\] ratios than [J0937$-$0626]{}.
![\[X/Fe\] ratios for each element in [J0937$-$0626 ]{}(red stars), [DES J0335$-$5403 ]{}(blue diamonds; @Ji2016), and BD$+80^{\circ}$ 245 (green squares; @Ivans2003 for Na, Si, La, and Eu, @Roederer2014b for all other elements). NLTE-corrected Na abundances for [J0937$-$0626 ]{}and [DES J0335$-$5403 ]{}are shown with open symbols; NLTE-corrected values are not shown for Mn and Co. The neutron-capture elements are on the right, separated by a dotted line. \[fig:DelXFe\]](J0937-0626DelXFe.eps)
To investigate the source of [J0937$-$0626]{}’s unusual abundance ratios, the procedure from @McWilliam2018, who performed a re-analysis of UMi COS 171, is followed. First, note that the removal of $\sim 0.6$ dex of Fe from [J0937$-$0626 ]{}would shift its $[\alpha/\rm{Fe}]$ ratios to normal values—this shift would lead to a lower metallicity of $[\rm{Fe/H}]=-2.50$. The specific yields from the event(s) that created [J0937$-$0626]{}’s unusual abundance pattern can then be investigated through a comparison with a star at $[\rm{Fe/H}]=-2.50$ that formed in the same environment. @McWilliam2018 used a more metal-poor star in UMi; however, such a star cannot be confidently identified for [J0937$-$0626]{}. Instead, the average abundance pattern of MW stars with $-2.6~\le~[\rm{Fe/H}]~\le~-2.4$ is utilized. To make this comparison as homogeneous as possible, the average abundances of the MW stars in this metallicity range from @Sakari2018b are used. These stars have been analyzed with the same techniques as [J0937$-$0626]{}, using the same $<$3D$>$, NLTE corrections to lines, the same model atmospheres, and the same line lists. (Though note that none of the stars in @Sakari2018b are as hot as [J0937$-$0626]{}.) These average values are also shown in Figures \[fig:alphaFe\] and \[fig:FePeakFe\]. Note that the NLTE corrections to the Mn and Co abundances will differ slightly between $[\rm{Fe/H}]=-1.86$ and $[\rm{Fe/H}]=-2.5$—however, the MPIA NLTE correction database indicates that the corrections are $\la 0.1$ dex higher at $[\rm{Fe/H}]=-2.5$ for both Mn and Co [@BergemannGehren2008; @Bergemann2010].
Figure \[fig:DelLogEps\] then shows the differences in the $\log\epsilon$ abundance between [J0937$-$0626 ]{}and the average MW values at $[\rm{Fe/H}]=-2.5$. This comparison indicates that one or more nucleosynthetic events have significantly enriched [J0937$-$0626 ]{}in Cr, Mn, Fe, and Ni, with minor enhancement in Ca, Sc, Ti, and Zn.
![Differences in $\log \epsilon$ abundance between [J0937$-$0626 ]{} and a typical MW star at $[\rm{Fe/H}] = -2.5$ (the Fe abundance that [J0937$-$0626 ]{}would need to have normal $[\alpha/\rm{Fe}]$ ratios). and are chosen to represent Ti and Cr, to minimize NLTE effects. Note that LTE abundances are shown, though both Mn and Co require NLTE corrections; however, these corrections differ by $\la -0.1$ between stars at $[\rm{Fe/H}] = -1.86$ and $[\rm{Fe/H}] = -2.5$, according to the MPIA NLTE database. \[fig:DelLogEps\]](J0937-0626DelLogEps.eps)
Potential Explanations for Light, $\alpha$, and Fe-peak Abundances {#subsec:Scenarios}
------------------------------------------------------------------
[J0937$-$0626 ]{}therefore either shows an enhancement in some Fe-peak elements relative to the light and $\alpha$-elements, or a relative deficiency in the light and $\alpha$-elements. Three possible scenarios to explain this abundance pattern are considered. First, the abundances could reflect an evolutionary effect, such as radiative levitation. Secondly, the entire abundance pattern could be representative of enrichment from a single object. Finally, the abundance profile could be due to multiple progenitors, as a consequence of extended star formation. These possibilities are addressed below.
Radiative levitation in hot horizontal branch stars has been shown to enhance Fe-peak abundances by large amounts (up to 3 dex; @Behr2003). Reproducing the observed \[X/Fe\] ratios would require levitation only for the Fe-peak (and neutron-capture) elements. Furthermore, significant abundance differences have only been observed in the hottest horizontal branch stars, with temperatures above $\sim~11,000$ K [@Lovisi2012; @Tailo2017]; at 5875 K, [J0937$-$0626 ]{}is not expected to experience significant radiative levitation. Such effects are also not seen in other field horizontal branch stars (e.g., HD 222925; @Roederer2018b). This scenario therefore seems unlikely to explain the abundance pattern in [J0937$-$0626]{}.
Enrichment by a single source has also been invoked as an explanation for $\alpha$-poor very metal-poor stars. Standard core-collapse supernovae are unlikely to produce sufficiently low \[$\alpha$/Fe\] ratios to match those in [J0937$-$0626]{}; however, more exotic supernovae can create unusual abundance signatures. @Aoki2014 found that a pair-instability (PISN) supernova could explain the abundance signature of SDSS J0018$-$0939, a star at $[\rm{Fe/H}]~\sim~-2.5$. They based this conclusion on the star’s low \[$\alpha$/Fe\] (with the exception of Si), \[C/Fe\], and \[Co/Fe\] ratios, as well as its strong odd-even effect (contrasting abundances in odd vs. even elements), though they do note that the predicted odd-even effect is stronger than observed. @Nagasawa2018 also explore the possibility that a PISN supernova enriched their three stars in Hor I (with metallicities ranging from $[\rm{Fe/H}]=-2.8$ to $-2.5$), finding that none of the PISN models can perfectly reproduce the abundance pattern. They note that the models do not match the observed Fe-peak abundances (particularly their solar \[Co/Fe\] ratios) and predict a strong odd-even effect that is not observed. [J0937$-$0626 ]{}also has low \[$\alpha$/Fe\] and \[Co/Fe\]—however, the @HegerWoosley2002 [@HegerWoosley2010] PISN models do not match all the abundance ratios. In particular, a PISN supernova cannot produce enough Sc or Zn, and produces a stronger odd-even effect than observed. It therefore seems unlikely that [J0937$-$0626 ]{}was enriched by a single PISN.
@Nishimura2017 and @TsujimotoNishimura2018 have noted that neutrino-heating in magneto-rotational supernovae may be a viable site for Zn production (along with Fe, Co, and Ni) in metal-poor environments. Using samples of stars in the Milky Way, @TsujimotoNishimura2018 argue that a high \[Zn/Mg\] ratio at very low metallicity indicates a high frequency of magneto-rotational supernovae; at higher metallicities, a high \[Zn/Mg\] may also reflect the onset of Type Ia supernovae (see their Figure 3). Given its high $[\rm{Zn/Mg}]=-0.05$ (compared to an average $[\rm{Zn/Mg}]~=~-0.40$ at $[\rm{Fe/H}]=-1.9$), it is tempting to speculate that [J0937$-$0626 ]{}may have been enriched by a magneto-rotational supernova; however, by $[\rm{Fe/H}] = -2.5$, the contributions from magneto-rotational supernovae are already expected to be decreasing. Instead, @TsujimotoNishimura2018 argue that the Zn enhancement in [DES J0335$-$5403 ]{}and a star in the Draco dwarf galaxy is due to Type Ia supernovae. Detailed yields from magneto-rotational supernovae are necessary to fully address this possibility.
@Kobayashi2014 have also argued that low \[$\alpha$/Fe\] in extremely metal-poor stars can be the nucleosynthetic result of $\sim10-20$ M$_{\sun}$ core-collapse supernovae or hypernovae. They suggest that hypernovae would produce a high $[\rm{Zn/Fe}]>0.3$, which is not observed in [J0937$-$0626]{}. The abundance pattern for $\sim10-20$ M$_{\sun}$ supernovae also does not quite match the pattern in [J0937$-$0626]{}, particularly the high $\alpha$ and Zn and the pattern of Fe-peak elements. Core-collapse supernovae on their own are therefore not a likely source of the abundance patterns in [J0937$-$0626]{}. It is also worth noting that [J0937$-$0626 ]{}is more metal-rich than the other targets whose abundances were explained by a single progenitor.
A more likely explanation for the abundance patterns in Figures \[fig:DelXFe\] and \[fig:DelLogEps\] is that [J0937$-$0626]{}’s host environment experienced extended star formation and chemical enrichment, with core-collapse supernovae building the metallicity up to $[\rm{Fe/H}] = -2.5$ before a second event produced a significant amount of Cr, Mn, Fe, and Ni and a smaller amount of Ca, Sc, Ti, and Zn. The most likely option is enrichment from a Type Ia supernova, which are known to produce Fe-peak elements (e.g., @Iwamoto1999 [@Badenes2003; @Badenes2008]). The precise yields depend on parameters such as white dwarf mass, metallicity, and the physics of the explosion.
Figures \[fig:IaYields\] and \[fig:subChIaYields\] show comparisons between [J0937$-$0626]{}’s \[X/Fe\] abundance ratios and Ia yields, added to the background MW average at $[\rm{Fe/H}]\sim -2.5$. The Chandrasekhar-mass “DDTa” model from @Badenes2003 [@Badenes2008] is shown in Figure \[fig:IaYields\], using the yields from @McWilliam2018, for five different metallicities. (Note that the other models overpredict \[Mn/Fe\].) Though the agreement for $Z=0.0025$ is generally decent, none of these metallicities can perfectly reproduce the pattern in the Fe-peak elements. At all metallicities, this Ia model also overproduces Si and Ca and (except for the highest metallicity model) underproduces Ti. Figure \[fig:subChIaYields\] then shows sub-Chandrasekhar mass models from E. Bravo, with the yields from @McWilliam2018, for two white dwarf masses: 1.06 M$_{\sun}$ and 1.15 M$_{\sun}$. For UMi COS 171, @McWilliam2018 found that a sub-Chandrasekhar mass model provided a better fit to the abundances, particularly the low \[Mn/Fe\] and \[Ni/Fe\]. Indeed, there is decent agreement with Mn and Co in [J0937$-$0626 ]{}for both sub-Chandrasekhar masses. As with the Ia case, both sub-Chandrasekhar models overpredict the amount of Si and Ca.
Though no model perfectly matches the pattern in [J0937$-$0626]{}, its abundances are generally consistent with enrichment from a Type Ia supernova, possibly one with a sub-Chandrasekhar mass. However, the precise Ia yields depend on the physical conditions of the models (e.g., explosion energy). Similarly, if the background composition of [J0937$-$0626]{}’s birth environment was different from the MW composition (e.g., if the Si and Ca abundances were lower at $[\rm{Fe/H}]~=~-2.5$), then these predicted yields would also change. The general enhancement in Fe-peak elements supposts enrichment from a Type Ia supernova.
It is worth noting that any of the proposed scenarios for enrichment in Fe-peak elements are unlikely to have created a significant number of neutron-capture elements. Neither Type Ia or PISN supernovae will create $r$-process elements (e.g., @HegerWoosley2002), while standard core-collapse supernovae have been ruled out as a significant source of $r$-process elements (e.g., @ArconesThielemann2013). Though magneto-rotational supernovae have been identified as possible sources of both Fe-peak and $r$-process elements, @Nishimura2017 showed that the supernovae that produce significant amounts of Fe, Ni, and Zn do not produce much Eu, and vice versa. The enhancement of neutron-capture elements therefore likely requires enrichment by a separate event.
![Comparisons of “DDTa” Ia yields from @Badenes2003 [@Badenes2008 as reported by @McWilliam2018] with the [J0937$-$0626 ]{}abundances (red stars). The Ia yields are added to the average MW abundances at $[\rm{Fe/H}] \sim -2.5$ (black circles) so that the \[Fe/H\] is increased to [J0937$-$0626]{}’s value. The five lines show the Ia yields for five different metallicities. Note that the yield patterns change when the physics of the explosion are altered; the “DDTa” model represents the best fit to [J0937$-$0626]{}’s abundances. \[fig:IaYields\]](DDTaPlots.eps)
r-Process Enhancement and Patterns {#subsec:Patterns}
----------------------------------
Unlike many of the light, $\alpha$, and Fe-peak elements, [J0937$-$0626 ]{}shows solar or supersolar \[X/Fe\] ratios for the neutron-capture elements. At $[\rm{Eu/Fe}] = 0.85\pm0.06$, [J0937$-$0626 ]{}is an $r$-I star; its low $[\rm{Ba/Eu}]=-0.77\pm0.07$ implies that its Eu enhancement is due to the $r$-process. Note that red horizontal branch stars have been discovered to be $r$-process enhanced [@Roederer2014n], including HD 222925 [@Roederer2018b], so this is not a unique feature of [J0937$-$0626]{}. Sr, Y, and Zr are also roughly solar, while La, Ce, Pr, Nd, Gd, and Dy are enhanced. Figure \[fig:rProcPatterns\] shows that [J0937$-$0626]{}’s neutron-capture abundance pattern is generally consistent with the $r$-process residual in the Sun and two stars in Reticulum II (though, like Ret II, Sr, Y, and Zr in [J0937$-$0626 ]{}are slightly lower than the solar residual, as discussed in @Ji2016). The pattern is inconsistent with the solar $s$-process. The low \[Sr/Ba\] also indicates that significant contributions from the weak $s$-process in rapidly rotating massive stars (e.g., @Chiappini2011 [@Frischknecht2012; @Cescutti2013; @Frischknecht2016]) are unlikely. The upper limit in Th also implies $\log\epsilon(\rm{Th/Eu})~<~-0.3$.
Figure \[fig:NCFe\] shows [J0937$-$0626]{}’s slightly enhanced \[Ba/Fe\] and strongly enhanced \[Eu/Fe\], relative to the “normal” and “Fe-enhanced” MW field stars. The $r$-process enhancement in [J0937$-$0626 ]{}makes it unlike most of the other “Fe-enhanced” metal-poor stars, whose low \[X/Fe\] ratios persist through the neutron-capture elements (including BD$+80^{\circ}$ 245; see Figure \[fig:DelXFe\]).[^10] Instead, the $r$-process enhancement in [J0937$-$0626 ]{}more closely resembles [DES J0335$-$5403]{}, the Ret II star, which is also an $r$-I star.
Section \[subsec:MWComp\] demonstrated that many of the \[X/Fe\] ratios could be brought into agreement with Milky Way stars by removing 0.6 dex of Fe. The removal of 0.6 dex of Fe would increase the \[X/Fe\] ratios of the $r$-process elements, as shown by the maroon star in Figure \[fig:NCFe\]. If the $r$-process event occurred prior to the Fe-peak event, then [J0937$-$0626 ]{}would have been an $r$-II star if the Fe-peak event had not occurred. Furthermore, if [J0937$-$0626 ]{}originated in a dwarf galaxy (see Section \[subsec:Kinematics\]), this dwarf galaxy would likely have contained a population of highly $r$-process-enhanced stars, similar to Ret II. Indeed, though none have been linked to [J0937$-$0626]{}, many of the $r$-II stars in the MW have been kinematically identified as probable captures from dwarf galaxies [@Roederer2018; @Sakari2018a; @Sakari2018b], hinting that many $r$-II stars have originated in $r$-process-enhanced dwarf galaxies like Ret II.
Kinematics {#subsec:Kinematics}
----------
Figure \[fig:Toomre\] shows a Toomre diagram of MW field stars from [*Gaia*]{} DR2 (using the halo stars within 1 kpc from @Koppelman2018), along with distinctions between prograde and retrograde orbits. Its velocities (derived with the `gal_uvw` code[^11]) show that [J0937$-$0626 ]{}resides in the MW halo, with a retrograde orbit. Previous work has found that a significant number of MW $r$-I and $r$-II stars have retrograde orbits [@Roederer2018; @Sakari2018a; @Sakari2018b] and may have been accreted from satellite galaxies. [J0937$-$0626]{}’s kinematics also suggest that it may also have been accreted from a satellite galaxy.
@Koppelman2018 and @Roederer2018 have also identified specific groups of stars with similar kinematics which may have originated in the same galaxy. Recently, @Helmi2018 argued that the majority of the retrograde stars from Koppelman et al.’s analysis are due to a single merger event from a galaxy with a mass slightly higher than the Small Magellanic Cloud, which they named Gaia-Enceladus. They also found that the Gaia-Enceladus stars have slightly lower \[$\alpha$/Fe\] ratios than MW stars (also see @NissenSchuster2010 and @Hayes2018). [J0937$-$0626 ]{}lies approximately in the correct kinematic space for Gaia-Enceladus stars; however, its \[Fe/H\] and \[$\alpha$/Fe\] ratios are lower than the majority of the Gaia-Enceladus stars. It is still possible that [J0937$-$0626 ]{}was brought in by the same merger event if it experienced inhomogeneous mixing within the larger galaxy (similar to the scenario proposed by @Venn2012 for Carina). Full orbital calculations will also be essential for identifying [J0937$-$0626]{}’s birth site and locating other stars from the same environment.
@Roederer2018 examined the kinematics of 35 $r$-II stars with high-quality [*Gaia*]{} data, and identified several groups with similar orbits and metallicities. This technique could be used to identify other stars from the same birth environment as [J0937$-$0626]{}: its chemistry and kinematics should be similar to other $r$-I stars from the same birth environment; similarly, if that environment was enriched in $r$-process elements before the event that created the Fe-peak enrichment, [J0937$-$0626 ]{}should have similar kinematics as more metal-poor $r$-II stars. RPA discoveries of more $r$-I and $r$-II stars, combined with future [*Gaia*]{} data, will identify other stars that could have originated in the same environment as [J0937$-$0626]{}.
Conclusions {#sec:Conclusions}
===========
[RAVE J093730.5$-$062655]{}$\;$ is a moderately $r$-process-enhanced ($[\rm{Eu/Fe}]~=~+~0.85\pm0.1$), metal-poor ($[\rm{Fe/H}]~=~-1.86$) horizontal branch star on a retrograde orbit in the MW halo that was identified by the RPA. Most of its \[X/Fe\] abundance ratios are distinct from those of typical MW field stars, particularly its subsolar $[\alpha/\rm{Fe}]$ (e.g., $[\rm{Mg/Fe}]~=~-~0.25\pm0.04$, $[\rm{Ca/Fe}]~=~-0.18\pm0.03$), light element ($[\rm{Na/Fe}]~=~-0.25\pm0.13$) and some Fe-peak ratios (e.g., $[\rm{Ni/Fe}]~=~-0.18\pm0.06$). [J0937$-$0626 ]{}seems to have the abundance pattern typical of a “normal” MW star at $[\rm{Fe/H}]\sim-2.5$ that was diluted by ejecta from an event that created $\sim 0.6$ dex of Fe-peak elements. Although none of the models perfectly fit the abundance patterns in [J0937$-$0626]{}, the best candidate for this Fe-peak enrichment is a Type Ia supernova.
[J0937$-$0626]{}’s $r$-process enrichment is unlikely to have been caused by a Type Ia supernova. Instead, its birth environment may have been enhanced in $r$-process elements prior to the enrichment from the Type Ia supernova; [J0937$-$0626 ]{}therefore could have been an $r$-II star were it not for the occurrence of the Type Ia supernova. In this sense, [J0937$-$0626 ]{}may be similar to the metal-rich $r$-I star in Reticulum II. Ultimately, [J0937$-$0626]{}’s chemical abundances and kinematics indicate that it was likely accreted from a satellite dwarf galaxy. [J0937$-$0626]{}’s host galaxy may have been responsible for depositing other stars into the MW halo, possibly even more metal-poor $r$-II stars. Additional discoveries of $r$-I and $r$-II stars by the RPA, combined with proper motions and parallaxes from [*Gaia*]{}, will enable specific subgroups to be identified in the future.
The authors thank the current and previous observing specialists on the 3.5-m telescope at Apache Point Observatory for their continued help and support. The authors thank the anonymous referee for helpful comments which improved this manuscript. The authors also thank Anish Amarsi and Karin Lind for providing the NLTE grids and assisting with their usage, and Helmer Koppelman for providing the 1 kpc [*Gaia*]{} sample.
C.M.S. and G.W. acknowledge funding from the Kenilworth Fund of the New York Community Trust. V.M.P., T.C.B., R.E., A.F., and I.U.R. acknowledge partial support from grant PHY 14-30152 (Physics Frontier Center/JINA/CEE), awarded by the US National Science Foundation. I.U.R. acknowledges additional support from NSF grants AST 16-13536 and AST 18-15403. BKG acknowledges the support of STFC, through the University of Hull’s Consolidated Grant ST/R000840/1. EKG gratefully acknowledges support via Sonderforschungsbereich SFB 881 “The Milky Way System” (subproject A5) of the German Research Foundation (DFG). Funding for RAVE has been provided by: the Leibniz-Institut für Astrophysik Potsdam (AIP); the Australian Astronomical Observatory; the Australian National University; the Australian Research Council; the French National Research Agency; the German Research Foundation (SPP 1177 and SFB 881); the European Research Council (ERC-StG 240271 Galactica); the Istituto Nazionale di Astrofisica at Padova; The Johns Hopkins University; the National Science Foundation of the USA (AST-0908326); the W. M. Keck foundation; the Macquarie University; the Netherlands Research School for Astronomy; the Natural Sciences and Engineering Research Council of Canada; the Slovenian Research Agency; the Swiss National Science Foundation; the Science & Technology Facilities Council of the UK; Opticon; Strasbourg Observatory; and the Universities of Groningen, Heidelberg and Sydney. The RAVE web site is at <https://www.rave-survey.org>.
This research is based on observations obtained with the Apache Point Observatory 3.5-meter telescope, which is owned and operated by the Astrophysical Research Consortium. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. This work has also made use of data from the European Space Agency (ESA) mission [*Gaia*]{} (<http://www.cosmos.esa.int/gaia>), processed by the [*Gaia*]{} Data Processing and Analysis Consortium (DPAC, <http://www.cosmos.esa.int/web/gaia/dpac/consortium>). Funding for the DPAC has been provided by national institutions, in particular the institutions participating in the [*Gaia*]{} Multilateral Agreement.
[^1]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
[^2]: <https://github.com/alexji/moog17scat>
[^3]: <http://kurucz.harvard.edu/linelists.html>
[^4]: <https://github.com/vmplacco/linemake>
[^5]: <http://kurucz.harvard.edu/sun.html>
[^6]: <http://inspect-stars.com/>
[^7]: <http://nlte.mpia.de>
[^8]: Note that [J0937$-$0626 ]{}is more iron- and $\alpha$-poor than the low-$\alpha$ disk stars (e.g., @Helmi2018), as will be discussed in Section \[subsec:Kinematics\].
[^9]: The low Sc in [J0937$-$0626 ]{}is also somewhat reminiscent of the metal-poor bulge stars observed by @CaseySchlaufman2015, though those stars have otherwise normal \[X/Fe\] ratios.
[^10]: Note that though there are $r$-process-enhanced stars in UMi, COS 171 is not $r$-process-enhanced.
[^11]: <https://github.com/segasai/astrolibpy/blob/master/astrolib/gal_uvw.py>
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'M. Zechmeister'
- 'M. Kürster'
- 'M. Endl'
bibliography:
- 'paper.bib'
date: 'Received / Accepted'
subtitle: A search for terrestrial planets in the habitable zone of M dwarfs
title: 'The M dwarf planet search programme at the ESO VLT + UVES [^1] [^2] '
---
Introduction
============
High-precision differential radial velocity (RV) measurements of stellar reflex motions induced by an orbiting companion have so far been the most successful method to discover extrasolar planets and to characterise their orbital properties. Originally, RV planet search programmes have largely concentrated on main sequence stars of spectral types late-F through K. That only a comparatively small number of M dwarfs were included comes from their faintness, which requires large telescopes to perform high-precision RV measurements of a few m/s. For an understanding of the formation and abundance of extrasolar planets it is important to determine the presence and orbital characteristics of planets around stars of as many different types as possible, and especially around this most abundant type of star.
Even if M dwarfs are faint and require large telescopes, however, they have two advantageous characteristics when searching for terrestrial exoplanets in the habitable zones (HZ) with radial velocity (RV) methods. Compared to solar-like stars, (i) they have a lower mass ($M=0.1-0.6M_{\odot}$), and (ii) the habitable zone is close-in around this cooler and less luminous type of star ($L=0.0008-0.06L_{\odot}$, @Guinan09). By habitable zone we understand the region that allows liquid water on the planet surface as described in @Kasting93.
For M dwarfs, the HZ is typically 0.03 – 0.4 AU. The RV amplitude induced due to a planet by the Doppler effect is $$K=\sqrt{\frac{G}{1-e^{2}}}\frac{m\sin i}{\sqrt{(M+m)a}}=28.4\mathrm{m/s}\cdot\frac{m\sin i}{M_{\mathrm{Jup}}}\left(\frac{M_{\odot}}{M}\frac{\mathrm{AU}}{a}\right)^{1/2}.\label{eq:Amplitude}$$ It increases for closer distances $a$ (shorter periods) and lower stellar mass $M$. Thus the RV amplitude induced by a planet in the HZ of an M dwarf is higher than that of a solar-like star. M dwarfs are also ideal targets for astrometric follow-up due to their lower mass, as well as for transit observations. In combination with the RV method, astrometry allows the resolution of the $\sin i$-ambiguity and true masses to be obtained.
Star b c d e References
-------- ------------------------ ------------------------ ------------------ ----------------- ------------------------------------------------------
GJ 876 $2.53M_{\mathrm{Jup}}$ $0.79M_{\mathrm{Jup}}$ $7.53M_{\oplus}$ [\[]{}1[\]]{}[\[]{}2[\]]{}[\[]{}3[\]]{}[\[]{}4[\]]{}
GJ 581 $15.6M_{\oplus}$ $5.1M_{\oplus}$ $8.2M_{\oplus}$ $1.9M_{\oplus}$ [\[]{}5[\]]{}[\[]{}6[\]]{}[\[]{}7[\]]{}
GJ 436 $21M_{\oplus}$ [\[]{}8[\]]{}
GJ 317 $0.71M_{\mathrm{Jup}}$ [\[]{}9[\]]{}
GJ 674 $11.1M_{\oplus}$ [\[]{}10[\]]{}
GJ 849 $0.82M_{\mathrm{Jup}}$ [\[]{}11[\]]{}
GJ 176 $8.4M_{\oplus}$ [\[]{}12[\]]{}[\[]{}13[\]]{}
GJ 832 $0.64M_{\mathrm{Jup}}$ [\[]{}14[\]]{}
: \[Tab:Mdwarfplanets\]M dwarfs with known planets and their masses discovered with the RV method.
Note: The masses are minimum masses with the exception for GJ 436 and GJ 876. [\[]{}1[\]]{} @Delfosse98, [\[]{}2,3[\]]{} @Marcy98 [@Marcy01], [\[]{}4[\]]{} @Rivera05, [\[]{}5[\]]{} @Bonfils05, [\[]{}6[\]]{} @Udry07, [\[]{}7[\]]{} @Mayor09, [\[]{}8[\]]{} @Butler04, [\[]{}9[\]]{} @Johnson07, [\[]{}10[\]]{} @Bonfils07, [\[]{}11[\]]{} @Butler06, [\[]{}12[\]]{} @Forveille09, [\[]{}13[\]]{} @Butler09, [\[]{}14[\]]{} @Bailey09.
At present there are only a few M dwarfs known to have planets, as summarized in Table \[Tab:Mdwarfplanets\]. The M4V star GJ 876 has two Jovian planets orbiting in a 2:1 resonance (@Marcy01; see also @Benedict02 for an astrometric determination of the mass of the outer planet), and a third planet in this system has been found by @Rivera05. The M3V star GJ 581 is another multiple system with three Neptune-type planets. The latest detection of a fourth planet in this system by @Mayor09 is a new highlight in planet search with RVs. With only $1.9M_{\oplus}$, GJ 581 e is the lowest-mass planet found so far with the RV method. The same work also resulted in a revised period for the outer planet GJ 581 d, placing it inside the habitable zone. The planet around the M2.5V star GJ 436 first discovered with the RV method [@Butler04] also turned out to be a transiting one [@Gillon07]. So far it is the only known transiting planet around an M dwarf. Most recently, @Bailey09 has discovered the first long-period planet around an M dwarf (GJ 832). A planet with a mass of $24.5M_{\oplus}$ ($P=10.24\,$d) around GJ 176 announced by @Endl08 is rejected by @Forveille09 and @Butler09. Instead, both groups find evidence of another planet with a shorter period (8.78 d) and a minimum mass of $8.4M_{\oplus}$ [@Forveille09] and $12M_{\oplus}$ @Butler09.
Several radial velocity (RV) surveys of M dwarfs have resulted in few or no detections, indicating a lower frequency of planets compared to solar-like stars; e.g., @Endl06 monitored 90 M dwarfs (including the first data for 21 stars from our sample) without a planet detection. The sample studied by @Cumming08 surveying 110 M dwarfs contained only two planet hosting M dwarfs (GJ 876 and GJ 436). HARPS guaranteed time project \#3 had 50 nights on 120 M dwarfs within 11pc, and so far it has revealed only three planet-hosting M dwarfs (GJ 876, GJ 581, and GJ 674).
M dwarfs are an ideal testing ground for competing models of the formation of gas giants. While the classical core-accretion model has severe problems with forming Jupiter-mass planets in the less massive protoplanetary disks even around M dwarfs (e.g. @Laughlin04), the competing gravitational instability model can also efficiently form Jovian-type companions around M dwarfs [@Boss06]. @Ida05 even predict a higher frequency of icy giant planets with masses comparable to Neptune in short periodic orbits for M dwarfs than for G type stars.
Recent results of microlensing surveys (e. g. OGLE-2005-BLG-390b, $5.5M_{\oplus}$, @Beaulieu06; OGLE-2005-BLG-169b, @Gould06; presumed M dwarfs) may also indicate that low-mass planetary companions might be abundant around M dwarfs.
Targets and Observations
========================
Our sample consists of 40 M dwarfs and one M giant[^3], which are listed in Table \[Tab:Sample\]. All M dwarfs are brighter than $V\lesssim12.2\,\mathrm{mag}$ and nearby within a distance of 37pc (33 M dwarfs even within 20pc). Their spectral types range from M0 to M5. The stellar masses $M$ were derived from the mass-luminosity relation by @Delfosse00 using the absolute brightness in K band. As an indicator of activity we selected the X-ray luminosity as detected in the ROSAT all-sky survey [from @Huensch99]. Detected as X-ray sources are Barnard’s star ($L_{x}=0.1$), GJ 1 (0.6), GJ 190 (0.9), GJ 229 (1.3), and the most active Proxima Cen (1.7). The rest of the sample was not detected by ROSAT, implying that these stars are inactive.
The observations started in 2000, initially with 20 targets. Typically, three consecutive spectra per night were taken for each star with exposure times of 90 – 900s depending on the object brightness. In April 2004 (JD=2453100) the sample was complemented by 21 additional stars, while the monitoring of HG 7-15 was ended. Since then the number of spectra per night was reduced to one, with the exception of Barnard’s star, GJ 160.2, GJ 821, and Proxima Cen.
Our data of the first 20 stars (+ GJ 510) taken before mid-2005 were already included in the study by @Endl06. Here we present the full data set for all 41 stars as observed until March 2007. A detailed study of the full Proxima Cen data set can be found in @Endl08b. A study of the pre-2002.75 data set of Barnard’s star is given by @Kuerster03.
Star Spec Type $M$ [\[]{}$M_{\odot}$[\]]{}
------------------- ----------- ------- -------- -----------------------------
Barnard (GJ 699) M4Ve 9.54 1.82 0.16
GJ 1 M1.5 8.57 4.34 0.45
GJ 27.1 M0.5 11.42 23.99 0.53
GJ 118 M2.5 10.70 11.65 0.36
GJ 160.2 M0V 9.69 23.12 0.69
GJ 173 M1.5 10.35 11.10 0.48
GJ 180 M2 12.50 12.12 0.43
GJ 190 M3.5 10.31 9.27 0.44
GJ 218 M1.5 10.72 15.03 0.50
GJ 229 M1/M2V 8.14 5.75 0.58
GJ 263 M3.5 11.29 16.02 0.55
GJ 357 M2.5V 10.85 9.02 0.37
GJ 377 M3 11.44 16.29 0.52
GJ 422 M3.5 11.66 12.67 0.35
GJ 433 M1.5 9.79 8.88 0.48
GJ 477 M1 11.08 18.99 0.54
GJ 510 M1 11.05 16.74 0.49
GJ 620 M0 10.25 16.44 0.61
GJ 637 M0.5 11.36 15.88 0.41
GJ 682 M3.5 10.96 5.08 0.27
GJ 739 M2 11.14 14.09 0.45
GJ 817 M1 11.48 19.17 0.43
GJ 821 M1 10.87 12.17 0.44
GJ 842 M0.5 9.74 11.99 0.58
GJ 855 M0.5 10.74 19.15 0.60
GJ 891 M2V 12.20 16.08 0.35
GJ 911 M0V 10.88 24.26 0.63
GJ 1009 M1.5 11.16 17.98 0.56
GJ 1046 M2.5+v 11.62 14.07 0.40
GJ 1100 M0 11.48 28.93 0.57
GJ 3020 M2.5 11.54 22.78 0.62
GJ 3082 M0 11.10 16.56 0.47
GJ 3098 M1.5Vk: 11.21 17.86 0.50
GJ 3671 M0 11.20 17.74 0.50
GJ 3759 M1V 10.95 16.97 0.49
GJ 3916 M2.5V 11.25 15.10 0.49
GJ 3973 M1.5Vk: 10.94 18.23 0.54
GJ 4106 M2 10.82 110.50 0.55
GJ 4293 M0.5 10.90 25.06 0.57
HG 7-15 M1V 10.85 37.31 0.78
Prox Cen (GJ 551) M5.5Ve 11.05 1.30 0.12
: \[Tab:Sample\]Targets with their spectral type, visual magnitude $V$, distance $d$ [@vanLeeuwen07], and stellar mass $M$ derived from the K-band mass-luminosity relation by @Delfosse00.
{height="0.96\textheight" width="1.\linewidth"}
{width="1\linewidth"}
The observations were carried out with the UVES spectrograph at the VLT-UT2, directly fed via image slicer \#3 that redistributes the light from a 1$\arcsec\times$1$\arcsec$ aperture along a 0.3$\arcsec$ wide slit. This resulted in a resolving power of $R=100\,000-120\,000$. The red arm of UVES was employed and a wavelength coverage 495 – 704nm over 37 orders was selected. An iodine cell was used for precise wavelength calibration and modelling of the instrumental profile. Only the range of 500 – 600nm was used to determine the RVs. This range is rich in iodine lines.
The data reduction included bias subtraction, flat-fielding, Echelle straylight subtraction, wavelength calibration, and barycentric correction [see also @Endl08b]. The data modelling with the “AUSTRAL” code to obtain the RV is described in @Endl00. We achieve an RV precision of 2m/s for bright stars. In practice, photon noise limits the RV precision for faint M dwarfs and correspondingly the errors are larger for those stars that were observed with lower S/N. Radial velocities and also their errors (see @Kuerster03 for discussion) were combined into nightly averages. They are shown in Figs. \[Fig:RV1\] and \[Fig:RV2\] (the RV data are available as online material).
Data analysis
=============
Secular acceleration
--------------------
Even if a star moves undisturbed with a constant space velocity $v$, it can show a change in its radial velocity (RV). This secular or perspective acceleration was first measured by @Kuerster03 for Barnard’s star. In polar coordinates the radial and tangential velocity component, respectively, are (Fig. \[Fig:SecAcc\]) $$v_{\mathrm{r}}=-v\cos\varphi\quad\text{and}\quad v_{\mathrm{t}}=v\sin\varphi.$$ The differentiation of $v_{\mathrm{r}}$ with respect to time $t$ yields the secular acceleration ($v=\mathrm{const}.$) $$\dot{v}_{\mathrm{r}}=\frac{\mathrm{d}v_{\mathrm{r}}(t)}{\mathrm{d}t}=v\sin\varphi\cdot\dot{\varphi}=v_{\mathrm{t}}\dot{\varphi}.$$ The tangential velocity $v_{\mathrm{t}}=d\cdot\mu$ depends on the distance $d$ of the star and its proper motion $\mu$, which can also be identified with the time derivative of the angular position $\dot{\varphi}=\mu$. Therefore the instantaneous secular acceleration is given by $$\frac{\mathrm{d}v_{\mathrm{r}}(t)}{\mathrm{d}t}=\frac{v_{\mathrm{t}}^{2}}{d}=\mu^{2}d=22.98\frac{\mathrm{m/s}}{\mathrm{yr}}\frac{(\mu_{\alpha}^{2}+\mu_{\delta}^{2})\cdot\mathrm{yr}^{2}/\mathrm{arcsec}^{2}}{\pi/\mathrm{mas}}.\label{eq:SecuAcc}$$ where $\mu_{\alpha}$ and $\mu_{\delta}$ are the proper motion in right ascension and declination, respectively. It only depends on the proper motion $\mu$ and parallax $\pi$, which are easily accessible from the Hipparcos catalogue. Note that the knowledge of $v_{\mathrm{r}}$ and $v$ is *not* explicitly required for the prediction of the instantaneous secular acceleration.
![\[Fig:SecAcc\]Change in the radial velocity $v_{r}$ due to constant motion $v$ (secular acceleration $\dot{v}_{r}$).](12479f03.eps){width="1\linewidth"}
One has to take this effect into account, especially for high proper motion stars, to avoid the misleading conclusion that this RV change is a disturbance by a companion. A feature of the secular acceleration is that this effect is only geometrical, not physical, and always positive ($\dot{v}_{\mathrm{r}}\ge0$). With the high astrometric precision of Hipparcos, this effect can be predicted with high accuracy for nearby stars.
----------- ------------------- ------------------- ----------------- ------------------
Star
Barnard -798.58$\pm$1.72 10328.12$\pm$1.22 548.31$\pm$1.51 4.497$\pm$0.012
GJ 1 5634.68$\pm$0.86 -2337.71$\pm$0.71 230.42$\pm$0.90 3.711$\pm$0.015
GJ 27.1 485.57$\pm$2.81 -223.13$\pm$2.11 41.69$\pm$2.80 0.157$\pm$0.011
GJ 118 978.61$\pm$2.27 633.21$\pm$2.64 85.87$\pm$1.99 0.364$\pm$0.008
GJ 160.2 51.90$\pm$1.24 -780.04$\pm$1.57 43.25$\pm$1.61 0.325$\pm$0.012
GJ 173 -225.75$\pm$1.94 -192.57$\pm$1.88 90.10$\pm$1.74 0.022$\pm$0.000
GJ 180 408.07$\pm$2.49 -642.82$\pm$2.06 82.52$\pm$2.40 0.161$\pm$0.005
GJ 190 502.99$\pm$1.32 -1399.76$\pm$1.52 107.85$\pm$2.10 0.471$\pm$0.009
GJ 218 768.13$\pm$1.39 -123.17$\pm$1.54 66.54$\pm$1.43 0.209$\pm$0.005
GJ 229 -137.09$\pm$0.50 -713.66$\pm$0.81 173.81$\pm$0.99 0.070$\pm$0.000
GJ 263 -123.73$\pm$3.68 -814.92$\pm$2.87 62.41$\pm$3.16 0.250$\pm$0.013
GJ 357 136.67$\pm$1.53 -989.13$\pm$1.41 110.82$\pm$1.92 0.207$\pm$0.004
GJ 377 -1096.84$\pm$2.11 647.29$\pm$2.63 61.39$\pm$2.55 0.607$\pm$0.025
GJ 422 -2466.98$\pm$2.87 1180.09$\pm$2.17 78.91$\pm$2.60 2.178$\pm$0.072
GJ 433 -72.51$\pm$1.49 -851.92$\pm$0.88 112.58$\pm$1.44 0.149$\pm$0.002
GJ 477 -101.39$\pm$2.28 -697.57$\pm$1.75 52.67$\pm$3.05 0.217$\pm$0.013
GJ 510 -452.04$\pm$2.55 -104.79$\pm$1.55 59.72$\pm$2.43 0.083$\pm$0.003
GJ 620 -348.40$\pm$2.35 -675.73$\pm$1.87 60.83$\pm$2.06 0.218$\pm$0.007
GJ 637 -480.59$\pm$1.17 -529.12$\pm$1.86 62.97$\pm$1.99 0.186$\pm$0.006
GJ 682 -708.98$\pm$2.55 -937.40$\pm$1.88 196.90$\pm$2.15 0.161$\pm$0.002
GJ 739 153.46$\pm$2.91 -495.17$\pm$1.98 70.95$\pm$2.56 0.087$\pm$0.003
GJ 817 -918.58$\pm$3.68 -2038.13$\pm$2.50 52.16$\pm$2.92 2.202$\pm$0.123
GJ 821 713.47$\pm$2.82 -1994.64$\pm$0.95 82.18$\pm$2.17 1.255$\pm$0.033
GJ 842 888.07$\pm$1.31 -125.56$\pm$1.16 83.43$\pm$1.77 0.222$\pm$0.005
GJ 855 587.72$\pm$2.17 -376.75$\pm$1.32 52.22$\pm$2.17 0.214$\pm$0.009
GJ 891 717.43$\pm$3.40 22.12$\pm$2.69 62.17$\pm$3.27 0.190$\pm$0.010
GJ 911 -42.36$\pm$3.32 128.87$\pm$2.44 41.22$\pm$2.64 0.010$\pm$0.001
GJ 1009 62.95$\pm$2.45 -196.98$\pm$2.46 55.62$\pm$2.32 0.018$\pm$0.001
GJ 1046 1395.67$\pm$1.72 547.16$\pm$2.52 71.06$\pm$3.23 0.727$\pm$0.033
GJ 1100 118.29$\pm$2.67 -498.47$\pm$1.30 34.57$\pm$2.79 0.174$\pm$0.014
GJ 3020 -44.73$\pm$4.05 -237.17$\pm$3.70 43.89$\pm$4.39 0.030$\pm$0.003
GJ 3082 104.14$\pm$2.04 311.53$\pm$1.81 60.38$\pm$1.81 0.041$\pm$0.001
GJ 3098 -589.19$\pm$1.86 -887.83$\pm$1.24 55.98$\pm$1.91 0.466$\pm$0.016
GJ 3671 -603.54$\pm$1.64 -296.43$\pm$1.36 56.38$\pm$2.04 0.184$\pm$0.007
GJ 3759 -391.22$\pm$1.50 -411.01$\pm$1.71 58.94$\pm$2.40 0.126$\pm$0.005
GJ 3916 -332.19$\pm$2.90 -352.83$\pm$2.62 66.21$\pm$3.18 0.082$\pm$0.004
GJ 3973 -9.48$\pm$2.45 -221.00$\pm$1.72 54.86$\pm$2.18 0.020$\pm$0.001
GJ 4106 31.55$\pm$4.63 -104.82$\pm$3.59 9.05$\pm$3.70 0.030$\pm$0.012
GJ 4293 198.08$\pm$2.42 -113.57$\pm$2.08 39.90$\pm$3.04 0.030$\pm$0.002
HG 7-15 176.02$\pm$2.85 5.73$\pm$1.79 26.80$\pm$2.05 0.027$\pm$0.002
Prox Cen -3775.75$\pm$1.63 765.54$\pm$2.01 771.64$\pm$2.60 0.442$\pm$0.002
GJ 451 A 4003.98$\pm$0.37 -5813.62$\pm$0.23 109.99$\pm$0.41 10.411$\pm$0.039
GJ 9511 B -999.75$\pm$1.29 -3542.60$\pm$1.13 35.14$\pm$1.48 8.861$\pm$0.373
GJ 9511 A -997.47$\pm$1.20 -3543.55$\pm$1.03 34.65$\pm$1.28 8.988$\pm$0.332
GJ 191 6505.08$\pm$0.98 -5730.84$\pm$0.96 255.66$\pm$0.91 6.756$\pm$0.024
GJ 9371 264.99$\pm$2.23 -3157.36$\pm$2.30 42.79$\pm$2.70 5.391$\pm$0.340
----------- ------------------- ------------------- ----------------- ------------------
Table \[Tab:SecuAcc\] lists the prediction for secular acceleration for all of our sample stars as derived from Eq. (\[eq:SecuAcc\]). Some stars (e.g. Barnard’s star, GJ 1, and Proxima Cen) have such high proper motion that the secular acceleration can be measured (4.5, 3.7, and 0.4m/s/yr, respectively). This effect is depicted with a dashed line for Proxima Cen and for all stars with $\dot{v}_{\mathrm{r}}>1\,\mathrm{m/s/yr}$ in Fig. \[Fig:RV1\]. The secular acceleration was subtracted before subsequent analysis of the RV data.
We note that, even though Barnard’s star has the highest (angular) proper motion $\mu$ and the dependence is quadratic on $\mu$, it is *not* the star with the highest secular acceleration. When inspecting high proper motion stars from the Hipparcos catalogue [@vanLeeuwen07] we found four stars that have a higher secular acceleration because of their smaller parallax (see lower part of Table \[Tab:SecuAcc\]), namely: GJ 451 (Groombridge 1830, G8Vp + M5.5V), GJ 9511 (K2V + K2Vfe), GJ 191 (Kapteyn’s star, M1V), and GJ 9371 (sdK4).
Tests for variability and trends
--------------------------------
Following the recipe outlined by @Endl02, we performed several statistical tests to identify variability and RV trends in our data. First, we asked the question for each star of whether the observed variability or rms $\sigma$ is significantly higher than the mean measurement error $\overline{\sigma_{RV}}$ using the $F$-Test (and $F=\frac{\sigma^{2}}{\overline{\sigma_{RV}}^{2}}$ as $F$-value)[^4].
--------- ----- ------ ----------------- ----------------- ------- ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------- ----------------------------- -- ------------------------- ---------
Star $N$ $\chi_{\mathrm{slope}}^{2}$ $P(F_{\mathrm{slope}})$ Comment
[\[]{}m/s[\]]{} [\[]{}m/s[\]]{} [\[]{}m/s/yr[\]]{}
Barnard 75 2358 3.3 2.7 0.065 114 **[0.0022]{} & -0.688 & 99.4 & **[0.022]{} & **[0.0038]{}\
GJ 1 & 24 & 2151 & 2.5 & 2.4 & 0.9 & 27.3 & 0.24 & -0.454 & 25.9 & **[0.26]{} & 0.56\
GJ 27.1 & 30 & 2177 & 6.1 & 6.1 & 0.95 & 41.2 & 0.066 & 0.122 & 41.1 & **[0.052]{} & 0.32\
GJ 118 & 26 & 2266 & 6.5 & 5.8 & 0.56 & 40.5 & 0.026 & 1.99 & 23.8 & **[0.47]{} & **[0.00083]{}\
GJ 160.2 & 33 & 2325 & 8.1 & 7.7 & 0.79 & 41.1 & 0.13 & 1.36 & 35.5 & **[0.26]{} & 0.067\
GJ 173 & 12 & 897 & 5.3 & 3.1 & 0.094 & 30.6 & **[0.0013]{} & 1.61 & 28.7 & 0.0014 & 0.86\
GJ 180 & 24 & 2325 & 3.8 & 4.1 & 0.71 & 29.4 & 0.17 & 0.418 & 27.5 & **[0.19]{} & 0.46\
GJ 190 & 2 & 4 & 861.8 & 80.7 & 0.12 & 97.4 & **[$\bf{<10^{-7}}$]{} & -1.12$\cdot10^{5}$ & 0 & $<10^{-7}$ & **[$\bf{<10^{-7}}$]{} & SB2\
GJ 218 & 9 & 896 & 3.1 & 3.1 & 0.98 & 8.49 & 0.39 & 1.03 & 7.96 & **[0.34]{} & 0.96\
GJ 229 & 32 & 2325 & 5.5 & 2.8 & **[0.00036]{} & 139 & **[$\bf{<10^{-7}}$]{} & 1.43 & 95.6 & $<10^{-7}$ & **[0.0017]{}\
GJ 263 & 2 & 82 & 117.6 & 78.6 & 0.75 & 1.19 & 0.27 & 743 & 0 & $<10^{-7}$ & **[$\bf{<10^{-7}}$]{} & SB2\
GJ 357 & 30 & 2321 & 5.3 & 3.2 & **[0.0096]{} & 59.9 & **[0.00064]{} & 0.393 & 57.5 & 0.00084 & 0.58\
GJ 377 & 14 & 1089 & 6.7 & 3.2 & 0.014 & 40 & **[0.00014]{} & -0.284 & 39.8 & 7.7$\cdot10^{-5}$ & 0.36\
GJ 422 & 15 & 1112 & 4.0 & 3.4 & 0.55 & 16.8 & 0.27 & 0.599 & 16.2 & **[0.24]{} & 0.99\
GJ 433 & 54 & 2554 & 4.4 & 3.6 & 0.16 & 80 & **[0.0097]{} & 0.284 & 78.4 & **[0.011]{} & 0.61\
GJ 477 & 8 & 389 & 3486.0 & 4.3 & $\bf{1.6\cdot10^{-19}}$ & 5.58$\cdot10^{6}$ & **[$\bf{<10^{-7}}$]{} & 6.99$\cdot10^{3}$ & 4.52$\cdot10^{5}$ & $<10^{-7}$ & **[0.00034]{} & SB1\
GJ 510 & 23 & 1115 & 5.6 & 3.5 & 0.039 & 54 & **[0.00016]{} & 2.64 & 37.9 & **[0.013]{} & 0.014\
GJ 620 & 5 & 422 & 7.3 & 4.4 & 0.36 & 13.1 & 0.011 & -5.1 & 11 & **[0.012]{} & 0.98\
GJ 637 & 17 & 1099 & 6.4 & 4.4 & 0.16 & 27 & 0.041 & 0.305 & 26.9 & **[0.03]{} & 0.38\
GJ 682 & 20 & 1134 & 4.0 & 2.3 & 0.024 & 53.8 & $\bf{3.6\cdot10^{-5}}$ & 1.62 & 40.8 & 0.0016 & 0.056\
GJ 739 & 19 & 1070 & 4.4 & 3.1 & 0.13 & 48.8 & **[0.00012]{} & -2.09 & 37.4 & 0.003 & 0.072\
GJ 817 & 25 & 1551 & 4.9 & 4.3 & 0.54 & 32.2 & 0.12 & 0.184 & 32.1 & **[0.098]{} & 0.42\
GJ 821 & 35 & 1516 & 5.0 & 3.8 & 0.12 & 53.7 & 0.017 & -0.268 & 53.5 & **[0.013]{} & 0.56\
GJ 842 & 17 & 926 & 6.7 & 4.2 & 0.065 & 44.6 & **[0.00016]{} & -1.2 & 43.5 & 0.00013 & 0.91\
GJ 855 & 22 & 1561 & 5.8 & 4.5 & 0.24 & 28.7 & 0.12 & -1.14 & 24.9 & **[0.21]{} & 0.19\
GJ 891 & 25 & 2178 & 7.5 & 5.1 & 0.068 & 48.4 & **[0.0023]{} & 0.556 & 47.6 & 0.0019 & 0.93\
GJ 911 & 17 & 2136 & 14.9 & 7.7 & 0.012 & 25.2 & 0.067 & 0.389 & 25.1 & **[0.049]{} & 0.38\
GJ 1009 & 22 & 2177 & 5.3 & 4.0 & 0.23 & 47.8 & **[0.00074]{} & 0.0316 & 47.8 & 0.00046 & 0.055\
GJ 1046 & 14 & 766 & 1248.3 & 3.6 & $\bf{1.9\cdot10^{-30}}$ & 1.59$\cdot10^{6}$ & **[$\bf{<10^{-7}}$]{} & 1.09$\cdot10^{3}$ & 8.8$\cdot10^{5}$ & $<10^{-7}$ & 0.018 & SB1 (BD)\
GJ 1100 & 12 & 897 & 9.3 & 5.1 & 0.061 & 37 & **[0.00011]{} & -1.67 & 36.3 & 7.4$\cdot10^{-5}$ & 0.66\
GJ 3020 & 13 & 749 & 298.8 & 9.0 & $\bf{5\cdot10^{-16}}$ & 7.62$\cdot10^{3}$ & **[$\bf{<10^{-7}}$]{} & -287 & 2.29$\cdot10^{3}$ & $<10^{-7}$ & **[0.00074]{} & SB1\
GJ 3082 & 10 & 761 & 6.2 & 4.2 & 0.27 & 17.5 & 0.041 & 1.19 & 17.1 & **[0.029]{} & 0.66\
GJ 3098 & 9 & 733 & 9.1 & 4.7 & 0.079 & 26.7 & **[0.00081]{} & 7.05 & 16 & **[0.025]{} & 0.13\
GJ 3671 & 12 & 1090 & 5.6 & 4.4 & 0.46 & 17.4 & 0.095 & 3.26 & 10.2 & **[0.43]{} & 0.046\
GJ 3759 & 11 & 1080 & 3.9 & 3.6 & 0.81 & 11.5 & 0.32 & 2.48 & 6.84 & **[0.65]{} & 0.071\
GJ 3916 & 6 & 406 & 2170.7 & 9.2 & $\bf{1.5\cdot10^{-11}}$ & 4.07$\cdot10^{5}$ & **[$\bf{<10^{-7}}$]{} & 4.1$\cdot10^{3}$ & 2.12$\cdot10^{3}$ & $<10^{-7}$ & $\bf{2\cdot10^{-5}}$ & SB1\
GJ 3973 & 5 & 420 & 6.8 & 3.6 & 0.25 & 12 & 0.017 & 3.3 & 10.8 & **[0.013]{} & 0.78\
GJ 4106 & 5 & 396 & 20.7 & 15.9 & 0.62 & 7.4 & 0.12 & 32.2 & 1.99 & **[0.58]{} & 0.13 & giant\
GJ 4293 & 14 & 875 & 8.7 & 5.6 & 0.13 & 32.3 & **[0.0021]{} & -0.114 & 32.3 & 0.0012 & 0.064\
HG 7-15 & 11 & 417 & 8.7 & 10.3 & 0.59 & 11 & 0.36 & 10.5 & 9.03 & **[0.43]{} & 0.39\
Prox Cen & 76 & 2555 & 3.6 & 2.3 & **[0.00028]{} & 183 & **[$\bf{<10^{-7}}$]{} & 0.703 & 159 & $<10^{-7}$ & **[0.0026]{}\
************************************************************************************************************
--------- ----- ------ ----------------- ----------------- ------- ----- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ -------------------- ----------------------------- -- ------------------------- ---------
The calculated probabilities $P(F)$ from the $F$-Test are listed in Table \[Tab:StatTests\] for each star. A low value of $P(F)$ (e.g. $<0.01$, i.e. 99% confidence, in bold face in Table \[Tab:StatTests\]) indicates that the observed scatter can probably not be explained with the measurement errors and that there is an excess variability or a trend. This is the case for the stars with a high rms (GJ 477, GJ 1046, GJ 3020, and GJ 3916), which seem to have a companion, probably in the brown dwarf or low-mass star regime (see also Fig. \[Fig:RV2\]). GJ 190 and GJ 263 also have a high sample variance. They may have a stellar companion bright enough to contaminate the spectrum. We deduce this from the large measurement error that would occur for a double-lined spectroscopic binary (SB2) because our data modelling is only designed for a single-lined spectroscopic binary (SB1). Also, because there are only two measurements for GJ 190 and GJ 263, they do not stand out in the $F$-statistics. Indeed GJ 263 has already been identified as a spectroscopic binary, and an adaptive optics image was presented by @Beuzit04. Among the stars with an rms smaller than 20m/s, only GJ 229, GJ 357, and Proxima Cen have an rms that is significantly greater than the measurement error.
A similar test for RV variability is to determine the goodness of fit for a constant model, i.e. calculating the $\chi^{2}$ above the weighted RV mean[^5] and deriving the probability from the $\chi^{2}$-distribution. These probability values $P(\chi_{\mathrm{const}}^{2})$ are mostly lower than $P(F)$, and therefore there would be more variable stars according to our $P<0.01$ criterion (Table \[Tab:StatTests\]).
As long-period planets can cause a trend in the RV, we also tested for this by weighted fitting of a linear slope. A high probability of the resulting $\chi_{\mathrm{slope}}^{2}$ indicates that this is an acceptable fit (on the contrary, a low probability indicates remaining variability). But it is also informative to compare $\chi_{\mathrm{slope}}^{2}$ with the $\chi_{\mathrm{const}}^{2}$ of the weighted mean via the $F$-value [@Cumming99] $$F_{\mathrm{slope}}=(N-2)\frac{\chi_{\mathrm{constant}}^{2}-\chi_{\mathrm{slope}}^{2}}{\chi_{\mathrm{slope}}^{2}}.$$ A high $F_{\mathrm{slope}}$-value indicates a fit improvement whereas a low probability $P(F_{\mathrm{slope}})$ shows this improvement to be significant and that it is probable not due to noise.
In Table \[Tab:StatTests\] the stars with a significant trend are marked in bold face. Besides the companion hosting M dwarfs GJ 477, GJ 3020, and GJ 3916, these are the stars Barnard’s star, GJ 118, GJ 229, and Proxima Cen. The significant trend of 1.4 m/s/yr for GJ 229 may be caused by the wide T7p brown dwarf companion GJ 229 B, as one can show with a rough estimate. Assuming a circular orbit and the projected separation of 45AU ($\alpha=7.7\arcsec$ @Nakajima95, $\pi=0.173\arcsec$) as the orbital radius to GJ 229, the orbital period is of the order of $P=400\,$yr. Then a 30$\, M_{\mathrm{Jup}}$ companion can cause a velocity amplitude of $K=170\,$m/s (Eq. (\[eq:Amplitude\])) and the maximum RV change is $\dot{RV}=K\frac{2\pi}{P}=2.7\,\mathrm{m/s/yr}$, about twice as large as the observed trend. However, the low $P(\chi_{\mathrm{slope}}^{2})$ indicates there is still some variability after trend subtraction.
Periodogram analysis
--------------------
To test for periodicities in the RV data, we computed the generalized Lomb-Scargle periodogram (GLS, @Zechmeister09), which is the equivalent of fitting sine waves including an offset. The adopted period search interval ranges from 2d to the time baseline $T$ of each data set. Note that 2d will mostly exceed the average Nyquist frequency. However, searching at frequencies higher than the Nyquist frequency is possible for irregular sampling [@Pelt09]. False-alarm probabilities (FAP) were determined by bootstrap randomization [e.g. @Kuerster97]. In this method random data sets are generated from an original data set by shuffling the RVs while retaining the observing times. For each random data set, the GLS periodogram was computed and searched for its maximum. The FAP is then given by that fraction of random data sets having a periodogram power higher than the original one. For each star we generated 1000 random data sets which resolves FAP $>10^{-3}$ and is sufficient to decide whether the FAP is below our threshold of 0.01.
Table \[Tab:Periodicites\] shows that GJ 4106 and GJ 1046 have a FAP marginally lower than 0.01. The probable brown dwarf companion to GJ 1046 with a minimum mass of 26.9$\, M_{\mathrm{Jup}}$ in an eccentric orbit ($e=0.28$) with a 168.8d period has been already published by @Kuerster08 based on this UVES data set. For the 365d period in Proxima Cen we refer to @Endl08b, who recently analysed the RV data and identified this period as a 1-year alias.
Based on data from the first 21/2yr, @Kuerster03 determined two RV periods of 32d and 45d for Barnard’s star with a FAP of 0.56% and 1.05%, respectively. Now, in the enlarged data set with a 61/2yr time baseline, the second period (45d; with an amplitude of 2.9m/s) has the highest periodogram peak, and its FAP is now less than 0.1%, i.e. more significant (see Fig. \[Fig:GLS\], second panel for the periodogram and Fig. \[Fig:Barnardphased\] for the RVs phased to this period). However, stellar activity is the probable cause of this variability and will be discussed in Sect. \[sub:Correlation-HRV\].
--------------------------------------------------------------------------------------------------------
Star $P$ [\[]{}d[\]]{} $\chi_{\mathrm{sin}}^{2}$
---------- ------------------- --------------------------- ---------------------------------------------
Barnard 44.9 72.5 $\mathbf{<10^{-4}}$
GJ 1 2.73 12.5 0.733
GJ 27.1 2.01 18 0.189
GJ 118 3.97 15.6 0.329
GJ 160.2 4.15 20.7 0.577
GJ 173 2.28 3.67 0.026
GJ 180 5.93 8.63 0.127
GJ 218 2.02 0.473 0.557
GJ 229 10.9 63.7 0.102
GJ 357 3.41 28.4 0.157
GJ 377 15.1 7.47 0.029
GJ 422 8.82 4.69 0.498
GJ 433 6.5 55.6 0.451
GJ 477 243 1.58$\cdot10^{5}$ 0.541
GJ 510 2.92 29.1 0.867
GJ 620 2.51 0.00667 0.637
GJ 637 8.54 8.54 0.245
GJ 682 89.3 20.1 0.178
GJ 739 2.34 16.3 0.334
GJ 817 2.36 15.1 0.743
GJ 821 12.6 32.9 0.393
GJ 842 92.6 10.8 0.219
GJ 855 16 12.6 0.668
GJ 891 30.5 22.1 0.36
GJ 911 2.35 9.85 0.909
GJ 1009 3.73 17.1 0.502
GJ 1046 174 1.23$\cdot10^{5}$ **[0.008]{}\
GJ 1100 & 3.79 & 5.27 & 0.477\
GJ 3020 & 5.14 & 3.03$\cdot10^{3}$ & 0.895\
GJ 3082 & 3.72 & 1.7 & 0.261\
GJ 3098 & 2.7 & 2.23 & 0.5\
GJ 3671 & 68.1 & 1.69 & 0.218\
GJ 3759 & 2.46 & 1.8 & 0.769\
GJ 3916 & 8.6 & 82.5 & 0.077\
GJ 3973 & 2.25 & 0.00154 & 0.105\
GJ 4106 & 2.41 & 0.000139 & **[0.009]{}\
GJ 4293 & 19.1 & 11 & 0.994\
HG 7-15 & 4.55 & 0.919 & 0.179\
Prox Cen & 365 & 101 & $\mathbf{<10^{-4}}$\
****
--------------------------------------------------------------------------------------------------------
: \[Tab:Periodicites\]Test for periodicities in the RV (GJ 190 and GJ 263 excluded): the best period $P$ with its $\chi^{2}$ and bootstrapped FAP.
![\[Fig:Barnardphased\]Radial velocity time series for Barnard’s star phased to the 44.9d period and the best-fitting sinusoid.](12479f05.ps)
\[sub:Upper-detection-limits\]Upper detection limits
----------------------------------------------------
A Jupiter-mass planet in a circular orbit with a radius of 1AU around a $0.4M_{\odot}$-M dwarf would cause an RV amplitude of 45m/s (even higher for closer orbits; Eq. (\[eq:Amplitude\])) and would be easily detectable with our precision of typically a few m/s. However, most of our sample stars show low RV-variations and no indication of a planet. To determine which planets in circular orbits can be excluded, we calculated detection limits in the following way.
We considered the data as noise and simulated planetary signals by adding sine waves to the data with a range of trial frequencies and for 12 different equidistant phases. The time sampling remained unaffected. For the simulated data the generalised Lomb-Scargle periodogram power was calculated at the trial sine wave frequency where the peak of the signal is expected. If this power was below a power threshold (which corresponds to an FAP of 0.01, see note below), we increased the amplitude of the sine wave. The simulated planet is considered as detected, if the power is in all 12 phases (equal to or) higher than the threshold. The corresponding amplitude is considered as the 99% detection limit. In Appendix \[sec:Amp\_limits\] we provide an analytic solution to calculate this amplitude for a given power threshold and phase. This procedure was performed for a number of trial frequencies resulting in amplitude limits which finally were translated into mass upper limits (Fig. \[Fig:DetectLimits\]) using Eq. (\[eq:Amplitude\]) and the stellar masses from Table \[Tab:Sample\].
It should be noted that the periodogram power $p$ is a measure/quantity for the FAP. By applying bootstrap randomisation to the original data we get a FAP vs. power relation $p(FAP)$. No assumptions are necessary about the number of independent frequencies. The power value for FAP=0.01 was also taken as the threshold for the simulated data. This modification to the method by @Endl02 bypasses bootstrapping the simulated data again and is therefore more efficient and allows a dense sampling of the frequency.
Here we assume that the $p(FAP)$ relation does not change much when adding a sine wave, because the time sampling and number of measurements does not change. The only thing that can happen is that the rms (or $\chi^{2}$ above the mean) of the simulated data changes (increases) by a certain factor. But as the normalized power is invariant when the measurements are scaled the resulting effect is small. A comparison of the detections limits given in Fig. \[Fig:DetectLimits\] with those from @Endl08b for Proxima Cen shows quite similar results. Figure \[Fig:DetectLimits\] also shows the steep increase that generally appears for periods longer than the time base (see also @Cumming04 [@Nelson98]).
Instead of again searching the whole frequency range, only the power at the original sine frequency was calculated, since one can expect to find the simulated signal there. This saves computational effort and is more conservative, because we exclude spurious detections that exceed the power threshold with a lower amplitude at alias or noise frequencies.
Figure \[Fig:DetectLimits\] shows the results of our detection limit calculation. It can be seen that for Barnard’s star or Proxima Cen, i.e. stars with low masses and many measurements, the detection limit reaches down to a few Earth masses for close-in circular orbits and even within their habitable zones (HZ). Both stars have a priori frequencies with a FAP<0.01. These frequencies were excluded from the detection limit calculation (Barnard’s star: 36.1 – 36.4d and 44.6 – 45.1d; Proxima Cen: 295.0 – 313.11d and 347.1 – 392.5d). The HZ is depicted for each star derived from Fig. 15 of @Kasting93, whereas we used the stellar masses from Table \[Tab:Sample\]. For an M dwarf with a mass of 0.3$M_{\odot}$, the HZ is beyond 0.1AU. For several stars we can exclude planets with a few ten Earth masses in their HZ and Jupiter-mass planets ($1M_{\mathrm{Jup}}=318M_{\mathrm{Earth}}$) up to a few AU.
{height="0.96\textheight" width="1\linewidth"}
\[sub:Correlation-HRV\]Correlation between RV and H$\alpha$ index?
------------------------------------------------------------------
Stellar activity can affect the measured RV. The H$\alpha$ line is an indicator of stellar activity (the only available one in the UVES spectra). In Proxima Cen as an active flare star, the H$\alpha$ line is in emission in contrast to low-activity Barnard’s star. However, the H$\alpha$ line is variable in both cases. Therefore investigating the correlation between RV and variations in the H$\alpha$ line may be useful for correcting RV for stellar activity.
@Kuerster03 report a correlation of this type for Barnard’s star. We checked this again with the now available longer data set for Barnard’s star, as well as for the other stars. As a measure of the variability of the H$\alpha$ line we adopt the definition of the H$\alpha$-index by @Kuerster03: $$I=\frac{\overline{F_{0}}}{0.5(\overline{F_{1}}+\overline{F_{2}})}\label{eq:index}$$ where $\overline{F_{0}}$ is the mean flux in the range of [\[]{}-15.5km/s,+15.5km/s[\]]{} around the H$\alpha$ line ($\lambda=656.28\,$nm) and $\overline{F_{1}}$ and $\overline{F_{2}}$ are the mean flux two reference bandpasses ([\[]{}-700km/s,-300km/s[\]]{} and [\[]{}600 km/s, 1000 km/s[\]]{}, respectively) used for normalization. This index is a kind of filling-in of the H$\alpha$ line (i.e for emission $I\ga1$) and is related to equivalent width (see Appendix \[sec:index-EW\]). Following @Kuerster03, we also computed a CaI index [\[]{}+441.5 km/s, 472.5 km/s[\]]{} for comparison. The CaI Line ($\lambda=6572.795\,$Å) is expected to be stable.
Figure \[Fig:H-t\_barnard\] shows the variation in the H$\alpha$ index for Barnard’s star with time. In Fig. \[Fig:H-RV\_barnard\] the RV is plotted against the H$\alpha$ index. The flare event has already been described in @Kuerster03, does not seem to have any effect on the RV, and is excluded from the correlation analysis. No further flare was detected in the longer data set. We calculated a correlation coefficient of $r=-0.42$ for the whole data set. Because the new data (JD>2452600) have a less pronounced anti-correlation with a correlation coefficient of $r=-0.25$, the correlation decreases compared to the old data set where $r=-0.50$ [@Kuerster03], but the anti-correlation is still present and becomes more significant due to the longer baseline.
We performed a period search (Fig. \[Fig:GLS\], top panel) for the H$\alpha$ index of Barnard’s star (again flare event is excluded). With the new data, a 1000d period dominates the GLS periodogram (upper panel). There is also some power at 44.5 d that was found to be the dominant period in the RV data (second panel). Therefore it is probable that the 45-day RV period is caused by stellar activity rather than by a planet (third panel). When we subtract the correlation from the RV data, the rms reduces slightly from 3.35 m/s to 3.09 m/s. The FAP of this period for the corrected RV data is only 5.5% compared to $<0.01$% of the uncorrected RV data (Table \[Tab:Periodicites\]). A similar analysis for the other stars[^6] only yields a significant correlation for GJ 433, GJ 821, and GJ 855 (Figs. \[Fig:H-RV\_gj433\]–\[Fig:H-RV\_gj855\]). All stars show no significant RV-CaI index correlation.
![\[Fig:H-t\_barnard\]Time series of the H$\alpha$-index for Barnard’s star.](12479f07.ps)
![\[Fig:H-RV\_barnard\]Correlation of the H$\alpha$-index with the RV for Barnard’s star (dots with error bars, $r_{\mathrm{H}\alpha}=0.42$). For comparison the CaI-index (crosses).](12479f08.ps)
![\[Fig:H-RV\_gj433\]Correlation of the H$\alpha$- and CaI-index with the RV for GJ 433 ($r_{\mathrm{H}\alpha}=-0.40$).](12479f09.ps)
![\[Fig:H-RV\_gj821\]Correlation of the H$\alpha$- and CaI-index with the RV for GJ 821 ($r_{\mathrm{H}\alpha}=-0.49$).](12479f10.ps)
![\[Fig:H-RV\_gj855\]Correlation of the H$\alpha$- and CaI-index with the RV for GJ 855 ($r_{\mathrm{H}\alpha}=0.62$).](12479f11.ps)
Finally, we investigated the question of whether more active stars show more RV excess scatter. Figure \[Fig:RV-H\_scatter\] compares the RV scatter (rms from Table \[Tab:StatTests\]) and the relative H$\alpha$ line index scatter for all M dwarfs from our sample and demonstrates that no correlation can be seen.
![\[Fig:RV-H\_scatter\] Comparison of the RV scatter (rms) and H$\alpha$ variability for all M dwarfs with an rms $<20\,$m/s. No correlation can be seen. The very active star Proxima Cen shows only very low RV excess.](12479f12.ps)
Discussion
==========
We browsed the available literature and catalogues for known binarity of our M dwarfs, which may have an impact on planet formation. Well known are the substellar companions to GJ 229 [@Nakajima95] and GJ 263 [@Beuzit04]. Proxima Cen constitutes a widely separated common proper motion pair with $\alpha$ Cen A+B.
For GJ 477 and GJ 3916, binarity is indicated in the Hipparcos catalogue. GJ 477 has the double/multiple systems flag “X” (stochastic solution, probable astrometric binaries with a short period). GJ 3916 is listed with the flag “G” (acceleration or higher order terms). The large amplitudes seen in the RVs for both stars confirm this.
Other stars with an entry are: GJ 620 (“G”), GJ 4106 (“G”), and GJ 433 (“O”, i.e. orbital solution). For these stars we found only low RV variation. This is at least controversial for GJ 433 (HIP 56528). The announced period is 500d [see also @Bernstein97] and the amplitude is expected to be several hundred m/s.[^7] At 1AU our detections limit reaches down to 0.2$M_{\mathrm{Jup}}$ (Fig. \[Fig:DetectLimits\]).
To our knowledge the rest of our sample so far has no discovered companions. Companions are explicitly excluded with near-infrared speckle interferometry by @Leinert97 for GJ 1 and GJ 682 ($\Delta K=-4.5\,$mag at 1-10AU), as well as GJ 891 with infrared coronagraphic imaging by @McCarthy04 (>30$M_{\mathrm{Jup}}$ at 140-1200AU).
@Kuerster03 discussed how several stellar activity phenomena, such as spots, plages, or convective RV shifts, might affect the RV measurements. A linear RV-H$\alpha$ anti-correlation that is present in Barnard’s star could be a result of a convective redshift caused by plages that suppress blueshifted convective flows. We find such an anticorrelation for GJ 433 and GJ 821, while GJ 855 exhibits a positive correlation (Figs. \[Fig:H-RV\_gj433\]–\[Fig:H-RV\_gj855\]). @Bonfils07 presented with GJ 674 (M2.5V) an example where the RV and H$\alpha$ variations seem to be phase-shifted. The correlation thus differs from a linear one and looks like a loop. One would expect such behaviour for a plage rotating with the surface.
Flare events, such as the one observed in Barnard’s star, do not seem to take part in the correlation. This agrees with the fact that no significant and strong correlation is observed in the flaring M dwarf Proxima Cen [see also @Endl08b].
Conclusion
==========
Within the sensitivity provided by our RV precision of a few m/s we have not detected any planets around our sample stars. Most of the M dwarfs exhibit only low RV variations and some of them have a measurable secular acceleration due to their high proper motion.
We have discussed two effects on the RV that are not caused by planetary companions. First, the secular acceleration, is a perspective effect and causes always a positive RV trend. This effect can be corrected easily. Vise versa, this can be seen as an independent measurement of the astrometric quantity $\frac{\mu^{2}}{\pi}$, i.e. a confirmation for the ratio of squared proper motion and parallax (the absolute RV is not required). The second effect are RV variations caused by stellar activity. This is likely the case for the 45d period we found in the RV data of Barnard’s star. The RVs correlate with the H$\alpha$ index. Such a correlation was found here only for a few M dwarfs and therefore no conclusions can be drawn in general.
As a by-product of our survey we have identified 6 M dwarfs with low-mass companions, four of them (GJ 477, GJ 1046 and GJ 3020, and GJ 3916) are brown dwarf or low-mass stellar candidates and two are spectroscopic binaries (SB2: GJ 263 and GJ 190). Follow-up RV observations will yield the orbital parameters and the lower limits for the companion masses $m\sin i$.
Our detection limits demonstrate that we can exclude giant planets with $1M_{\mathrm{Jup}}$ up to 1AU for half of our M dwarfs and that we are in principle capable of discovering planets with a few Earth masses in the habitable zones of M dwarfs with VLT+UVES. For this purpose an adequate number of measurements is needed to find low amplitudes of the order of the achieved high RV precision. The low frequency of Jupiter-mass planets around M dwarfs requires a large sample. The given detection limits are based on the search for single planets in circular orbits. These limits would be higher for eccentric orbits or multi-planet systems.
Even if planet detections are more spectacular, it is also important to report non-detections, which are required to estimate the planet frequency. Our non-detections of planets support the increasing observational evidence of a lower frequency of Jupiter-mass planets around M dwarfs. @Endl06 estimated a frequency of $\approx1$% or less up to 1 AU orbital radius for Jupiter-mass planets around M dwarfs compared to 2.5% for solar like stars. This comparison cannot be done yet for low-mass planets because they are much harder to detect (and even more so for G stars), which introduces observational biases. However low-mass planets seem to be quite frequent around M dwarfs [@Bonfils07].
Expansion of current M dwarf planet searches will allow more precise determination of the true frequency of giant planetary companions to this type of stars and lower the detection limits. Photometric surveys for transits like the recently started project MEarth [@Irwin08] monitoring 2000 nearby M dwarfs and microlensing projects will also contribute. The discoveries of planetary systems around GJ 876 [@Delfosse98; @Marcy98; @Marcy01; @Rivera05] and GJ 581 [@Udry07] show that planets do exist around M dwarfs. This promises further discoveries of low-mass planets in the future with high-precision RV surveys.
We thank the ESO OPC and the DDTC for the generous allocation of observing time. We are also grateful to all of ESO staff who helped with the preparation of the service mode observations, carried them out, or processed, verified, and distributed the data. A number of people have contributed in many ways to make this survey happen. Our thanks go to Artie P. Hatzes, Stéphane Brillant, William D. Cochran, Sebastian Els, Thomas Henning, Andreas Kaufer, Sabine Reffert, Florian Rodler, Frédéric Rouesnel, and Steve S. Saar. This material is based on work supported by the National Aeronautics and Space Administration under Grants NNG04G141G, NNG05G107G issued through the Terrestrial Planet Finder Foundation Science program and Grant NNX07AL70G issued through the Origins of Solar Systems Program.
Relation between index and equivalent width\[sec:index-EW\]
===========================================================
The equivalent width is defined as $$EW=\sum_{i}\delta\lambda_{i}\frac{F_{C_{i}}-F_{i}}{F_{C_{i}}}$$ where $\delta\lambda_{i}$ is the width of the $i$-th pixel in wavelength, and $F_{i}$ and $F_{C_{i}}$ are the flux and continuum flux in the $i$-th pixel, respectively. The equivalent width $EW$ is measured in terms of wavelengths. In a normalized spectrum the continuum flux is constant $F_{C_{i}}=F_{C}$ resulting in $$EW=\Delta\lambda\left[1-\frac{1}{\Delta\lambda}\sum\delta\lambda_{i}\frac{F_{i}}{F_{C}}\right]$$ where $\Delta\lambda=\sum\delta\lambda_{i}$ is the considered wavelength range.
The index $I$ as defined (for H$\alpha$) in Eq. (\[eq:index\]) is, on the other hand, a dimensionless measure. It is normalized by the mean flux $\overline{F}=\frac{1}{\Delta\lambda}\sum_{i}\delta\lambda_{i}F_{i}$ taken from reference ranges instead of the continuum, which is sometimes difficult to estimate, in particular for M dwarfs with their ubiquitous absorption lines.
In the case that the reference regions are estimated as continuum, i.e. $\overline{F_{1}}=\overline{F_{2}}=F_{C}$, the index becomes $I=\frac{1}{F_{C}}\frac{1}{\Delta\lambda}\sum_{i}\delta\lambda_{i}F_{i}$ and is related to $EW$ as $$I=1-\frac{EW}{\Delta\lambda}.$$ Note that an absorption line is indicated by $EW>0$ ($0<I<1$) and an emission line by $EW<0$ ($I>1$).
\[sec:Amp\_limits\]Response of the GLS periodogram when adding a sine wave
==========================================================================
The definition of the generalized Lomb-Scargle periodogram (GLS) is (we use the notation introduced in @Zechmeister09) $$p_{y}(\omega)=\frac{1}{YY}\cdot\frac{SS\cdot YC^{2}+CC\cdot YS^{2}-2CS\cdot YC\cdot YS}{CC\cdot SS-CS^{2}}$$ whereas abbreviations are weighted covariances for data $y_{i}$, sine, and cosine terms (e.g. $YC=\sum w_{i}y_{i}\cos\omega t_{i}-\sum w_{i}y_{i}\cdot\sum w_{i}\cos\omega t_{i}=\sum w_{i}(y_{i}-\overline{y})\cos\omega t_{i}$; $w_{i}$ are normalised weights).
When introducing a parameter $\tau$ defined by $\tan2\tau=\frac{2CS}{CC-SS}$ and replacing $t_{i}$ by $\tau_{i}=t_{i}-\tau$ (resulting in $CS_{\tau}=0$) the GLS can be written as (see @Zechmeister09 for details) $$p_{y}(\omega)=\frac{1}{YY}\left[\frac{YC_{\tau}^{2}}{CC_{\tau}}+\frac{YS_{\tau}^{2}}{SS_{\tau}}\right]$$ in a form very similar to classical Lomb-Scargle periodogram. While the first formulation can save some computational effort, the use of the second is more elegant for our purpose.
By adding a sine wave with frequency $\omega_{0}$ to the data $y_{i}$, we generate new data $x_{i}=y_{i}+a\cos\omega_{0}t_{i}+b\sin\omega_{0}t_{i}$ or $x_{i}=y_{i}+a_{\tau}\cos\omega_{0}\tau_{i}+b_{\tau}\sin\omega_{0}\tau_{i}$. This results in a new periodogram $p_{x}(\omega)$ for the new data set $x$ $$\begin{aligned}
p_{x}(\omega) & =\frac{1}{XX}\left[\frac{XC_{\tau}^{2}}{CC_{\tau}}+\frac{XS_{\tau}^{2}}{SS_{\tau}}\right].\end{aligned}$$ Because the times are not changed, this neither affects the parameter $\tau$ or the sums that only depend on the time sampling ($CC_{\tau}$ and $SS_{\tau}$ or $CC$, $SS$, $CS$, and $D$, respectively). When adding the sine wave the mean changes $$\begin{aligned}
\overline{x} & =\sum w_{i}x_{i}=\sum w_{i}y_{i}+a\sum w_{i}\cos\omega_{0}t_{i}+b\sum w_{i}\sin\omega_{0}t_{i}\\
& =\overline{y}+aC_{0}+bS_{0}.\end{aligned}$$ The sums $YC$ and $YS$ change as follows $$\begin{aligned}
XC & =\sum w_{i}(x_{i}-\overline{x})\cos\omega t_{i}=YC+aCC_{0}+bCS_{0}\nonumber \\
XS & =\sum w_{i}(x_{i}-\overline{x})\sin\omega t_{i}=YS+aC_{0}S+bSS_{0}\nonumber \\
XX & =\sum w_{i}(x_{i}-\overline{x})^{2}\nonumber \\
& =\sum w_{i}(y_{i}-\overline{y}+a(\cos\omega_{0}t_{i}-C_{0})+b(\sin\omega_{0}t_{i}-S_{0}))^{2}\nonumber \\
& =YY+a^{2}C_{0}C_{0}+b^{2}S_{0}S_{0}+2aYC_{0}+2bY_{0}S_{0}+2abC_{0}S_{0}.\label{eq:XX}\end{aligned}$$ As described in Sect. \[sub:Upper-detection-limits\], it is more conservative to scan the power only at $\omega=\omega_{0}$ instead of the whole frequency range. This bypasses the calculation of sums comprising two different frequencies and is an enormous simplification. The sums of simulated data can be expressed by the sums for the original data. Therefore is not necessary to repeat the summation for the simulated data. When using the notation with $\tau$ this becomes $XC_{\tau}=YC_{\tau}+a_{\tau}CC_{\tau}$ and $XS_{\tau}=YS_{\tau}+a_{\tau}SS_{\tau}$ because $CS_{\tau}=0$.
The power response at the frequency where the sine wave was added is $$\begin{aligned}
p_{x}(\omega) & =\frac{1}{XX}\left[\frac{YC_{\tau}^{2}}{CC_{\tau}}+a^{2}CC_{\tau}+2aYC_{\tau}+\frac{YS_{\tau}^{2}}{SS_{\tau}}+b^{2}SS_{\tau}+2bYS_{\tau}\right]\\
& =\frac{1}{XX}\left[p_{y}YY+XX-YY\right]=1-\frac{(1-p_{y})YY}{XX}.\end{aligned}$$
Now we are interested in the amplitude $A$ or variance $XX$ that is required for a given phase $\varphi$ to produce a desired power threshold $p_{x}(\omega)$, which corresponds to an FAP. The required variance is $XX=YY\frac{1-p_{y}}{1-p_{x}}$.
We obtain the required amplitude by solving the quadratic equation resulting from Eq. (\[eq:XX\]) ($a=A\cos\varphi$, $b=A\sin\varphi$) $$\begin{aligned}
0 & =a^{2}CC+b^{2}SS+2aYC+2bYS+2abCS-(XX-YY)\\
& =A^{2}(CC\cos^{2}\varphi+SS\sin^{2}\varphi+2CS\cos\varphi\sin\varphi)\\
& \quad+2A(YC\cos\varphi+YS\sin\varphi)-(XX-YY)\\
& =\alpha A^{2}+2\beta A-\gamma\end{aligned}$$ with the substitutions $\alpha=CC\cos^{2}\varphi+SS\sin^{2}\varphi+2CS\cos\varphi\sin\varphi$, $\beta=YC\cos\varphi+YS\sin\varphi$, and $\gamma=XX-YY=YY\frac{p_{x}-p_{y}}{p_{x}}$. This leads us to the amplitude $$A(\omega,\varphi)=-\frac{\beta}{\alpha}\underset{(-)}{+}\sqrt{\left(\frac{\beta}{\alpha}\right)^{2}+4\frac{\gamma}{\alpha}},$$ which we consider as the amplitude detection limit for a fixed phase and for a given power threshold $p_{x}$ and which can be expressed by GLS sums for the original data. Probing a set of phases $\varphi$, we finally choose $\max_{\varphi}A(\omega,\varphi)$.
The second of the two solution is rejected, because we demand positive amplitudes ($A>0$, $\varphi\in[0,360\degr$). The terms $\alpha$ and $\gamma$ are always positive: $\gamma=YY\frac{p_{x}-p_{y}}{p_{x}}>0$ (as long as $p_{x}-p_{y}>0$) and $\alpha=CC\cos^{2}\varphi+SS\sin^{2}\varphi+2CS\cos\varphi\sin\varphi=CC_{\tau}\cos^{2}\varphi+SS_{\tau}\sin^{2}\varphi>0$.
[^1]: Based on observations collected at the European Southern Observatory, Paranal Chile, ESO programmes 65.L-0428, 66.C-0446, 267.C-5700, 68.C-0415, 69.C-0722, 70.C-0044, 71.C-0498, 072.C-0495, 173.C-0606, 078.C-0829.
[^2]: Radial velocity data are available in electronic form at the CDS via anonymous ftp to cdsarc.u-strasbg.fr (130.79.128.5) or via http://cdsweb.u-strasbg.fr/cgi-bin/qcat?J/A+A/
[^3]: GJ 4106 is listed in the Catalogue of Nearby Stars with a parallax of $84\pm17\,$mas (11.9pc). However, with the Hipparcos parallax GJ 4106 should be an M giant.
[^4]: We use the one-tailed $F$-test because we are not interested in cases of error overestimation.
[^5]: When each measurement has the same error $\sigma_{RV}$, one gets $$\chi_{\mathrm{red}}^{2}=\frac{\chi_{\mathrm{const}}^{2}}{N-1}=\frac{1}{(N-1)\sigma_{RV}^{2}}\sum(RV_{i}-\overline{RV})^{2}=\frac{\sigma^{2}}{\sigma_{RV}^{2}}=F.$$
[^6]: The 6 six stars with companions were excluded.
[^7]: The parameters for the circular orbit ($e=0$, $\omega=0$) are $P=500\pm32\,$d, $T_{0}=2\,448\,402\pm28\,$d, $a_{0}=4.27\pm2.04\,$mas, $i=54\pm2$, and $\Omega=346\pm22\deg$. With the parallax $\pi=112.58\pm1.44$ the semi major axis of the photocentre is $a_{0}=0.0379\,$AU. Assuming $a_{0}\approx a_{1}$ the amplitude would be $K_{1}=\sin i\frac{2\pi a_{0}}{P}=668\,$m/s for the circular orbit. With $m_{1}=0.48M_{\odot}$ the secondary mass is $m_{2}=19.8M_{\mathrm{Jup}}$ calculated from the mass function $f(m)=\frac{(m\sin i)^{3}}{(m+M)^{2}}=\frac{P}{2\pi G}(K\sqrt{1-e^{2}})^{3}$, i.e. $m\sin i=K\cdot\sqrt[3]{(m+M)^{2}\frac{P}{2\pi G}}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is thought that the emergence of the “nightmare scenario” at the LHC could be a serious crisis for particle physics that could require radical new concepts and even a major paradigm change. A root cause may have been exaggeration of the significance of asymptotic freedom, leading to the historically profound mistake of formulating new short-distance extensions of the Standard Model while ignoring both serious infra-red problems and central elements of long-distance physics. In fact, pursuit of the uniquely unitary Critical Pomeron leads to a possible gauge theory origin for the Standard Model that is both radical and paradigm changing, but also explains many mysteries. A bound-state S-Matrix embedded in a unique weak coupling massless SU(5) field theory emerges. The states and interactions of the Standard Model are enhanced, and the underlying SU(5) unification suppressed, by a wee parton divergence phenomenon involving wee gauge bosons coupled to S-Matrix massless fermion anomalies. Confinement, chiral symmetry breaking, the parton model, electroweak symmetry breaking, dark matter, and neutrino masses, all appear to be present. Most significantly, perhaps, there is a Higgs boson but, as seen experimentally at the LHC, there is no new short-distance physics. The only new physics is electroweak-scale QCD interactions due to color sextet quarks.'
---
-0.4in 8.5in 1=0.5in -0.5in
plus 1000pt minus 1000pt \#1 \#1[= to]{}\#1[ to]{} \#1 \#1 \#1[ to 2]{} \#1
\#1[ to 1in]{} \#1
\#1\#2[[\#1\#2]{}]{} \#1\#2\#3[[ 2 \#1\#2 \#3]{}]{} \#1[\#1|]{} \#1[| \#1]{} \#1[\#1]{} \#1\#2[\#1|. \#2 ]{} \#1[/]{} \#1[| \#1|]{} \#1[\#1 ]{} \#1 \#1\#2[[0=1=1>0.51-.500-.50-.5110>1 .50-.51]{}]{} \#1\#2 \#1\#2 \#1 \#1[{.]{} \#1 \#1[=$^{#1}$=]{} = \#1 \#1,[by1 =1 \[\#1\]==,\[\#1\]=]{} \#1[-\#1-]{} \#1[(\#1)]{} \#1/\#2
\#1[$^{#1}$ ]{} \#1
\#1
\#1\#2\#3[[**\#1**]{}, \#2 (19\#3)]{} \#1[\[\#1\]]{} \#1 \#1 \#1 \#1[(\#1)]{}
The Nightmare Scenario and the Origin of the Standard Model.
“We Got it Wrong ...How did we misread the signals? ... What to Do?”
Alan R. White[^1]
0.6cm
Argonne National Laboratory
9700 South Cass, Il 60439, USA.
A Deep Crisis Requiring Radical New Concepts?
=============================================
In his overview talk[@DaGr] at Strings 2013, David Gross discussed the “nightmare scenario” in which the Standard Model Higgs boson is discovered at the LHC but no other new short-distance physics, in particular no signal for SUSY, is seen. He called it the “extreme pessimistic scenario” but also said it was looking more and more likely and (if it is established) then, he acknowledged
[“We got it wrong.” “How did we misread the signals?” “What to do?”.]{}
He said that if it comes about definitively the field, and string theorists in particular, will suffer badly. He said that it will be essential for theorists who entered the field most recently to figure out where previous generations went wrong and also to determine what experimenters should now look for.
In the following, I will argue that a root cause has been the exaggeration of the significance of the discovery of asymptotic freedom that has led to the historically profound mistake of trying to go forward by simply formulating new short-distance theories, supersymmetric or otherwise, while simultaneously ignoring both deep infra-red problems and fundamental long-distance physics.
In his recent “Welcome” speech[@NTur] at the Perimeter Institute, Neil Turok expressed similar concerns to those expressed by Gross. He said that
“All the [*{beyond the Standard Model}*]{} theories have failed ... Theoretical physics is at a crossroads right now ... [*{there is}*]{} a very deep crisis.”
He argued that nature has turned out to be simpler than all the models - grand unified, super-symmetric, super-string, loop quantum gravity, etc, and that string theorists, especially, are now utterly confused - with no predictions at all. The models have failed, in his opinion, because they have no new, simplifying, underlying principle. They have complicated the physics by adding extra parameters, without introducing any simplifying concepts.
The needed simplifying principle may simply be the fundamental long-distance requirement of full high-energy unitarity. The only known solution is the Critical Pomeron[@cri] that occurs uniquely (I have argued) in a bound-state S-Matrix that is embedded in QUD[^2] - a very weak coupling (almost conformal) massless SU(5) field theory[@kw], [@arw1]-[@arw8], Remarkably, it seems that the S-Matrix states and interactions are those of the Standard Model. Consistency with the underlying SU(5) unification is achieved, not by a short-distance extension of the theory, but instead by the infra-red enhancement of the Standard Model interactions by a “wee parton vacuum” of anomalous wee gauge bosons coupled to S-Matrix massless fermion anomalies. There is a Higgs boson but there is no new short-distance physics - just as is seen experimentally at the LHC,. The only new physics is electroweak-scale QCD interactions due to color sextet quarks (that could be discovered at the LHC) and there is no GUT scale.
The dynamical requirements of confinement, chiral symmetry breaking, the parton model and electroweak symmetry breaking all appear to be included in the bound-state S-Matrix, with dark matter and neutrino masses also present. Consequently, it seems that all the well-known Standard Model problems, including those which commonly motivate proposals for new short-distance physics, could be solved by the wee-parton anomaly solution of the long-distance problem of high-energy unitarity. Within the current theory paradigm, this is a radical proposition. Nevertheless, it may both simplify and unify the Standard Model, while dramatically changing expectations for new physics.
If it becomes accepted that supersymmetry and, by inference, string theory are not invoked by nature in the high-energy extension and/or the unification of the forces of the Standard Model, then an enormous part of the research effort in theoretical high-energy physics, over the last few decades, will have been critically wasted. Also, as a consequence, vast resources will have been wasted on corresponding experimental searches. Perhaps, the worst consequence is for the future. As currently planned, the incredible power and potential of the LHC will be used at absurdly high luminosity, where only the existence of very rare events due to new short-distance physics (such as more exotic SUSY theories) could possibly be discovered. Most likely, this physics does not exist and all that will be achieved is the extension of the nightmare scenario to even higher energy.
In addition, if new electroweak scale long distance physics exists, as I am arguing for, the wasted opportunity will reach truly historic dimensions. This physics could beautifully underly and unify the Standard Model but it is within a dynamical and philosophical framework that is very different to the current theory paradigm. Not surprisingly, therefore, it’s unambiguous discovery would require the LHC accelerator and detectors to be operated in a very different manner.
Very unfortunately, also, the technical formalism of multi-regge theory that I use is unfamiliar to almost all of the current, short-distance educated, high-energy physics community. Moreover, an enormous amount of development is needed to provide a detailed implementation of my (so far only outlined) construction of the QUD bound-state S-Matrix.
How Did it Happen?
==================
It is extremely important to understand how the current situation has come about and, even more important, to determine what might be done to remedy it. That Gross seems to have been the first leader in the field to publicly address the growing concerns is particularly remarkable since I will ascribe much of the “wrong direction” of the field to the unequivocal advocacy by him and his contemporaries of a narrow (in the extreme) focus on short-distance physics, following the discovery of asymptotic freedom. Asymptotic freedom is a crucial field theory property that is responsible for strong interaction deep-inelastic scaling and, more generally, for the very existence of quantum field theories in the euclidean short distance region. However, the strong interaction nirvana that was anticipated to follow the discovery has not materialized.
The hope was raised that the strong interaction, and hence all interactions, could be calculated and describe experiments performed in a broad short distance region of phase-space. Indeed it might seem, at first sight, that this hope has been realized. However, it is at a level that is entirely phenomenological. By appealing to the parton model, physicists have become accustomed to focusing on short-distance physics and to thinking of quarks and gluons as physical particles. The calculation of perturbative QCD amplitudes has then allowed extensive “beyond the Standard Model” searches to be undertaken. Unfortunately, the validity of all of this thinking is completely undercut by the necessary introduction of a wide range of supplementary (parton model) assumptions and parameter dependent concepts that prohibit any pretense that the calculations involved are providing unique, fundamental, predictions.
There is no derivation of the parton model and, consequently, there is no proof of the existence of parton distribution functions and certainly no possibility to calculate them. The current phenomenological application of the parton model involves the experimental determination of parton distributions from existing strong interaction data. The assumptions made are self-consistent only if it is assumed that QCD is well-understood in principle, even if not in detail, and so only new short-distance physics remains to be discovered. If there is new electroweak scale QCD physics to be discovered, of the kind that I will discuss, then the conventional parton model assumptions and procedure will clearly be inadequate.
On a deep level, the role and historical significance of asymptotic freedom has actually been that of a seductive siren enticing physicists to a shipwreck on the rocks of a belief that short distance physics can be a self-sufficient foundation for all future progress. Focussing entirely on new short-distance theories, supersymmetric or otherwise, while putting aside serious theoretical infra-red problems and ignoring significant experimental and theoretical long-distance physics results, has been foolhardy at best and, at worst, may be a major cause of the nightmare scenario. Understanding both the origin of the parton model and it’s coexistence with confinement is surely essential for long-term progress and demands top priority. It will be clear from the physics I describe that negelected long-distance physics could play a crucial role, with the origin of the parton model and confinement closely intertwined.
In Section 4, I will focus on an article written by Gross, entitled “Asymptotic freedom and the emergence of QCD”, that is an unrelenting advocacy of large transverse momentum physics. It was first published[@djg1] in 1992 and then republished[@djg2], without any changes, under the new title “Twenty Five Years of Asymptotic Freedom” in 1998. With added material and some editing, it was republished[@djg3] under the title “Asymptotic freedom and QCD: A historical perspective” and finally, it was again republished[@djg4], with further modifications, as his Nobel lecture “The discovery of asymptotic freedom and the emergence of QCD”. Perhaps unfairly, I will focus on the initial article as representative of the “We got it wrong” element of the history of asymptotic freedom.
The aim of the article was to provide both an outline of the discovery of asymptotic freedom and a broad historical description of the background research environment. Regrettably, in addition to appropriately emphasizing the significance of asymptotic freedom, there are also a number of inaccurate, misleading, and even wrong, arguments provided to make the case for a research program (theoretical and experimental) based solely on large transverse momentum physics, to the exclusion of other more complex, more difficult, and ultimately I believe, more fundamental directions.
A basic (and surely revealing) problem is apparent in the opening paragraph. In the second sentence, Gross says “history is written by the victorious”. He saw his research as part of a battle between quantum field theory and S-Matrix theory and, as he describes it, the battle was intense, long-lasting, and at times acrimonius. It was perceived as a fight for the very survival of quantum field theory and, indeed, it is possible that today’s nightmare scenario has come to pass as a direct, and profound, consequence of the all or nothing level of the fight. In a battle, all the spoils go to the victor and, apparently, to Gross and many others this meant that the discovery of asymptotic freedom had swept field theory to victory and vanquished all research programs based on, or even having any connection to, S-Matrix theory. The omnipotence of field theory was established!
The most serious problem threatening the existence of field theories had always been the uncontrollable (wildly divergent) behavior of the perturbation expansion in the large momentum region. In an asymptotically free theory this problem is removed and so does not prevent the existence of such theories. However, all the problems are transferred to the infra-red region and so, as has to be strongly emphasized, a jump to the existence of a full non-perturbative quantum field theory containing, in particular, correlation functions with all the conventionally assumed properties, is enormous and much, much, more (beyond asymptotic freedom) would be needed. The problems involved can not be addressed within the framework of short-distance field theory and, in fact, they are so serious that no four-dimensional field theory has ever been shown to exist and, most likely I would argue, never will be.
Gross also talks disparagingly about being taught, as a graduate student, that [*Field Theory = Feynman Rules*]{} and argues that this attitude was diverting attention from more important non-perturbative issues. Yet still today, in reality, the Feynman integrals are the only well defined formulation of a non-abelian theory such as QCD. What non-perturbative quantities they represent (if any) has still to be determined. The Feynman path integral, which is universally assumed to be the desired non-perturbative formulation, is deceptively alluring and is surely a very powerful formal tool. Unfortunately, because of the four-dimensional infinite volume divergence and the undefined nature of the function space implied by the Gribov copy problem, it is very unlikely to actually exist. Indeed, there is a “Catch-22” element to this issue. To prove the existence of the infinite volume limit requires an effective infra-red lagrangian without massless particles, but such a lagrangian would have to be derived from the (previously defined) infinite volume integral. It is important to note that the extensive, and largely successful, physical applications of lattice gauge theory always use finite volume approximations to the path-integral which, in effect, approximate finite momentum contributions to perturbative feynman integrals and which, in principle at least, could be Borel summable.
The Way Forward?
================
It would surely be much more satisfying, ultimately, to understand that nature has not been fickle in failing to take advantage of the beautiful elegance of supersymmetric field theories. Rather it may have not done so simply because such theories retain infra-red problems that actually prevent them from providing the basic necessity for a particle theory, namely a massive spectrum of physical states with a unitary S-Matrix. As I discuss further in Section 5, the perturbative infra-red problems of an asymptotically-free gauge theory are even more serious than the ultra-violet problems that have been solved. A priori, an infinite number of vacuum condensate parameters is introduced by the wild divergence of the perturbation series and there is no evidence that the conventional formulation of confinement, even if it could be proved, would be sufficient to allow a physical S-Matrix to exist. It can not be assumed that serious infra-red problems of this kind simply take care of themselves and that physicists need only calculate “safe, infra-red finite” short-distance cross-sections.
I will argue that the requirement of a consistent, unitary, particle S-Matrix is so strong that it is satisfied only in very specific circumstances in which the dynamics actually excludes the existence of off-shell correlation functions. To arrive at this conclusion it is necessary to accept the seriousness of all of the infra-red problems that are inherent in the formulation of quantum field theories containing massless fields. It is also essential to incorporate the theoretical results and understanding of high-energy unitarity acquired in the context of forward physics experimental results that have motivated my search for, and discovery of, the Critical Pomeron in a gauge theory. As I discuss in the next Section, in his article, Gross foolishly derides these experimental results as having provided no insight into the dynamics of the strong interaction. If I am right, they provide the ultimate key to the origin of the Standard Model.
Multi-regge S-Matrix theory provides a vehicle for handling elaborate multiparticle Feynman diagrams in the generalized infinite momentum kinematics of multi-regge limits. Remarkably[^3], perhaps, the emergence of full bound-state amplitudes via infra-red divergences can be studied. As I will outline in later Sections, I have used this formalism to argue that the existence of a bound-state unitary S-matrix with the desired forward physics properties, including unitary Critical Pomeron high-energy behavior, could be a very special property of QUD - the unique SU(5) massless field theory that I discussed in the last Section. The infinite-momentum dynamics involves massless fermions in a fundamental manner. Combinations of infra-red divergent anomalous gauge bosons coupled to massless fermion chirality transitions, produce “universal wee parton” gauge bosons in all S-Matrix multi-regge amplitudes. It is deeply significant that the anomalies responsible for the chirality transitions are only present in on-shell infinite-momentum amplitudes so that, as a result, the formation of bound-states and S-Matrix interactions is inexorably linked. The outcome is a very special version of QCD and ultimately, I believe, the origin of the Standard Model.
It will be truly, truly, ironic if the physics that I describe does actually provide the way forward from the nightmare scenario. Most likely, the S-Matrix is the only well-defined non-perturbative element of QUD and I believe that this is what the Standard Model is reproducing. Unification of the interactions is achieved, not by a short-distance extension of the theory, but instead by the underlying wee parton structure present in the S-Matrix as a consequence of long-distance anomaly dynamics. Moreover, the very different dynamical role played by fermions (via anomalies) and gauge bosons (via infra-red divergences) makes it apparent that this dynamics could not be present in a supersymmetric theory. Indeed, it may be that the field theory is only well-defined in the short-distance region where asymptotic freedom is operative and that, outside of this region, only the particle S-Matrix is well-defined. This has very important scientific and philosophical implications, as we discuss later. Certainly, the misleading and counterproductive nature of Gross’s arguments (described further in the next Section) that short-distance physics should be the focus of experiments is clear and the resulting historical misdirection glaring.
The simplification provided by QUD, as well as the radical philosophical and conceptual changes implied, surely goes in the direction anticipated by Turok. Obviously, the scientific and aesthetic importance of an underlying and unifying non-supersymmetric massless field theory for the Standard Model can not be exaggerated. QUD is self-contained and is either entirely right, or simply wrong! The only new physics is a high mass color sextet quark sector of the strong interaction that gives electroweak symmetry breaking and dark matter. Moreover, many other fundamental puzzles appear to be explained.
Only Short-Distance Physics Matters!
====================================
In this Section I will discuss David Gross’s article in some detail. I will list a few quotes (in the order in which they appear) that I think best illustrate the attitudes and priorities that (I believe) have contributed to the occurrence of the nightmare scenario.
1. [*“Field theory was in disgrace; S-Matrix theory was in full bloom. .. A powerful dogma emerged - that field theory was fundamentally wrong, especially in its application to the strong interactions ... was to be replaced by S-matrix theory; a theory based on general principles, such as unitarity and analyticity, ...”*]{}
That a battle is to be fought for the survival of field theory is crystal clear. As far as possible, after victory, all elements of S-Matrix theory, including much very powerful large distance formalism, will be dismissed as irrelevant dogma.
2. [*“The basic dynamical idea was that there was a unique S-Matrix that obeyed these principles. ... This is of course false. We now know that there are an infinite number of consistent S-Matrices that satisfy all the sacred principles. One can take any non-Abelian gauge theory, with any gauge group, and many sets of fermions (as long as there are not too many to destroy asymptotic freedom).”*]{}
The claim of non-uniqueness of the unitary S-Matrix is unbelievably misleading and, most likely, completely wrong. As I have already emphasized, gauge theory S-Matrices can only be calculated perturbatively and the D=4 expansions for every field (and string) theory are wildly divergent and, almostly certainly, can not be summed. There is certainly no non-perturbative formulation of any theory that can derive S-Matrix amplitudes - let alone discuss unitarity.
A need to worry about the existence of a unitary S-Matrix would surely have been a severe constraint on the wild proliferation of short-distance field theories (discussed in the next Section) that have been formulated without any concern for potential infra-red problems. The claimed non-uniqueness of the S-Matrix, by Gross, was intended to discredit the basic ingredient of the bootstrap program that was the early dynamical centerpiece of S-Matrix theory. However, although the bootstrap program has long since disappeared, uniqueness of the unitary S-Matrix may deeply intertwine with the origin of the Standard Model, as I have already suggested and will return to later.
3. [*“theorists and experimentalists reinforced each other’s conviction that the secret of the strong interactions lay in the high-energy behavior of scattering amplitudes at low momentum transfer. ... prompted by the regularities that were discovered at low momentum transfer, theorists developed an explanation based on the theory of Regge poles. This was the only strong interaction dynamics that was understood, for which there was a real theory. Therefore theorists concluded that Regge behavior must be very important and forward scattering experiments were deemed to be the major tool of discovery.”*]{}
This is, perhaps, the strongest contributor in the category of “How did we misread the signals?” It is both ignorant and scandalously arrogant. All the beautiful experimental results on the regge behavior of forward cross-sections, most deeply the isolated regge pole nature of pomeron exchange, are dismissed as resulting from theorists and experimentalists having nothing else to talk to each other about. In addition, fundamental theoretical results based on dispersion/regge theory analysis of unitarity are to be thrown away as irrelevant.
In fact, as a crucial outcome of the disdained experiments, it was definitively established that, in first approximation, the pomeron is a factorizing single Regge pole with no resonances on the trajectory. Because later, higher energy, experiments did not measure a variety of scattering processes, and because the forward region was very inadequately explored, it became possible for this inconvenient truth to be universally, and persistently, ignored (by BFKL enthusiasts in particular). However, recent LHC results have now beautifully confirmed the existence of the lower-energy diffraction peak at the highest energy.
=5.5in
Figure 1. TOTEM Measurement of the Forward Elastic Slope, (a) Data, (b) The Linear Extrapolation
The plot of measurements of the forward elastic slope shown in Fig. 1(a) is taken from a recent TOTEM paper[@TOT]. It illustrates the point perfectly. The TOTEM measurement (red dot) is stunningly accurate and, as is shown by the hand-drawn line in Fig. 1(b), very clearly lies on a linear extrapolation of the lower energy results. The linear energy dependence of the slope is, of course, a direct property of a Regge pole amplitude. Also evident in Fig. 1(a) is the inaccuracy of the intermediate energy results, reflecting the lack of priority given to the experiments involved. Without the TOTEM result, the freedom of invention for forward physics theorists (at the highest energies) would be largely unfettered!
As I will return to later, an approximate regge pole pomeron is deeply important for the existence of the parton model in QCD. It is very fortunate that a small minority of experimenters have continued to pursue the forward physics that has been pushed out of the mainstream by the short-distance mania that Gross and company have incited!
It will be very apparent by the end of this article that reproducing the forward physics seen in experiments is the key to understanding the present mysteries of QCD. All that can be said, apparently, from Gross’s lofty viewpoint is that
4. [*“... the explanation of Regge behavior remains an interesting, unsolved and complicated problem for QCD.”*]{}
All of the elaborate, and extensive, results on the reggeization of both gauge bosons and fermions in non-abelian gauge theories are considered to be irrelevant. Not, surprisingly I will also return to this subject later.
5. [[*“It was not at all realized by theorists that the secret of hadronic dynamics could be revealed by experiments at large momentum transfer that probed the short distance structure of hadrons.”*]{}]{}
6. [*“Only much later, after the impact of the deep inelastic scattering experiments that had been ridiculed by many as unpromising, was it understood that the most informative experiments were those at large momentum transfers that probe short or light-like distances.”*]{}
The almost universal, uncritical and unqualified, acceptance of these last statements has probably been as damaging as any other consequence of the discovery of asymptotic freedom. Taken literally, they are\[key\] wildly untrue. With no derivation of the parton model and the associated parton distribution functions, no derivation of confinement, and no derivation of “non-perturbative” contributions, initially precise short-distance calculations are submerged in a phenomenological sea. As a result, the only unambiguous information produced by short-distance experiments is that the strong interaction can be described by a non-abelian gauge theory. The details of which gauge group, which fermions, the mass spectrum, all come from long-distance experiments and would be very difficult to derive directly from the comparison of short distance calculations with strong interaction experiments.
7. [*“Thus the discovery of asymptotic freedom greatly reassured one of the consistency of four-dimensional quantum field theory. ... We are very close to having a rigorous mathematical proof of the existence of asymptotically free gauge theories in four dimensions – at least when placed into a finite box to tame the infrared dynamics that produces confinement. As far as we know, QCD by itself is a totally consistent theory at all energies.”*]{}
The introduction of the finite (euclidean) box is viewed as little more than a technical convenience. Apparently, all that remains is for confinement, which is not understood at all, to be elaborated sufficiently to allow the box to be taken away. As I will discuss more in the next Section, this is an incredible (head in the sand) attitude towards all that would have to be done to allow consistency of QCD at all energies to be claimed.
Some (well-known) quotes from Feynman seem particularly relevant, with the second perhaps addressing the origin of the nightmare scenario most directly.
*“The first principle is that you must not fool yourself and you are the easiest person to fool.”*
“It is in the admission of ignorance and the admission of uncertainty that there is a hope for the continuous motion of human beings in some direction that doesn’t get confined, permanently blocked, as it has so many times before in various periods in the history of man.”
The Infra-red Mysteries of QCD Can be Shelved!
==============================================
While the discovery of asymptotic freedom has clearly led to the idea that the focus of physics should be on short-distance processes, it has also produced an even greater acceptance of the idea that all the rest of physics is both unfathomable and unproductive to study. At the same time, paradoxically but very conveniently, it has become widely believed that QCD is sufficiently well understood that the following list of problems can be “shelved” while the march towards the discovery of new short distance physics proceeds.
1. [There is no proof, and not even an argument, that the euclidean functional integral exists, and so there is no known way to define QCD as a quantum field theory, outside of perturbation theory.]{}
2. [The infra-red renormalons determine that the perturbation expansion is wildly divergent and is not Borel summable, except when the maximum number of massless quarks allowed by asymptotic freedom is present. In this special case, which is at the core of my discussion later in the paper, there is an infra-red fixed-point which eliminates the renormalons. As a result, the theory is almost conformal and so lies in the “conformal window”. This is generally thought to imply that it can not have a massive particle spectrum. However, it actually could have an S-Matrix with a spectrum including massive particles[*- provided there are no corresponding off-shell amplitudes.*]{}]{}
3. [There is no derivation of the parton model. Assuming, nevertheless, that it exists, the factorization theorems needed for self-consistency apply only to deep-inelastic leading-order perturbation theory. They do not apply, at all, to hadron scattering.]{}
4. [There is no understanding of the confinement of color and chiral symmetry breaking. Most importantly, there is no understanding of how these properties can coexist with the parton model. ]{}
5. [There is no derivation of parton distribution functions and so, necessarily, their formulation and application is entirely phenomenological. Consequently, as I have discussed[@arw6] in previous papers the presence or absence of new, electroweak scale, physics in strong interaction cross-sections can be a matter of choice! Morover, infra-red finiteness has apparently become the only criterion that must be satisfied for physical cross-sections to be derived from QCD.]{}
The above QCD problems are universally acknowledged as unresolved. I would also add two more that have emerged from the unfathomable region of small transverse momentum physics and that also pose a major challenge to the conventional understanding of QCD. They are closely related.
- [There are no glueballs seen experimentally, even though they are widely believed to be a general consequence of QCD. This suggests that the nature of confinement is more selective than conventionally thought. ]{}
- [As discussed in the previous Section, the forward scattering experiments disdained by Gross established that, in first approximation, the pomeron is a factorizing single Regge pole. Recent LHC results have beautifully confirmed this result. It is a complete contradiction with the conventional description of near forward physics via the BFKL pomeron.]{}
It seems to have become accepted that this long list of infra-red problems, to which should be added the most important of general principles - non-perturbative unitarity, are irrelevant in the search for a physical theory. The paradigm has become that theories should first be discovered by their ability to fit experimental facts. If necessary, consistency can be looked for afterwards. Indeed, the conventional wisdom has become that, because of it’s success in describing experiments, the Standard Model must be a well-defined quantum field theory that, necessarily, must have all of the desired properties of a full, non-perturbative, theory[^4] - with a massive unitary S-Matrix automatically included.
Amazingly, as I believe it will be seen in retrospect, it is also commonly thought that, while the Standard Model fits the experimental data, it is just one of an infinity of renormalizable field theories that nature could have chosen. From this perspective, it has been assumed that progress beyond the Standard Model will be determined by new experimental phenomena that can be similarly fitted by an enlarged field theory. Since (it has also been thought that) long distance physics is sufficiently well understood via QCD, the new phenomena should appear at short distance. Given Gross’s assurance, theorists that have speculated about new physics that might appear have not been fettered by any fear that the long-distance physics demand for unitarity of the physical S-matrix could make any significant selection amongst candidate theories.
As I hope will be crystal clear by the end of this article, I am arguing that progress requires understanding both why QCD is the only gauge theory with all the necessary requisites to describe a unitary strong interaction and how it has to be modified to make it a fully consistent theory.
Until the advent of the nightmare scenario, the general belief has been that there will be a unifying quantum field theory that includes the Standard Model and which (if quantum gravity is to be included) will embed in an (ever more elusive) string theory. It is rarely acknowledged that the extra (far from trivial) assumption has to be made that unification is possible without conflict between the intrinsic non-perturbative applicability of QCD (involving confinement) and the perturbative applicability of the electroweak sector. Although there is no explicit understanding of how it could happen, it is thought that a transition from perturbative physics to non-perturbative confinement physics can simply be a consequence of the evolution of couplings with the scale involved.
As I noted in the previous Section, Gross considered that regge behavior is merely [*“an interesting, unsolved, and complicated problem for QCD”*]{}. This has been a widely held view that, I would argue, could not be further from the truth. The regge region is where the connection between perturbative and “non-perturbative” physics should be explicitly evident, since a mixture of small and large momenta is involved. Moreover, multi-regge theory provides the central framework for our discussion of the existence of a unitary S-Matrix in the following. Regge-region (reggeon) unitarity is deeply related to all the fundamental problems in the formulation of QCD and is central in our construction of a fully unitary gauge theory.
In fact, the nightmare scenario implies there is simply no experimental support for the viewpoint, long held by many, that the more difficult dynamical problems in QCD can be put aside and replaced by the guiding principle/paradigm that progress will come via inspired guesses for missing short-distance physics, combined with experimental verification via predicted rare processes. It should also be emphasized that there is no historical precedent supporting such a viewpoint.
As I have already implied, my solution to this situation satisfies, perhaps, Turok’s requirement that new simplifying principles be involved. Insistence on high-energy unitarity of the S-Matrix, as a principle, may actually uncover the desired extension of the Standard Model. Moreover, an extensive revision, both practical and philosophical, of the current theory paradigm is necessarily involved.
Why is S-Matrix Unitarity a Deep Requirement?
=============================================
During the barren years for quantum field theory, that produced the development of S-Matrix theory (before the discovery of asymptotic freedom!), basic questions concerning the necessary and sufficient elements of a quantum theory of particle physics were intensely discussed. The conclusion was that the minimum requirement is a unitary S-Matrix that describes the scattering of particles with a massive spectrum. The additional superstructure of off-shell Green’s function that a field theory provides (at least in perturbation theory) is both unphysically detectable and unnecessary - unless, as is often discussed nowadays, a short-distance unification with quantum gravity is desired.
Paradoxically, it is the unnecessary superstructure of off-shell short-distance amplitudes (in an infra-red cut-off field theory) that asymptotic freedom has shown the existence of. In fact, maximizing the physical consequences of asymptotic freedom requires, not only the simultaneous existence of a physical spectrum of particles and an infra-red finite S-Matrix, but also some form of parton model that allows asymptotic physical amplitudes to be expressed in terms of elementary field theory amplitudes,
The possibility that a field theory S-Matrix could exist without corresponding off-shell amplitudes has not been commonly envisaged and so cavalier assumptions that confinement will provide wave-functions coupling physical states to off-shell amplitudes are often made. Surprisingly, perhaps, I will argue that the existence of a parton model is a very special property that is connected with the origin of both the Reggeon Field Theory Critical Pomeron and the Standard Model and is in conflict with the existence of off-shell physical amplitudes.
In the midst of the various publications of David Gross’s article, I published a review[@arw9] describing some of the major results of S-Matrix theory and included a discussion of the Critical Pomeron and how it might connect to a non-abelian gauge theory. Here are some quotes suggesting that the disdained pomeron physics might point the way towards the origin of a unitary S-Matrix.
1. [*“... the Critical Pomeron is the summit of abstract S-Matrix Theory. It satisfies all known unitarity constraints ... provides a uniquely attractive possibility for ... an S-Matrix satisfying the maximum strength postulate.”*]{}
To potentially match with large transverse momentum gauge theory amplitudes, a regge pole pomeron must have unit intercept, i.e. it must satisfy the maximum strength postulate.
2. [*“If the Critical Pomeron is the only high-energy solution of unitarity that can match with asymptotic freedom then perhaps there is a uniqueness property for the strong-interaction S-Matrix ...”*]{}
3. [*But, why should Critical Pomeron asymptotic behavior be unique? Why ... should the pomeron be only a single regge pole plus multipomeron cuts?*]{}
It was introduced just because this is exactly what forward scattering experiments (that Gross said gave no information) tell us is the case!
4. [*A single regge pole ... uniquely has the factorization properties needed ... a universal wee-parton distribution in hadrons. ... the maximal applicability of short-distance perturbation theory. ... may well be essential to produce a completely finite (and unitary) S-Matrix.*]{}
5. [*Could the full S-Matrix including the electroweak interaction be unique?*]{}
6. [*... the uniqueness of the S-Matrix determines the underlying gauge theory, before ... gravity, ... would be strongly counter to today’s prevailing philosophy.”*]{}
The existence of a short-distance field theory may be essential, not only for the large momentum finiteness of the S-Matrix, but also, as discussed in [@arw9], for local analyticity properties. In later Sections I will outline how the unique SU(5) massless theory, QUD, appears. In the final Section, I will discuss why, in line with the above quotes, the properties that select this particular field theory may be essential for the existence of a physical S-Matrix.
As I have already noted in Section 4, it is currently accepted, almost without question, that “non-perturbative” QCD and all similar unbroken non-abelian gauge theories should be well-defined by the euclidean path integral. This is taken to imply that there must be a physical S-Matrix and that, moreover, the physical states appear as intermediate states in off-shell Green’s functions (derived from the path integral) of appropriate operators. Although there is no evidence to support this hypothesis, the considerable phenomenological success of “non-perturbative QCD” formalisms, particularly lattice QCD, implies there must be some approximate truth in the assumptions. Nevertheless, at the level where we are concerned with whether a particular theory is uniquely chosen by nature, it is important to emphasize that approximations are being made and that there are significant assumptions involved. (As I have emphasized earlier, lattice applications always use a finite volume approximation - that can, in principle, be related back to feynman diagram contributions.)
That the S-Matrix can be obtained from “non-perturbative” off-shell Green’s functions does not appear to be essential for any of it’s basic properties. The global analyticity domains that are normally thought to be a consequence of an off-shell field theory probably follow from the construction of physical high-energy amplitudes via the perturbation expansion. In fact, when the fields are massless and bound states related to infra-red anomalies are involved there is probably no general reason to expect a connection between Green’s functions and the S-Matrix.
In general, there is not even a formal property of a non-abelian gauge theory path integral which implies that a unitary, bound-state, S-Matrix can be derived via Green’s functions. Even worse, as I noted earlier, because of infra-red problems, the path integral itself is, most likely, not well-defined. Since there are no “non-perturbative” methods for constructing gauge theory S-Matrix amplitudes that do not, effectively, appeal to the formal euclidean functional integral, to seriously discuss whether a unitary S-Matrix exists in a general gauge theory is a highly non-trivial problem.
The Supercritical Pomeron and QCD
=================================
Although motivated directly by the forward scattering experiments, it was not clear what the deep significance of a regge pole critical pomeron might be. Since a Reggeon Field Theory renormalization group fixed-point is involved it might seem that there could be a link with the asymptotic freedom calculations. However, the pomeron was thought to involve complicated composite degrees of freedom that have no simple connection to the underlying gauge fields. Nevertheless, there was a mystery. While the regge pole pomeron that appears in experiments does not appear perturbatively in any gauge theory, all the elementary fields are associated with reggeizing particles. Naively, this might suggest that (in contradiction with confinement) the pomeron should be related to a non-abelian gauge boson and baryons should be gauge theory fermions. I will describe why, remarkably, this suggestion might be much closer to reality than conventional expectations would imply.
That the Critical Pomeron is selective in it’s association with an underlying field theory is seen directly via the approach from the supercritical side[@arw10]. The even signature of the pomeron requires a pure imaginary triple pomeron vertex that makes the effective action non-hermitian. Consequently, there is no supercritical minimum and only supercritical stationary points exist. As a result, a graphical expansion containing a pomeron field condensate is not straightforwardly obtained. Instead, a “supercritical pomeron condensate” is introduced as a zero transverse momentum component of the scattering states - via multi-regge theory. There is then a supercritical regge pole pomeron, together with an exchange degenerate vector regge pole. There are also singular vector exchange interactions due to the wee parton “pomeron condensate”. In effect, the rapidity divergences of the bare pomeron are replaced by transverse momentum divergences due to the vector particle.
The vector reggeon immediately suggests that the supercritical pomeron might be found by starting from the reggeon diagrams of color superconducting QCD[^5]. However, a smooth (asymptotically free) connection to unbroken QCD is realistically possible only if color sextet quarks produce electroweak symmetry breaking. A further constraint is that the pomeron condensate has to originate from gluon infra-red divergences that, to avoid the exponentiation of reggeization, must couple to anomalies that appear only when massless quarks are present.
Because the contribution of arbitrarily high order feynman diagrams is included, the analysis of multiparticle multi-reggeon diagrams that I have outlined in previous papers, systematically including the anomalies, is potentially the most extensive study of the infra-red divergences of QCD feynman diagrams that has yet been formulated. I consider multiparticle amplitudes within which bound-state scattering amplitudes can occur. Triangle anomaly diagrams occur in reggeon interactions that connect different (rapidity and transverse momentum) reggeon channels. Zero mass quarks generate anomaly poles in the triangle diagrams via chirality transitions that are zero momentum Dirac sea shifts of positive to negative (or vice versa) zero energies. As a result, chirality transitions determine how reggeon states in different channels couple, but do not contribute to the dynamical formation of individual reggeons.
I consider color superconducting QCD in the di-triple-regge region, where gluon divergences coupled to anomaly poles appear simultaneously in bound-state and interaction channels. An analysis involving reggeon interaction kernels shows that almost all infra-red divergences are exponentiated by reggeization, except for an overall divergence, which is subtracted to define physical amplitudes. A color confining “parton model” appears in which “anomalous wee gluons”, produced by multi-reggeon generalizations of the well-known anomaly current, provide vacuum-like universal wee partons. Anomaly poles coupled to the wee gluons both produce chiral Goldstone particle poles[^6] and also couple the wee gluons in distinct reggeon channels.
The color compensation by anomalous wee gluons beautifully resolves the mystery of the connection between a regge pole pomeron and gauge theory reggeization. $SU(2)$ anomalous gluons have $\tau=-1 = -C$ and so a $\tau=+1$ supercritical pomeron appears composed of an SU(2) singlet massive gauge boson reggeon plus anomalous wee gluons. It is exchange degenerate with a massive gluon reggeon, just as in supercritical RFT. The anomalous, $C= -1$, color charge parity of the pomeron is directly linked to the chiral symmetry breaking nature of the anomaly pole bound-states.
If all the quarks are massless, there is an infra-red fixed-point and the anomalous wee gluon divergence remains as arbitrarily higher-order reggeon diagrams are included. It effectively produces the desired supercritical pomeron condensate and correctly reproduces the supercritical pomeron interactions. Potentially, the final result is a confining, chiral symmetry breaking, spectrum generated in superconducting massless quark QCD.
Very importantly, the condensate remains after the restoration of the full SU(3) color symmetry and so it is present in both the physical pomeron and the physical bound-states of the critical theory. In effect, the supercritical phase involves color symmetry breaking of the condensate as well the production of a massive vector reggeon.
The Critical Pomeron in Massless QCD$_S$
========================================
The need for a massless quark infra-red fixed point determines that the Critical Pomeron can occur in QCD only when the maximum number of massless quarks is present. The only “semi-realistic” possibility, that we already arrived at by wanting to start from superconducting QCD, is massless QCD$_S$ in which there are six massless triplet quarks and two massless sextet quarks. Because of the infra-red fixed point there are no infra-red renormalons and the perturbation expansion is much less divergent - without the array of multi-gluon condensates normally produced. Consequently the possibility that perturbation theory (sums) can produce meaningful results is much improved. “Non-perturbative” physics could be provided, in principle at least, by topological (multi-instanton) contributions. The theory is “almost conformal”, with the infra-red fixed-point implying that if off-shell correlation functions exist (to which a renormalization group scaling transformation can be applied) they can have only scale-invariant intermediate states. Therefore, if there are massive physical states in massless QCD$_S$, off-shell correlation functions containing these states can not exist.
At first sight, there are many experimentally desirable features of the states that appear.
-1
These features would be realized if a QCD$_S$ bound-state S-Matrix exists. However, because of the large array of chiral symmetries, there would necessarily be a large multiplicity of massless Goldstone bosons that would create, probably insuperable, infra-red problems for the existence of such an S-Matrix. Moreover, that the Critical Pomeron appears only in massless QCD$_S$ makes it’s appearance in a massive hadron theory seem very unlikely, if not impossible! Fortunately, this conflict is resolved by the embedding of QCD$_S$ in QUD, as described in the next Section.
A priori, the critical behavior involves zero momentum quark/antiquark chirality transitions (Dirac sea shifts), due to initial SU(2) anomalous wee gluons, becoming random dynamical fluctuations associated with general anomalous wee gluons within the full SU(3) group. Note that, since the chirality-transition anomalies are a multi-regge S-Matrix phenomenon that can not produce off-shell correlation functions, reproducing the same physics at finite momentum would surely be very challenging. Since the wee gluon configurations all couple to multi-fermion instanton interactions we could perhaps, very loosely and for conceptual purposes only, think of the instanton interactions as responsible for the wee parton “vacuum” within which perturbative interactions operate. In practise, of course, this would be impossible to demonstrate directly.
The Critical Pomeron $~\longleftrightarrow~$ QUD
================================================
A remarkable result emerges when we consider combining the electroweak interaction with the Critical Pomeron. We discover a unique theory[@kw; @arw1] that is again massless, asymptotically free, and saturated with fermions, but has left-handed couplings to all fermions. Requiring asymptotic freedom and no anomaly, massless QCD$_S$ and the electroweak interaction embed uniquely in QUD, i.e. SU(5) gauge theory with left-handed massless fermions in the $5 \oplus 15 \oplus40 \oplus 45^*$ representation.
Under $ SU(3)\otimes SU(2)\otimes
U(1)$ -0.7
There is an infra-red fixed-point but now there are no exact chiral symmetries and so all bound-states should acquire masses. In fact, QUD has all the additional structure needed to generate, via massless fermion anomaly dynamics, a bound-state S-Matrix that reproduces the full Standard Model. The only additional elements are a dark matter sector and neutrino masses, both of which are extremely welcome.
There are three “generations” of both elementary leptons and elementary triplet quarks, and the theory is vector-like with respect to SU(3)xU(1)$_{em}$. SU(2)xU(1) is not quite right but, in the S-Matrix constructed[@arw1; @arw2; @arw6] via multi-regge theory, all elementary fermions are confined and only Standard Model interactions and states emerge. The color sextet sector provides “sextet pions” that produce electroweak symmetry breaking and sextet baryons - with the sextet neutron and antineutron providing stable, massive, dark matter particles. There is also a color octet sector that is responsible, via large $k_{\perp}$ anomalies, for the generation structure of the physical states.
Infra-Red Analysis of QUD Reggeon Diagrams
------------------------------------------
In QCD$_S$ the chirality transitions do not conflict with the vector gauge symmetry. In QUD they break the [**non-vector**]{} part of the gauge symmetry. The QUD reggeon diagrams are initially well-defined, if all reggeons are given masses. 24 and 5$\oplus$5$^*$ scalars give initial masses to all fermions and their decoupling leaves chirality transitions that break SU(5) to SU(3)$_C\otimes$U(1)$_{em}~$ [**in reggeon anomaly vertices only.**]{} Using 5$\oplus$5$^*$ scalars for the gauge bosons gives a smooth massless limit (via complementarity). Their successive decoupling gives reggeon global symmetries $~$
$\rightarrow SU(2)_C~, \rightarrow SU(4),~
\lambda_{\perp} \to \infty,~ \rightarrow SU(5)$
The last scalar is asymptotically free and so the $\lambda_{\perp} \to \infty$ limit can be taken between the SU(4) and SU(5) limits.
An elaborate analysis, involving all reggeon interaction kernels, shows how the complexity of the resulting anomalous wee gauge bosons increases with each infra-red limit. After the SU(2)$_C$ limit, the wee gluons of color superconducting QCD appear. With $\lambda_{\perp} \neq \infty $ many fermion loops violate Ward identities and in the next SU(4) limit left-handed gauge boson interactions (except those mixing with sextet pions) are eliminated. Also left-handed gauge bosons do not contribute as wee gauge bosons. Massive vector bosons, with the flavor symmetry of the sextet pions remain. After $\lambda_{\perp} \to \infty$ and the SU(5) limit, the Critical Pomeron and the massless photon appear together. Also, color octet chiral anomalies at $ k_{\perp} = \infty$ produce bound-states in Standard Model generations.
QUD Interactions and States
---------------------------
The final “universal wee partons” are combinations of vector coupling anomalous wee gauge bosons in the adjoint SU(5) representation. “Standard Model” vector interactions between bound-states appear, that preserve vector SU(3)xU(1)$_{em}$ and couple via anomalies.
-0.5
The elementary QUD coupling is kept very small by the infra-red fixed-point and so, for physical values of the Standard Model couplings to emerge, it is crucial that the infinite sums of wee gauge boson anomaly color factors enhance couplings, with $$~\alpha_{\scriptscriptstyle QCD} ~> ~\alpha_{\scriptscriptstyle em}~>>~ \alpha_{\scriptscriptstyle QUD} \sim
\frac{1}{120}$$ For bound-states, anomaly vertex mixing, combined with fermion and wee parton color factors, also produces a wide range of mass scales.
Three Standard Model generations of physical hadrons and leptons appear via octet anomalies that remain at infinite light-cone momentum after the full SU(5) symmetry is restored. When described in terms of fermions only, all bound-states are SU(5) singlets composed of five elementary fermions, two of which are color octets forming an “octet pion” large $k_{\perp}$ anomaly pole contribution. In addition to the color octets, lepton bound states contain three elementary leptons with two producing an anomaly pole. The electron is very close to elementary because the anomaly pole disturbs the Dirac sea minimally. The muon has the same constituents, but in a different anomaly pole dynamical configuration that will obviously generate a significant mass. Very significantly, the very small QUD coupling should be the origin of desirably small neutrino masses. Anomaly color factors imply $$M_{\scriptstyle hadrons} >> M_{\scriptstyle leptons} >> M_{\scriptstyle \nu 's} ~\sim~
\alpha_{\scriptstyle QUD}$$
Two QUD triplet quark generations give Standard Model hadrons - that mix appropriately. The physical b quark is a mixture[@arw6] of all three QUD generations. Sextet pions produce electroweak symmetry breaking by mixing with left-handed vector bosons. Sextet neutrons {“neusons”} are stable (the sextet proton {“proson”} is unstable) and so will provide dark matter - with many desirable properties[@arw6]. Top quark physics is very different from the Standard Model. However, because the sextet ${\eta}$ reproduces the Standard Model final states (at an electroweak scale mass !!) it is hard, experimentally, to distinguish the difference[@arw6]. Mixing of triplet and sextet states gives two mixed-parity scalars[@arw4; @arw5] - the $\eta_3$ and the $\eta_6$. The $\eta_3$ could be the “Higgs boson” discovered at the LHC. The $\eta_6$ may have been seen at the LHC - at the experimental $t\bar{t}$ threshold (where $t$ is the Standard Model top quark). “Tree-unitarity” suggests the combined $\eta_3$ and $\eta_6$ couplings should reproduce[@arw5] the Standard Model electroweak couplings of the Higgs boson.
There is a significant number of experimental phenomena[@arw2; @arw3; @arw6], already existing, that provide suggestive (if not definitive) evidence for the QCD sextet quark sector. To see this physics definitively at the LHC, the luminosity should be turned right down[@arw3] and the maximum possible energy attained. The detectors should also be modified to cover the maximal possible rapidity range at moderate transverse momentum. It would then be possible to detect the (relatively) large cross-section production of multiple vector bosons, across a wide rapidity range, that is currently missed[@arw3; @arw6] by the current focus on very large $p_{\perp}$, central rapidity, small cross-section physics. If the Higgs boson is indeed the $\eta_3$ then it will also be produced along with multiple vector bosons. It could even be that the production of dark matter neuson/antineuson pairs would be seen. Unfortunately, it is likely to take a long time for the needed turnaround in outlook to come to pass.
The Unique Unitary S-Matrix?
============================
The infra-red problem of constructing physical states that produce a unitary S-Matrix may actually be more difficult than, and the solution more special and at least as fundamental as, the solution of the ultra-violet problem of a field theory via asymptotic freedom. I have been led to a very beautiful proposition. The relevant entity for particle physics is the bound-state S-Matrix of a very special, small $\beta$-function, massless field theory that may only be evident in (and, therefore, need only exist as a quantum field theory in) short-distance perturbation theory. Mass generation becomes an S-Matrix property which is, effectively, separated from the problem of having a sufficiently well-defined short-distance field theory. This could have the great advantage that (as a matter of principle) there would be no need to confront the overwhelmingly difficult, and so far elusively intractable, problem[@jw] of constructing a full, non-perturbative, quantum field theory (with or without a mass gap) in four dimensions. Note that besides my construction via multi-regge theory, there is no other formalism capable of constructing bound-state scattering amplitudes. Without this ability it would not have been possible to envision the existence of an S-Matrix within a field theory with the properties of QUD.
There are many theoretical virtues for QUD as the origin of the Standard Model, including the following.
-0.5
It would, surely, be incredible if the Standard Model, with all of it’s complexity, has the underlying simplicity that I am suggesting. Nevertheless, all the necessary ingredients are present and if the predicted effects of the sextet sector are eventually seen clearly at the LHC, I doubt that the radical nature (with respect to the current theory paradigm) of what I am proposing will impede the rapid rise of interest in QUD that will surely ensue. It is important to emphasize again that, in principle, there is no freedom for variation in QUD. It is an “all or nothing” explanation of the origin of the Standard Model. Moreover, it predicts that
[**the “nightmare scenario”- a “Higgs boson” produced without new short-distance physics - will occur at the LHC !**]{}
We can ask, of course, why the underlying massless field theory has to be QUD. My answer would obviously be that I demand the appearance of the Critical Pomeron. However, I can alsoe phrase this requirement in more general terms. The infra-red fixed-point small $\beta$-function is required, firstly for the persistence of the scaling wee gluon interactions that enhance infra-red fermion anomaly interactions, and secondly to allow the color-superconductivity starting point that resolves the quantization ambiguities associated with Gribov copies and Gauss’s law. The vector interaction non-abelian gauge group has to be as large as SU(3) to produce, via scaling wee gluon interactions, a universal wee gluon distribution that can carry vacuum properties. This property is surely essential for the existence of an infinite momentum “parton model” that allows asymptotically free perturbation theory to produce an ultra-violet finite S-Matrix.
If the vector gauge group is larger than SU(3), the anomalous wee gluon scaling interactions are more complicated and the universal wee parton property is lost. The SU(3) gauge group can, however, be extended by left-handed interactions, that aquire a mass via the infra-red anomalies, since the only effect is to also generate bound state masses. This is a beneficial effect in that it alleviates potential S-Matrix infra-red problems. Asking that this extension generates masses for all bound-states while introducing no short-distance anomaly then brings us close to, if not directly to, QUD. Therefore, I believe that although I have funneled my discussion through the Critical Pomeron, in fact all of the properties needed to obtain a well-defined particle S-Matrix may come together to uniquely select QUD. If so,
[**the Standard Model could be reproducing the Unique, Unitary, S-Matrix !!!**]{}
[99]{}
David Gross, Outlook/Vision presentation at Strings 2013, Seoul, June 2013. Video only - http://www.youtube.com/embed/vtXAwk1vkmk.
Neil Turok, Perimeter Institute Welcome speech, September, 2013. Video only - http://pirsa.org/displayFlash.php?id=13080001.
A. A. Migdal, A. M. Polyakov and K. A. Ter-Martirosoyan, [*Zh. Eksp. Teor. Fiz.*]{} [**67**]{}, 84 (1974); H. D. I. Abarbanel and J. B. Bronzan, [*Phys. Rev.*]{} [**D9**]{}, 2397 (1974).
K. Kang and A. R. White, [*Int. J. Mod. Phys.*]{} [**A2**]{}, 409 (1987).
A. R. White, “Could a Weak Coupling Massless SU(5) Theory Underly the Standard Model S-Matrix, arXiv:1009.4850, 2011qcb..conf..496W, (2010)
A. R. White, “Past, Present, and Future Multi-Regge Theory” arXiv:1403.4454W (2014).
A. R. White, “The Potential Importance of Low Luminosity and High Energy at the LHC”, arXiv;1310.6649 (2013).
A. R. White, “A Higgs-Like Scalar With Mixed-Parity and Standard-Model-Like Couplings Could be Further Evidence for Underlying Massless SU(5) Unification”, arXiv:1301.5628 (2013).
A. R. White, “A Pseudoscalar Resonance That Could Resemble the Higgs”, arXiv:1206.0192 (2012).
A. R. White, “New Physics at the Tevatron and the LHC May Relate to Dark Matter Visible in UHE Cosmic Rays”, arXiv:1106.5662 (2o11).
A. R. White, “The “Crisis in Fundamental Physics” - Will the LHC Pomeron End it?”, arXiv:0803.1151 (2008).
A. R. White, “The LHC Pomeron and Unification of the Standard Model - a Bound-State S-Matrix Within a Fixed-Point Field Theory?”, arXiv:0708.1306 (2007).
D. J. Gross, “Asymptotic freedom and the emergence of QCD” Published in In \*Stanford 1992, The rise of the standard model\*, hep-ph/9210207.
D. J. Gross, “Twenty five years of asymptotic freedom”, [*Nucl. Phys. Proc. Suppl.*]{} [**74**]{}, 426 (1999), hep-th/9809060.
D. J. Gross, “Asymptotic freedom and QCD: A historical perspective”, [*Nucl. Phys. Proc. Suppl.*]{} [**135**]{}, 193 (2004).
D. J. Gross, “The discovery of asymptotic freedom and the emergence of QCD” [*Proc. Nat. Acad. Sci.*]{} [**102**]{}, 9099 (2005) 9099, [*Rev. Mod. Phys.*]{} [**77**]{}, 837 (2005).
TOTEM Collaboration, G. Antchev et al. “Measurement of proton-proton elastic scattering and total cross-section at sqrt s = 7 TeV” CERN-PH-EP-2012-239, EPL 101 (2013) 21002.
“Quantum Yang-Mills Theory”, A. M. Jaffe, E. Witten, Clay Mathematics Institute Millenium Prize Problem (2000).
A. R. White, “The Past and Future of S-matrix Theory”, Published in In \*Pike, R. (ed.) et al.: Scattering, vol. 2\* 1483, hep-ph/0002303.
A. R. White, “Analytic Multi-Regge Theory and the Pomeron in QCD I: the Supercritical Pomeron”, [*Int. J. Mod. Phys.*]{} [**A6**]{}, 1859 (1991).
[^1]: [email protected]
[^2]: QUD $\equiv$ Quantum Uno/Unification/Unique/Unitary/Underlying Dynamics
[^3]: The multi-regge formalism may be alone in allowing the study, in four dimensions, of gauge theory bound-state amplitudes.
[^4]: -1
[^5]: Color superconducting QCD reggeon diagrams potentially resolve the quantization problems of Gribov copies and Gauss’ law, via zero momentum longitudinal gluons.
[^6]: Via the triangle diagram, a Goldstone anomaly pole is both a simple quark/antiquark state, with one of the pair having zero momentum and zero negative energy or, alternatively, it is a physical reggeon state containing physical quark/antiquark reggeons plus anomalous wee gluons.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A Monte-Carlo study is presented using ground based measurements of the electromagnetic part of showers initiated in the atmosphere by high energetic cosmic rays to reconstruct energy and mass of primary particles with energies above 300TeV. With two detector arrays measuring the Cherenkov light and the particle densities as realized in the HEGRA experiment the distance to the shower maximum and the lateral development of air showers can be coarsely inferred. The measurable shower properties are interpreted to determine energy and energy per nucleon of the primary particle.'
---
=12.6cm =1.3cm =1.5cm
\#1[$_{\mbox{\rule{0cm}{0.2cm}#1}}$]{} \#1[\_]{}
**A New Method to Reconstruct the Energy**
**and Determine the Composition of Cosmic Rays**
**from the Measurement of Cherenkov Light**
**and Particle Densities in Extended Air Showers**
A.Lindner, [*Uni.Hamburg, II.Inst.f.Exp.Physik,*]{}
*LuruperChaussee149, D-22761Hamburg, Germany, E-Mail:[email protected]*
Introduction
============
This paper describes a new method to derive the elemental composition and the energy spectrum of cosmic rays for energies above a few hundred TeV from measurements of Cherenkov light and particle densities at ground level. The main idea is to determine the distance between the detector and the shower maximum, which is used to correct experimental observables for fluctuations in the shower developments. The depth of the shower maximum itself turns out to be a coarse measurement of the energy per nucleon of the nucleus hitting the atmosphere. The paper is organized as follows:\
After describing the experimental observables which will be used to reconstruct characteristics of the primary particle, the event simulation is sketched and features of the longitudinal and lateral shower development in the atmosphere are considered. The following three sections deal with reconstruction of the position of the shower maximum, of the energy per nucleon and of the primary energy. Finally methods to determine the chemical composition are described followed by the summary and conclusions.
The Experiment and Observables
==============================
Although the method described in this paper was developed primarily for the HEGRA experiment it can equally well be applied to any installation registering Cherenkov light and charged particles of extended air showers (EAS). The method can be easily generalized to all experimental setups which allow the determination of the distance to the shower maximum. However some properties of the HEGRA experiment need to be mentioned to understand details discussed in the following sections.\
The experiment HEGRA is a multi-component detector system described in detail elsewhere [@hegra] for the measurement of extended air showers (EAS). At a height of 2200m a.s.l. it covers an area of 180${\rm \cdot 180\,m^2}$. In this paper only the scintillator array of 245 huts with a grid spacing of 15m including denser part near the center and the so called AIROBICC array of 72 open photomultipliers measuring the Cherenkov light of air showers on a grid with 30m spacing also with a central concentration, are used. The energy threshold (demanding a signal from at least 14 scintillator or 6 AIROBICC huts) lies at 20TeV for proton and 80TeV for iron induced showers.\
\[observables\] The measured particle density in the plane perpendicular to the shower axis is fitted by the NKG formula [@nkg]. In the fit a fixed Moliere radius of 112m is used. The shape parameter [*age*]{} and the integral number of particles [*Ne*]{} result from the fitting procedure. As the HEGRA scintillators are covered with 5mm of lead (which suppresses the detection of low energy electrons but allows the measurement of photons after pair production in the lead) the values obtained for [*age*]{} and [*Ne*]{} cannot be compared to simple expectations from the cascade theory: [*age*]{} maybe smaller than 1 although the measurement takes place well behind the shower maximum while the shower size [*Ne*]{} is generally larger than for measurements without lead on the scintillator detectors.\
The Cherenkov light density is only analyzed in the interval 20m ${\rm <}$ r ${\rm <}$ 100m from the shower core due to technical reasons although HEGRA in principle could sample the Cherenkov light density up to 200m. In the range between 20 and 100m the Cherenkov light density can be well described by an exponential $${\rm \rho_C(r)={\it a} \cdot exp(r\cdot {\it slope})}.$$ As in the NKG fit two parameters are obtained from the analysis of the Cherenkov light: the shape parameter [*slope*]{} and the total number of Cherenkov photons reaching the detector level between 20 and 100m core distance [*L(20-100)*]{}.
Simulation
==========
EAS in the energy range from 300TeV to 10PeV were simulated using the code CORSIKA 4.01 [@cors]. The model parameters of CORSIKA were used with their default values and the fragmentation parameter was set to “complete fragmentation”. This results in a complete disintegration of the nucleus after the first interaction. Showers induced by the primary proton, ${\rm \alpha}$, oxygen and iron nuclei were calculated.\
The number of generated Cherenkov photons corresponds to a wavelength interval of ${\rm [340-550\,nm]}$. In the main this paper assumes perfect measurements of the number of particles and Cherenkov photons and a perfect shower core determination in order to concentrate on the physical principles and limitations of the methods to be described. To study the influence of the realistic experimental performance the events were passed through a carefully checked detector simulation [@volker] (performed with measured response functions) and reconstructed with the same program as applied to real data. Here each event was used 20 times to simulate different core positions inside and impact points outside the experimental area, which nevertheless fire sufficient huts of the arrays to fulfill the trigger conditions.\
In total 1168 events were generated with CORSIKA4.01 with zenith angles of 0,15,25 and 35$^0$ at discrete energies between 300TeV and 10PeV.
The Development of Showers in the Atmosphere
============================================
Some basic characteristics of the EAS simulated with CORSIKA4.01 are summarized here. Features independent or sensitive to the mass of the nucleus hitting the atmosphere are described. These will allow the reconstruction of primary energy and mass from the observables mentioned above.
The longitudinal Shower Development {#long}
-----------------------------------
Shown on the left of Figure\[longdev\] are the mean longitudinal developments of 300TeV proton and iron induced air showers, where electrons and positrons above an energy of 3MeV were counted. This will be subsequently called, the shape of the longitudinal shower development. For each shower the maximum (defined as the point in the shower development with the maximal number of particles) was shifted to zero before averaging. Afterwards the mean distribution was normalized to the mean particle number at the shower maximum.
=10.55cm
\[longdev\]
Concerning the shape of the longitudinal development behind the shower maximum no systematic differences depending on the primary particle are visible. The right plot in Figure\[longdev\] shows the change in the longitudinal development with increasing primary energy. The shapes broaden independently of the primary particle. Both plots may be explained by a lucky combination of two effects:
1. As visible from simulated proton showers at 300TeV and 5PeV the longitudinal shower shape gets broader with increasing energy.
2. After the first interaction an iron induced shower can be described as a superposition of nucleon induced subshowers. Each of them have different subshower maxima fluctuating around a mean value. To achieve the distributions in Figure\[longdev\] the maxima of the whole EAS (and not the maxima of the subshowers) were overlayed. Therefore a Fe shower appears to be broader than a proton shower of the same energy per nucleon.\
In CORSIKA 4.01 (other simulations have to be tested) both effects combine in such a way that the longitudinal shape of the EAS behind the maximum becomes independent of the mass of the primary nucleon for the same primary energy.\
The mean atmospheric depths of the maxima depend on the energy per nucleon E/A (see section\[detea\]) and are subjected to large fluctuations. Figure\[detenuc\] shows the corresponding correlation: the column density traversed by a shower up to its maximum, named depth of maximum in the following (calculable from the distance and the zenith angle), is correlated with E/A for all different simulated primaries from p to Fe and all zenith angles.
=10.55cm
\[detenuc\]
With the “complete fragmentation” option in our simulations the correlation follows a linear function. From Figure\[detenuc\] an elongation rate of approximately 82g/cm$^2$/log$_{10}$(E/E$_0$) is derived: $${\rm depth(max)\,=\,\left[ (335 \pm 3) \,+\,(82 \pm 2) \cdot log_{10}
\left(\frac{E/A}{TeV}\right)\right]\,g/cm^2 .}
\label{eqenuc}$$ If the depth of the shower maximum is measured the energy per nucleon E/A can be inferred. Due to statistical fluctuations in the shower development and therefore in the depth of the shower maximum (Figure\[detenuc\]) the resolution for E/A is modest. The spread decreases with increasing nucleon number A as the EAS of a complex nucleus consists of many overlapping nucleon induced subshowers so that the whole EAS exhibits less fluctuations than the individual subshowers. The resolution improves slightly also with rising E/A because more interactions take place until the shower maximum is reached. Figure\[detenuc\] (right) shows a parameterization of the fluctuations of the depth of the shower maxima.\
It is interesting to note that for a specific primary particle and energy the number of particles in the shower maximum Ne(max) is independent of the depth of the maximum. Therefore this number differs between proton and iron induced showers of the same primary energy even if the shower maxima are accidently at the same position.\
The most important characteristics of the longitudinal shower development discussed above and which will be used for the reconstruction of primary energy and mass in the next sections are:
- The longitudinal shower development behind the shower maximum does not depend on the mass of the primary particle and only slightly on the primary energy.
- The mean depth of the shower maximum is determined only by the energy per nucleon E/A.
- Fluctuations in the position of the shower maximum decrease with increasing nucleon number and slightly with increasing energy.
The lateral Shower Development {#showerlat}
------------------------------
In hadronic interactions the typical transverse momentum stays roughly constant with energy. Therefore the lateral spread of hadronic showers should decrease with increasing energy per nucleon as the ratio of transverse to longitudinal momentum gets smaller in the early part of the shower development where the energies of the interacting particles are still comparable to the primary energy. In principle this effect could be measured for a known distance to the shower maximum i.e. by comparing the number of Cherenkov photons reaching the detector level relatively close to the core with the number of all photons detectable at the ground level or by analyzing [*age*]{}. In this way a separation of heavy and light primaries turns out to be possible at energies below 1PeV. At energies in the knee region nearly no differences between proton and iron showers remain. Obviously here E/A even for iron showers becomes so large that any influence of the hadronic transverse momentum is washed out and the lateral shape of the shower is dominated by scattering processes and interactions of particles of relatively low energies in the later part of the shower development.
Reconstruction of the Position of the Shower Maximum {#recmax}
====================================================
This section deals with the reconstruction of the distance between the shower maximum and the detector. It can be determined from the shape of the lateral Cherenkov light density ([*slope*]{}) and in principle also from the shape parameter [*age*]{} of the particle distribution at detector level.\
\[recmaxcl\] As already noticed by Patterson and Hillas [@hillas] for showers with energies above 1PeV the distance between the detector and the maximum of an EAS can be inferred from the lateral distribution of the Cherenkov light within about 100m core distance. This is possible, because Cherenkov light emitted at a specific height shows a specific lateral distribution at detector level. The light from the early part of the shower development, where the energies of the particles are still very high so that scattering angles are very small, is concentrated near 120m (the so called Cherenkov ring). Cherenkov light produced closer to the detector level hits the ground closer to the shower core. The measurable Cherenkov light density of one EAS is the sum of all contributions from all heights, where lateral distributions from different heights enter with amplitudes corresponding to the number of Cherenkov light emitting particles in the different heights. Hence the shape of the measurable lateral light density distribution depends on the longitudinal shower development. If the shower maximum approaches the observation level more light is produced close to the detector reaching the ground near the shower core. Consequently the lateral Cherenkov light density in the range up to 100m core distance drops the steeper the closer the shower maximum approaches the detector.\
In Figure\[maxcor\] the correlation between the distance to the shower maximum and the parameter [*slope*]{} derived from the Cherenkov light distribution is plotted for different primary nuclei and zenith angles up to 35 degree for primary energies of 0.3 and 5PeV.
=10.55cm
\[maxcor\]
The distance to the shower maximum can be accurately determined from [*slope*]{} independent of the type of the primary particle and zenith angle[^1] The correlation between distance and [*slope*]{} depends slightly on the primary energy due to the changing longitudinal shower development. Figure\[maxcor\] shows simple polynomial fits describing the correlation rather well. The dependence of the fit parameters on log(E) were again parameterized with polynomials resulting in a two dimensional function of [*slope*]{} and log(E) to determine the distance to the shower maximum. The systematic uncertainties in the reconstruction of the distances depending on particle type or primary energy are less than 5% increasing a little for zenith angles of 35$^0$. Such a systematic error for large zenith angles \[sysangel\] is expected as these showers develop longer at high altitude where the threshold energy for electrons to produce Cherenkov light is higher than for showers with vertical incidence.\
Two shower properties contribute to the accuracy of the determination of the shower maximum. For a given primary energy the resolution improves with decreasing distance between detector and shower maximum and with increasing number of nucleons.
=10.55cm
\[distrms\]
Accidentally both contributions behave in such a manner that for fixed zenith angle and primary energy the resolution becomes independent of the mass of the primary particle within the statistical errors of the event sample. Corresponding results are plotted in Figure\[distrms\]. The right part of Figure\[distrms\] shows the reason for the finite resolution of the distance determination with [*slope*]{}. Showers, where the distance is underestimated, do not decay as fast as an average shower behind the shower maximum. Consequently [*slope*]{} is smaller than expected and the distance is reconstructed too small. Showers with overestimated distances exhibit a shorter longitudinal extension behind the shower maximum. It is interesting to note that the length of the shower behind the maximum is anti correlated with the number of particles in the maximum: the faster the decay after the maximum the more particles arise in the maximum.\
It is worthwhile to note that other approaches, like the reconstruction of the shower maximum from measurements of the time profile of the Cherenkov light pulses at core distance beyond 150m [@volker] show different limitations as different shower properties than with [*slope*]{} are measured. In principle it is also possible to determine the distance to the shower maximum with [*age*]{}, but the experimental resolution of an [*age*]{} measurement cannot compete with a [*slope*]{} determination. An improvement is only possible with a much denser coverage with active detector components than in present experiments (typically around 1%). In addition [*age*]{} depends on the primary particle type so that an unbiased measurement of the position of the shower maximum is not possible at energies below 1PeV.
Determination of Energy per Nucleon E/A {#detea}
=======================================
=6.5cm =4.3cm \[enucrms\]
With known distance to the shower maximum and known zenith angle the penetration depth of the shower into the atmosphere until it reaches the maximum can be inferred. E/A can be estimated using the correlation shown in Figure\[detenuc\]. The resolution for E/A is limited by the natural fluctuations of the shower maxima positions (Figure\[detenuc\]) and by principle accuracy limits of the [*slope*]{} method (Figure\[distrms\]). Of course further experimental errors may contribute in addition. In Figure5 the different contributions (to be added quadratically) are compared. The uncertainty for proton induced showers is always dominated by statistical fluctuations of the shower developments, whereas for iron showers at low energies the intrinsic uncertainty of the [*slope*]{} method contributes significantly.
At energies around the “knee” the accuracy is always limited by variations of the shower developments.
Reconstruction of primary Energy
================================
With a HEGRA type of detector two different methods may be used to reconstruct the primary energy:
1. Interpreting the shower size at detector level measured by the scintillator array corresponds to a determination of the leakage out of the “atmospheric calorimeter”. These “tailcatcher” data allow for an accurate energy reconstruction if combined with information on the shower development. This ansatz can be applied as the shape of the longitudinal development behind the shower maximum does not depend on the primary particle (see section\[long\]) and the distance to the shower maximum can be measured independently of the mass of the primary nucleus.
2. The measurement of the amount of Cherenkov light makes use of the atmosphere as a fully active calorimeter. Although in principle superior to the “tailcatcher” approach this idea suffers from the large extension of the Cherenkov light pool. Depending on the distance to the shower maximum only 20 to 55% of all Cherenkov photons reach the detector within 100m core distance. In contrast to this the particle density drops very fast with increasing distance to the shower core so that a detector of the HEGRA size is sufficient to determine the number of all particles reaching the ground level.
Both algorithms described in the present paper can roughly be divided into two steps: first the distance of the shower maximum to the detector (derived from the shape of the lateral Cherenkov light density distribution) is used to correct for different longitudinal shower developments. As only experimental quantities measuring the electromagnetic part of the air shower are considered here it follows naturally, that only the energy deposited in the electromagnetic cascade can be reconstructed directly. In a second step a correction for the non measured energy has to be performed. This correction depends on E/A only, which is determined from the depth of the shower maximum as described in the previous section.\
The following plots and parameterizations only take into account showers which reach their maximum at least 50g/cm${\rm^2}$ above the detector, because otherwise one can hardly decide whether a shower reaches its maximum above detector level at all. The treatment of showers arriving at detector level before reaching their maximum has to be considered separately.
Energy Reconstruction from Particle and Cherenkov Light Measurement {#ene}
-------------------------------------------------------------------
In this section the primary energy will be reconstructed in the following manner:
1. With the known distance to the shower maximum the number of particles in the shower maximum Ne(max) is derived from the measurement at detector level [*Ne*]{}.
2. Ne(max) is proportional to the energy deposited in the electromagnetic cascade.
3. From the energy per nucleon E/A and Ne(max) the total primary energy is inferred.
{#section .unnumbered}
[l]{}\[0pt\][7cm]{}
=6.5cm =4.3cm
The first step for reconstructing the primary energy is the determination of Ne(max) from the observables [*Ne*]{} and [*slope*]{}. As the scintillator huts of the HEGRA experiment are covered with 1${\rm X_0}$ of lead the total number of particles is not measured directly. In simulations the ratio of measured particles and the number of particles before the lead was found to depend only on the distance to the shower maximum but not on primary energy nor on the nucleon number of the primary particle. Therefore no systematic uncertainties in the energy reconstruction originate from the lead coverage. Due to the conversion of photons in the lead the measured number of particles is larger than Ne(max) for showers reaching their maximum close to the detector. In Figure\[eemcor\] the ratio of [*Ne*]{} to [Ne(max)]{} is correlated with [*slope*]{} measuring the distance between detector and shower maximum. Using [*slope*]{} in this correlation instead of the distance permits the handling of all energies in one correlation:
$\xi$([*slope*]{})
: is applied to determine Ne(max) from the observed [*Ne*]{} at detector level and from [*slope*]{} measuring the distance to the shower maximum.
No change of the correlation for different primary energies was observed (compare table\[tabene\] and the related discussion in the text). No systematic differences between different primary particles are visible. In the second step the primary energy is determined from Ne(max) and E/A. Two sub steps are necessary here: first the energy contained in the electromagnetic part of the EAS has to be derived, followed by an E/A dependent correction to determine the primary energy. The electromagnetic energy is proportional to Ne(max) for a fixed shape of the longitudinal shower development. Because the shape changes slightly with primary energy E both Ne(max) and E are necessary to determine the electromagnetic energy. During our simulations the total amount of electromagnetic energy in an EAS (defined as the sum of all ${\rm \pi^0}$ energies) was not recorded so that the ratio of electromagnetic energy to the primary energy and the ratio of Ne(max) to the electromagnetic energy could not be calculated directly for the simulated events. Hence arbitrary factors may be multiplied to the two correction functions below as long as their product is kept constant. The functions are:
$\xi$(E)
: takes into account the change of the longitudinal shower development with E.
$\xi$(E/A)
: is used to correct from the electromagnetic energy to the total energy E of the primary nucleus and will be discussed a little more detailed.
The fraction of the primary energy which goes into the electromagnetic part of the shower rises with increasing energy as the probability for hadrons to perform subsequent interaction with the production of additional neutral pions increases with the hadron energy. For a nucleus the fraction of the electromagnetic energy should depend only on E/A. Following the results of [@eem] a function was fitted to the correlation of Ne(max)/E with E/A using the generated MC events: $${\rm \xi\scm{em}(E/A)\,=\,
\frac{Ne(max)}{E \cdot \xi\scm{lon}(E)}\,=\,
\left[ 1.- \left( \frac{E/A}{33\,GeV} \right)
^{-0.181}\right]\,TeV^{-1}\ .}
\label{eqgab}$$ The ratio of this correction for protons to iron at 300TeV amounts to 1.34 decreasing to 1.16 at 5PeV. The difference between proton and iron showers is larger than derived by extrapolating the results in [@eem] to the mean atomic number of air because the fraction of the total energy deposited in the electromagnetic cascade is different in air showers compared to showers developing in solid state calorimeters: in air the interaction length for charged pions is comparable to their decay length so that the competition between pion decay and secondary interaction with subsequent production of neutral pions (feeding the em. cascade by their decay to two photons) lowers the fraction of energy deposited in the electromagnetic cascade. Now the primary energy is calculable by $${\rm E\,=\,\frac{Ne(max)}{\xi\scm{lon} \cdot \xi\scm{em}}
\,=\,\frac{{\it Ne}}{\xi\scm{dis} \cdot \xi\scm{lon} \cdot \xi\scm{em}}\ \ \ .}
\label{eqenemax}$$ Due to the energy dependence of the correlation between [*slope*]{} and distance to the maximum and of Ne(max)/E (both due to a slightly changing shape of the longitudinal shower development) energy and distance to the shower maximum cannot strictly be determined separately but have to be calculated iteratively. However the energy dependencies are small. Therefore in the calculation as a start value a parameterization of the correlation of distance and [*slope*]{} neglecting any energy dependence is used to derive a first distance value. With this an energy estimation is calculated and with this energy a new distance. This distance in turn gives a new energy value again. After two iterations usually neither the distance nor the energy results change further.\
The application of the whole procedure to simulated events results in systematic uncertainties on the order of 5%. Several contributions to the energy resolution for iron and proton showers are listed in table\[tabene\].
Method Fe 300TeV Fe 5PeV Prot. 300TeV Prot. 5PeV
---------------------------- ---------------- --------------- -------------- --------------
Ne(max) $(6 \pm 1)$% $(4 \pm 1)$% $(15\pm 1)$% $(9 \pm 1)$%
Ne at detector,
dist. from MC $ (12 \pm 1)$% $(6\pm 2)$% $(17\pm 1)$% $(7\pm 1)$%
[*Ne*]{} from fit,
dist. from MC $(20\pm 2)$% $(7 \pm 2)$% $(18\pm 1)$% $(11\pm 2)$%
[*Ne, slope*]{} from fits,
E/A from MC $(22\pm 2)$% $(7\pm 2)$% $(15\pm 1)$% $(11\pm 2)$%
[*Ne, slope*]{} from fits,
E/A reconstr. $(31\pm 3)$% $(12 \pm 4)$% $(25\pm 2)$% $(11\pm 2)$%
: [Rms values of different quantities contributing to the energy resolution which can be achieved with the method using [*Ne*]{} and [*slope*]{}. In the first two rows the identity of the primary particle is used from the simulations. “Ne(max)”, “Ne at detector” and “dist” are MC quantities, [*Ne*]{} and [*slope*]{} experimental observables. E/A denotes the energy per nucleon, which in the last line of the table is reconstructed from [*slope*]{} and the zenith angle.]{}
\[tabene\]
For 300TeV iron showers most of the uncertainties stem from the NKG fit to the scintillator data with subsequent fluctuations in [*Ne*]{} and from a modest resolution for E/A (compare Figure5). At 300TeV the ratio of reconstructed to generated energy exhibits a tail to large values which originates from the large uncertainties in the determination of E/A and the relatively large corrections depending on E/A as shown in equation\[eqgab\]. Already at an energy of 500TeV the tail to high energies nearly disappears resulting in a rms value of 20%. For energies of 1PeV and larger the resolution amounts to roughly 10%. The much improved energy resolution is achieved due to better [*Ne*]{} and E/A determinations and a smaller correction depending on E/A.\
The energy resolution for proton showers improves from 25% at 300TeV to about 10% at 5PeV. Even a direct measurement of Ne(max) and an unambiguous identification of its mass would not improve the energy resolution very much compared to the reconstruction using only experimental observables.
Energy Reconstruction from Cherenkov Light alone {#ecl}
------------------------------------------------
As in the previous section the energy can be reconstructed by replacing [*Ne*]{} by [*L(20-100)*]{}. The fraction of the Cherenkov light in the interval from 20 to 100m depends on the distance to the shower maximum (geometry) and on E/A, because the lateral spread of an EAS decreases with decreasing ratio of transverse to longitudinal momentum in the interactions. The following effects were taken into account for the energy reconstruction with Cherenkov light only:
$\zeta$([*slope*]{},E):
: The fraction of light contained in [*L(20-100)*]{} depends on the distance to the shower maximum. Due to the differences in the longitudinal shower development an additional small energy dependence was also taken into account (corresponding to $\xi$ and $\xi$ of the previous section).
$\zeta$(E/A):
: For a given distance the fraction in [*L(20-100)*]{} depends on E/A also. As in the previous section ($\xi$(E/A)) this originates from the energy fraction deposited in the electromagnetic cascade of an EAS, but the lateral extension of the Cherenkov light pool also contributes.
$\zeta$(height):
: The threshold for electrons to produce Cherenkov light varies from 38MeV at a height of 300g/cm$^2$ to 21MeV at the detector level of 793g/cm$^2$. Therefore the amount of Cherenkov light generated in the atmosphere depends on the height of the shower maximum.
The correction depending on energy per nucleon is given explicitly below. $${\rm \zeta\scm{em}(E/A)}\,=\,
{\rm
\frac{{\it L(20-100)}\,/\, 1.3 \cdot 10^7}
{E \cdot \zeta\scm{dis} \cdot \zeta\scm{den}}}
\, =\,
{\rm
\left[ 1.- \left( \frac{E/A}{178\,GeV} \right)
^{-0.180}\right]\,TeV^{-1}.}
\label{eqlgab}$$ This correction is larger than equation\[eqgab\] because of the correction for the Cherenkov light beyond 100m distance. The remaining systematic uncertainties after all corrections for different primary particles, energies and zenith angles are less than 10%, a little worse compared to the energy reconstruction with [*Ne*]{}. The reasons are the larger E/A dependencies. The energy resolution ranges from 45% (35%) at 300TeV to 8% (11%) at 5PeV for iron (proton) induced showers.
Comparison of both Energy Reconstructions
-----------------------------------------
Assuming perfect detectors the light in the interval from 20 to 100m core distance can be reconstructed with an error smaller than 1%. Clearly this is superior to the measurement of [*Ne*]{}. However the obtainable energy resolution at low energies is limited by uncertainties in the distance and E/A reconstruction. In spite of the accurate measurement of the Cherenkov light the final resolution at 300TeV is even worse than using [*Ne*]{} mainly because the correction depending on E/A is larger. At high energies fluctuations in the shower development concerning its shape and the fraction of energy deposited in the electromagnetic cascade limit the energy resolution, which could only be improved by an accurate determination of the non electromagnetic component. It should be noted however that energy resolutions around 10% as obtained here for 5PeV are already much better than needed for most applications.\
The energy resolutions for both methods suffer mainly from the same uncertainties in distance and E/A determination. Therefore no significant improvement can be obtained by combining both measurements. However the two methods to determine the primary energy are not equally sensitive to experimental errors. For example the energy reconstructed with Cherenkov light only is much less influenced by faulty [*slope*]{} measurements than the reconstruction with [*Ne*]{}. Assuming that the light density at 100m distance from the shower core is measured correctly while [*slope*]{} is reconstructed incorrect, the energy determined with [*Ne*]{} is incorrect by nearly a factor of two if [*slope*]{} is changed by 20% whereas the energy derived from Cherenkov light only is shifted by less than 10%. This different behavior is explained by the fact, that a too large (small) [*slope*]{} leads to too small (large) corrections $\xi$ and $\zeta$, but in addition too large (small) [*slopes*]{} give too small (large) values for [*L(20-100)*]{} partly compensating the effect of the wrong $\zeta$ correction.\
Determining the Chemical Composition {#chemie}
====================================
With the reconstruction of the shower development described in the previous section not only the energy is inferred independently of the mass of the primary nucleus but also the nucleon number of the hadron hitting the atmosphere can be determined coarsely by only measuring the electromagnetic part of the EAS. In the following sections first the energy per nucleon derived from the longitudinal shower development and then the properties of the lateral shower extensions will be analyzed for their sensitivity on the nucleon number of the primary particle. The third section combines all information concerning the chemical composition deduced here from the four observables [*Ne, age, slope*]{} and [*L(20-100)*]{} for 300TeV showers as an example.
The Nucleon Number from the longitudinal Shower Development
-----------------------------------------------------------
With procedures to determine the primary energy and an estimation of the energy per nucleon E/A from the position of the shower maximum as described in section\[detea\] it is straightforward to calculate the nucleon number: $${\rm log_{10}(A) \, \equiv \,
log_{10}\left(\frac{energy}{energy/nucleon}\right)
.}
\label{loga}$$ Figure\[massrec\] displays the reconstructed log$_{10}$(A) values for proton and iron showers of 0.3 and 5PeV. The other primaries were omitted in order to keep a clear picture.
=10.55cm
\[massrec\]
Further results are summarized in table\[tabmass\]. The reconstructed mean values correspond within statistical errors to the expectation values. For the energy reconstruction the method with [*Ne*]{} was used, but the same results are obtained with the second method of determining the primary energy.\
Light and heavy primaries
--------- -- ------------------ ------------------ ------------------ ------------------
Primary RMS 300TeV RMS MC RMS 5PeV RMS MC
Proton $1.19 \pm 0.09 $ $1.00 \pm 0.07 $ $0.82 \pm 0.13 $ $0.77 \pm 0.12 $
Helium $0.83 \pm 0.08 $ $0.70 \pm 0.07 $ $0.42 \pm 0.12 $ $0.30 \pm 0.09 $
Oxygen $0.63 \pm 0.08 $ $0.52 \pm 0.07 $ $0.26 \pm 0.08 $ $0.22 \pm 0.06 $
Iron $0.53 \pm 0.06 $ $0.35 \pm 0.04 $ $0.38 \pm 0.11 $ $0.29 \pm 0.09 $
--------- -- ------------------ ------------------ ------------------ ------------------
: [The mean and rms values of the distributions of the reconstructed log$_{10}$(A) values. “MC” symbolizes the result obtained by using the generated MC energy and the depth of the shower maximum directly from MC. The numbers given in the “MC” columns therefore show the contributions from fluctuations in the longitudinal shower development only. Differences to the fits shown in Figure5 originate from the summation over all zenith angles in this table.]{}
\[tabmass\]
can be distinguished by their different mean values and by their different spreads. The spread of the ${\rm log_{10}(A)}$ distributions is dominated by the statistical fluctuations of the depth of the shower maximum with subsequent uncertainties in the E/A determination (see Figures\[detenuc\] and 5). Even a perfect energy determination would hardly improve the separation of different primary particles.
Composition Analysis from the lateral Shower Development {#complat}
--------------------------------------------------------
In section\[showerlat\] differences concerning observable lateral extensions of EAS which depend on E/A have been touched briefly. [*Age*]{} can be used to estimate E/A if the energy of the primary particle and the distance to the shower maximum is known. Figure\[compae\] (left) compares [*age*]{} for different primaries of 300TeV energy. To use this discrimination the expectation value of [*age* ]{} for proton induced showers was parameterized: $${\rm { age}(p)\, =\, 1.42 - 0.10 \cdot log_{10}(E/TeV)
+ 18.0 \cdot {\it slope}\ .}
\label{agep}$$ For each reconstructed shower the actual [*age*]{} is then compared to the expectation for primary protons. At 300TeV iron primaries show a mean value [*age/age(p)*]{} of 1.20 with a rms of 0.17 while the mean for protons lies at 1.00 as expected with a rms of 0.15.\
EAS of different nuclei can also be distinguished by their lateral shower developments as measurable by comparing the number of Cherenkov photons within and beyond 100m core distance. Unfortunately it is very difficult to measure the low density Cherenkov light up to a few hundred meters distance from the shower core with great precision. However using the energy reconstruction methods developed in this paper an indirect measurement of the light beyond 100m is possible: if the energy is reconstructed only with Cherenkov light an E/A dependent correction (eq.\[eqlgab\]) has to be applied to take into account the changing fraction of Cherenkov light measurable below 100m and the fraction of the primary energy deposited in the electromagnetic cascade. Only the latter point has to be corrected if the energy reconstruction is done with the help of [*Ne*]{} (eq.\[eqgab\]). Therefore omitting all E/A corrections in both energy reconstruction methods (resulting in E$^*$(Cl) and E$^*$([*Ne*]{})) and then comparing E$^*$(Cl)/E$^*$([*Ne*]{}) provides an indirect estimation of the amount of Cherenkov light at large distances. This is equivalent to compare the number of Cherenkov photons between 20 and 100m core distance to [*Ne*]{} taking into account the distance to the shower maximum and density effects for the production of Cherenkov light. In Figure\[compae\] (right) the energy ratios are plotted. A clear separation is visible for 300TeV showers.
=10.55cm
\[compae\]
An analysis of the chemical composition can profit from measuring the lateral extent of the electromagnetic cascade of an EAS at energies below approximately 1PeV. At higher energies the lateral extensions no longer depend on the primary mass in a measurable way. With the observables used in this paper the chemical composition around the “knee” can only be derived from the longitudinal shower development.
Chemical Composition from a combined Analysis of the longitudinal and lateral Shower Development
-------------------------------------------------------------------------------------------------
The sample of 300TeV showers was used to compare the sensitivity of the different parameters discussed in the last two sections on the mass of the primary nucleus. If cuts in different quantities are applied so that 90% of the iron showers are selected the following fractions of proton shower remain:
- Cut in log$_{10}$(A): 20%, $\bullet$ cut in [*age*]{}/age(p) or E$^*$(Cl)/E$^*$([*Ne*]{}): 40%.
The most sensitive parameter is derived from the longitudinal shower development, but also a discrimination between heavy and light nuclei can be derived from the lateral shower extensions. To combine the information from the longitudinal shower development and the two comparisons referring to lateral extension of the EAS the probability densities for observing a specific log$_{10}$(A), E$^*$(Cl)/E$^*$([*Ne*]{}) or [*age/age(p)*]{} value were parameterized for primary p, ${\rm \alpha}$, O and Fe nuclei of 300TeV from the MC library. Following equation\[eqallco\] (similar for other primaries than iron) a combined probability is calculated:\
${\rm With \ \rho_{Fe}\, =\, prob({\it log_{10}(A)},Fe) \cdot
prob({\it E^*(Cl)/E^*(Ne)},Fe) \cdot prob({\it age/age(p)},Fe)}$ $${\rm prob(Fe)\,=\,
\frac{\rho_{Fe}}{\rho_{p}+\rho_{\alpha}+\rho_{O}+\rho_{Fe}}}
\label{eqallco}$$ Table\[taballco\] lists the fractions for nuclei of different masses which are obtained by selecting 90% or 50% of all proton or iron showers.
Primary
--------- ----- ------- ----- -------
Proton 90% 50% 8% $<$1%
Helium 80% 17% 12% 1.5%
Oxygen 43% 3% 60% 24%
Iron 5% $<$1% 90% 50%
: [The remaining fraction of primaries with energies of 300TeV after selecting 90% or 50% of the proton (iron) showers with cuts in prob(p) or prob(Fe).]{}
\[taballco\]
Clearly an analysis of the chemical composition improves if measurements of the longitudinal and lateral shower developments are combined. Light and heavy particles can be separated rather well. However with the four observables used here primary nuclei with masses similar to oxygen can be separated only in a statistical sense but not on an event by event basis. It seems to be difficult to distinguish between primary protons and ${\rm \alpha}$ particles.\
Systematics {#wo}
-----------
Studies of systematic effects related to the CORSIKA code (stepwidth of the EGS part), atmospheric transmission, the fragmentation of the primary nucleus and the influence of different models to simulate the high energy interactions of the CR will be described in a forthcoming publication. In general the observables analyzed here show up to be much less model dependent than hadronic shower properties or muon distributions, mainly because the development of a shower behind its maximum determines the Cherenkov light and particle measurement, as considered throughout this paper.
Summary and Conclusions
=======================
In this paper methods were presented to determine energy and mass of charged cosmic rays from ground based observations of the electromagnetic cascade of air showers. From the slope of the lateral Cherenkov light density in the range of 20 to 100m core distance the position of the shower maximum can be inferred without knowledge of the nucleon number of the primary particle. This leads to an unbiased determination of the energy per nucleon and, combined with the shower size at detector level or the number of registered Cherenkov photons, to a measurement of the primary energy. Thus a measurement of the energy spectrum and a coarse determination of the chemical composition are possible without any a priori hypotheses.\
With the observables considered in the present paper the energy resolution for primary nuclei is limited to approximately 30% at 300TeV improving to 10% at 5PeV due to natural fluctuations in the shower development. Further improvements of these results are only possible if accurate measurements of the non electromagnetic components of EAS are added.\
At energies below 1PeV, where results from EAS measurements can be compared to direct data from balloon flights, the sensitivity of the analysis of air showers by observing Cherenkov light and particles at detector level can be substantially improved by combining the results related to the longitudinal shower development with parameters derived from the lateral extension. This allows detailed tests of the described method to determine the chemical composition and the energy spectrum of cosmic rays.\
One main characteristic of deriving energy and mass of primary nuclei from observations of the electromagnetic component of extensive air showers with the observables used here is the fact, that it are mainly the longitudinal shower development behind the shower maximum, the number of particles at the maximum and the penetration depth of the shower until it reaches the maximum, which determine the results. While the first two items do not vary much for different models describing the development of air showers the last item is more model dependent. In order to achieve results being as model independent as possible it is very desirable to combine the method described in this paper with complementary measurements. Analyses of the early stage of the shower development, of the hadronic component of EAS or detailed studies of the shower core may be considered for this purpose.\
[**ACKNOWLEDGMENTS**]{}\
The author would like to thank the HEGRA members for their collaboration. Especially I am very grateful to V.Haustein, who performed most of the MC simulations and determined the chemical composition of CR with a different technique, to G.Heinzelmann for many detailed, constructive proposals and improvements as well as for his general support, and to R.Plaga, who pioneered the analysis of charged CR in the HEGRA collaboration, for motivations and detailed discussions. Special thanks to the authors of CORSIKA for supplying us with the simulation program and their support. I thank Gerald Lopez for his careful reading of the text and for providing valuable suggestions. This work was supported by the BMBF (Germany) under contract number 052HH264.
[900]{} Aharonian,F., et al. (HEGRA collab.), ${\rm 24^{th}}$ ICRC Rome [**1**]{}, 474 (1995) Kamata,K. and Nishimura,J., Prog.Theoret.Phys., Suppl[**6**]{} (1958)\
Greisen,K. Ann.Rev.Nucl.Sci.[**10**]{},63 (1960) Capdevielle,J.N. et al., KFK Report [**4998**]{} (1992)\
Knapp,J., Heck,D., KfK Report [**5196B**]{} (1993) Haustein,V., doctoral thesis at Univ.ofHamburg (1996), in preparation Patterson,J.R., Hillas,A.M., J.Phys.G.[**9**]{},1433 (1983) Gabriel,T.A., et al., Nucl.Instrum.Meth.[**A338**]{},336 (1994)
[^1]: neglecting atmospheric absorption
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present two practical and widely applicable methods, including some criteria and a general procedure, for detecting Brunnian property of a link, if each component is known to be unknot. The methods are based on observation and handwork. They are used successfully for all Brunnian links known so far. Typical examples and extensive experiments illustrate their efficiency. As an application, infinite families of Brunnian links are created, and we establish a general way to construct new ones in bulk.'
address:
- 'School of Mathematical Sciences, Peking University'
- 'School of Mathematical Sciences, Peking University'
author:
- Sheng Bai
- Weibiao Wang
date: '2020 June. 16'
title: New criteria and Constructions of Brunnian Links
---
Introduction {#sect:introduction}
============
Detecting unlink is generally considered as difficult at least as detecting unknot, which is a famous problem in knot theory. The most common methods are link invariants. These rich and powerful tools, though, often become hard to compute by hand when links are large. A question that has always attracted the authors is:
Can we establish an intuitive and rigorous method to detect unlinks, under some assumptions, such as some components are known to be unknotted?
It is still unknown whether HOMFLY polynomial detects unlink (also whether Jones polynomial detects unknot, c.f. [@EKT]). If one expects a negative answer and tries to seek for a counterexample, one will need to prove the desired link is nontrivial. At this point, an intuitive method might be more enlightening than looking for other invariants.
We are concerned in this paper with Brunnian links.
\[def:brunn\] A link $L$ of $n$ $(>1)$ components in $\mathbb{R}^3$ is *Brunnian* if it satisfies the following properties:
1. every sublink with $(n-1)$ components is trivial;
2. $L$ itself is nontrivial.
Analyzing a link, we may naturally consider its nontrivial sublinks, the minimal ones among whom are exactly knots and Brunnian links. In this sense, they are of natural importance in study of links. For theoretic results of Brunnian links the reader is referred to [@MS; @MY; @SH]. The first examples were introduced by H. Brunn[@B] in 1892 but it was not until 1961 their Brunnian property was proved by DeBrunner[@HDb]. Since then, various authors constructed many kinds of examples, which, as far as we know, all appear in [@D.Rolfsen; @BFJRVZ; @BS; @BCS; @L; @J; @F; @M; @MM; @TKa]. Recently, Nils A. Baas et al.[@B10; @B12; @BC; @BCS; @BS] constructed several infinite families of Brunnian links, vastly expanding our knowledge in this area. However, in these papers, their Brunnian property is verified by general invariants, not taking advantage of the Brunnian property and often difficult in handwork.
The purpose of this paper is to give convenient and general approach to detecting Brunnian property of links. We focus on the following question:
Suppose a link $L$ possesses the property (ST), how can we check the property (NT)?
In fact, if all component of a link $L$ are known to be unknots, we can detect its Brunnian property, by applying the method for this question to all sublinks with $k$ components and induction on $k$.
We propose two methods, named Arc-method and Circle-method, including several criteria and a general procedure. The novel methods, significantly different from using invariants, depend crucially on observation and are worked by hand. They are practical and turn out to work well for all known Brunnian links. The efficiency will be demonstrated by typical examples. Arc-method is very fast when it works. The power of Circle-method will be reflected in large series of links and the flexibility of Circle-method leads us to construct new Brunnian links. We will exhibit infinitely many infinite families of new Brunnian links, far more than the existing ones.
(186.30, 81.46)(0,0) (0,0)[![\[fig:borromean\] Borromean ring](borromean.pdf "fig:")]{} (5.67,11.98)[$C_2$]{} (84.32,11.98)[$C_3$]{} (23.27,69.52)[$C_1$]{} (97.09,49.05)[$C_1$]{} (104.09,11.98)[$C_2$]{} (171.29,11.98)[$C_3$]{}
(186.30, 81.46)(0,0) (0,0)[![\[fig:borromean\] Borromean ring](borromean "fig:")]{} (5.67,11.98)[$C_2$]{} (84.32,11.98)[$C_3$]{} (23.27,69.52)[$C_1$]{} (97.09,49.05)[$C_1$]{} (104.09,11.98)[$C_2$]{} (171.29,11.98)[$C_3$]{}
Let us illustrate our natural idea with an example. The two links in Fig. \[fig:borromean\] are both Borromean rings. We use the right one. Take the gray disk $D$ bounded by component $C_1$, which intersects $C_2$ at two points. Connect these two points by an arc on $D$. Each component of $C_2- D$ together with the red arc forms a circle which can not bound a disk avoiding $C_3$ due to linking number. Arc-criterion then asserts Borromean ring is nontrivial.
Preliminaries {#sect:preliminaries}
=============
We work in smooth category. Without special explanation, we always consider links in $\mathbb{R}^3$. The notations $C,L,D$ (maybe with subscripts) are used always to denote an unknot, a link and an embedded disk respectively. The interior of a manifold is denoted by ${\rm Int}(\cdot)$. Each arc is assumed to be simple, and intersections are always assumed to be compact and transverse.
The following proposition is of fundamental importance for us, which follows immediately from Lemma \[lem:basic\].
\[prop:basic\] Let $L$ be a link satisfying property (ST) and $L_1$ be any proper sublink of $L$. Then $L$ satisfies (NT) if and only if $L_1$ can not bound mutually disjoint disks in the complement of $L-L_1$.
\[lem:basic\] Let $L=\cup _{i=1}^n C_i$ be an $n$-component link satisfying (ST). Then $L$ is trivial if and only if $C_1$ bounds a disk in the complement of $L-C_1$.
The “only if" part is obvious. For the converse, let $D_1$ be a closed disk with $\partial D_1=C_1$ such that ${\rm Int}(D_1)\cap L=\emptyset$. Retract $D_1$ to be small enough, and move it far away by isotopy. By (ST), $C_i, i=2,...,n$ bounds disjoint disks, and those disks can be assumed to be far from the now $D_1$. We can then move $D_1$ to the original position, while using isotopy extension to $\mathbb{R}^3 \backslash \cup_{i=2}^n C_i$. So $C_i, i=1,...,n$ bounds mutually disjoint disks and by definition $L$ is trivial.
We now describe our basic tool and notions. Suppose $L=\cup_{i=1}^n C_i$ ($n>1$) is an $n$-component link with property (ST). As a tool, choose a proper subset $I \subset \{1,2,\cdots,n\}$ and take disjoint disks $D_i (i\in I)$ bounded by $C_i (i\in I)$ with $C_i =\partial D_i$. Then $L$ is divided into three disjoint sublinks, $L_I \triangleq \cup_{i \in I} C_i$, $L_J \triangleq \cup_{j \in J} C_j$, and $L_0 \triangleq L- L_I -L_J$, where $L_J$ consists of the components intersecting $\cup_{i \in I} {\rm Int}(D_i)$, and $L_0$ may be empty. If $L_J =\emptyset$, by Proposition \[prop:basic\], $L$ is trivial. So we always assume $L_J \neq\emptyset$.
[test disk and thicken disk]{}: For each $i\in I$, we call $D_i$ a *test disk*. Now $\cup_{i \in I} D_i$ intersects $L_J$ with finitely many points $\{ p_t \}_{t\in T}$. Take an embedded cylinder $\mathbf{D_i} \cong D \times [0,1]$ such that $D \times \{\frac{1}{2}\} =D_i$, and $\mathbf{D_i}$ intersects $L_J$ with several disjoint arcs each of which takes one $p_t$ as interior point (see Fig. \[fig:testcomplex\]). Choose $\mathbf{D_i} (i\in I)$ to be “thin” enough so that they are mutually disjoint. We call $\mathbf{D_i}$ a *thicken disk* of $D_i$. It has two sides: $$\begin{aligned}
D_i^+ \triangleq D \times \{ 1 \}; D_i^- \triangleq D \times \{ 0 \}. \nonumber\end{aligned}$$
[test complex and test subcomplex]{}: We call $U \triangleq \cup_{i \in I} \mathbf{D_i} \cup L$ a *test complex*. A subset $U'\subset U$ is called a *test subcomplex of $U$* if it consists of some thicken disks and a closed subset of $L$.
Arc-method {#sect:arc}
==========
Arc-Criterion {#subsect:criterionA}
-------------
Let $L$ be a link satisfying (ST) and divided into sublinks $L_I$, $L_J$, and $L_0$ with respect to test disks $\{D_i\}_{i\in I}$ as in Preliminaries. Notice that for any $j \in J$, $C_j-\cup_{i \in I}{\rm Int}( \mathbf{D_i})$ is a disjoint union of arcs.
\[def:ear\] If an arc component $e$ of $C_j-\cup_{i \in I}{\rm Int}( \mathbf{D_i})$ has two endpoints both on $D_i^\sigma$ ($\sigma$ can be + or $-$), then $e$ is an *ear* of $C_j$ on $D_i^\sigma$.
For an ear $e$ of $C_j$ on $D_i^\sigma$ and a test subcomplex $U'$ containing $e$, if there exists an arc $\alpha \subset D_i^\sigma$ (avoiding $L$) connecting the endpoints of $e$, such that $e \cup \alpha$ bounds a disk $D_e$ with ${\rm Int}(D_e)\subset \mathbb{R}^3-U'$, then we say $e$ is *compressible for $U'$* and call $D_e$ a *compressing disk of $e$ for $U'$*. When $U'=L$, the arc $\alpha$ is called an *incredible arc of $e$*.
\[criterionA\](**Arc-criterion**) Suppose an $n$-component link $L=\cup_{i=1}^n C_i$ satisfies (ST) and is divided into sublinks $L_I$, $L_J$, and $L_0$ with respect to test disks $\{D_i\}_{i\in I}$ as in Preliminaries. If for some $j \in J$, there is at most one ear of $C_j$ compressible for $L$, then $L$ is Brunnian.
The criterion is a direct conclusion of the following proposition. We call it Arc-criterion for it involves incredible arcs to detect Brunnian property, and Circle-criterion is named for incredible circles (see Section \[sect:circle\]).
\[prop:A\] Suppose an $n$-component link $L=\cup_{i=1}^n C_i$ satisfies (ST) and is divided into sublinks $L_I$, $L_J$, and $L_0$ with respect to test disks $\{D_i\}_{i\in I}$ as in Preliminaries. If $L$ is trivial, then for each $j \in J$, there are at least 2 ears of $C_j$ compressible for $L$.
By Proposition \[prop:basic\], if $L$ is trivial, then $L_J$ bounds a disjoint union of disks $\sqcup_{j \in J} D_j$ with ${\rm Int}(D_j)\cap L=\emptyset$, for $\forall j\in J$. Fix $j\in J$ and we see that $D_j\cap \cup_{i\in I} D_i$ is a disjoint union of circles and arcs. Since $D_j\cap L_I=\emptyset$, the endpoints of the arcs are all on $C_j$, and all these arcs cut $D_j$ into some regions, each isomorphic to a disk. There are at least two such regions, say $R_1,R_2$, outermost on $D_j$ (i.e., the boundary of $R_s(s=1,2)$ is a union of an arc on $C_j$ and an arc component of $D_j\cap\cup_{i\in I}D_i$, and $\mathrm{Int}(R_s)\cap \cup_{i\in I} C_i=\emptyset$), as shown in Fig. \[fig:region\]. The boundaries of these regions provide the ears and their incredible arcs.
(85.33, 85.33)(0,0) (0,0)[![\[fig:region\] Outermost regions on $D_j$](region.pdf "fig:")]{}
(85.33, 85.33)(0,0) (0,0)[![\[fig:region\] Outermost regions on $D_j$](region "fig:")]{}
The proposition provides a necessary and sufficient condition for $L$ to be trivial. In fact, if an ear $e$ on $D_i^+$ or $D_i^-$ is compressible for $L$ with compressing disk $D_e$, we can push $e$ along $D_e$, across $D_i$ to eliminate two intersection points of $C_j\cap D_i$. It induces an isotopy of $L$, and $L_I,L_J$ may get adjusted with respect to the test disks $D_i,i\in I$. Repeat it inductively. If finally $L_J$ becomes empty, by Proposition \[prop:basic\] $L$ is trivial, otherwise Criterion \[criterionA\] tells it is Brunnian. Therefore, given a link consisting of unknots, theoretically we can check every sublink with Arc-criterion inductively and finally decide whether it is Brunnian.
Example analysis {#subsect:examplesA}
----------------
Suppose a link satisfies (ST). To prove it is Brunnian by Arc-method, we shall take appropriate test disks and then verify the incompressibility of enough ears by investigating all candidates for their incredible arcs. We present some examples here to show the way it works and the efficiency it possesses.
\[example:communi\]
![Example \[example:communi\][]{data-label="fig:communi"}](communi.pdf){height="3.6cm"}
The first link in Fig. \[fig:communi\] has 6 components, and it is easy to see that each sublink with 5 components is trivial. Now we apply Arc-criterion to show it is a Brunnian link. Take the test disk colored in gray and $L_I=C_2,L_J=C_1,L_0=\bigcup_{k=3}^6 C_k$. There are only two ears, denoted $e_1,e_2$. The candidate for incredible arc of $e_1$ is unique up to isotopy, colored in red and denoted $\alpha$. By Criterion \[criterionA\], we only need to show the circle $C_1'\triangleq e_1\cup\alpha$ can not bound a disk in the complement of $L_0$, and by Proposition \[prop:basic\], it is equivalent to that the second link in Fig. \[fig:communi\] is Brunnian. Inductively, it suffices to show the rightmost link is Brunnian, which is obvious.
(270.67, 91.98)(0,0) (0,0)[![\[fig:gBr\] Generalized Borrommean ring](gBr.pdf "fig:")]{} (8.12,13.75)[$C_1$]{} (7.96,49.62)[$C_2$]{} (7.44,76.01)[$C_3$]{} (185.47,76.01)[$C_3$]{} (219.17,51.85)[$e$]{} (222.42,69.55)
(270.67, 91.98)(0,0) (0,0)[![\[fig:gBr\] Generalized Borrommean ring](gBr "fig:")]{} (8.12,13.75)[$C_1$]{} (7.96,49.62)[$C_2$]{} (7.44,76.01)[$C_3$]{} (185.47,76.01)[$C_3$]{} (219.17,51.85)[$e$]{} (222.42,69.55)
(Fig. \[fig:gBr\]) Take the gray test disk. By symmetry, we just consider one ear, denoted $e$. Delete all other ears, then up to isotopy, there is only one way, as the arc $\alpha$, to connect the two endpoints of $e$ on the disk. The circle $e\cup\alpha$ has linking number 1 with $C_3$, thus $e$ is incompressible for the link.
For an ear $e$ on $D_i^\sigma$ ($\sigma$ is + or $-$), connect the endpoints by an arc $\alpha$ in $D_i^\sigma$. If $e \cup \alpha$ is a nontrivial knot, we say $e$ is *knotted*, otherwise *unknotted*. It is independent of the choice of $\alpha$. Obviously, a compressible ear must be unknotted.
\[example:Brunnian braid\] (Fig. \[fig:F\]) Take the gray test disk. There are four ears, denoted $a,b,c,d$. With the same argument as in Example \[example:gBr\], $b,d$ are incompressible. And $a$ is knotted as shown on the right, hence incompressible.
(316.73, 75.37)(0,0) (0,0)[![\[fig:F\] Example \[example:Brunnian braid\]](F.pdf "fig:")]{} (114.83,52.34)[$a$]{} (174.97,52.34)[$b$]{} (174.97,11.00)[$d$]{} (205.03,52.34)[$c$]{} (18.45,62.12)[$C_1$]{} (80.08,62.12)[$C_2$]{} (6.80,9.50)[$C_3$]{} (220.06,52.34)[$a$]{}
(316.73, 75.37)(0,0) (0,0)[![\[fig:F\] Example \[example:Brunnian braid\]](F "fig:")]{} (114.83,52.34)[$a$]{} (174.97,52.34)[$b$]{} (174.97,11.00)[$d$]{} (205.03,52.34)[$c$]{} (18.45,62.12)[$C_1$]{} (80.08,62.12)[$C_2$]{} (6.80,9.50)[$C_3$]{} (220.06,52.34)[$a$]{}
Example \[example:communi\] is Exercise 8 of [@D.Rolfsen] Chapter 3 section F, also in [@TK]. Exercise 7 of [@D.Rolfsen] Chapter 3 section F and Fig.1.10 in [@BS] can be proved in the same manner.
Example \[example:gBr\] is Tait series of Fig.2 in [@J], the simplest one of which is Borremean ring. [@J] shows lots of *Borromean links* and its Fig.4, 5 and 6, which generalize Borromean ring in different ways, can be proved to be Brunnian similarly. The first 6 links of Fig.7 and the first 6 links of Fig.8 in [@J] are more complicated but Arc-criterion works as well.
Example \[example:Brunnian braid\] is Fig.5 in [@F], constructed from Brunnian braids.
Compared with original methods in these papers, Arc-method is clearly more concise and faster in verifying Brunnian property. However, there are also cases, for example Brunn’s chain (Fig. \[fig:milnorlink\]), where Arc-method fails in practice. The obstruction is that there may be infinitely many ways to connect two points on a disk with marked points, see Fig. \[fig:auxiliaryarc\]. So if there are more than two intersection points on the same side of a test disk, there will be infinitely many candidates for incredible arcs of one ear, and Arc-method may get useless in practice. Deleting some intersection points, as used in Example \[example:gBr\], may help but not always. To overcome this limitation, another method, Circle-method, develops.
(341.28, 81.37)(0,0) (0,0)[![\[fig:auxiliaryarc\] Different ways to connect two points](auxiliaryarc.pdf "fig:")]{}
(341.28, 81.37)(0,0) (0,0)[![\[fig:auxiliaryarc\] Different ways to connect two points](auxiliaryarc "fig:")]{}
Circle-method {#sect:circle}
=============
Quasi-tangles {#subsect:terminologyC}
-------------
Suppose an $n$-component link $L=\cup_{i=1}^n C_i$ satisfies (ST) and is divided into sublinks $L_I$, $L_J$, and $L_0$ as in Preliminaries, with respect to test disks $\{D_i\}_{i\in I}$. We have thicken disks $\{\mathbf{D_i}\}_{i\in I}$ and the test complex $U$.
Fix an index $i\in I$ and a test subcomplex $U'$ containing $\mathbf{D_i}$. Recall that $\mathbf{D_i}$ has two sides $D_i^+,D_i^-$, so for each connected component of $U'-\mathbf{D_i}$, say $X$, one of the following four cases holds:
Case $+$: $\overline{X}\cap D_i^+\neq\emptyset,\,\overline{X}\cap D_i^-=\emptyset$;
Case $-$: $\overline{X}\cap D_i^+=\emptyset,\,\overline{X}\cap D_i^-\neq\emptyset$;
Case $\pm$: $\overline{X}\cap D_i^+\neq\emptyset,\,\overline{X}\cap D_i^-\neq\emptyset$;
Case $0$: $\overline{X}\cap D_i^+=\emptyset,\,\overline{X}\cap D_i^-=\emptyset$,
where $\overline{X}$ represents the closure of $X$ in $\mathbb{R}^3$. Denote the sets consisting of such $X$’s in the four cases by $K_i^+,K_i^-,K_i^\pm,K_i^0$ respectively. For any subset $Q$ of them, we always denote $\bigcup_{X\in Q}\overline{X}$ by $|Q|$.
\[def:quasitangle\] Let $\sigma$ be $+$ or $-$. For $U'$, a subset $Q \subset K_i^\sigma\cup K_i^0$ is a *quasi-tangle* on $D_i^\sigma$, if one of the following two conditions holds:
1. $Q \cap K_i^\sigma \neq \emptyset$ and $L\cap D_i^\sigma - |Q| \neq \emptyset$;
2. $Q$ only contains an ear on $D_i^\sigma$.
\[rmk\] We require $L\cap D_i^\sigma - |Q| \neq \emptyset$ in case that we, in practice, get caught in the task (see Flowchart \[fig:flowchart\]) to prove $C_i$ itself can not bound an open disk in the complement of $U$, which is exactly the origin problem by Proposition \[prop:basic\] and makes our method meaningless.
(315.36, 155.75)(0,0) (0,0)[![\[fig:testcomplex\] A test subcomplex and its quasi-tangles](test-subcomplex.pdf "fig:")]{} (20.63,97.12)[$a$]{} (119.72,106.13)[$b$]{} (76.86,138.60)[$c$]{} (102.55,97.12)[$e$]{} (33.20,31.96)[$d$]{} (122.03,31.96)[$f$]{} (8.06,78.05)[$D_1^+$]{} (8.06,47.89)[$D_1^-$]{} (5.67,62.34)[$\mathbf{D_1}$]{} (29.01,109.90)[$\mathbf{D_2}$]{} (66.72,9.63)[$\mathbf{D_3}$]{} (143.71,136.20)[Thicken disks: $\mathbf{D_1,D_2,D_3}$.]{} (143.71,66.61)[Ear on $D_1^+$: $e$.]{} (143.71,122.54)[$K_1^+=\{ a\cup c\cup \mathbf{D_2},e \}$,]{} (143.71,108.45)[$K_1^-=\emptyset$,]{} (143.71,94.35)[$K_1^ \pm=\{ d\cup f\cup \mathbf{D_3} \}$,]{} (143.71,80.26)[$K_1^ 0=\{ b \}$.]{} (143.71,53.27)[Quasi-tangles on $D_1^+$: ]{} (150.75,39.51)[$\{ a\cup c\cup \mathbf{D_2} \}, \{ e \},$]{} (150.75,25.42)[$\{ a\cup c\cup \mathbf{D_2}, e \}, \{a\cup c\cup \mathbf{D_2}, b \}, \{ e, b \}, $]{} (150.75,11.82)[$\{ a\cup c\cup \mathbf{D_2}, e, b \}.$]{}
(315.36, 155.75)(0,0) (0,0)[![\[fig:testcomplex\] A test subcomplex and its quasi-tangles](test-subcomplex "fig:")]{} (20.63,97.12)[$a$]{} (119.72,106.13)[$b$]{} (76.86,138.60)[$c$]{} (102.55,97.12)[$e$]{} (33.20,31.96)[$d$]{} (122.03,31.96)[$f$]{} (8.06,78.05)[$D_1^+$]{} (8.06,47.89)[$D_1^-$]{} (5.67,62.34)[$\mathbf{D_1}$]{} (29.01,109.90)[$\mathbf{D_2}$]{} (66.72,9.63)[$\mathbf{D_3}$]{} (143.71,136.20)[Thicken disks: $\mathbf{D_1,D_2,D_3}$.]{} (143.71,66.61)[Ear on $D_1^+$: $e$.]{} (143.71,122.54)[$K_1^+=\{ a\cup c\cup \mathbf{D_2},e \}$,]{} (143.71,108.45)[$K_1^-=\emptyset$,]{} (143.71,94.35)[$K_1^ \pm=\{ d\cup f\cup \mathbf{D_3} \}$,]{} (143.71,80.26)[$K_1^ 0=\{ b \}$.]{} (143.71,53.27)[Quasi-tangles on $D_1^+$: ]{} (150.75,39.51)[$\{ a\cup c\cup \mathbf{D_2} \}, \{ e \},$]{} (150.75,25.42)[$\{ a\cup c\cup \mathbf{D_2}, e \}, \{a\cup c\cup \mathbf{D_2}, b \}, \{ e, b \}, $]{} (150.75,11.82)[$\{ a\cup c\cup \mathbf{D_2}, e, b \}.$]{}
Let $Q$ be a quasi-tangle on $D_i^\sigma$ for $U'$. If there exists a circle $C \subset D_i^\sigma$ bounding disks $D_C,D_Q$, such that $D_C\subset D_i^\sigma,{\rm Int}(D_Q)\subset\mathbb{R}^3-U'$ and $D_C \cup D_Q$ bounds a 3-ball $B$ with $\overline{{\rm Int}(B)\cap U'} = |Q|$, then we say $Q$ is *compressible for $U'$*. The circle $C$ is called an *incredible circle of $Q$ for $U'$*, and $D_Q$ is called a *compressing disk for $U'$*.
We point out that, if a quasi-tangle $Q$ is compressible for $U$ and $|Q \cap K_i^+ |$ is an ear $e$ of $C_j$, then $e$ is unknotted. In fact, otherwise $e$ would be a connected summand of the unknot $C_j$, which is impossible. As a consequence, it is easy to see that an ear $e$ is compressible for $U$ if and only if the quasi-tangle $\{e\}$ is compressible for $U$.
Criterion {#subsect:criterionC}
---------
\[criterionC\](**Circle-criterion**) Let $L=\cup_{i=1}^n C_i$ be an $n$-component link with property (ST). Suppose $L$ is divided into sublinks $L_I$, $L_J$, and $L_0$ with respect to test disks $\{D_i\}_{i\in I}$ as in Preliminaries and $U$ is the test complex. If each quasi-tangle is not compressible for $U$, then $L$ is Brunnian.
Assume to the contrary that $L$ is trivial. By Proposition \[prop:basic\], there are disjoint disks $\sqcup_{j \in J} D_j\subset \mathbb{R}^3 - L_I-L_0$ bounded by $L_J$ with $C_j =\partial D_j$. Consider the intersection $S\triangleq\cup_{j \in J} D_j\cap\cup_{i\in I}D_i$. It is a disjoint union of finitely many circles and arcs, and the endpoints of arc components are all on $L_J$. Any circle component $C\subset D_i\cap S$ $(i\in I)$ bounds a disk $D_i ^C$ on $D_i$. Denote $S_C\triangleq D_i ^C\cap S$.
We can modify $\{D_j\}_{j\in J}$ so that $S_C$ includes at least one arc component for every circle component $C\subset S$. In fact, if $S_C$ has no arc component, then there is a circle component $C'\subset S_C$ innermost on $D_i$, that is, $S_{C'}=C'$. Thus we can replace the disk $C'$ bounds on that $D_j$ by $D_i ^{C'}$ and lift or lower it a little to eliminate $C'$ (see Fig. \[fig:innermost\]). Inductively we get what we want.
![Eliminating an innermost circle[]{data-label="fig:innermost"}](innermost.png){height="2cm"}
Moreover, modify $\{D_j\}_{j\in J}$ so that any circle component $C\subset S\cap D_i$ $(i\in I)$ is not outermost, that is, $S\cap D_i -S_C \neq \emptyset$. Otherwise $C\cup C_i$ bounds an annulus $A$, so $C$ can be eliminated by replacing a tubular neighbourhood of $C$ on $D_j$ by an annulus surrounding nearly outside $A$, as in Fig. \[fig:outermost\]. Now fix such disks $\{D_j\}_{j\in J}$ and still denote $\cup_{j \in J} D_j\cap\cup_{i\in I}D_i$ by $S$.
![Eliminating an outermost circle[]{data-label="fig:outermost"}](outermost.png){height="1.7cm"}
For any $j\in J$, all arc components of $D_j\cap S$ cut $D_j$ into some regions, each isomorphic to a disk. There are at least 2 such regions, say $R_1,R_2$, outermost on $D_j$, i.e., they include no arc component of $S$. If $R_1\cap S=\emptyset$, then $\partial R_1 \cap C_j$ is an ear compressible for $U$ with compressing disk $R_1$. Otherwise, there exists a circle $C\subset R_1\cap S$ innermost on $D_j$. Suppose $C\subset D_i$ $(i\in I)$ bounds $D_i ^C,D_j ^C$ on $D_i,D_j$ respectively, and $B$ is the 3-ball bounded by the sphere $D_i ^C\cup D_j ^C$. As $S_C$ includes an arc component of $S$, $B$ contains a quasi-tangle which is thus compressible for $U$. In either case, there is a contradiction.
This criterion, as well as Arc-criterion, provides a necessary and sufficient condition for $L$ to be trivial. In fact, if there is a compressible quasi-tangle on $D_i^+$ or $D_i^-$, we can compress it across the thicken disk $\mathbf{D_i}$ by an isotopy of $L$, to eliminate it. Do it inductively. $L$ is trivial if and only if finally all intersections are eliminated.
Generally, let $L$ be a link with each component unknotted. Suppose each proper sublink of $L$ is trivial, otherwise consider its proper sublinks first. Then we can follow Flowchart \[fig:flowchart\] to decide whether it is Brunnian. We consider quasi-tangles of cardinality $1,2,\cdots$ in order so that Sub-criterion \[prop:subcomplex\] and \[prop:boundaryconnectedsum\] can be applied to reduce the task.
=\[rectangle,rounded corners, minimum width = 2cm, minimum height=1cm,text centered, draw = red, line width=1pt\] = \[trapezium, trapezium left angle=70, trapezium right angle=110, minimum width=1cm, minimum height=1cm, text centered, draw=black, inner sep=12pt\] = \[rectangle, minimum width=2cm, minimum height=1cm, text centered, draw=black\] = \[diamond, aspect = 4, text centered, draw=black\] = \[->,>=stealth\]
(start)[Start]{}; (link)[Link $L=\cup_{i=1}^n C_i$ satisfying (ST)]{}; (test)[Fix test disks $\{D_i\}_{i\in I}$ and test\
complex $U = \coprod_{i \in I} \mathbf{D_i} \cup L$.]{}; (noLJ)[$L_J$ is empty?]{}; (sets1)[$s=1$]{}; (trivial)[$L$ is trivial.]{}; (qt)[Consider all quasi-tangles of cardinality $s$.]{}; (stop)[Stop]{}; (exist)[Do they exist?]{}; (sets2)[$s=s+1$]{}; (compressible)[All incompressible for $U$?]{}; (eliminate)[Eliminate a compressible quasi-tangle\
by pushing it through the corresponding thicken disk;\
equivalently, adjust the test disks and test complex.]{}; (Brunnian)[$L$ is Brunnian.]{}; (point) at (5.5cm, -3cm);
(start) – (link); (link) – (test); (test) – (noLJ); (noLJ) – node\[right\][No]{}(sets1); (noLJ) -| node\[above\][Yes]{}(trivial); (trivial) – (stop); (sets1) – (qt); (qt) – (exist); (exist) – node\[above\] [No]{} (Brunnian); (Brunnian) – (stop); (exist) – node\[right\] [Yes]{} (compressible); (compressible) -| node\[below\] [Yes]{} (sets2); (sets2) |- (qt); (compressible) – node\[right\] [No]{} (eliminate); (eliminate) -| (point); (point) – (test);
There may be only one compressible quasi-tangle for an unlink with respect to some test complex. For instance, see Fig. \[fig:uniqueqt\], where the test disk is colored grey and the only compressible quasi-tangle is in the red box.
![An unlink with only one compressible quasi-tangle[]{data-label="fig:uniqueqt"}](uniqueqt.png){height="5cm"}
Auxiliary criteria {#subsect:auxiliary}
------------------
Given a link $L$ with (ST), to prove (NT) by Circle-criterion, we need to choose some text complex $U$ and show all quasi-tangles are incompressible for $U$. It is sometimes no easy task. We provide four helpful criteria here. The first one can be easily proved by contradiction, and proofs of the others will be given in Subsection \[subsect:proof\].
\[prop:subcomplex\](**Containment-criterion**) Let $U'_1$ and $U'_2$ be test subcomplexes with $U'_1 \subset U'_2$. If a quasi-tangle $Q_1$ is incompressible for $U'_1$, then for any quasi-tangle $Q_2$ with $|Q_1|\subset |Q_2|\subset U'_2$ and $|Q_2-Q_1| \subset U'_2 - U'_1$, $Q_2$ is incompressible for $U'_2$.
Fix a test subcomplex $U'$ and without loss of generality, suppose $Q$ is a quasi-tangle on $D_1^+$. Take a disk $D_0\subset D_1^+$ such that $(D_1^+ - D_0)\cap L =\emptyset$ and consider the test subcomplex $U'_Q \triangleq \mathbf{D_1} \cup |Q|$. Then $\partial D_0$ bounds a disk $D$ outside $U'_Q$ such that $D_0 \cup D$ bounds a 3-ball $B$ including $|Q|$. Lemma \[lem:tangleball\] in Subsection \[subsect:proof\] shows $B$ is unique for $Q$ in some sense and thus we also call the pair $(B,Q)$ a quasi-tangle.
\[def:unsplit\] A quasi-tangle $(B,Q)$ is *split*, if there exists a properly embedded disk $D$ in $B$ avoiding $|Q|$, and a partition $Q=Q_1\cup Q_2$, such that $\partial D$ intersects $\partial D_0$ with two points, and $Q_1,Q_2$ are both non-empty and included in the different components of $B-D$. (Here $Q_1$ or $Q_2$ is not necessary to be a quasi-tangle.) $(B,Q)$ is *unsplit* if it is not split.
\[prop:boundaryconnectedsum\](**Split-criterion**) Let $U'$ be a test subcomplex and $Q$ be a quasi-tangle for $U'$. Suppose $Q$ is split with respect to the partition $Q=Q_1\cup Q_2$. If $Q_1$ is an incompressible quasi-tangle for $U'$, then $Q$ is also incompressible.
Therefore, in Criterion \[criterionC\], it suffices to show every unsplit quasi-tangle is incompressible for $U$.
Next we give an equivalent condition for a quasi-tangle to be compressible. Fix a test subcomplex $U'$ and consider a quasi-tangle on $D_1^+$, say $(B,Q)$. See Fig. \[fig:trivializedsubstitution\]. Take a 3-ball $B_T$ in the closure of $\mathbb{R}^3-U'$ such that $B_T \cap U'$ is a disk $D_T\subset D_1^+$. There is an orientation-preserving homeomorphism $h:B\to B_T$ that maps $B\cap D_1^+$ to $D_T$. The complex $\overline{U'-|Q|} \cup h(|Q|)$ is called a *trivial substitution of $Q$ for $U'$*. Up to isotopy, it is unique and independent of the choices of $h,B,B_T$.
![Trivial substitution[]{data-label="fig:trivializedsubstitution"}](trivialsubstitution.png){height="2.5cm"}
\[cor:incompressibility\](**Substitution-criterion**) Let $Q$ be a quasi-tangle for test subcomplex $U'$. If the trivial substitution of $Q$ for $U'$ is not isotopic to $U'$ in $\mathbb{R}^3$, then $Q$ is incompressible for $U'$.
Instead of investigating each unsplit quasi-tangle, we can also consider all candidates for incredible circles. If we can somehow exclude their potential, then by Circle-criterion, (NT) is proved. Generally there are infinitely many such candidate circles, and we need to narrow our focus.
\[prop:incrediblecircle\](**Criterion for incredible circles**) Let $C$ be an incredible circle of an unsplit quasi-tangle $Q$ for the test complex $U$. Fix a test subcomplex $U'$ containing $|Q|$ and suppose $C'$ is an incredible circle of $Q$ for $U'$. If there is no compressible quasi-tangle on $D_1^+$ for the test subcomplex $\overline{U'-|Q|}$, then $C$ is isotopic to $C'$ on $D_1^+$ in $U'$.
Example analysis {#subsect:examplesC}
----------------
We have established Circle-method, consisting of Circle-criterion, a flowchart, and four sub-criteria, in the previous two subsections. Some typical examples demonstrate efficiency of the criteria.
\[example:gWl\] (Fig. \[fig:gWl\]) Take the test disk colored in gray and test subcomplex $U'$ on the right. There are two unknotted ears $e_1,e_2$, and only two quasi-tangles $\{e_1\},\{e_2\}$. Connect the endpoints of the ears by arcs $\alpha_1,\alpha_2$ on the corresponding sides, then we see $lk ( e_1 \cup \alpha_1 , e_2 \cup \alpha_2 ) =1$, thus neither quasi-tangle is compressible for $U'$. By Sub-criterion \[prop:subcomplex\] and Circle-criterion, the link is Brunnian.
(140.49, 93.04)(0,0) (0,0)[![\[fig:gWl\] Generalized Whitehead link](gWl.pdf "fig:")]{} (88.77,58.15)[$e_1$]{} (118.90,33.30)[$e_2$]{} (103.06,76.16)[$\alpha_1$]{} (103.06,15.91)[$\alpha_2$]{}
(140.49, 93.04)(0,0) (0,0)[![\[fig:gWl\] Generalized Whitehead link](gWl "fig:")]{} (88.77,58.15)[$e_1$]{} (118.90,33.30)[$e_2$]{} (103.06,76.16)[$\alpha_1$]{} (103.06,15.91)[$\alpha_2$]{}
The third, fourth and last picture in [@D.Rolfsen] Chapter 5 section E Exercise 5 generalize Whitehead link in different ways. The fourth is Example \[example:gWl\] and the other two can be proved by Arc-criterion. The link occurs in Chapter 7 section C Theorem 5 in [@D.Rolfsen], and Fig. 1.5 in [@BS] are also Brunnian, proved similarly by Circle-criterion.
\[example:Milnor\] Milnor link and Brunn’s chain, as shown in Fig. \[fig:milnorlink\].
For Milnor link, take the gray test disk and by Arc-method, we only need to show the ear $e$ is incompressible. Connecting the two endpoints by arc $\alpha$, we obtain another Milnor link with one less component and it suffices to prove the new link is Brunnian. In this way we can begin recursion until Hopf link.
For Brunn’s chain, take the test disk colored in gray and fix sublinks $L_I,L_J,L_0$ as in Section \[sect:preliminaries\]. The unknotted ears $e_1,e_3$ are symmetric, and so are $e_2,e_4$. Delete $e_3$ and $e_4$, and connect the endpoints of $e_1,e_2$ by arcs $\alpha_1,\alpha_2$ on corresponding sides respectively, then $e_1 \cup \alpha_1,e_2 \cup \alpha_2$ and $L_3$ form a Milnor link. So neither $e_1$ nor $e_2$ is compressible. According to the definition, each quasi-tangle contains only one of $e_1,e_2,e_3,e_4$ and some components of $L_3$. One may check easily that such a quasi-tangle is split. So by Sub-criterion \[prop:boundaryconnectedsum\], every quasi-tangle is incompressible.
(325.65, 185.42)(0,0) (0,0)[![\[fig:milnorlink\] Milnor link and Brunn’s chain](Milnorlink.pdf "fig:")]{} (64.41,104.32)[$e_1$]{} (84.14,72.55)[$e_3$]{} (181.11,104.32)[$e_1$]{} (179.10,70.88)[$e_2$]{} (181.11,84.26) (194.48,50.82) (69.76,79.41)[$e_4$]{} (59.73,49.31)[$e_2$]{} (41.54,172.85)[$e$]{} (188.83,157.84)[$e$]{} (186.95,132.65)
(325.65, 185.42)(0,0) (0,0)[![\[fig:milnorlink\] Milnor link and Brunn’s chain](Milnorlink "fig:")]{} (64.41,104.32)[$e_1$]{} (84.14,72.55)[$e_3$]{} (181.11,104.32)[$e_1$]{} (179.10,70.88)[$e_2$]{} (181.11,84.26) (194.48,50.82) (69.76,79.41)[$e_4$]{} (59.73,49.31)[$e_2$]{} (41.54,172.85)[$e$]{} (188.83,157.84)[$e$]{} (186.95,132.65)
Milnor link occurs in Fig. 7 in [@M], Fig. 2.25 in [@BS], and Brunn’s chain appears in [@B; @HDb], also in [@D.Rolfsen] Chapter 7 section J, and Fig. 2.22 and 2.28 in [@BS].
The quick proofs above illustrate how Sub-criterion \[prop:subcomplex\] and \[prop:boundaryconnectedsum\] simplify the task in Flowchart \[fig:flowchart\], while the next two examples show the way Sub-criterion \[cor:incompressibility\] and \[prop:incrediblecircle\] work for complicated links.
(350.85, 82.56)(0,0) (0,0)[![\[fig:Lf1\] Example \[example:Lf1\]](Lf1.pdf "fig:")]{} (89.40,7.97)[$C_1$]{} (7.89,51.44)[$C_2$]{} (98.91,51.44)[$C_3$]{} (236.12,51.44)[$C$]{} (126.46,31.61)[$e_1$]{} (188.41,31.61)[$e_2$]{} (327.42,51.44)[$C_3$]{}
(350.85, 82.56)(0,0) (0,0)[![\[fig:Lf1\] Example \[example:Lf1\]](Lf1 "fig:")]{} (89.40,7.97)[$C_1$]{} (7.89,51.44)[$C_2$]{} (98.91,51.44)[$C_3$]{} (236.12,51.44)[$C$]{} (126.46,31.61)[$e_1$]{} (188.41,31.61)[$e_2$]{} (327.42,51.44)[$C_3$]{}
\[example:Lf1\] (Fig. \[fig:Lf1\]) Choose the test subcomplex $U'$ shown in the middle. There are only two ears $e_1$ and $e_2$ on different sides of the test disk, and by Sub-criterion \[prop:boundaryconnectedsum\] we only need to consider these two ears. In the complement of the trivial substitution of $\{e_2\}$ for $U'$, $C_3$ can bound a disk. However, there is a nontrivial link as a subset of $U'$, shown on the right, which is similar to Milnor link and implies $C_4$ can not bound a disk in the complement of $U'$. Thus by Sub-criterion \[cor:incompressibility\], $\{e_2\}$ is incompressible for $U'$. As for $\{e_1\}$, we see the green circle is the unique candidate for its incredible circle up to isotopy according to Sub-criterion \[prop:incrediblecircle\]. So it suffices to show the circle $C$ can not bound a disk in the complement of $U'$, which is again implied by the nontrivial link on the right.
Fig. 1 and 2 in [@L] are just similar to Example \[example:Lf1\].
\[example:An\] (Fig. \[fig:A5\]) The link $L_F$ in [@HDb], also $A_n$ of Exercise 15 in Chapter 3 section F in [@D.Rolfsen]. Proof for the case $n=2, 3$ need only to be changed a little.
(347.58, 124.22)(0,0) (0,0)[![\[fig:A5\] The link $A_5$](A5.pdf "fig:")]{} (9.76,91.65)[$C_1$]{} (9.76,24.16)[$C_2$]{} (83.66,7.97)[$C_3$]{} (88.72,57.91)[$C_4$]{} (77.25,109.87)[$C_5$]{} (129.06,25.74)[$e_1$]{} (143.14,110.20)[$e_2$]{} (179.09,31.83) (58.16,57.86)[$A_5$]{} (287.60,57.86)[$A_3$]{} (144.69,67.04)[$e_4$]{} (182.26,90.48)[$e_3$]{} (293.31,31.83) (126.68,86.53)
(347.58, 124.22)(0,0) (0,0)[![\[fig:A5\] The link $A_5$](A5 "fig:")]{} (9.76,91.65)[$C_1$]{} (9.76,24.16)[$C_2$]{} (83.66,7.97)[$C_3$]{} (88.72,57.91)[$C_4$]{} (77.25,109.87)[$C_5$]{} (129.06,25.74)[$e_1$]{} (143.14,110.20)[$e_2$]{} (179.09,31.83) (58.16,57.86)[$A_5$]{} (287.60,57.86)[$A_3$]{} (144.69,67.04)[$e_4$]{} (182.26,90.48)[$e_3$]{} (293.31,31.83) (126.68,86.53)
Take the gray test disk and denote the test complex by $U$. By symmetry and Sub-criterion \[prop:boundaryconnectedsum\], we only need to prove ears $e_1,e_2$ are both incompressible for $U$.
First take test subcomplex $U'_1=U-e_2-e_4$. Then up to isotopy, there is only one arc connecting $\partial e_1$ on $D_1^+$. The union of this arc and $e_1$ is isotopic to the red circle $C$. Notice that there is a link $A_{n-2}$ as a subspace of $U'_1$, and it is easily proven that $C$ is homotopically nontrivial in the complement of $A_{n-2}$ (c.f [@D.Rolfsen] p.69). Thus $e_1$ is incompressible for $U'_1$.
For $e_2$, take test subcomplex $U'_2=U-e_4$, and let $N$ be a regular neighborhood of the arc $\alpha$ on $D_1^+- \partial e_1$ connecting $\partial e_2$. Then $e_1$ is still incompressible and $\{e_2\}$ is compressible with incredible circle $\partial N$. Sub-criterion \[prop:incrediblecircle\] asserts $\partial N$ is the unique candidate for incredible circle of $\{ e_2 \}$ for $U$. However, $e_2 \cap \alpha$ is isotopic to $C$ in the complement of $A_n$. Thus $e_2$ is incompressible.
Sub-criterion \[cor:incompressibility\] and \[prop:incrediblecircle\] apply to Fig. 4 in [@MM], the examples in Exercise 4 and the 6th link in Exercise 5 in [@D.Rolfsen] Chapter 5 section E, Fig. 2(b) in [@BFJRVZ], Fig. 2 in [@BCS], Fig. 3 in [@F], and Fig. 6 in [@M] as well.
Proofs and some basic facts {#subsect:proof}
---------------------------
\[lem:tangleball\] With the notation in the paragraph before Definition \[def:unsplit\], suppose $\partial D_0$ bounds another disk $\tilde{D}$ outside $U'_Q$, and denote the corresponding 3-ball including $|Q|$ by $\tilde{B}$. Then there is a homeomorphism $f:(B \cup \mathbf{D_1}, U'_Q )\to (\tilde{B} \cup \mathbf{D_1}, U'_Q )$ which is fixed on $U'_Q$.
If $\mathbb{R}^3$ is replaced by $\mathbb{S}^3$, this lemma is trivial and the homeomorphism is an isotopy. However in $\mathbb{R}^3$, the homeomorphism is in general not an isotopy and this lemma needs to be verified with a caution. We put the proof in the Appendix.
Assume $Q$ is compressible for $U'$ with compressing disk $D_Q$. Then $\partial D_Q$ bounds a disk $D_0$ on $D_1^+$ and the sphere $D_Q\cup D_0$ bounds a 3-ball $B\subset\mathbb{R}^3$. As $Q$ is split, there exists a properly embedded disk $D$ in $B$ which divides $Q$ into $Q_1$ and $Q_2$. Moreover, by an isotopy of $D$ in $B$, we may assume $\partial D$ consists of an arc on $D_0$ and another one on $D_Q$. Then the union of $D$ and half of $D_Q$ is a compressing disk of $Q_1$ for $U'$, a contradiction.
Sub-criterion \[cor:incompressibility\] follows immediately from the following proposition.
\[prop:trivializedsubstitution\] With the notation in the paragraph before Sub-criterion \[cor:incompressibility\], the quasi-tangle $Q$ is compressible for $U'$ if and only if there is an isotopy $H : U'\times[0,1]\to \mathbb{R}^3$ between $U'$ and the trivial substitution $\overline{U'-|Q|} \cup h(|Q|)$, such that $H(\cdot,0)=\mathrm{id}_{U'}$, and $H(\cdot,1)$ coincides with identity on $(U'-( |Q|\cup \mathbf{D_1}))\cup D_1^-$ and $h$ on $|Q|$.
If such an isotopy exists, extend it to an ambient isotopy $\mathscr{H}:\mathbb{R}^3\times [0,1]\to\mathbb{R}^3$. As $U'$ is isomorphic to $(U'-|Q|) \cup h(|Q|)$, we see $(\mathscr{H}(\cdot,1))^{-1}(\partial B_T-{\rm Int}(D_T))$ is a compressing disk of $Q$ in the complement of $U'$.
For the “only if" part, we may assume $Q$ is compressible for $U'$ with compressing disk $D_Q$. Let $B_Q$ be the 3-ball containing $|Q|$, bounded by $D_Q$ and part of $D_1^+$. Then $B_Q\cup U'$ and $B_T\cup\overline{U'-|Q|}$ are isotopic. Any orientation-preserving homeomorphism of $(B_T, D_T)$ is isotopic to identity and the isotopy can be extended. Thus there is an isotopy from $U'$ to the trivial substitution. Moreover, each isotopy involved above can be made with respect to all the restrictions.
Suppose $C,C'$ are both incredible circles of $Q$ in $U'$ and $C'$ is not isotopic to $C$ on $D_1^+$. Minimize the cardinality of $C \cap C'$ with an isotopy of $C'$ on $D_1^+$. As $C$ and $C'$ are not isotopic, now they still intersect. Let $D,D'$ be compressing disks in the complement of $U'$ with $\partial D=C,\partial D'=C'$. The intersection $D \cap D'$ consists of finitely many circles and arcs. The circle components can be eliminated from an innermost one on $D$, by surgery of $D$ along a subdisk in $D'$. So we suppose $D\cap D'$ is a disjoint union of finitely many arcs with endpoints in $C\cap C'$.
Now choose one of these arcs, say $\alpha$, such that $D-\alpha$ has a component not intersecting $D'$. Denote the closure of this component by $D_\alpha$. Also, $D'-\alpha$ is a disjoint union of two disks, and we denote their closure by $D'_\alpha, \tilde{D}'_\alpha $. Now $\partial D_\alpha \cup \partial D'_\alpha -{\rm Int} (\alpha)$ is a circle and bounds a disk $D_0$ on $D_1^+$. Let $B$ be the 3-ball bounded by the sphere $D_0\cup D_\alpha\cup D'_\alpha$. If $L \cap D_0=\emptyset$, we can push $D_\alpha$ across $D'_\alpha$ to eliminate $\alpha$, and the cardinality of $C \cap C'$ is reduced by 2, which contradicts our assumption. Therefore, $L\cap D_0\neq\emptyset$, and there is a quasi-tangle $Q_\alpha$ with $|Q_\alpha |\subset B\cap U'$ which is compressible for $U'$. Without loss of generality, we can assume $Q_\alpha\cap Q=\emptyset$. In fact, if $Q_\alpha\cap Q \neq\emptyset$, then $Q_\alpha=Q$ for $Q$ is unsplit, and we can replace $D'_\alpha$ in our analysis by the other disk $\tilde{D}'_\alpha$. So $Q_\alpha$ is compressible for $U'-|Q|$, which is a contradiction.
We conclude this subsection with some basic facts on Brunnian links, which will be used in the next section.
\[lem:2pts\] Let $L$ be a link and $D$ be a disk with $D \cap L =\{ p_1 , p_2 \}$. If there is another embedded disk $D'$ with $\partial D = \partial D'$ and $D' \cap L =\emptyset$, then there exists an embedded sphere $S^2$ such that $S^2 \cap L =\{ p_1 , p_2 \}$.
Take a small regular neighbourhood $N$ of $L$. Disturb $D'$ such that $D\cap D' \cap N = \emptyset$. Then $D \cap D'$ is a disjoint union of finitely many circles. Let $C$ be an intersection circle innermost on $D'$, i.e., the disk $D'_C\subset D'$ bounded by $C$ satisfies ${\rm Int}(D'_C)\cap D=\emptyset$. We see that $C$ also bounds a disk $D_C$ on $D$, and if one of $p_1,p_2$ is in $D_C$, then both are in $D_C$. If $p_1,p_2\notin D_C$, we can replace $D$ by $(D-D_C)\cup D'_C$ and disturb it a little to eliminate the intersection circle $C$. In this way, finally for every circle $C\subset D\cap D'$, we have $p_1,p_2\in D_C$. Then we choose the circle $C_0\subset D\cap D'$, such that ${\rm Int}(D_{C_0})\cap D'=\emptyset$. The union $ D_{C_0}\cup D'_{C_0}$ is a sphere we are seeking.
\[lem:connectedsum\] Brunnian links are prime. That is, there are no nontrivial connected sum decompositions for them.
Let $L$ be a Brunnian link, and $S^2$ is an embedded sphere in $\mathbb{R}^3$ that intersects $L$ at only two points $p_1,p_2$. Up to isotopy, there is only one way to connect $p_1,p_2$ by an arc on $S_2$. Let $\alpha$ be the arc and $B$ denote the closed 3-ball bounded by $S^2$. To use proof by contradiction, suppose neither of $L_1\triangleq \alpha\cup L\cap B$ and $L_2\triangleq \alpha\cup (L- B)$ is an unknot. Then each of $L_1,L_2$ has less components than $L$. Note that $p_1,p_2$ must on the same component of $L$. Denote the component by $C$. As $C$ is a trivial knot, $C_1\triangleq \alpha\cup C\cap L_1, C_2\triangleq \alpha\cup C\cap L_2$ are also trivial. Therefore, $L_1$ is isotopic to the link $(L_1-C_1)\cup C$, which is a proper sublink of $L$. Thus $L_1$ bounds disjoint disks in $B$ and similarly, $L_2$ bounds disjoint disks in $\mathbb{R}^3-{\rm Int}(B)$. Now we see that $L$ bounds disjoint disks in $\mathbb{R}^3$, which is a contradiction.
\[prop:2pts\] Let $L=\cup_{i=1}^n C_i$ be a Brunnian link. Assume there is a circle $C\subset \mathbb{R}^3-L$ that bounds a disk $D$ with $D \cap L = D \cap C_1 =\{ p_1 , p_2 \}$. Suppose $\alpha\subset D$ is an arc connecting $p_1,p_2 $, and the two components of $C_1-\{p_1,p_2\}$ combined with $\alpha$ form circles $C'_1$ and $C''_1$ respectively. If neither $(L-C_1)\cup C'_1$ nor $(L-C_1)\cup C''_1$ is isotopic to $L$, then any disk bounded by $C$ intersects $L$ at two or more points.
If the linking number $lk(C, C_1 )$ equals 2, the conclusion holds. Otherwise $lk(C, C_1 )=0$. Suppose $C$ bounds a disk $D'$ with $D' \cap L =\emptyset$, then by Lemma \[lem:2pts\], there is an embedded sphere $S^2$ such that $S^2 \cap L =\{ p_1 , p_2 \}$. By Lemma \[lem:connectedsum\], either $(L-C_1)\cup C'_1$ or $(L-C_1)\cup C''_1$ is isotopic to $L$.
Series of large experiments {#sect:baas}
===========================
Nils A. Baas et al. constructed great many families of Brunnian links in [@BCS; @BS], using basic block to synthesis large Brunnian links like surfaces (see Fig. \[fig:Baasblock\]), which generalize Brunn’s chain. To show they are Brunnian, Circle-method will demonstrate significant advantages, compared with the original proofs by HOMFLY polynomial. We deal with two typical series in their construction:
![Baas’ block[]{data-label="fig:Baasblock"}](Baasblock.pdf){height="2cm"}
**Tube$(m,n)$** where $m,n \in \mathbb{N}^*$: a tube with $m$ rows and $n$ columns as shown in Fig. \[fig:Baas2rows\] (1) and \[fig:Baas3rows\] (1).
**Carpet$(m,n,p)$** where $m,n,p \in \mathbb{N}^*$ and $m<n$: a flat regular $p$-gon annulus in $(n-m+1)$ tiers where the innermost and outermost tiers have $mp$ and $np$ components respectively (Fig. 15 in section 3.3 in [@BCS]). This link has $\frac{(m+n)(n-m+1)}{2} p$ components.
We say a subset of $\mathbb{R}^3$ is *split* if there is an embedded sphere in its complement that separates it into two proper subspaces. Brunnian links are not split.
\[example:Baas3rows\] Tube$(3,4)$, Fig. \[fig:Baas3rows\](1).
(233.32, 207.36)(0,0) (0,0)[![\[fig:Baas3rows\] Tube$(3,4)$](Baas3rows.pdf "fig:")]{} (37.76,176.54)[$a$]{} (37.54,193.61)[$b$]{} (5.67,145.96)[(1)]{} (5.67,44.85)[(2)]{}
(233.32, 207.36)(0,0) (0,0)[![\[fig:Baas3rows\] Tube$(3,4)$](Baas3rows "fig:")]{} (37.76,176.54)[$a$]{} (37.54,193.61)[$b$]{} (5.67,145.96)[(1)]{} (5.67,44.85)[(2)]{}
Take the test disks colored in gray and denote the test complex by $U$. By symmetry, there are only two kinds of ears, $a$ and $b$, whose endpoints lie on the “back” side and the “front” side respectively. If we delete all the eight ears of kind $a$ from $U$, we obtain a test subcomplex $U'$. And $U'$ has a subset, the black part of the lower figure, which is a Brunn’s chain. So $U'$ is not split. Meanwhile, the trivial substitution of $b$ for $U'$ is split. Thus $b$ is incompressible for $U'$ by Sub-criterion \[cor:incompressibility\]. The union of $a$ and any arc connecting its endpoints on the disk, is isotopic to the red circle in the complement of the Brunn’s chain, which can not bound a disk according to Proposition \[prop:2pts\]. So $a$ is also incompressible.
\[example:Baas2rows\] Tube$(2,3)$, Fig. \[fig:Baas2rows\](1).
(325.81, 171.55)(0,0) (0,0)[![\[fig:Baas2rows\] Tube$(2,3)$](Baas2rows.pdf "fig:")]{} (113.64,157.80)[$a$]{} (106.14,143.76)[$b$]{} (84.12,141.58)[$c$]{} (103.96,125.61)[$d$]{} (76.68,86.61)[$(1)$]{} (240.96,86.61)[$(2)$]{} (240.96,7.90)[$(4)$]{} (76.68,7.90)[$(3)$]{}
(325.81, 171.55)(0,0) (0,0)[![\[fig:Baas2rows\] Tube$(2,3)$](Baas2rows "fig:")]{} (113.64,157.80)[$a$]{} (106.14,143.76)[$b$]{} (84.12,141.58)[$c$]{} (103.96,125.61)[$d$]{} (76.68,86.61)[$(1)$]{} (240.96,86.61)[$(2)$]{} (240.96,7.90)[$(4)$]{} (76.68,7.90)[$(3)$]{}
To show the Brunnian property of the link (1), we first claim that the black part of (2), denoted $L_{(2)}$, is a Brunnian link. Then take the gray test disk in (1) and denote the test complex by $U$. By Sub-criterion \[cor:incompressibility\], we only consider the unknotted ears $a,b,c,d$. Take test subcomplex $U'=U-a-c$, which takes $L_{(2)}$ as a subset. As we have claimed, $L_{(2)}$ is Brunnian thus not split, which implies $U'$ is neither split. The incompressibility of $b,d$ follows Sub-criterion \[cor:incompressibility\] and the fact that the trivial substitution of either $b$ or $d$ for $U'$ is split. For ear $a$, take an arbitrary arc connecting $\partial a$ on the disk. The union of $a$ and the arc is isotopic to the red circle in the complement of $L_{(2)}$, which can not bound a disk in the complement of $L_{(2)}$ according to Proposition \[prop:2pts\]. Therefore $a$ is incompressible for $U'$. Similarly, $c$ is incompressible (see the green circle). Now we only need to prove our claim that $L_{(2)}$ is Brunnian. Similarly, take the gray test disk in (2) and it suffices to show the link (3) is Brunnian. And we finally reduce to that the link (4) is Brunnian, which is a Brunn’s chain with 2 components.
Similarly to Example \[example:Baas2rows\] and \[example:Baas3rows\], one can verify the Brunnian property for Tube$(m,n)$, Brunnian annulus and Brunnian torus of any rows and columns (Fig. 12, 13 and 14 in section 3.3 in [@BCS]).
\[example:Baas3sheaves\] Carpet$(1,3,4)$, Fig. \[fig:Baas3sheaves\](1).
(299.53, 459.20)(0,0) (0,0)[![\[fig:Baas3sheaves\] Carpet$(1,3,4)$](Baas3sheaves.pdf "fig:")]{} (165.94,432.59)[$a$]{} (177.38,432.59)[$c$]{} (182.28,416.87)[$b$]{} (191.26,445.45)[$d$]{} (217.92,307.04)[$(2)$]{} (33.38,297.02)[$e$]{} (33.38,282.12)[$f$]{} (56.24,282.12)[$g$]{} (56.65,252.31)[$h$]{} (174.61,284.18)[$i$]{} (187.02,259.70)[$j$]{} (28.29,135.19)[$i$]{} (40.71,110.71)[$j$]{} (70.91,307.04)[$(1)$]{} (70.91,156.89)[$(3)$]{} (217.92,156.89)[$(4)$]{} (70.91,7.90)[$(5)$]{} (217.92,7.90)[$(6)$]{}
(299.53, 459.20)(0,0) (0,0)[![\[fig:Baas3sheaves\] Carpet$(1,3,4)$](Baas3sheaves "fig:")]{} (165.94,432.59)[$a$]{} (177.38,432.59)[$c$]{} (182.28,416.87)[$b$]{} (191.26,445.45)[$d$]{} (217.92,307.04)[$(2)$]{} (33.38,297.02)[$e$]{} (33.38,282.12)[$f$]{} (56.24,282.12)[$g$]{} (56.65,252.31)[$h$]{} (174.61,284.18)[$i$]{} (187.02,259.70)[$j$]{} (28.29,135.19)[$i$]{} (40.71,110.71)[$j$]{} (70.91,307.04)[$(1)$]{} (70.91,156.89)[$(3)$]{} (217.92,156.89)[$(4)$]{} (70.91,7.90)[$(5)$]{} (217.92,7.90)[$(6)$]{}
To prove the Brunnian property of the link (1), we claim that the link (3) is Brunnian. Then take the gray test disks as shown in (2) and denote the test complex by $U_1$. There are 4 ears on each test disk and we just consider $a,b,c,d$ on test disk $D_1$. In test subcomplex $U_1-c$, we can connect the endpoints of $a$ by a unique arc on the disk. The union of the arc and $a$ has linking number with a component (in fact there are two) of link (1), thus $a$ is incompressible for $U_1$. Now suppose there is a quasi-tangle $Q\ni a$ compressible for $U_1$ with compressing disk $D_{Q}\subset \mathbb{R}^3- U_1$ and incredible circle $C\triangleq\partial D_{Q}\subset D_1^+$. If $Q\neq\{a\}$, then $U_1-a$ is split by the sphere $D_{Q}\cup C\cup D_C$, where $D_C$ is the disk on $D_1^+$ bounded by $C$ (To be exact, we shall push $D_C$ out of $\mathbf{D_1}$ a little). But the Brunnian link (3) can be viewed as a subset of $U_1-a$, which implies $U_1-a$ is not split and leads to a contradiction. $c$ is symmetric to $a$. Now consider $b,d$ and take test subcomplex $U'_1\triangleq U_1-a-c$. The trivial substitution of $b$ or $d$ for $U'_1$ is split. On the other hand, there is a subset of $U'_1$, the link (3) again, which is not split. So $U'_1$ is not split and by Sub-criterion \[cor:incompressibility\], $b,d$ are incompressible for $U'_1$. According to Sub-criterion \[prop:boundaryconnectedsum\], no quasi-tangle that contains $b$ or $c$ is compressible for $U_1$. So the link (1) is Brunnian.
It remains to prove the Brunnian property of the link (3). Similarly we first claim the black part of (4), denoted $L_{(4)}$, is a Brunnian link. Then take the gray test disks in (3) and denote the test complex by $U_3$. Notice that $L_{(4)}$ can be viewed as a subset of $U_3$. We only consider the four unknotted ears $e,f,g,h$. In test subcomplex $U_3-h$, there is a unique arc connecting the endpoints of $f$ on the disk. The union of the arc and $f$ is isotopic to the red circle in the complement of $L_{(4)}$. According to Proposition \[prop:2pts\] it can not bound a disk in the complement of $L_{(4)}$, thus $f$ is incompressible for $U_3$. Considering $U_3-f$ and the green circle similarly, we see $h$ is also incompressible. For ear $e$ and $g$, take test subcomplex $U'_3\triangleq U_3-f-h$. The trivial substitution of $e$ or $h$ for $U'_3$ is split, while $L_{(4)}\subset U'_3$ implies $U'_3$ is not split. So $e,h$ are incompressible by Sub-criterion \[cor:incompressibility\]. Also, quasi-tangles are all incompressible for $U_3$.
Finally we prove $L_{(4)}$ is Brunnian by taking the gray test disks in (4) and test subcomplex $U_5$. We only need to show the ears $i,j$ are incompressible. The trivial substitution of $i$ or $j$ for $U_5$ is split. However, the link (6), which is a subset of $U_5$ and in fact a Brunn’s chain of 4 components (see Example \[example:Milnor\]), implies that $U_5$ is not split. Again by Sub-criterion \[cor:incompressibility\] we finish the proof.
The Brunnian property of general Carpet$(m,n,p)$ and Brunnian solid (Fig. 16 in section 3.3 in [@BCS]) can be verified in a similar manner.
We now see Circle-method has three distinct features. First, it is simpler than using link invariants in that it is worked by hand. This advantage grows with the size of links. Second, it is highly efficient. For instance, during the proofs in Example \[example:Baas2rows\] and \[example:Baas3sheaves\], we also show the Brunnian property of some new links as by-products. Third, it is very flexible. If we make some “open end” of a component go around the adjacent “interior bend” from the other side, i.e., conduct 4 crossing changes, or twist some double strands, the new link is still Brunnian, as we would only need a little change in the proof while retaining the whole framework. This leads to further construction of new Brunnian links.
Constructions of New Brunnian links {#sect:construction}
===================================
It should be pointed out that the constructions below will give only an indication of its potential for constructing Brunnian links. For instance, Fig. \[fig:jade-pendant\] illustrates a grand Brunnian link by twining Carpet$(2,3,4)$ and “adding" copies of the true lover’s knot and links $7_2^2,8_1^4$ in Rolfsen’s list [@D.Rolfsen]. One can also try to modify other known Brunnian links to create new ones.
![Jade pendant[]{data-label="fig:jade-pendant"}](Jade-pendant.pdf){height="11cm"}
Based on each link Carpet$(m,n,p)$, we now construct several infinite families of new Brunnian links. So there are infinitely many infinite families as $m,n,p$ vary. The proof for the original link, without any change in text, shows they are Brunnian. The next three families are based on Carpet$(2,3,4)$.
1. **Snake.** Choose a segment of double strands on one component, and twine it around other components as in Fig. \[fig:snake\]. As shown in the dashed red box, we can twine it arbitrarily many times once twining happens, giving infinitely many Brannian links.
2. **Fountains.** Twining a segment of double strands on each component around an adjacent component arbitrarily many times (as in the dashed red box) towards specific directions as in Fig. \[fig:fountains\] gives a new sequence of Brannian links.
3. **Cirrus.** As shown in Fig. \[fig:cirrus\], every component twines itself in various ways. Besides, we “add" a tangle in the middle, which may be chosen quite arbitrarily.
![Fountains[]{data-label="fig:fountains"}](Snake.pdf){height="6cm"}
![Fountains[]{data-label="fig:fountains"}](Fountains.pdf){height="6cm"}
![Cirrus[]{data-label="fig:cirrus"}](Cirrus.pdf){height="8cm"}
Recall that the proofs in Example \[example:Baas2rows\] and \[example:Baas3sheaves\] bring new Brunnian links. We can utilize them as well to create new links. We give a series here.
**Wheel.** Fig. \[fig:otherclass\] illustrates an infinite family of new Brunnian links by twining the components of $L_{(4)}$ in Fig. \[fig:Baas3sheaves\].
![Wheel[]{data-label="fig:otherclass"}](otherclass.png){height="8cm"}
Concluding Remarks
==================
We have established general approaches to detecting Brunnian property of links, and used them to create new Brunnian links in bulk. In fact, Arc-method and Circle-method apply for more general problems.
![[]{data-label="fig:fractalborromean"}](nonbrunnian.pdf){height="4cm"}
![[]{data-label="fig:fractalborromean"}](Brunner.pdf){height="4cm"}
![[]{data-label="fig:fractalborromean"}](fractalborromean.pdf){height="4cm"}
1\. Our method can be extended directly to show a link is nontrivial, if it has some unknot components so that we can take test disks, without considering whether each component is unknotted. For example, the test disk in Fig. \[fig:notBrunnian\] shows fast this link is nontrivial. Especially, our methods work to detect generalized Brunnian properties (c.f. [@HDb; @J; @TKa; @TK]). For instance, Arc-method shows Fig. \[fig:brunner\] is nontrivial, which was originally proved by Alexander invariants[@HDb].
2\. Given a link, considering its Brunnian sublinks helps to show it is not split. Notice that Brunnian links are unsplit and if two unsplit links share a common component, then the union of them is not split either. Thus, for instance, Fig. \[fig:fractalborromean\] (from [@J]) is not split.
3\. In a 3-dimensional manifold $M^3$, an *unlink* is defined by the boundary of disjoint union of embedded disks. So we can define Brunnian links in $M^3$ as well. It can be verified that Arc-method still holds. As a sphere in $M^3$ (even if irreducible) generally does not bound a ball on specific side, to generalize Circle-method into $M^3$, we need change the definition of “compressible” for quasi-tangles with care. This modification is relatively easy when $M^3$ is a submanifold of $\mathbb{S}^3$.
[10]{} Baas, N. A., , The European Physical Journal Special Topics 178.1 (2009): 25-44.
Baas, N. A., , International Journal of General Systems 42.2 (2013): 137-169.
Baas, N. A. and Seeman, N. C., , Journal of Mathematical Chemistry 50.1 (2012): 220-232.
Baas, N. A., Seeman, N. C., and Stacey, A., , Journal of mathematical chemistry 53.1 (2015): 183-199.
Baas, N. A., Fedorov, D. V., Jensen, A. S., et al., , Physics of Atomic Nuclei 77.3 (2014): 336-343.
Baas, N. A. and Stacey, A., , arXiv:1602.06450v1 \[math.AT\] (2016).
Bai, S. and Ma, J. M., , arXiv:1906.01253v1 \[math.GT\] (2019).
Brunn, H., , Math-Phys. Klasse, 22 (1892): 77-99.
Debrunner, H., , Duke Mathematical Journal 28.1 (1961): 17-23.
Duzhin, F. and Wong, S. M. Z., , Journal of Knot Theory and Its Ramifications 23.03 (2014): 1420002.
Eliahou, S., Kauffman, L. H. and Thistlethwaite, M. B., , Topology 42.1 (2003): 155-169.
Jablan, S. V., , Forma 14.4 (1999): 269-277.
Kanenobu, T., , Archiv der Mathematik 44.4 (1985): 369-372.
Kanebobu, T., , Journal of the Mathematical Society of Japan 38.2 (1986): 295-308.
Li, G. S., , Proceedings of the American Mathematical Society 126.5 (1998): 1557-1563.
Milnor, J., , Annals of Mathematics 59.2 (1954): 177-195.
Melikhov, S. A. and Mikhailov, R. V., , arXiv:math/0201022v1 \[math.GT\] (2002).
Mangum, B. S. and Stanford, T., , Algebraic & Geometric Topology 1.1 (2001): 143-152.
Miyazawa, H. A. and Yasuhara, A., , Topology and its Applications 153.11 (2006): 1643-1650.
Rolfsen, D., , Berkeley, CA: Publish or Perish, 1976.
Appendix
========
(133.66, 124.26)(0,0) (0,0)[![\[fig:tangleball\] $\overline{B_1}=\mathbb{S}^3-{\rm Int}(\mathbf{D}_1)$](tangleball.pdf "fig:")]{} (69.20,36.98)[$\infty$]{} (5.67,72.70)[$D_1^+$]{} (5.67,47.35)[$D_1^-$]{} (35.19,88.74)[$C$]{} (68.65,108.00)[$Q$]{} (103.21,78.80)[$D$]{} (108.35,107.35)[$D_0$]{} (78.81,31.55)
(133.66, 124.26)(0,0) (0,0)[![\[fig:tangleball\] $\overline{B_1}=\mathbb{S}^3-{\rm Int}(\mathbf{D}_1)$](tangleball "fig:")]{} (69.20,36.98)[$\infty$]{} (5.67,72.70)[$D_1^+$]{} (5.67,47.35)[$D_1^-$]{} (35.19,88.74)[$C$]{} (68.65,108.00)[$Q$]{} (103.21,78.80)[$D$]{} (108.35,107.35)[$D_0$]{} (78.81,31.55)
Consider it in $\mathbb{S}^3 = \mathbb{R}^3 \cup \{ \infty \}$, then $B_1\triangleq\mathbb{S}^3 -\mathbf{D_1}$ and $B_2\triangleq B_1-B$ are both 3-balls, see Fig. \[fig:tangleball\]. Connect $\infty$ with a point $p\in D$ by an arc in $B_2$, and let $V$ be its regular neighborhood in $B_1$ such that $V\cap Q=\emptyset$ and $V\cap D$ is a regular neighborhood of $p$ on $D$. Let $B^*=B\cup V$. Then there is a homeomorphism $f_0:(B \cup \mathbf{D_1}, U'_Q )\to (B^* \cup \mathbf{D_1}, U'_Q )$ which is identity on $U'_Q$ and supported by a regular neighborhood of $V$. Moreover, the disk $\partial B^*\cap B_2$ is parallel to $\partial \mathbf{D}_1-D_0$ hence unique up to an isotopy in $B_1$ which fixes $Q$. For $\tilde{B}$ there is a corresponding 3-ball $\tilde{B}^*$. Extending the isotopy from $\partial B^*\cap B_2$ to $\partial \tilde{B}^*\cap B_2$, we obtain a homeomorphism $g:(B^* \cup \mathbf{D_1}, U'_Q )\to (\tilde{B}^* \cup \mathbf{D_1}, U'_Q )$ which is identity on $U'_Q$. $f$ can be chosen as the composition $f_0,g$ and the inverse of $f_0':(\tilde{B} \cup \mathbf{D_1}, U'_Q )\to (\tilde{B}^* \cup \mathbf{D_1}, U'_Q )$ which is defined similarly to $f_0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $T_\beta$ be the $\beta$-transformation on $[0,1)$ defined by $$T_\beta(x)=\beta x\text{ mod }1.$$ We study the Diophantine approximation of the orbit of a point $x$ under $T_\beta$. Precisely, for given two positive functions $\psi_1,~\psi_2:\mathbb{N}\rightarrow\mathbb{R}^+$, define $$\mathcal{L}(\psi_1):=\left\{x\in[0,1]:T_\beta^n x<\psi_1(n),\text{ for infinitely many $n\in\mathbb{N}$}\right\},$$ $$\mathcal{U}(\psi_2):=\left\{x\in [0,1]:\forall~N\gg1,~\exists~n\in[0,N],~s.t.~T^n_\beta x<\psi_2(N)\right\},$$ where $\gg$ means large enough. We compute the Hausdorff dimension of the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$. As a corollary, we estimate the Hausdorff dimension of the set $\mathcal{U}(\psi_2)$.'
address:
- 'Department of Mathematics, Soochow University, Suzhou, 215006, China'
- 'Laboratoire d’Analyse et de Mathématiques Appliquées, Université Paris-Est Créteil Val de Marne, Créteil, 94010, France'
- 'Department of Mathematics, South China University of Technology, Guangzhou, 510000, China'
- 'Laboratoire d’Analyse et de Mathématiques Appliquées, Université Paris-Est Créteil Val de Marne, Créteil, 94010, France'
author:
- Wanlou Wu
- Lixuan Zheng
title: 'Dimension theory of Diophantine approximation related to $\beta$-transformations'
---
Introduction
============
Diophantine approximation, which originaly asks how closely can a given irrational number be approximated by a rational number $p/q$ with denominator $q$ no larger than a fixed positive integer $q_0$, has been widely studied by mathematicians. In 1842, Dirichlet [@D1842] proved the following theorem.\
[**Dirichlet Theorem**]{} Given two real numbers $\theta,~Q$ with $Q\geq 1$, there is an integer $n$ with $1\leq n\leq Q$ such that $$\lVert n \theta\rVert<Q^{-1},$$ where $\lVert\xi\rVert$ denotes the distance from $\xi$ to the nearest integer.
Dirichlet Theorem is called a *uniform approximation theorem* in [@W12 pp.2]. A weak form of Dirichlet Theorem, called an *asymptotic approximation theorem* in [@W12 pp.2], which was often refered to as a corollary of Dirichlet Theorem in the litterature has already existed in the book of Legendre [@L2009 1808, pp.18-19] (using a continued fraction fact): for any real number $\theta$, there are infinitely many $n\in\mathbb{N}$ such that $$\lVert n\theta\rVert<n^{-1}.$$ For the general case, Khintchine in 1924 [@K1924] showed that for a positive function $\psi:\mathbb{N}\rightarrow\mathbb{R}^+$, if $x\mapsto x\psi(x)$ is non-increasing, then $$\mathcal{L}_\psi:=\left\{\theta\in\mathbb{R}:\lVert n\theta\rVert<\psi(n),\text{ for infinitely many }n\in\mathbb{N}\right\}$$ has Lebesgue measure zero if the series $\sum\psi(n)$ converges and has full Lebesgue measure otherwise. In the case where the set has Lebesgue measure zero, it is natural to calculate the Hausdorff dimension of $\mathcal{L}_\psi$. The first result on the Hausdorff dimension of $\mathcal{L}_\psi$ dates back to Jarník-Bosicovitch Theorem [@B34; @J29]. It was shown that the set $$\left\{\theta\in\mathbb{R}:\lVert n\theta\rVert<\dfrac{1}{n^\tau},\text{ for infinitely many }n\in\mathbb{N}\right\}$$ has Hausdorff deminsion $\dfrac{2}{1+\tau}$, for any $\tau>1$.
In analogy with the classical Diophantine approximation, Hill and Velani [@HV1995] studied the approximation properties of the orbits of a dynamical system and introduced the so called *shrinking target problems*: for a measure preserving transformation $T:M\rightarrow M$ on a manifold $M$, what is the size (Lebesgue measure, Hausdorff dimension) of the set $$\left\{x\in M:T^nx\in B(n),\text{ for infinitely many $n\in\mathbb{N}$}\right\},$$ where $B(n)=B(x_0, r(n))$ is a ball centred at $x_0$ with radius $r(n)$($r(n)\rightarrow 0$)? They answered the case where $T$ is an expanding rational map of the Riemann sphere $\overline{\mathbb{C}}=\mathbb{C}\cup\{\infty\}$.
In this papper, we are interested in the approximation properties of the orbits of $\beta$-transformations. The $\beta$-transformation $T_\beta~(\beta>1)$ on $[0,1)$ is defined by $$T_\beta(x):=\beta x-\lfloor\beta x\rfloor,$$ where $\lfloor\cdot\rfloor$ is the integer part function. For any positive function $\psi:\mathbb{N}\rightarrow\mathbb{R}^+$, define the set of *$\psi$-well asymptotically approximable* points by $x_0$ as $$\mathcal{L}(\psi,x_0):=\{x\in[0,1]:\lvert T_\beta^n x-x_0\rvert<\psi(n),\text{ for infinitely many $n\in\mathbb{N}$}\}.$$ By [@P67 Theorem 2A, B, C], the set $\mathcal{L}(\psi,x_0)$ has Lebesgue measure zero if and only if the series $\sum\psi(n)$ converges. Shen and Wang [@SW2013 Theorem 1.1] established the following result on the Hausdorff dimension of $\mathcal{L}(\psi,x_0)$.\
*[**Theorem SW**]{}([@SW2013 Theorem 1.1]) For any real number $\beta>1$ and any point $x_0\in[0,1]$, one has $${\rm dim}_H \left(\mathcal{L}(\psi, x_0)\right)=\dfrac{1}{1+v},\quad\text{where $v:=\liminf\limits_{n\rightarrow\infty}\dfrac{-\log_\beta\psi(n)}{n}$}.$$*
Parallel to the asymptotic approximation theorem, it is also worth of studying the uniform approximation properties as in Dirichlet Theorem. The uniform Diophantine approximation related to $\beta$-transformations was studied by Bugeaud and Liao [@YLiao2016]. For $x\in[0,1)$, let $$\nu_\beta(x):=\sup\left\{v\geq 0:T^n_\beta x<(\beta^n)^{-v},\text{ for infinitely many $n\in\mathbb{N}$}\right\},$$ $$\hat{\nu}_\beta(x):=\sup\left\{v\geq 0:\forall~N\gg 1,~T^n_\beta x<(\beta^N)^{-v}\text{ has a solution $n\in[0,N]$}\right\}.$$ Bugeaud and Liao [@YLiao2016] proved the following theorem.\
*[**Theorem BL**]{} ([@YLiao2016 Theorem 1.4]) For any $v\in(0,+\infty)$ and any $\hat{v}\in(0,1)$, if $v<\hat{v}/(1-\hat{v})$, then the set $$\left\{x\in[0,1]:\nu_\beta(x)= v\right\}\cap\left\{x\in[0,1]:\hat{\nu}_\beta(x)\geq \hat{v}\right\}$$ is empty. Otherwise, $${\rm dim}_H \left(\{x\in[0,1]:\nu_\beta(x)= v\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)=\hat{v}\}\right)=\dfrac{v-\hat{v}-v\hat{v}}{(1+v)(v-\hat{v})}.$$*The exponents $\nu_\beta$ and $\hat{\nu}_\beta$ were introduced in [@AB10](see also [@B2012 Ch.7]). They are strongly related to the run-length function of $\beta$-expansions (see [@Z18]). The aim of this paper is to study the Diophantine approximation sets in [@YLiao2016] when the approximation speed function $n\mapsto\beta^{-nv}$ is replaced by a general positive function. More precisely, fix two positive functions $\psi_1,~\psi_2:\mathbb{N}\rightarrow\mathbb{R}^+$, and define $$\mathcal{L}(\psi_1):=\left\{x\in [0,1]:T_\beta^n x<\psi_1(n),\text{ for infinitely many $n\in\mathbb{N}$}\right\},$$ $$\mathcal{U}(\psi_2):=\left\{x\in [0,1]:\forall~N\gg 1,~T^n_\beta x<\psi_2(N)\text{ has a solution $n\in[0,N]$}\right\}.$$ We will estimate the Hausdorff dimension of the sets $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ and $\mathcal{U}(\psi_2)$. Let $$\underline{v}_1:=\liminf_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_1(n)}{n},\qquad\overline{v}_1:=\limsup_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_1(n)}{n};$$ $$\underline{v}_2:=\liminf_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n)}{n},\qquad\overline{v}_2:=\limsup_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n)}{n}.$$ If $\underline{v}_1<0$, by the definition of $\underline{v}_1$, there is a sequence $\{n_j\}$ such that $$\lim_{j\rightarrow\infty}\dfrac{-\log_\beta\psi_1(n_j)}{n_j}=\underline{v}_1.$$ Then, for $\varepsilon>0$ small enough, there exists an integer $j_0$ such that $$1<\beta^{-n_j(\underline{v}_1+\varepsilon)}\leq\psi_1(n_j),~\text{for any $j\geq j_0$}.$$ By the fact $T^n_\beta x<1$, for any $x\in[0,1)$ and any $n\in\mathbb{N}$, for any $x\in[0,1)$, we have $T^{n_j}_\beta x<1<\psi_1(n_j)$. This implies $$[0,1)\subseteq\mathcal{L}(\psi_1).$$ On the other hand, if we take all the integers $n_i$ with the following property $$\psi_2(n_i)>1,~\text{for $i=1,2,3\cdots$}$$ then for any $x\in[0,1)$ and any integer $n\in[1,n_i]$, we have $T^n_\beta x<1<\psi_2(n_i)$. Thus, we can replace $\psi_2(n)$ by the function $$\widetilde{\psi}_2(n)=
\begin{cases}
\psi_2(n), &\text{ if $n\neq n_i$}\\
1, &\text{ if $n=n_i$}
\end{cases},~i=1,2,\cdots$$ The size (Lebesgue measure, Hausdorff dimension) of the sets $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ and $\mathcal{U}(\psi_2)$ are the same as that of the sets $\mathcal{L}(\psi_1)\cap\mathcal{U}(\widetilde{\psi}_2)$ and $\mathcal{U}(\widetilde{\psi}_2)$. Therefore, in this paper, we always assume $\underline{v}_1\geq0$ and $\underline{v}_2\geq0$. We establish the following theorems on the Hausdorff dimension of the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$.
\[Special-case\]
1. If $\underline{v}_1=\overline{v}_1=\underline{v}_2=\overline{v}_2=0$, then $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=1;$$
2. If $\underline{v}_2=\infty$ and $0\leq\underline{v}_1\leq\overline{v}_1\leq\infty$, then $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable.
3. If $\underline{v}_1=\infty$ and $0\leq\underline{v}_2\leq\overline{v}_2\leq\infty$, then ${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=0$.
For Item (1), the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is not necessary of full Lebesgue measure. In fact, if the series $\sum\psi_1(n)$ converges, by [@P67 Theorem 2A, B, C], $$m\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=0,$$ where $m(A)$ denotes the Lebesgue measure of $A$. However, the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ can also be of full Lebesgue measure. For example, if $\psi_1(n)=\psi_2(n)=1/n$, according to Dmitry, Konstantoulas, and Florian [@KKR19 Theorem 1.1], $$m\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=1.$$
For Item $(3)$, if $\underline{v}_2=\infty$, then $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable. If $1<\underline{v}_2<\infty$, then $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is empty (see Lemma \[empty\]). If $0<\underline{v}_2\leq1$, then $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is uncountable (see Proposition \[uncountable\]).
\[Nonspecial\] If $\underline{v}_2>1$, then $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable. If $\overline{v}_1/(2+\overline{v}_1)\leq\underline{v}_2\leq1<\overline{v}_2$, then $$0\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ If $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2\leq\overline{v}_2\leq1$ and $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2$, then $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~ \left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ If $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$ and $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2\leq1$, then $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$ If $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2\leq\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, then $$\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}\leq{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ If $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$ and $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, then $$\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}\leq{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
We remark that Theorems \[Special-case\] and \[Nonspecial\] give all the cases. We also estimate the Hausdorff dimension of $\mathcal{U}(\psi_2)$.
\[Uniform\] If $\underline{v}_2>1$, then $\mathcal{U}(\psi_2)$ is countable. If $\underline{v}_2\leq1<\overline{v}_2$, then $$0\leq{\rm dim}_H \left(\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ If $\overline{v}_2\leq1$, then $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H \left(\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$
We will show in Examples \[Exa1\], \[Exa2\], \[Exa3\], \[Exa4\], \[Exa5\], \[Exa6\] and \[Exa7\], that the upper and lower bound of the Hausdorff dimension in Theorems \[Nonspecial\] and \[Uniform\] can be all reached. When $\underline{v}_1=\overline{v}_1=0$, we have the result as Theorem \[Belong\].
\[Belong\] Assume $\underline{v}_1=\overline{v}_1=0$. If $\underline{v}_2>0$, then $\mathcal{U}(\psi_2)\subseteq \mathcal{L}(\psi_1)$.
Our paper is organized as follows. We recall some classical results of the theory of $\beta$-expansion in Section $2$. Theorems \[Special-case\] and \[Nonspecial\] are proved in Section $3$. Section $4$ establishes Theorems \[Uniform\] and \[Belong\]. In Section $5$, we give examples to show that the estimations in Theorems \[Nonspecial\] and \[Uniform\] are sharp.
$\beta$-expansions
==================
The notion of $\beta$-expansion was introduced by Rényi [@Ren57] in 1957. For any $\beta>1$, the $\beta$-transformation $T_{\beta}$ on $[0,1)$ is defined by $$T_\beta x=\beta x-\lfloor\beta x\rfloor,$$ where $\lfloor\xi\rfloor$ denotes the largest integer less than or equal to $\xi$. Let $$\lceil\beta\rfloor=
\begin{cases}
\beta -1, & \text{if}~\beta \text{ is a positive integer},\\
\lfloor \beta \rfloor, & \text{otherwise}.
\end{cases}$$
The $\beta$-expansion of a number $x\in[0,1)$ is the sequence $\{\varepsilon_n\}_{n\geq 1}:=\{\varepsilon_n(x,\beta)\}_{n\geq 1}$ of integers from $\{0,1,\cdots,\lceil\beta\rfloor\}$ such that $$\label{E1}
x=\dfrac{\varepsilon_1}{\beta}+\dfrac{\varepsilon_2}{\beta^2}+\cdots+\dfrac{\varepsilon_n}{\beta^n}+\cdots,$$ where $$\varepsilon_1=\varepsilon_1(x,\beta)=\lfloor\beta x \rfloor,~\varepsilon_n=\varepsilon_n(x,\beta)=\lfloor\beta T^{n-1}_\beta x\rfloor,~\text{ for all $n\geq 2$}.$$ We also write $ d_\beta(x)=\left(\varepsilon_1,\cdots,\varepsilon_n,\cdots\right)$.
We can extend the definition of the $\beta$-transformation to the point $1$ as:$$T_\beta 1=\beta-\lfloor\beta \rfloor.$$ One can obtain $$1=\dfrac{\varepsilon_1(1,\beta)}{\beta}+\dfrac{\varepsilon_2(1,\beta)}{\beta^2}+\cdots+\dfrac{\varepsilon_n(1,\beta)}{\beta^n}+\cdots,$$ where $\varepsilon_1(1,\beta)=\lfloor\beta\rfloor,~\varepsilon_n=\lfloor\beta T^{n-1}_\beta 1\rfloor,~\text{for all $n\geq 2$}$. We also write $$d_\beta(1)=\left(\varepsilon_1(1,\beta),\cdots,\varepsilon_n(1,\beta),\cdots\right).$$ If $d_\beta(1)$ is finite, i.e., there is an integer $m>0$ such that $\varepsilon_m(1,\beta)\neq 0$ and $\varepsilon_i(1,\beta)=0$ for all $i>m$, then $\beta$ is called a *simple Parry number*. In this case, the infinite $\beta$-expansion of $1$ is defined as: $$(\varepsilon^\ast_1(\beta),\varepsilon^\ast_2(\beta),\cdots,\varepsilon^\ast_n(\beta),\cdots):=(\varepsilon_1(1,\beta),\varepsilon_2(1,\beta),\cdots,\varepsilon_m(1,\beta)-1)^\infty,$$ where $(\omega)^\infty$ denotes the periodic sequence $(\omega, \omega,\cdots)$. If $d_\beta(1)$ is infinite, then we define $$(\varepsilon^\ast_1(\beta),\varepsilon^\ast_2(\beta),\cdots,\varepsilon^\ast_n(\beta),\cdots):=(\varepsilon_1(1,\beta),\varepsilon_2(1,\beta),\cdots,\varepsilon_n(1,\beta),\cdots).$$
Endow the set $\{0,1,\cdots,\lceil\beta\rfloor\}^{\mathbb{N}}$ with the product topology and define the one-sided shift operator $\sigma$ as:$$\sigma\left((\omega_n)_{n\geq1}\right):=(\omega_{n+1})_{n\geq1},$$ for any infinite sequence $(\omega_n)_{n\geq1}$ in $\{0,1,\cdots,\lceil\beta\rfloor\}^{\mathbb{N}}$. The lexicographical order $<_{lex}$ on $\{0,1,\cdots,\lceil\beta\rfloor\}^{\mathbb{N}}$ is defined as: $$\omega=(\omega_1,\omega_2,\cdots)<_{lex}\omega'=(\omega'_1,\omega'_2,\cdots),$$ if $\omega_1<\omega'_1$ or if there is an integer $k\geq 2$ such that for all $1\leq i< k$, $\omega_i=\omega'_i$ but $\omega_k<\omega'_k$. Denote by $\omega\leq_{lex}\omega'$ if $\omega<_{lex}\omega'$ or $\omega=\omega'$.
A finite word $(\omega_1,\omega_2,\cdots,\omega_n)$ is called $\beta$-admissible, if there is $x\in[0,1]$ such that the $\beta$-expansion of $x$ begins with $(\omega_1,\omega_2,\cdots,\omega_n).$ An infinite sequence $(\omega_1,\omega_2,\cdots,\omega_n,\cdots)$ is called $\beta$-admissible, if there is $x\in[0,1]$ such that the $\beta$-expansion of $x$ is $(\omega_1,\omega_2,\cdots,\omega_n,\cdots).$
Denote by $\Sigma_\beta$ the set of all infinite $\beta$-admissible sequences and $\Sigma^n_\beta$ the set of all $\beta$-admissible sequences with length $n$. The $\beta$-admissible sequences are characterized by Parry [@P1960] and Rényi [@Ren57].
\[Admissible\] Let $\beta>1$,
(1) ([@P1960 Lemma 1]) A word $\omega=(\omega_n)_{n\geq 1}\in\Sigma_\beta$ if and only if $$\sigma^k(\omega)\leq_{lex}(\varepsilon^\ast_1(\beta),\varepsilon^\ast_2(\beta),\cdots,\varepsilon^\ast_n(\beta),\cdots),~\text{for all $k\geq 0$}.$$
(2) ([@P1960 Lemma 3]) For any $x_1,~x_2\in[0,1]$, $x_1<x_2$ if and only if $$d_\beta(x_1)<_{lex}d_\beta(x_2).$$
(3) ([@P1960 Lemma 4]) For any $\beta_2>\beta_1>1$, one has $$\Sigma^n_{\beta_1}\subseteq\Sigma^n_{\beta_2},\quad\Sigma_{\beta_1}\subseteq\Sigma_{\beta_2}.$$
\[cardinality\]([@Ren57 Theorem 2]) For any $\beta>1$, one has $$\beta^n\leq\sharp \Sigma^n_\beta\leq\dfrac{\beta^{n+1}}{\beta-1},$$ where $\sharp$ denotes the cardinality of a finite set.
For every $(\omega_1,\cdots,\omega_n)\in\Sigma^n_\beta$, we call $$I_n(\omega_1,\cdots,\omega_n):=\{x\in[0,1]:d_\beta(x)\text{ starts with }(\omega_1,\cdots,\omega_n)\}$$ an *$n$-th order basic interval* with respect to $\beta$. Denote by $I_n(x)$ the $n$-th order basic interval containing $x$. The basic intervals are also called *cylinders* by some authors. It is crucial to estimate the lengths of the basic intervals. We will use the key notion of full basic intervalas follows (see [@LW2008; @FW2012]).
For any $(\omega_1,\cdots,\omega_n)\in\Sigma^n_\beta$, a basic interval $I_n(\omega_1,\cdots,\omega_n)$ is said to be full if its length is $\beta^{-n}$.
\[full\] ([@FW2012 Lemma 3.1] and [@SW2013 Lemma 2.5])
For any $(\omega_1,\cdots,\omega_n)\in\Sigma^n_\beta$, the following statements are equivalent:
(1) $I_n(\omega_1,\cdots,\omega_n)$ is a full basic interval;
(2) $T^n_\beta I_n(\omega_1,\cdots,\omega_n)=[0,1)$;
(3) For any $\omega'=(\omega'_1,\cdots,\omega'_m)\in\Sigma^m_\beta$, the concatenation $$(\omega_1,\cdots,\omega_n,\omega'_1,\cdots,\omega'_m)\in\Sigma^{n+m}_\beta, \text{ i.e., is $\beta$-admissible.}$$
\[fullc\] ([@SW2013 Corollary 2.6])
(1) If $(\omega_1,\cdots,\omega_{n+1})$ is a $\beta$-admissible sequence with $\omega_{n+1}\neq0$, then $$I_{n+1}(\omega_1,\cdots,\omega'_{n+1})$$ is full for any $0\leq \omega'_{n+1}<\omega_{n+1}$.
(2) For every $\omega\in\Sigma^n_\beta$, if $I_n(\omega)$ is full, then for any $\omega'\in\Sigma^m_\beta$, one has $$\lvert I_{n+m}(\omega,\omega')\rvert=\lvert I_n(\omega)\rvert\cdot\lvert I_m(\omega')\rvert=\dfrac{\lvert I_m(\omega')\rvert}{\beta^n}.$$
(3) For any $\omega\in\Sigma^n_\beta$, if $I_{n+m}(\omega,\omega')$ is a full basic interval contained in $I_n(\omega)$ with the smallest order, then $$\lvert I_{n+m}(\omega,\omega')\rvert\geq\dfrac{\lvert I_n(\omega)\rvert}{\beta}.$$
Next, we define a sequence of numbers $\beta_N$ approaching to $\beta$ as follows. Let $\{\varepsilon^\ast_k(\beta):k\geq 1\}$ be the infinite $\beta$-expansion of $1$. Let $\beta_N$ be the unique real solution of the equation $$\label{ED1}
1=\dfrac{\varepsilon^\ast_1(\beta)}{z}+\cdots+\dfrac{\varepsilon^\ast_N(\beta)}{z^N}.$$ Therefore, $\beta_N<\beta$ and the sequence $\{\beta_N:N\geq 1\}$ increases and converges to $\beta$ when $N$ tends to infinity.
\[length\] ([@SW2013 Lemma 2.7]) For every $\omega\in\Sigma^n_{\beta_N}$ viewed as an element of $\Sigma^n_\beta$, one has $$\dfrac{1}{\beta^{n+N}}\leq\lvert I_n(\omega_1,\cdots,\omega_n)\rvert\leq\dfrac{1}{\beta^n}.$$
Proofs of Theorems \[Special-case\] and \[Nonspecial\] {#sec2}
=======================================================
First, we give an easy fact which is useful for the proofs of Theorems \[Special-case\] and \[Nonspecial\].
\[fact\] For any $x\in[0,1)$, if there is an integer $n_0$ such that $T^{n_0}_\beta x=0$, then $$\nu_\beta(x)=\hat{\nu}_\beta(x)=\infty.$$
Now, we prove that if $\overline{v}_2=\infty$, then the Hausdorff dimensions of the sets $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ and $\mathcal{U}(\psi_2)$ are zero.
\[Pro\] If $\overline{v}_2=\infty$, then $\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)=\infty\}$. Thus, $${\rm dim}_H(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2))={\rm dim}_H(\mathcal{U}(\psi_2))=0.$$
For every $x\in\mathcal{U}(\psi_2)$, we distinguish two cases:
[**Case 1:**]{} There is an integer $n_0$ such that $T^{n_0}_\beta x=0$. By Fact \[fact\], we have $$\nu_\beta(x)=\infty.$$
[**Case 2:**]{} For any $n\in\mathbb{N}$, we always have $T^n_\beta x>0$. Since $x\in\mathcal{U}(\psi_2)$, there is $N_0\geq1$ such that for any $N\geq N_0$, there is an integer $n\in[0,N]$ such that $$0<T_\beta^nx<\psi_2(N).$$ Since $\limsup\limits_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n)}{n}=\overline{v}_2=\infty$, for any $L>0$ large enough, there is a sequence $\{n_i\}$ such that $\psi_2(n_i)\leq\beta^{-n_iL}$. Let $m_1:=\min\{n_i:~n_i\geq N_0\}$, there is an integer $j_1\in[0,m_1]$ such that $$0<T_\beta^{j_1}x<\psi_2(m_1)\leq\beta^{-m_1L}\leq\beta^{-j_1L}.$$ Take $m_2:=\min\left\{n_i>m_1:\beta^{-n_iL}<T_\beta^{j_1}x\right\}$. There is $j_2\in[0,m_2]$ such that $$0<T_\beta^{j_2}x<\psi_2(m_2)\leq\beta^{-m_2L}\leq\beta^{-j_2L}.$$ Since $T_\beta^{j_2}x<\psi_2(m_2)\leq\beta^{-m_2L}<T_\beta^{j_1}x$, $j_2\neq j_1$. Repeat this process, one can get a sequence of pairwise disjoint integers $\{j_i:i\geq1\}$ such that $$0<T_\beta^{j_i}x<\psi_2(m_i)\leq\beta^{-m_iL}\leq\beta^{-j_iL}.$$ Therefore, $\nu_\beta(x)\geq L$. By the arbitrariness of $L$, we have $\nu_\beta(x)=\infty$.
Hence, in all cases, we have $$\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)=\infty\}.$$ By [**Theorem SW**]{}, $${\rm dim}_H(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2))\leq{\rm dim}_H(\mathcal{U}(\psi_2))\leq{\rm dim}_H(\{x\in[0,1]:\nu_\beta(x)=\infty\})=0.$$
We discuss the relation between $\nu_\beta(x)$, $\hat{\nu}_\beta(x)$ and $\underline{v}_1$, $\overline{v}_1$, $\underline{v}_2$, $\overline{v}_2$, which are important to the proof of Theorem \[Nonspecial\].
\[SET\] For $0\leq\underline{v}_1\leq\overline{v}_1<\infty$ and $0\leq\underline{v}_2\leq\overline{v}_2<\infty$, one has
1. $\{x\in[0,1]:\nu_\beta(x)>\overline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)>\overline{v}_2\}\subseteq\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2);$
2. $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)\geq \underline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$
$(1)$ For any $x$ with $\nu_\beta(x)>\overline{v}_1$ and any $\varepsilon>0$ small enough, there is a sequence $\{n_i\}$ such that $T_\beta^{n_i}x<\beta^{-n_i(\overline{v}_1+\varepsilon)}$. By the definition of $\overline{v}_1$, for the above $\varepsilon$, there is an integer $i_0$ such that $$\psi_1(n_i)>\beta^{-n_i(\overline{v}_1+\varepsilon)},\quad\text{for any $i\geq i_0$}.$$ Then, $$T_\beta^{n_i}x<\beta^{-n_i(\overline{v}_1+\varepsilon)}<\psi_1(n_i),~\text{for any $i\geq i_0$}.$$ Thus, $x\in\mathcal{L}(\psi_1)$. Therefore, $$\{x\in[0,1]:\nu_\beta(x)>\overline{v}_1\}\subseteq\mathcal{L}(\psi_1).$$ By similar discussion, $\{x\in[0,1]:\hat{\nu}_\beta(x)>\overline{v}_2\}\subseteq \mathcal{U}(\psi_2).$ Thus, $$\{x\in[0,1]:\nu_\beta(x)>\overline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)>\overline{v}_2\}\subseteq\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2).$$
$(2)$ For any $x\in\mathcal{L}(\psi_1)$, there is a sequence $\{n_i\}$ such that $$T_\beta^{n_i}x<\psi_1(n_i).$$ By the definition of $\underline{v}_1$, for any $\varepsilon>0$, there is an integer $i_0$ such that $$T_\beta^{n_i}x<\psi_1(n_i)<\beta^{-n_i(\underline{v}_1-\varepsilon)},~\text{for any $i\geq i_0$}.$$ Thus, $\nu_\beta(x)\geq\underline{v}_1-\varepsilon$. Therefore, $$\mathcal{L}(\psi_1)\subseteq\{x\in[0,1]:\nu_\beta(x)\geq \underline{v}_1-\varepsilon\}.$$ By the arbitrariness of $\varepsilon$, one can obtain $$\mathcal{L}(\psi_1)\subseteq\cap_{\varepsilon>0}\{x\in[0,1]:\nu_\beta(x)\geq\underline{v}_1-\varepsilon\}=\{x\in[0,1]:\nu_\beta(x)\geq\underline{v}_1\}.$$ By similar discussion, $\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$ Thus, $$\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)\geq\underline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$
To prove Theorem \[Special-case\], we characterize the set of all points $x$ with $\nu_\beta(x)=\infty$ and $\hat{\nu}_\beta(x)=\infty$.
\[infinite\] If $\nu_\beta(x)=\infty$ and $\hat{\nu}_\beta(x)=\infty$, then one has $$(1)~\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}\subseteq\{x\in[0,1]:\nu_\beta(x)=\infty\};$$ $$(2)~\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}=\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}.$$
By Fact \[fact\], the statement (1) and the inclusion $$\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}\subseteq\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}$$ are obvious. What is left is to show $$\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}\subseteq\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}.$$ By contrary, for any $x$ with $\hat{\nu}_\beta(x)=\infty$, we suppose $$x\notin\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}.$$ Then $T^n_\beta x>0$ for every $n\in\mathbb{N}$. Denote the $\beta$-expansion of $x$ by $$x=\dfrac{a_1}{\beta}+\dfrac{a_2}{\beta^2}+\cdots+\dfrac{a_n}{\beta^n}+\cdots,$$ where $a_i\in\{0,\cdots,\lceil\beta\rfloor\}$, for all $i\geq1$. We can take two increasing sequences $\left\{n'_i:i\geq 1\right\}$ and $\left\{m'_i:i\geq1\right\}$ with the following properties:
(1) For every $i\geq1$, one has $$a_{n'_i}>0,\quad a_{n'_i+1}=\cdots=a_{m'_i-1}=0,\quad a_{m'_i}>0.$$
(2) For every $a_n=0$, there is an integer $i$ such that $n'_i<n<m'_i$.
By the choices of $\left\{n'_i: i\geq 1\right\}$ and $\left\{m'_i:i\geq1\right\}$, for every $i\geq1$, one has $n'_i<m'_i<n'_{i+1}$. Since $\hat{\nu}_\beta(x)>0$, one has $$\limsup_{i\rightarrow\infty}(m'_i-n'_i)=\infty.$$ Take $n_1=n'_1$ and $m_1=m'_1$. Suppose that $m_k,~n_k$ have been defined. Let $i_1=1$ and $i_{k+1}:=\min\{i>i_k:m'_i-n'_i> m_k-n_k\},~{\rm for }~k\geq1.$ Then, define $$n_{k+1}:=n'_{i_{k+1}},\quad m_{k+1}:=m'_{i_{k+1}}.$$ Note that $\limsup_{i\rightarrow\infty}(m'_i-n'_i)=\infty$, the sequence $\{i_k:k\geq1\}$ is well defined. By this way, we obtain the subsequences $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$ of $\left\{n'_i:i\geq 1\right\}$ and $\left\{m'_i:i\geq1\right\}$, respectively, such that the sequence $\{m_k-n_k: k\geq1\}$ is non-decreasing. As the similar discussion in [@YLiao2016], one has $$\hat{\nu}_\beta(x)=\liminf_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}\leq1.$$ This contradicts our assumption $\hat{\nu}_\beta(x)=\infty$. Thus, we have proved $$\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}\subseteq\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}.$$ Therefore, $$\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}=\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\},$$ which implies that the set $\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}$ is countable.
\[empty\] The set $\{x\in[0,1]:1<\hat{\nu}_\beta(x)<\infty\}$ is empty.
This follows from the proof of Item $(2)$ of Lemma \[infinite\].
\[uncountable\] The sets $$\{x\in[0,1]:\nu_\beta(x)=\infty\},\quad \{x\in[0,1]:\hat{\nu}_\beta(x)=1\}$$ are uncountable.
For any real number $a>1$, we give a correspondence: $$\Psi(a)\mapsto 10^{\left\lfloor 2^a\right\rfloor}10^{\left\lfloor 2^{a^{2^2}}\right\rfloor}10^{\left\lfloor2^{a^{3^2}}\right\rfloor}1\cdots10^{\left\lfloor2^{a^{k^2}}\right\rfloor}1\cdots.$$ The infinite string is a $\beta$-expansion of some $x\in[0,1)$. Denote this $x$ by $x_a$. Then, we can obtain a correspondence: $$\Phi(a)\mapsto x_a$$ from $\{a: a>1\}$ to $\{x_a\}_{a>1}$. One can check $\nu_\beta(x_a)=\infty$ and $\hat{\nu}_\beta(x_a)=1$. Therefore, $$\{x_a \}_{a>1}\subseteq\{x\in[0,1]:\nu_\beta(x)=\infty\},\quad\{x_a\}_{a>1}\subseteq\{x\in[0,1]:\hat{\nu}_\beta(x)=1\}.$$
For different $a_1>1$ and $a_2>1$, there is a $k_0\in\mathbb{N}^+$ such that $$\left|2^{a_1^{k^2}}-2^{a_2^{k^2}}\right|>1, \text{ for any }k\geq k_0.$$ Thus, $\Psi(a_1)\neq\Psi(a_2)$. Then, $\Phi(a_1)\neq\Phi(a_2)$. Hence, the cardinality of $\{a: a>1\}$ is less than or equal to that of $\{x_a\}_{a>1}$. Similarly, the cardinality of $\{x_a: a>1\}$ is less than or equal to that of $\{x\in[0,1]:\nu_\beta(x)=\infty\}$ ($\{x\in[0,1]:\hat{\nu}_\beta(x)=1\}$). Since $\{a: a>1\}$ is uncountable, $\{x\in[0,1]:\nu_\beta(x)=\infty\}$ and $\{x\in[0,1]:\hat{\nu}_\beta(x)=1\}$ are uncountable.
$(1)$ If $\underline{v}_1=\overline{v}_1=\underline{v}_2=\overline{v}_2=0$, then $$\lim_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_1(n)}{n}=0,\qquad\lim_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n)}{n}=0.$$ Thus, for any positive integer $m$ large enough, there is an integer $n_0>0$ such that $$\beta^{-n/m}\leq\psi_1(n),\quad \beta^{-n/m}\leq\psi_2(n),\quad \text{for any $n\geq n_0$}.$$ Therefore, $$\{x\in[0,1]:\nu_\beta(x)\geq1/m\}\subseteq\mathcal{L}(\psi_1),\quad\{x\in[0,1]:\hat{\nu}_\beta(x)=1/m\}\subseteq\mathcal{U}(\psi_2).$$ By the fact $\{x\in[0,1]:\hat{\nu}_\beta(x)=1/m\}\subseteq\{x\in[0,1]:\nu_\beta(x)\geq1/m\}$, one has $$\{x\in[0,1]:\hat{\nu}_\beta(x)=1/m\}\subseteq\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2).$$ According to [@YLiao2016 Theorem 1.5], we have $$1=\sup_{m\in\mathbb{N}^+}{\rm dim}_H\left(\{x\in[0,1]:\hat{\nu}_\beta(x)=1/m\}\right)\leq{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right).$$ Thus, ${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=1.$
$(2)$ If $\underline{v}_2=\infty$, then $$\lim_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n)}{n}=\infty.$$ For any $L>0$ large enough, there is an integer $n_0$ such that $$\psi_2(n)\leq\beta^{-nL},~\text{ for any $n\geq n_0$}.$$ Therefore, for any $x\in\mathcal{U}(\psi_2)$, one has $\hat{\nu}_\beta(x)\geq L$. By the arbitrariness of $L$, $\hat{\nu}_\beta(x)=\infty$. Thus, $\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}$. By Lemma \[infinite\] (2), $$\mathcal{U}(\psi_2)=\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}=\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}.$$ According to Lemma \[infinite\] (1), $$\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)=\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}.$$ Since $\{x\in[0,1]:\hat{\nu}_\beta(x)=\infty\}$ is countable, $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable.
$(3)$ If $\underline{v}_1=\infty$, by [**Theorem SW**]{}, $${\rm dim}_H(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2))\leq{\rm dim}_H(\mathcal{L}(\psi_1))=0.$$
Now, we prove Theorem \[Nonspecial\]. We divide the proof into three propositions as Propositions \[Thm3.1\], \[Thm3.2\], \[Thm3.3\].
\[Thm3.1\] For any $0\leq\underline{v}_2\leq\overline{v}_2<\infty$, one has $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{1+\overline{v}_2}.$$
By the definition of $\overline{v}_2$, take a subsequence $\{n_k:k\geq1\}$ such that $$\lim_{k\rightarrow\infty}\dfrac{-\log_\beta\psi_2(n_k)}{n_k}=\overline{v}_2.$$ Then, for any $\varepsilon>0$, there is an integer $k_0$ such that $$\beta^{-n_k(\overline{v}_2+\varepsilon)}\leq\psi_2(n_k)\leq\beta^{-n_k(\overline{v}_2-\varepsilon)},~\text{for any $k\geq k_0$}.$$ For every $x\in\mathcal{U}(\psi_2)$, by the same argument as Proposition \[Pro\], we have $$\mathcal{U}(\psi_2)\subseteq\left\{x\in[0,1]:T_\beta^n x <\beta^{-n(\overline{v}_2-\varepsilon)},\text{ for infinitely many $n\in\mathbb{N}$}\right\}.$$ By [@SW2013 Theorem 1.1], ${\rm dim}_H\left(\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{1+\overline{v}_2}$. Then, $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq{\rm dim}_H\left(\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{1+\overline{v}_2}.$$
\[Thm3.2\] If $\underline{v}_2>1$, then the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable. If $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2\leq1$, then $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$ If $\underline{v}_2\leq \underline{v}_1/(2+\underline{v}_1)$, then $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq \dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
By Lemma \[SET\] $(2)$, $$\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)\geq\underline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$ The argument on the upper bound of the Hausdorff dimension of the set $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ can be obtained by a natural covering of the set $$\{x\in[0,1]:\nu_\beta(x)\geq \underline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$ According to Theorem \[Special-case\] $(2)$ and $(3)$, we only need to consider the case $\nu_\beta(x)\in[\underline{v}_1,\infty)$ and $\hat{\nu}_\beta(x)\in[\underline{v}_2,\infty)$. For any $x\in\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$, there is a number $v_\beta\in[\underline{v}_1,\infty)$ such that $$x\in\mathbb{B}:=\{x\in[0,1]:\nu_\beta(x)=v_\beta\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$ Denote its $\beta$-expansion by $$x=\dfrac{a_1}{\beta}+\dfrac{a_2}{\beta^2}+\cdots+\dfrac{a_n}{\beta^n}+\cdots,$$ where $a_i\in\{0,\cdots,\lceil\beta\rfloor\}$, for all $i\geq1$. Since $\nu_\beta(x)<\infty$, $T_\beta^nx>0$ for all $n\geq0$. By the same way as Lemma \[infinite\], we take the maximal subsequences $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$ of $\left\{n'_i:i\geq 1\right\}$ and $\left\{m'_i:i\geq1\right\}$, respectively. Notice $\beta^{n_k-m_k}<T^{n_k}_\beta x<\beta^{n_k-m_k+1}$. We have the following claim.
$v_\beta=\limsup\limits_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_k},\quad \underline{v}_2\leq\liminf\limits_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}.$
Without loss of generality, we assume $$\limsup\limits_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_k}=c_1,\quad\liminf\limits_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}=c_2.$$ First, we show $v_\beta=c_1$. For any $\varepsilon>0$, there is an integer $k_0>0$ such that $$m_k-n_k\leq n_k(c_1+\varepsilon),\quad\text{for any $k\geq k_0$}.$$ Since $T^{n_k}_\beta x>\beta^{n_k-m_k}$, we have $T^{n_k}_\beta x>\beta^{n_k-m_k}\geq\beta^{-n_k(c_1+\varepsilon)}$. In general, for any $n\geq n_{k_0}$, there is an integer $k\geq k_0$ such that $n_k\leq n<n_{k+1}$. By the choice of $\{n_k\}$, we have $$T^n_\beta x>T^{n_k}_\beta x>\beta^{-n_k(c_1+\varepsilon)}>\beta^{-n(c_1+\varepsilon)}.$$ It means $v_\beta=v_\beta(x)<c_1+\varepsilon$. On the other hand, by the definition of $c_1$, taking subsequence $\{n_{k_i}\}$ and $\{m_{k_i}\}$ such that $$\dfrac{m_{k_i}-n_{k_i}}{n_{k_i}}\geq c_1-\varepsilon,$$ one has $T^{n_{k_i}}_\beta x<\beta^{n_{k_i}-m_{k_i}+1}\leq\beta^{-n_{k_i}(c_1-\varepsilon)+1}$. Thus, $v_\beta=\nu_\beta(x)\geq c_1-\varepsilon$. By the arbitrariness of $\varepsilon$, $v_\beta=c_1$.
Next, we will prove $\underline{v}_2\leq c_2$. By the definition of $\underline{v}_2$, for any $\varepsilon>0$, there is an integer $n_0=n_0(\varepsilon)>0$ such that $$\psi_2(n)\leq\beta^{-n(\underline{v}_2-\varepsilon)},\quad\text{for any $n\geq n_0$}.$$ By the definition of $c_2$, one can take a subsequence $\{k_i:i\geq 1\}$ such that $$\lim_{i\rightarrow\infty}\dfrac{m_{k_i}-n_{k_i}}{n_{k_{i+1}}}=c_2.$$ For the above $\varepsilon>0$, there is an integer $i_0=i_0(\varepsilon)>0$ such that $$m_{k_i}-n_{k_i}\leq n_{k_{i+1}}(c_2+\varepsilon),\quad\text{for any $i\geq i_0$}.$$ By contrary, suppose $c_2<\underline{v}_2$. Then, for any $\varepsilon\in(0,(\underline{v}_2-c_2)/4)$ and any integer $J\geq K:=\max\{n_0(\varepsilon),~n_{i_0(\varepsilon)}\}$, there is an integer $n_{k_{i+1}}>J$ such that for any integer $n\in[1,n_{k_{i+1}}]$, one has $$T^n_\beta x>T^{n_{k_i}}_\beta x>\beta^{n_{k_i}-m_{k_i}}\geq\beta^{-n_{k_{i+1}}(c_2+\varepsilon)}>\beta^{-n_{k_{i+1}}(\underline{v}_2-\varepsilon)}\geq\psi_2(n_{k_{i+1}}).$$ This contracts the fact $x\in\mathcal{U}(\psi_2)$. Therefore, $$\underline{v}_2\leq c_2=\liminf\limits_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}.$$
Now, we consider $$\label{E2}
v_\beta=\limsup_{n\rightarrow\infty}\dfrac{m_k-n_k}{n_k}=\limsup_{n\rightarrow\infty}\dfrac{m_k}{n_k}-1,$$
$$\label{E3}
\underline{v}_2\leq\liminf_{n\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}\leq\liminf_{n\rightarrow\infty}\dfrac{m_k-n_k}{m_k}=1-\limsup_{n\rightarrow\infty}\dfrac{n_k}{m_k}.$$
Since $\left(\limsup\dfrac{n_k}{m_k}\right)\cdot\left(\limsup\dfrac{m_k}{n_k}\right)\geq 1$, one has $$\label{in1}
v_\beta\geq\dfrac{\underline{v}_2}{1-\underline{v}_2},\quad \underline{v}_2\leq\dfrac{v_\beta}{1+v_\beta}.$$
If $\underline{v}_2>1$, then $\underline{v}_2>v_\beta/(1+v_\beta)$, for any $v_\beta\geq v_1$. This contradicts (\[in1\]). Thus, $\mathbb{B}$ is empty. By Lemma \[infinite\], $$\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)=\cup_{n=1}^{\infty}\cup_{\omega\in\Sigma^n_\beta}\{x\in[0,1]:d_\beta(x)=(\omega,0^\infty)\}.$$ Hence, $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable.
If $\underline{v}_2\leq1$, by the inequality (\[in1\]), for any $v_\beta<\underline{v}_2/(1-\underline{v}_2)$, the set $\mathbb{B}$ is empty. Therefore, we consider the case $v_\beta\geq\underline{v}_2/(1-\underline{v}_2)$. Take a subsequence $\{k_i:i\geq 1\}$ such that the supremum of (\[E2\]) is obtained. For abbreviation, we continue to write $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$ for the subsequence $\{n_{k_i}:i\geq 1\}$ and $\{m_{k_i}:i\geq1\}$, respectively. Given $0<\varepsilon<\underline{v}_2/2$, for $k$ large enough, one has $$\label{E4}
(v_\beta-\varepsilon)n_k\leq m_k-n_k\leq(v_\beta+\varepsilon)n_k,$$ $$\label{E5}
m_k-n_k\geq(\underline{v}_2-\varepsilon)n_{k+1}.$$ By inequality (\[E4\]), one has $$(1+v_\beta-\varepsilon)m_{k-1}\leq(1+v_\beta-\varepsilon)n_k\leq m_k.$$ Therefore, the sequence $\{m_k:k\geq1\}$ increases at least exponentially. Since $n_k\geq m_{k-1}$ for every $k\geq2$, the sequence $\{n_k:k\geq1\}$ also increases at least exponentially. Thus, there is a positive constant $C$ such that $k\leq C\log_\beta n_k$. Combining (\[E4\]) and (\[E5\]), one obtains $$(\underline{v}_2-\varepsilon)n_{k+1}\leq(v_\beta+\varepsilon)n_k.$$ Thus, for $k$ large enough, there is an integer $n_0$ and a postive real number $\varepsilon_1$ small enough such that the sum of all lengths of the blocks of $0$ in the prefix of length $n_k$ of the infinite sequence $a_1a_2\cdots$ is at least equal to $$\begin{aligned}
\sum^k_{i=1}(\underline{v}_2-\varepsilon)n_i-n_0&=&n_k(\underline{v}_2-\varepsilon)\left\lgroup \dfrac{1-\left(\dfrac{\underline{v}_2-\varepsilon}{v_\beta+\varepsilon}\right)^k}{1-\left(\dfrac{\underline{v}_2-\varepsilon}{v_\beta+\varepsilon}\right)}\right\rgroup-n_0\\ &\geq& n_k\left(\dfrac{v_\beta \cdot \underline{v}_2}{v_\beta-\underline{v}_2}-\varepsilon_1\right).\end{aligned}$$ Among the digits $a_1\cdots a_{m_k}$, there are $k$ blocks of digits which are free. Denote their lengths by $l_1,\cdots,l_k$. Let $\varepsilon_2=\dfrac{(v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2)\varepsilon_1}{v_\beta-\underline{v}_2}$, one has $$\sum_{i=1}^kl_i\leq n_k-n_k\left(\dfrac{v_\beta\cdot \underline{v}_2}{v_\beta-\underline{v}_2}-\varepsilon_1\right)= n_k(1+\varepsilon_2)\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot\underline{v}_2}{v_\beta-\underline{v}_2}.$$ By Theorem \[cardinality\], there are at most $\beta\cdot\beta^{l_i}/(\beta-1)$ ways to choose the block with length $l_i$. Thus, one has in total at most $$\left(\dfrac{\beta}{\beta-1}\right)^k\cdot\beta^{\sum_{i=1}^kl_i}\leq \left(\dfrac{\beta}{\beta-1}\right)^k\cdot\beta^{n_k(1+\varepsilon_2)(v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2)/(v_\beta-\underline{v}_2)}$$ possible choices of the digits $a_1\cdots a_{m_k}$. On the other hand, there are at most $k(k\leq C\log_\beta n_k)$ blocks of $0$ in the prefix of length $n_k$ of the infinite sequence $a_1a_2\cdots$. Since there are at most $n_k$ possible choices for their first index, one has in total at most $(n_k)^{C\log_\beta n_k}$ possible choices. Consequently, the set of those $x\in\mathbb{B}$ is covered by $$\left(\dfrac{\beta n_k}{\beta-1}\right)^{C\log_\beta n_k}\cdot\beta^{n_k(1+\varepsilon_2)(v_\beta-\underline{v}_2-v_\beta\cdot\underline{v}_2)/(v_\beta-\underline{v}_2)}$$ basic intervals of length at most $\beta^{-m_k}$. Moreover, by (\[E4\]) and by letting $\varepsilon_3=\varepsilon/(1+v_\beta)$, we have $\beta^{-m_k}\leq \beta^{-(1+v_\beta)(1-\varepsilon_3)n_k}$. Set $\varepsilon'=\max\{\varepsilon_2,~\varepsilon_3\}$. The set of those $x$ is covered by $$\left(\dfrac{\beta n_k}{\beta-1}\right)^{C\log_\beta n_k}\cdot\beta^{n_k(1+\varepsilon')(v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2)/(v_\beta-\underline{v}_2)}$$ basic intervals of length at most $\beta^{-(1+v_\beta)(1-\varepsilon')n_k}$. We consider the series $$\sum_{N\geq1}(N)^{C\log_\beta N}\beta^{N(1+\varepsilon')(v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2)/(v_\beta-\underline{v}_2)}\beta^{-(1+v_\beta)(1-\varepsilon')Ns}.$$ The critical exponent $s_0$ such that the series converges if $s>s_0$ and diverges if $s<s_0$ is given by $$s_0=\dfrac{1+\varepsilon'}{1-\varepsilon'}\cdot\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot\underline{v}_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}.$$ By a standard covering argument and the arbitrariness of $\varepsilon'$, the Hausdorff dimension of $\mathbb{B}':=\{x\in[0,1]: \nu_\beta(x)=\nu_\beta\}\cap\mathcal{U}(\psi_2)$ is at most equal to $${\rm dim}_H \left(\mathbb{B}'\right)\leq\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot\underline{v}_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}.$$
For $v_\beta\geq\underline{v}_2/(1-\underline{v}_2)$, fix $L$ large enough. We consider the set $$\mathbb{D}:=\{x\in[0,1]:v_\beta\leq \nu_\beta(x)<v_\beta+1/L\}\cap\mathcal{U}(\psi_2).$$ Repeat the above discussion, if $\underline{v}_2<1$, then $${\rm dim}_H \left(\mathbb{D}\right)\leq\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot v_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}+\dfrac{\underline{v}_2^2/L}{1-\underline{v}_2}.$$
If $\underline{v}_2=1$, then $v_\beta=+\infty$. By [**Theorem SW**]{}, ${\rm dim}_H \left(\mathbb{D}\right)=0$. If $\underline{v}_2<1$, since $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is a subset of $$\cup_{N=0}^{+\infty}\cup^L_{i=1}\left\{x\in[0,1]:v_1+N+(i-1)/L\leq\nu_\beta(x)<v_1+N+i/L\right\}\cap\mathcal{U}(\psi_2),$$ let $L$ tend to $+\infty$, one has $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\sup_{v_\beta\geq \underline{v}_2/(1-\underline{v}_2)}\left\{\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}\right\}.$$ Regard the right side as a function of $v_\beta$, if $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2$, then the maximum is attained for $v_\beta=2\underline{v}_2/(1-\underline{v}_2)$. Therefore, $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$ If $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$, then the maximum is attained for $v_\beta=\underline{v}_1$. Thus, $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq \dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
\[Thm3.3\] If $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2\leq1$, then $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$ If $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, then $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
By Lemma \[SET\] $(1)$, $$\{x\in[0,1]:\nu_\beta(x)>\overline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)>\overline{v}_2\}\subseteq\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2).$$ If $\overline{v}_2=1$, then ${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq0$ always holds. If $\overline{v}_2<1$, then we fix $\delta>0$ with $\overline{v}_2+\delta<1$, we consider the lower bound of the Hausdorff dimension of the set $$\mathbb{F}:=\{x\in[0,1]:\nu_\beta(x)=v_\beta+\delta\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\overline{v}_2+\delta\},$$ where $v_\beta\geq\overline{v}_1$ is a real number. By [**Theorem BL**]{}, if $\dfrac{\overline{v}_1+\delta}{\overline{v}_2+\delta}<\dfrac{1}{1-(\overline{v}_2+\delta)}$, then $\mathbb{F}$ is empty. Therefore, we consider the case $\dfrac{\overline{v}_1+\delta}{\overline{v}_2+\delta}\geq\dfrac{1}{1-(\overline{v}_2+\delta)}$. If $\overline{v}_2>0$, then there is $\delta_0>0$ such that for any $\delta\in(0,\delta_0]$, one has $$\dfrac{v_\beta+\delta}{\overline{v}_2+\delta}\geq\dfrac{\overline{v}_1+\delta}{\overline{v}_2+\delta}\geq\dfrac{1}{1-(\overline{v}_2+\delta)}>1.$$ For any $\delta\in(0,\delta_0]$, we will construct a Cantor subset $E_\delta$ of $\mathbb{F}$. Let $$n'_k=\left\lfloor\left(\dfrac{v_\beta+\delta}{\overline{v}_2+\delta}\right)^k\right\rfloor,\quad m'_k=\lfloor(1+v_\beta+\delta)n'_k\rfloor,\quad k=1,2,\cdots$$ If $\overline{v}_2=0$, let $$n'_k=k^k,\quad m'_k=\lfloor(1+ v_\beta+\delta)n'_k\rfloor,\quad k=1,2,\cdots$$ Making an adjustment, we can choose two subsequences $\{n_k\}$ and $\{m_k\}$ with $n_k<m_k<n_{k+1}$ for every $k\geq1$ such that $\{m_k-n_k\}$ is a non-decreasing sequence and $$\label{lowle}
\lim_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_k}=v_\beta+\delta,\qquad \lim_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}=\overline{v}_2+\delta.$$
Consider the set of real numbers $x\in[0,1)$ whose $\beta$-expansion $$x=\dfrac{a_1}{\beta}+\dfrac{a_2}{\beta^2}+\cdots+\dfrac{a_n}{\beta^n}+\cdots,$$ satisfies that for all $k\geq1$, $$a_{n_k}=1,~a_{n_k+1}=\cdots=a_{m_k-1}=0,~a_{m_k}=1,$$ $$a_{m_k+(m_k-n_k)}=a_{m_k+2(m_k-n_k)}=\cdots=a_{m_k+t_k(m_k-n_k)}=1,$$ where $t_k$ is the largest integer such that $m_k+t_k(m_k-n_k)<n_{k+1}$. Then, $$t_k\leq\dfrac{n_{k+1}-m_k}{m_k-n_k}\leq\dfrac{2}{\overline{v}_2+\delta},$$ for $k$ large enough. Therefore, the sequence $\{t_k:k\geq1\}$ is bounded. Fix $N$, let $\beta_N$ be the real number defined by the infinite $\beta$-expansion of $1$ as equality (\[ED1\]). We replace the digit $1$ for $a_{n_k},~a_{m_k}$ and $a_{m_k+i(m_k-n_k)}$ for any $1\leq i\leq t_k$ by the block $0^N10^N$. Fill other places by blocks belonging to $\Sigma_{\beta_N}$. Thus, we have constructed the Cantor type subset $E_\delta$. Since $\{t_k\}$ is bound, one has $$\lim_{k\rightarrow\infty}\dfrac{m_k-n_k-1+2N}{n_k+(4k-2)N+\sum_{i=1}^{k-1}2Nt_i}=\lim_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_k}=v_\beta+\delta,$$ $$\lim_{k\rightarrow\infty}\dfrac{m_k-n_k-1+2N}{n_{k+1}+(4k+2)N+\sum_{i=1}^k2Nt_i}=\lim_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_{k+1}}=\overline{v}_2+\delta.$$ According to the construction, the sequence $d_\beta(x)$ is in $\Sigma_{\beta_N}.$
$E_\delta\subseteq \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$.
Given $\varepsilon>0$, by (\[lowle\]), there exists an integer $k_0$ such that $$m_k-n_k\leq(v_\beta+\delta+\varepsilon)n_k,~ m_k-n_k\leq(\overline{v}_2+\delta+\varepsilon)n_{k+1},~\text{for any $k\geq k_0$}.$$ By the definitions of $\overline{v}_1$ and $\overline{v}_2$, there is an integer $n_0$ such that $$\beta^{-n(\overline{v}_1+\delta+\varepsilon)}\leq\psi_1(n),\quad \beta^{-n(\overline{v}_2+\delta+\varepsilon)}\leq\psi_2(n),~\text{for any $n\geq n_0$}.$$
Let $N_0=\max\{n_{k_0},~n_0\}$, for any $x\in E_\delta$ and any $n_k\geq N_0$, one has $$T^{n_k}_\beta x<\beta^{n_k-m_k+1}\leq\beta^{-n_k(v_\beta+\delta+\varepsilon-1/n_k)}\leq\beta^{-n_k(\overline{v}_1+\delta+\varepsilon-1/n_k)}\leq\psi_1(n_k).$$ It means $x\in\mathcal{L}(\psi_1)$. On the other hand, for $N\geq N_0$, there is an integer $i$ such that $n_{k+i}\leq N<n_{k+i+1}$. Therefore, $$T^{n_{k+i}}_\beta x<\beta^{n_{k+i}-m_{k+i}+1}\leq\beta^{-N(\overline{v}_2+\delta+\varepsilon-1/n_{k+i+1})}\leq\psi_2(N).$$ It means $x\in\mathcal{U}(\psi_2)$. Then, $x\in\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$. Therefore, $$E_\delta\subseteq \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2).$$
We distribute the mass uniformly when meet a block in $\Sigma_{\beta_N}$ and keep the mass when go through the positions where the digits are determined by construction of $E_\delta$. The Bernoulli measure $\mu$ on $E_\delta$ is defined as follows.
If $n<n_1$, define $\mu(I_n)=1/\sharp\Sigma^n_{\beta_N}$. If $n_1\leq n\leq m_1+4N$, define $\mu(I_n)=1/\sharp\Sigma^{n_1-1}_{\beta_N}$. If there is an integer $t$ with $0\leq t\leq t_1-1$ such that $$m_1+4N+(t+1)(m_1-n_1)+2Nt<n\leq m_1+4N+(t+1)(m_1-n_1)+2N(t+1),$$ define $$\mu(I_n)=\dfrac{1}{\sharp\Sigma^{n_1-1}_{\beta_N}}\cdot\dfrac{1}{\left(\sharp\Sigma^{m_1-n_1-1}_{\beta_N}\right)^{t+1}}.$$ If there is an integer $t$ with $0\leq t\leq t_1$ such that $$m_1+4N+t(m_1-n_1)+2Nt<n\leq c,$$ where $c:=\min\{n_2+4N+2Nt_1, m_1+4N+(t+1)(m_1-n_1)+2Nt\}$, define $$\mu(I_n)=\dfrac{1}{\sharp\Sigma^{n_1-1}_{\beta_N}}\cdot\dfrac{1}{\left(\sharp\Sigma^{m_1-n_1-1}_{\beta_N}\right)^t}\cdot\dfrac{1}{\sharp\Sigma^{n-(m_1+4N+t(m_1-n_1)+2Nt)}_{\beta_N}}.$$
For $k\geq2$, let $$l_k:=n_k+4(k-1)N+\sum^{k-1}_{i=1}2Nt_i,\quad h_k:=m_k+4kN+\sum^{k-1}_{i=1}2Nt_i,$$ $$p_k:=m_k-n_k-1,\quad q_k:=h_k+t_k(m_k-n_k)+2Nt_k.$$ If $l_k\leq n\leq h_k$, define $$\mu(I_n)=\dfrac{1}{\sharp\Sigma^{n_1-1}_{\beta_N}}\cdot\dfrac{1}{\prod^{k-1}_{i=1}\left(\sharp \Sigma^{p_i}_{\beta_N}\right)^{t_i}\cdot\left(\sharp\Sigma^{l_{i+1}-q_i-1}_{\beta_N}\right)}=\mu(I_{l_k})=\mu(I_{h_k}).$$ If there is an integer $t$ with $0\leq t\leq t_k-1$ such that $$h_k+(t+1)(m_k-n_k)+2Nt<n\leq h_k+(t+1)(m_k-n_k)+2N(t+1),$$ define $$\mu(I_n)=\mu(I_{h_k})\cdot\dfrac{1}{\left(\sharp\Sigma^{p_k}_{\beta_N}\right)^{t+1}}.$$ If there is an integer $t$ with $0\leq t\leq t_k$ such that $$h_k+t(m_k-n_k)+2Nt<n\leq\min\{l_{k+1}, h_k+(t+1)(m_k-n_k)+2Nt\},$$ define $$\mu(I_n)=\mu(I_{h_k})\cdot\dfrac{1}{\left(\sharp\Sigma^{p_k}_{\beta_N}\right)^t}\cdot\dfrac{1}{\sharp\Sigma^{n-(h_k+t(m_k-n_k)+2Nt)}_{\beta_N}}.$$
By the construction and Proposition \[full\], $I_{h_k}$ is full. For calculating the local dimension of $\mu$, we discuss different cases as follows.
[**Case $A$:**]{} If $n=h_k$, then $$\begin{aligned}
\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_{h_k})}{\log_\beta \lvert I_{h_k}\rvert}&=&\liminf_{k\rightarrow\infty}\dfrac{n_1-1+\sum\limits_{i=1}^{k-1}\left(t_ip_i+l_{i+1}-q_i-1\right)}{h_k}\cdot\log_\beta\beta_N\\&=&\liminf_{k\rightarrow\infty}\dfrac{n_1-1+\sum\limits_{i=1}^{k-1}\left(l_{i+1}-h_i-2Nt_i-1\right)}{h_k}\cdot\log_\beta\beta_N.\end{aligned}$$ Recall that $\{t_k:k\geq1\}$ is bounded and $\{m_k:k\geq1\}$ grows exponentially fast in terms of $k$, therefore, $$\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_{h_k})}{\log_\beta \lvert I_{h_k}\rvert}=\liminf_{k\rightarrow\infty}\dfrac{\sum_{i=1}^{k-1}\left(n_{i+1}-m_i\right)}{m_k}\log_\beta\beta_N.$$ By equalities (\[lowle\]), one has $$\lim_{k\rightarrow\infty}\dfrac{m_k}{n_k}=1+v_\beta+\delta, \lim_{k\rightarrow\infty}\dfrac{m_{k+1}}{m_k}=\dfrac{v_\beta+\delta}{\overline{v}_2+\delta},\lim_{k\rightarrow\infty}\dfrac{n_{k+1}}{m_k}=\dfrac{v_\beta+\delta}{(\overline{v}_2+\delta)(1+v_\beta+\delta)}.$$ According to Stolz-Cesàro Theorem, $$\begin{aligned}
\liminf_{k\rightarrow\infty}\dfrac{\sum\limits_{i=1}^{k-1}\left(n_{i+1}-m_i\right)}{m_k}&=&\liminf_{k\rightarrow\infty}\dfrac{n_{k+1}-m_k}{m_{k+1}-m_k}\\&=&\liminf_{k\rightarrow\infty}\dfrac{\dfrac{n_{k+1}}{m_k}-1}{\dfrac{m_{k+1}}{m_k}-1}=\dfrac{v_\beta-\overline{v}_2-(v_\beta+\delta) (\overline{v}_2+\delta)}{(1+v_\beta+\delta)(v_\beta-\overline{v}_2)}.\end{aligned}$$ Thus, $$\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_{h_k})}{\log_\beta \lvert I_{h_k}\rvert}=\dfrac{v_\beta-\overline{v}_2-(v_\beta+\delta) (\overline{v}_2+\delta)}{(1+v_\beta+\delta)(v_\beta-\overline{v}_2)}\cdot\log_\beta\beta_N.$$
[**Case $B$:**]{} For an integer $n$ large enough, if there is $k\geq2$ such that $l_k\leq n\leq h_k$, then $$\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_n)}{\log_\beta\lvert I_n\rvert}\geq \liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_n)}{\log_\beta\lvert I_{h_k}\rvert}=\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_{h_k})}{\log_\beta\lvert I_{h_k}\rvert}.$$
[**Case $C$:**]{} For $n$, if there is an integer $t$ with $0\leq t\leq t_k-1$ such that $$h_k+(t+1)(m_k-n_k)+2Nt<n\leq h_k+(t+1)(m_k-n_k)+2N(t+1),$$ then one has $$\mu(I_n)\leq\mu(I_{h_k})\cdot\beta^{-(t+1)p_k}_N.$$ Since $I_{h_k}$ is full, by Proposition \[fullc\], $\lvert I_n\rvert=\lvert I_{h_k}\rvert\cdot\lvert I_{n-h_k}(\omega')\rvert$, where $\omega'$ is an admissible block in $\Sigma^{n-h_k}_{\beta_N}$. By Lemma \[length\], $$\lvert I_n\rvert\geq\lvert I_{h_k}\rvert\cdot\beta^{-(n-h_k+N)}.$$ Hence, $$\dfrac{-\log_\beta\mu(I_n)}{-\log_\beta\lvert I_n\rvert}\geq\dfrac{-\log_\beta\mu(I_{h_k})+(t+1)p_k\log_\beta\beta_N}{-\log_\beta\lvert I_{h_k}\rvert+((t+1)p_k+N(2t+1))}\geq\dfrac{-\log_\beta\mu(I_{h_k})}{-\log_\beta\lvert I_{h_k}\rvert}\cdot\varphi(N),$$ where $\varphi(N)<1$ and $\varphi(N)$ tends to $1$ as $N$ tends to infinity. If there is an integer $t$ with $0\leq t\leq t_k$ such that $$h_k+t(m_k-n_k)+2Nt<n\leq\min\{l_{k+1}, h_k+(t+1)(m_k-n_k)+2Nt\},$$ then letting $l:=n-(h_k+t(m_k-n_k)+2Nt)$, one has $$\mu(I_n)\leq\mu(I_{h_k})\cdot\beta^{-tp_k-l}_N.$$ Since $I_{h_k}$ is full, by Proposition \[fullc\], $\lvert I_n\rvert=\lvert I_{h_k}\rvert\cdot\lvert I_{n-h_k}(\omega')\rvert$, where $\omega'$ is an admissible block in $\Sigma^{n-h_k}_{\beta_N}$. By Lemma \[length\], $\lvert I_{n-h_k}(\omega')\rvert\geq\beta^{-(n-h_k+N)}.$ Therefore, $$\lvert I_n\rvert\geq\lvert I_{h_k}\rvert\cdot\beta^{-(n-h_k+N)}.$$ Hence, $$\dfrac{-\log_\beta\mu(I_n)}{-\log_\beta\lvert I_n\rvert}\geq\dfrac{-\log_\beta\mu(I_{h_k})+(tp_k+l)\log_\beta\beta_N}{-\log_\beta\lvert I_{h_k}\rvert+(tp_k+l+t+N(2t+1))}\geq\dfrac{-\log_\beta\mu(I_{h_k})}{-\log_\beta\lvert I_{h_k}\rvert}\cdot\varphi(N).$$
Therefore, in all cases, $$\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_n)}{\log_\beta\lvert I_n\rvert}\geq\dfrac{v_\beta-\overline{v}_2-(v_\beta+\delta) (\overline{v}_2+\delta)}{(1+v_\beta+\delta)(v_\beta-\overline{v}_2)}\cdot\log_\beta\beta_N\cdot\varphi(N).$$ Given a point $x\in E_\delta$, let $r$ be a number with $\lvert I_{n+1}(x)\rvert\leq r<\lvert I_n(x)\rvert$. We consider the ball $B(x,r)$. By Lemma \[length\], every $n$-th order basic interval $I_n$ satisfies $\lvert I_n\rvert\geq \beta^{-(n+N)}$. Hence, the ball $B(x,r)$ interests at most $\lfloor 2\beta^N\rfloor+2$ basic intervals of order $n$. On the other hand, $$r\geq\lvert I_{n+1}(x)\rvert\geq\beta^{-(n+1+N)}=\beta^{-(1+N)}\cdot\beta^{-n}\geq\beta^{-(1+N)}\cdot\lvert I_n(x)\rvert.$$ Therefore, $$\liminf_{r\rightarrow0}\dfrac{\log_\beta\mu(B(x,r))}{\log_\beta r}=\liminf_{n\rightarrow\infty}\dfrac{\log_\beta\mu(I_n(x))}{\log_\beta\lvert I_n(x)\rvert}.$$ By the arbitrariness of $\delta\in(0,\delta_0]$, one has $$\liminf_{k\rightarrow\infty}\dfrac{\log_\beta\mu(I_n)}{\log_\beta\lvert I_n\rvert}\geq\dfrac{v_\beta-\overline{v}_2-v_\beta\cdot\overline{v}_2}{(1+v_\beta)(v_\beta-\overline{v}_2)}\cdot\log_\beta\beta_N\cdot\varphi(N).$$ Let $N$ tend to infinity, by Mass Distribution Principle [@FK90 pp.60], one has $$\label{LBE}
{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq \dfrac{v_\beta-\overline{v}_2-v_\beta\overline{v}_2}{(1+v_\beta)(v_\beta-\overline{v}_2)}.$$ Regarding the right side as a function of $v_\beta$ with $v_\beta\geq\underline{v}_2/(1-\underline{v}_2)$, if $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2$, then the maximum is attained for $v_\beta=2\overline{v}_2/(1-\overline{v}_2)$. Therefore, $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$ If $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, then the maximum is attained for $v_\beta=\overline{v}_1$. Thus, $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq \dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
If $\underline{v}_2>1$, by Proposition \[Thm3.2\], $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is countable. If $\underline{v}_1/(2+\underline{v}_1)\leq\underline{v}_2\leq1<\overline{v}_2$, by Proposition \[Thm3.1\], $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{1+\overline{v}_2}.$$ By Proposition \[Thm3.2\] and the definition of Hausdorff dimension, we have $$0\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$
If $\overline{v}_1/(2+\overline{v}_1)<\underline{v}_2\leq\overline{v}_2\leq1$ and $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2$, by Propositions \[Thm3.1\] and \[Thm3.2\], we also have $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ Since $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2$, according to Propositions \[Thm3.3\], $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right).$$ Then, $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~ \left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$
If $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$ and $\overline{v}_1/(2+\overline{v}_1)<\overline{v}_2\leq 1$, by Propositions \[Thm3.1\], \[Thm3.2\] and \[Thm3.3\], $$\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2\leq{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq \dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
If $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2\leq\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, combining Proposition \[Thm3.1\] with Proposition \[Thm3.2\], one has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ According to Proposition \[Thm3.3\], we have $${\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$ Therefore, $$\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}\leq{\rm dim}_H \left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$
If $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$ and $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, by Propositions \[Thm3.1\], \[Thm3.2\] and \[Thm3.3\], $$\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\cdot\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}\leq{\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
Proofs of Theorems \[Uniform\] and \[Belong\] {#sec1}
=============================================
In this section, we will give the proofs of Theorems \[Uniform\] and \[Belong\].
By Lemma \[SET\], one has $$\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\subseteq\{x\in[0,1]:\nu_\beta(x)\geq\underline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$ Replace the role of $\underline{v}_1$ by $v_\beta$, for $L$ large enough, we consider the set $$\mathbb{D}:=\{x\in[0,1]:v_\beta\leq\nu_\beta(x)< v_\beta+1/L\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)\geq\underline{v}_2\}.$$ By the similar discussions in Proposition \[Thm3.2\], if $\underline{v}_2>1$, then $\underline{v}_2\geq (v_\beta+1/L)/(1+v_\beta+1/L)$ for any $v_\beta$. Therefore, $\mathbb{D}$ is empty. Thus, $$\mathcal{U}(\psi_2)=\{x\in[0,1]:\hat{\nu}_\beta(x)=+\infty\}.$$ By Proposition \[Thm3.2\] and Lemma \[infinite\], $\mathcal{U}(\psi_2)$ is countable.
If $\underline{v}_2=1$, then $v_\beta=+\infty$. By [**Theorem SW**]{}, ${\rm dim}_H \left(\mathbb{D}\right)=0$. If $\underline{v}_2<1$, then for $v_\beta\geq\underline{v}_2/(1-\underline{v}_2)$, one has $${\rm dim}_H \left(\mathbb{D}\right)\leq\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot v_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}+\dfrac{\underline{v}_2^2/L}{1-\underline{v}_2}.$$ Since $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$ is a subset of $$\cup_{N=0}^{+\infty}\cup^L_{i=1}\{x\in[0,1]:\underline{v}_1+N+(i-1)/L\leq\nu_\beta(x)<v_1+N+i/L\}\cap\mathcal{U}(\psi_2),$$ let $L$ tend to $+\infty$, one has $${\rm dim}_H\left( \mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\sup_{v_\beta\geq \underline{v}_2/(1-\underline{v}_2)}\left\{\dfrac{v_\beta-\underline{v}_2-v_\beta\cdot \underline{v}_2}{(1+v_\beta)(v_\beta-\underline{v}_2)}\right\}.$$ Regarding the right side as a function of $v_\beta$ with $v_\beta\geq\underline{v}_2/(1-\underline{v}_2)$, we obtain the maximum at $v_\beta=2\underline{v}_2/(1-\underline{v}_2)$. Therefore, $${\rm dim}_H \left(\mathcal{U}(\psi_2)\right)\leq\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$ Combining with Proposition \[Thm3.1\], one has $$0\leq{\rm dim}_H \left(\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$
If $\overline{v}_2\leq1$, then $\underline{v}_2\leq\overline{v}_2\leq1$, one also has $${\rm dim}_H \left(\mathcal{U}(\psi_2)\right)\leq\min\left\{\dfrac{1}{1+\overline{v}_2},~\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2\right\}.$$ To obtain the lower bound of the Hausdor dimension of $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$, we will construct a Cantor type subset $E$ of $\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$. By Lemma \[SET\], $$\{x\in[0,1]:\nu_\beta(x)>\overline{v}_1\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)>\overline{v}_2\}\subseteq\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2).$$ We replace the role of $\overline{v}_1$ by $v_\beta$ with $v_\beta\geq\overline{v}_2/(1-\overline{v}_2)$. Fix $\delta>0$, consider $$\{x\in[0,1]:\nu_\beta(x)= v_\beta+\delta\}\cap\{x\in[0,1]:\hat{\nu}_\beta(x)=\overline{v}_2+\delta\}.$$If $\overline{v}_2=0$, then let $$n'_k=k^k,\quad m'_k=\lfloor(1+ v_\beta+\delta)n'_k\rfloor,\quad {\rm for \ }k=1,2,\cdots$$ If $\overline{v}_2>0$, then, let $$n'_k=\left\lfloor\left(\dfrac{v_\beta+\delta}{\overline{v}_2+\delta}\right)^k\right\rfloor,\quad m'_k=\lfloor(1+v_\beta+\delta)n'_k\rfloor,\quad k=1,2,\cdots$$ Arguing as in the proof of Proposition \[Thm3.3\], one has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq \dfrac{v_\beta-\overline{v}_2-v_\beta\overline{v}_2}{(1+v_\beta)(v_\beta-\overline{v}_2)}.$$ Regarding the right side as a function of $v_\beta$ and taking $v_\beta\geq\overline{v}_2/(1-\overline{v}_2)$ into account, we obtain the maximum at $v_\beta=2\overline{v}_2/(1-\overline{v}_2)$. Then, $${\rm dim}_H\left(\mathcal{U}(\psi_2)\right)\geq\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$
If $\underline{v}_1=\overline{v}_1=0$, then $$\lim_{n\rightarrow\infty}\dfrac{-\log_\beta\psi_1(n)}{n}=0.$$ If $\underline{v}_2>0$, by the definitions of $\underline{v}_2$ and $\underline{v}_1=\overline{v}_1=0$, for any $\varepsilon\in(0,\underline{v}_2/2)$, there is an integer $n_0$ such that for any $n\geq n_0$, one has $$\psi_2(n)\leq \beta^{-n(\underline{v}_2-\varepsilon)}<\beta^{-n\varepsilon}\leq \psi_1(n).$$
For any $x\in\mathcal{U}(\psi_2)$, by the same argument as Proposition \[Pro\], we have $$\mathcal{U}(\psi_2)\subseteq\mathcal{L}(\psi_1).$$
Examples
========
In this section, we will show that the upper and lower bounds of Theorems \[Nonspecial\] and \[Uniform\] can be all reached. Examples \[Exa1\], \[Exa2\] and \[Exa3\] explain that the upper bound estimation $\dfrac{1}{1+\overline{v}_2}$, $\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2$ and $\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}$ are reachable, respectively.
\[Exa1\] Let $\psi_1(n)=1$, for $n=1,2,\cdots$ and $$\psi_2(n)=
\begin{cases}
\beta^{-3n}, &\text{ if $n=k^k$}\\
1, &\text{ if $n\neq k^k$}
\end{cases},\quad k=1,2,\cdots$$ Then, $\underline{v}_1=\overline{v}_1=0$, $\underline{v}_2=0$, $\overline{v}_2=3$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{1}{4}=\dfrac{1}{1+\overline{v}_2}.$$
By Proposition \[Thm3.1\], we only need to show that $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq1/4.$$ Now, we construct a Cantor subset $E$ of $\mathcal{L}(\psi_1)\bigcap\mathcal{U}(\psi_2)$ such that $${\rm dim}_H(E)\geq 1/4.$$ Let $n_k=k^k$ and $m_k=4k^k$, for $k=1,2,\cdots$, we construct a Cantor subset $E$ by the same way as Proposition \[Thm3.3\], then $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq1/4=\dfrac{1}{1+\overline{v}_2}.$$
\[Exa2\] Let $\psi_1(n)=1$, for $n=1,2,\cdots$ and $$\psi_2(n)=
\begin{cases}
\beta^{-2n}, &\text{if $n=4^k$}\\
\beta^{-n/2}, &\text{if $n\neq 4^k$}
\end{cases},\quad k=1,2,\cdots$$ Then, $\underline{v}_1=\overline{v}=0$, $\underline{v}_2=1/2$, $\overline{v}_2=2$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{1}{9}=\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$
Since $\underline{v}_1/(2+\underline{v}_1)<\underline{v}_2$, by Proposition \[Thm3.2\], the proof is completed by showing $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\dfrac{1}{9}=\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$ Let $n_k=4^k$ and $m_k=3\cdot 4^k$, for $k=1,2,\cdots$, we construct a Cantor subset $E$ by the same way as Proposition \[Thm3.3\], then $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\dfrac{1}{9}=\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2.$$
\[Exa3\] For $k=1,2,\cdots$, let $$\psi_1(n)=
\begin{cases}
\beta^{-n/2}, & \text{ if $n=3^k$}\\
\beta^{-n}, & \text{ if $n\neq 3^k$}
\end{cases}, \ \
\psi_2(n)=
\begin{cases}
\beta^{-n/2}, &\text{ if $n=3^k$}\\
\beta^{-n/6,} &\text{ if $n\neq 3^k$}.
\end{cases}$$ Then, $\underline{v}_1=1/2$, $\overline{v}_1=1$, $\underline{v}_2=1/6$, $\overline{v}_2=1/2$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{1}{2}=\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
Since $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$, by Proposition \[Thm3.2\], one has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq \dfrac{\underline{v}_1-\underline{v}_2-v_1\cdot\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}=\dfrac{1}{2}.$$ It suffices to show ${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq1/2$. Let $n_k=4^k$ and $m_k=3\cdot 4^k$, for $k=1,2,\cdots$, we construct a Cantor subset $E$ by the same way as Proposition \[Thm3.3\], then $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\geq\dfrac{1}{2}=\dfrac{\underline{v}_1-\underline{v}_2-\underline{v}_1\underline{v}_2}{(1+\underline{v}_1)(\underline{v}_1-\underline{v}_2)}.$$
Examples \[Exa4\] and \[Exa5\] explain that the lower bound estimation $0$ and $\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2$ are reachable, respectively.
\[Exa4\] Let $\psi_1(n)=1$ for $n=1,2,\cdots$ and $$\psi_2(n)=
\begin{cases}
\beta^{-3n}, &\text{ if $n=4^k$}\\
\beta^{-n}, &\text{ if $n\neq 4^k$}
\end{cases}, \quad k=1,2,\cdots$$ Then, $\underline{v}_1=\overline{v}_1=0$, $\underline{v}_2=1$, $\overline{v}_2=3$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=0.$$
It remains to prove ${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq0$. In fact, by Proposition \[Thm3.2\], one has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\left(\dfrac{1-\underline{v}_2}{1+\underline{v}_2}\right)^2=\left(\dfrac{1-1}{1+1}\right)^2=0.$$
\[Exa5\]For $k=1,2,\cdots$, let $$\psi_1(n)=
\begin{cases}
\beta^{-3n}, & \text{ if $n=2k+1$}\\
\beta^{-10n/3}, & \text{ if $n=2k$}
\end{cases}, \ \
\psi_2(n)=
\begin{cases}
\beta^{-21n/32}, &\text{ if $n=2k+1$}\\
\beta^{-2n/3}, &\text{ if $n=2k$}.
\end{cases}$$ Then, $\underline{v}_1=3$, $\overline{v}_1=10/3$, $\underline{v}_2=21/32$ and $\overline{v}_2=2/3$. One has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{1}{25}=\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$
Since $\underline{v}_1/(2+\underline{v}_1)<\overline{v}_2$, by Proposition \[Thm3.3\], the proof is completed by showing $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{25}=\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$ In fact, since $\underline{v}_2\leq\underline{v}_1/(2+\underline{v}_1)$, by Proposition \[Thm3.2\], we have $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{25}=\left(\dfrac{1-\overline{v}_2}{1+\overline{v}_2}\right)^2.$$
The following two examples explain that the lower bound estimation $\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}$ is reachable. We consider the cases of $\underline{v}_1=\overline{v}_1$ and $\underline{v}_2=\overline{v}_2$, respectively.
\[Exa6\] Let $\psi_1(n)=\beta^{-n}$ for $n=1,2,\cdots$ and $$\psi_2(n)=
\begin{cases}
1, &\text{ if $n=2k+1$}\\
\beta^{-n/4}, &\text{ if $n=2k$}
\end{cases},~k=1,2,\cdots$$ Then, $\underline{v}_1=\overline{v}_1=1$, $\underline{v}_2=0$, $\overline{v}_2=1/4$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{1}{3}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
Since $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, by Proposition \[Thm3.3\], what is left is to show $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{3}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$ For any $x\in[0,1]$, denote its $\beta$-expansion by $$x=\dfrac{a_1}{\beta}+\dfrac{a_2}{\beta^2}+\cdots+\dfrac{a_n}{\beta^n}+\cdots,$$ where $a_i\in\{0,\cdots,\lceil\beta\rfloor\}$, for all $i\geq1$. Let $$a_{n'_i}>0,\quad a_{n'_i+1}=\cdots=a_{m'_i-1}=0,\quad a_{m'_i}>0.$$ If $x\in\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)$, since $v_1>0$, then one has $$\limsup_{i\rightarrow\infty}(m'_i-n'_i)=+\infty.$$ Arguing as in the proof of Proposotion \[Thm3.2\], we take the maximal subsequences $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$ of $\left\{n'_i:i\geq 1\right\}$ and $\left\{m'_i:i\geq1\right\}$, respectively, such that the sequence $\{m_k-n_k: k\geq1\}$ is non-decreasing. Notice $\beta^{n_k-m_k}<T^{n_k}_\beta x<\beta^{n_k-m_k+1}$, one has $$\limsup_{k\rightarrow\infty}\dfrac{m_k-n_k}{n_k}=1.$$ Since $x\in\mathcal{U}(\psi_2)$, there is an integer $k_0$ such that for any $k\geq k_0$, one has $m_k-n_k\geq\lfloor n_{k+1}/4\rfloor$. If not, for any $j\geq1$, there is an integer $k_j$ such that $m_{k_j}-n_{k_j}< n_{k_{j+1}}/4$. Since one of $n_{k_{j+1}}$ and $n_{k_{j+1}}+1$ is even, denote it by $l_{k_{j+1}}$, for any integer $n\in[1,l_{k_{j+1}}]$, one has $$T^n_\beta x>\beta^{n_{k_j}-m_{k_j}}>\beta^{-l_{k_{j+1}}/4}=\psi_2(l_{k_{j+1}}).$$ It contradicts $x\in\mathcal{U}(\psi_2)$.
Choose the subsequence $\{n_{k_i}:i\geq1\}$ and $\{m_{k_i}:i\geq1\}$ of $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$, respectively, such that $$\lim_{i\rightarrow\infty}\dfrac{m_{k_i}-n_{k_i}}{n_{k_i}}=1.$$ For simplicity, let $\{n_k:k\geq1\}$ and $\{m_k:k\geq1\}$ stand for $\{n_{k_i}:i\geq1\}$ and $\{m_{k_i}:i\geq1\}$, respectively. For any $\varepsilon>0$, there is an integer $k'$ such that for any $k\geq k'$, one has $$(1-\varepsilon)n_k\leq m_k-n_k\leq(1+\varepsilon)n_k,\quad m_k-n_k\geq\dfrac{n_{k+1}}{4}-2.$$ Arguing as in the proof of Proposition \[Thm3.2\], one has $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{1}{3}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
\[Exa7\] Let $$\psi_1(n)=
\begin{cases}
\beta^{-n/3}, &\text{ if $n=2k+1$}\\
\beta^{-2n/3}, &\text{if $n=2k$}
\end{cases},~k=1,2,\cdots$$ and $\psi_2(n)=\beta^{-2n/11}$ for $n=1,2,\cdots$. Then, $\underline{v}_1=1/3$, $\overline{v}_1=2/3$, $\underline{v}_2=\overline{v}_2=2/11$ and $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)=\dfrac{9}{20}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
Since $\overline{v}_2\leq\overline{v}_1/(2+\overline{v}_1)$, by Proposition \[Thm3.3\], what is left is to show $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{9}{20}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$ In fact, since $\underline{v}_2=\dfrac{2}{11}\leq\dfrac{1}{7}=\underline{v}_1/(2+\underline{v}_1)$, by Proposition \[Thm3.2\], $${\rm dim}_H\left(\mathcal{L}(\psi_1)\cap\mathcal{U}(\psi_2)\right)\leq\dfrac{9}{20}=\dfrac{\overline{v}_1-\overline{v}_2-\overline{v}_1\overline{v}_2}{(1+\overline{v}_1)(\overline{v}_1-\overline{v}_2)}.$$
[99]{} M. Amou and Y. Bugeaud, Exponents of Diophantine approximation and expansions in integer bases, [*J. Lond. Math. Soc. (2)*]{}, [**81**]{} (2010), no. 2, 297-316.
A. Besicovitch, Sets of Fractional Dimensions (IV): On Rational Approximation to Real Numbers, [*J. London Math. Soc.*]{}, [**9**]{} (1934), no. 2, 126-131.
Y. Bugeaud, Distribution modulo one and Diophantine approximation, Cambridge Tracts in Mathematics, vol. 193, [*Cambridge University Press, Cambridge*]{}, 2012.
Y. Bugeaud and L. Liao, Uniform Diophantine approximation related to $b$-ary and $\beta$-expansions, [*Ergodic Theory Dynam. Systems*]{}, [**36**]{} (2016), no. 1, 1-22.
L. Dirichlet, Verallgemeinerung eines Satzes aus der Lehre von den Kettenbrüchen nebst einigen anwendungen auf die Theorie der Zahlen, [*SB Preuss. Akad. Wiss*]{}, [**1842**]{} (1842), 93-95.
K. Dmitry, K. Ioannis and K. Florian. Diophantine approximation and run-length function on $\beta$-expansions, [*arXiv preprint arXiv:1904.08584*]{}, 2019.
K. Falconer, Fractal geometry, Mathematical foundations and applications, [*John Wiley & Sons, Ltd., Chichester*]{}, 1990.
A. Fan and B. Wang, On the lengths of basic intervals in beta expansions, [*Nonlinearity*]{}, [**25**]{} (2012), no. 5, 1329-1343.
R. Hill and S. Velani, The ergodic theory of shrinking targets, [*Invent. Math.*]{}, [**119**]{} (1995), no. 1, 175-198.
V. Jarník, Diophantische Approximationen und Hausdorffsches Mass, [*Rec. Math. Moscou*]{}, [**36**]{} (1929), 371-382.
A. Khintchine, Einige Sätze über Kettenbrüche, mit Anwendungen auf die Theorie der Diophantischen Approximationen, [*Math. Ann.*]{}, [**92**]{} (1924), no. 1, 115-125.
A. Legendre, Essai sur la théorie des nombres, Cambridge Library Collection, [*Cambridge University Press, Cambridge*]{}, 2009.
B. Li and J. Wu, Beta-expansion and continued fraction expansion over formal Laurent series, [*Finite Fields Appl.*]{}, [**14**]{} (2008), no. 3, 635-647.
W. Parry, On the $\beta$-expansions of real numbers, [*Acta Math. Acad. Sci. Hungar.*]{}, [**11**]{} (1960), 401–416.
W. Philipp, Some metrical theorems in number theory, [*Pacific J. Math.*]{}, [**20**]{} (1967), 109-127.
A. Rényi, Representations for real numbers and their ergodic properties, [*Acta Math. Acad. Sci. Hungar.*]{}, [**8**]{} (19570), 477-493.
L. Shen and B. Wang, Shrinking target problems for beta-dynamical system, [*Sci. China Math.*]{}, [**56**]{} (2013), no. 1, 91-104.
M. Waldschmidt, Recent advances in Diophantine approximation, Number theory, analysis and geometry [*Springer, New York*]{}, 2012.
L. Zheng, Diophantine approximation and run-length function on $\beta$-expansions, [*Journal of Number Theory*]{}, [**202**]{} (2019), 60-90.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[**Z. Néda** ]{}\
[ *Babes-Bólyai University, Dept. of Physics*]{}\
[ *str. Kogalniceanu 1, RO-3400 Cluj, Romania* ]{}\
- |
[*and* ]{}\
[ *University of Bergen, Dept. of Physics* ]{}\
[ *Allégaten 55, N-5007 Bergen, Norway*]{}
title: |
**CURIE TEMPERATURES\
FOR BINARY ISING FERROMAGNETS\
ON THE SQUARE LATTICE**
---
TEX, C Version 3.14t3
Abstract\
High-accuracy Swendsen and Wang Monte Carlo simulations were performed to study the Curie temperature of ferromagnetic, binary Ising systems on the square lattice. Our results are compared with mean-field like approaches. Based on these former theories, we give a new formula to estimate the Curie temperature of the system.
PACS number(s): 75.10H; 75.20E; 75.40M
§1. Introduction {#introduction .unnumbered}
================
Binary Ising systems were studied from both bond and site perspectives (Katsura and Matsubara 1974; Thorpe and McGurn 1978). In the bond-disordered model the lattice sites are considered to be equivalent and the interaction energies between the neighbouring sites are assigned randomly from a set of possible values. In the site-disordered model the lattice sites are randomly occupied by two different types of magnetic ions, $A$ and $B$, with spins $S_A$ and $S_B$, the interaction parameters between neighbouring spins being completly determined by their species. In this way there exist three type of exchange interactions $J_{AA}$, $J_{AB}$ and $J_{BB}$, between neighbouring spins in the system. The disorder can be considered either quenched or annealed. The annealed systems are much more easier handled by mean-field like approaches, and so there are better understood than the quenched ones. In spite of this, for practical applications the quenched systems are much more appropriate. This is the main reason why we proposed to limit the discussion only for the case of quenched systems.
In the case of only ferromagnetic interactions between the spins, binary Ising or Heisenberg models were used with succes to describe the magnetic properties of magnetic alloys (Vonsovski 1974; Luborsky 1980). When antiferromagnetic and ferromagnetic interactions compete, frustration appears, and the system becomes a Mattis-Luttinger type spin-glass model ( Binder and Young 1986; Tatsumi 1977-78).
For real physical cases the site-disordered models are much more characteristic, and so we proposed to study the Ising version of this model on the square lattice. We also considered the simplest case of $S_A=S_B=\frac{1}{2}$, and all exchange interactions of ferromagnetic type. The hamiltonian of the problem will be: $$H=-\sum_{<i,j>}[J_{AA}\cdot \delta_{iA} \cdot \delta_{jA} + J_{BB} \cdot
\delta_{iB} \cdot \delta_{jB} + J_{AB} \cdot (\delta_{iA} \cdot
\delta_{jB} + \delta_{iB} \cdot \delta_{jA})]\cdot S_i^z \cdot S_j^z,$$ where $\delta_{ix}=1$ if the spin $i$ is of type $x$, and $0$ otherwise. The sum in (1) is refering to all nearest neighbours of the lattice. In this paper we consider the two-dimensional version of the model. For results concerning the three-dimensional, real model we have just preliminary results (Neda 1994\[b\]). The first molecular-field approximations for the proposed systems were due to Vonsovskii (1940 and 1948). The frustrated systems were first studied by Aharony (1975) using renormalization-group technics and by Tatsumi (1977-78) with Monte Carlo simulations. The interesting case for us, with all the interactions of ferromagnetic type, was studied using a mean-field like approach by Kouvel (1969), and with the coherent potential approximation by Foo and Wu (1972). Mean-field theoretical approaches were also made in the works of Tahir-Kheli and Kawasaki (1977) , respective Thorpe and McGurn (1978). Ishikawa and Oguchi (1978) considered a Bethe-Peierls approach, and in the work of Honmura, Khater, Fittipaldi and Kaneyoshi (1982) we find an effective-field theory for the model. Monte Carlo studies for the critical temperature of binary, ferromagnetic Ising alloys in function of the relative species concentration and relative interaction energy between unlike ions were performed by Scholten (1985) on the square lattice. Scholten (1989) also studied the phase diagram for three-dimensional frustrated systems on simple-cubic lattices, including next-nearest neighbour interactions too. The phase diagrams of binary Ising ferromagnets were studied by Thorpe and McGurn (1978), both in the site-disordered and bond-disordered cases. They realized that these phase diagrams can be usefully cataloged in terms of the initial slope $\frac{\partial \ln{T_c}}{\partial q}$ of the transition temperature $T_c$ plotted in function of concentration $q$, at the two points $q=0$ and $q=1$. Using the perturbation theory, they also determined the initial slopes for two-dimensional systems. The phase diagrams for binary Ising systems with randomly distributed exchange parameters were studied by Kaneyoshi and Li (1987) using effective-field theory with correlations. In the book from Vonsovskii (1974) and in the paper from Luborsky (1980), one may find promising comparisions between experimental data and mean-field like predictions. Diluted systems ($J_{AB}=0$ and $J_{BB}=0$) also presented interest for physicists (Wu 1982; Belokon and Semkin 1992; Neda 1994\[a\]). Recently there has been much interest in systems of mixed $S_A$ and $S_B$ spins, where $S_A \neq S_B$ (Kaneyoshi 1989; Silva and Salinas 1991; Kaneyoshi, Jascur and Tomczak 1992; Zhang and Yang 1993).
Although Monte Carlo simulations were performed by Scholten (1985) on the considered model, there remained some not completly clarified questions even in the simplest two-dimensional and ferromagnetic case. The main problems are concerning the values of the critical exponents and the elaboration of a practically usable and general formula to estimate the critical temperature of the system. Our work is intended to complete Scholtens paper in some sense, studiing by a high-accuracy Monte Carlo simulation the Curie temperature of the system. We do this in a review context, comparing our simulation results with available theoretical formulas. In this manner we give a practically usable and easy method of approximating the Curie temperature of the system and illustrate the validity and limitations of different theoretical approaches.
§2. Used theoretical formulas {#used-theoretical-formulas .unnumbered}
=============================
The localized model of ferromgnetism involving nearest-neighbour exchange integrals has an attractive simplicity for describing some magnetic systems. Although this approach for the magnetism in metallic systems is not completly acceptable due to the partially itinerant nature of the magnetic electrons, the obtained results are usually in good agreement with experimental data. In the case of binary magnetic alloys we are in a similar situation. The localized model based on the Heisenberg or Ising hamiltonian (1) with nearest-neighbour exchange, or the molecular-field theories proved to be applicable in describing the variation of the critical temperature in function of the alloys composition.
The first formula based on the molecular-field approximation was derived, as we stated earlier, by Vonsovskii (1940; 1948) and used with success to describe transition temperatures of binary magnetic alloys. The proposed formula was: $$T_c(q)=T_c(A,A)-2 \cdot [T_c(A,A)-T_c(A,B)]\cdot q + [T_c(A,A)+T_c(B,B)-
2 \cdot T_c(A,B)] \cdot q^2 ,$$ where $T_c(A,A)$ and $T_c(B,B)$ are the Curie temperatures of the pure $A$ and $B$ systems, $T_c(A,B)$ is the Curie temperature for a pure system caracterized with all exchange interactions equal with the ones between the $A$ and $B$ magnetic ions ($J_{AB}$), $T_c(q)$ is the Curie temperature of the mixture, and $q$ is the concentration of the $B$ component.
We mention here that the critical temperature $T_c$ for an Ising system on the square lattice, caracterized with $J$ exchange interaction constants (considering just nearest-neighbour interactions) is given by $T_c\approx 2.2681 \cdot \frac{J}{k_B}$, with $k_B$ the Boltzmann constant.
Using a phenomenological model based on mean-field theory suitably modified, so that the individual atomic moments are allowed to vary in magnitude with their local environment, and considering only nearest-neighbour interactions Kouvel (1969) proposed the formula: $$\begin{aligned}
& T_c(q)=\frac{1}{2} \cdot [T_c(A,A) \cdot (1-q)+T_c(B,B) \cdot q] +
\nonumber \\
& + \{ \frac{1}{4}\cdot
[T_c(A,A) \cdot (1-q) - T_c(B,B) \cdot q]^2+ T_c(A,B)^2 \cdot q \cdot
(1-q) \} ^{\frac{1}{2}} .\end{aligned}$$
In the work of Foo and Wu (1972) the disordered composition dependent exchange interaction is treated in a coherent potential approximation (CPA). In the limit of weak scattering their method give the mean-field like results, but in the strong scattering limit they predict such effects as critical concentration for the appearance of ferromagnetism in the diluted models (Neda 1994\[a\]), which is not obtained in mean-field theories. They proposed the following cubic equation for $T_c(q)$ $$\begin{aligned}
& \alpha^2 \cdot T_c(q)^3 + \nonumber \\
& +[\alpha \cdot (T_c(A,A)+T_c(B,B)+T_c(A,B))- \alpha
\cdot (1+\alpha) \cdot <T_c>] \cdot T_c(q)^2+ \nonumber \\
& + [(1+\alpha) \cdot T_c(A,A) \cdot
T_c(B,B) \cdot T_c(A,B) \cdot < \frac{1}{T_c} > - \nonumber \\
& -\alpha \cdot (T_c(A,A) \cdot
T_c(B,B) + T_c(A,B) \cdot T_c(A,A) + T_c(A,B) \cdot T_c(B,B))] \cdot T_c(q) -
\nonumber \\
& -T_c(A,A) \cdot T_c(B,B) \cdot T_c(A,B)=0,\end{aligned}$$ where $$\alpha=\frac{z}{2}-1,$$ with $z$ the coordination number of the lattice (in our case $z=4$), and $$\begin{aligned}
& <T_c>=(1-q)^2 \cdot T_c(A,A) + 2 \cdot q \cdot (1-q) \cdot T_c(A,B) +
q^2 \cdot T_c(B,B) ,\\
& <\frac{1}{T_c}>= \frac{(1-q)^2}{T_c(A,A)}+ \frac{2\cdot q \cdot
(1-q)}{T_c(A,B)}
+ \frac{q^2}{T_c(B,B)}.\end{aligned}$$
We mention that there are also other, more evoluate possibilities of calculating the Curie temperature, based on the Ising model (1) of the system, such as mean-field like renormalization-group technics, series expansion and perturbation methods. Unfortunately these are all very technical ones, and do not give practically usable formulas.
§3. The computer simulation method {#the-computer-simulation-method .unnumbered}
==================================
The Monte Carlo simulations performed by Scholten on the proposed model were made by using the classical single spin-flip Metropolis algorithm (Metropolis, Rosenbluth and Teller 1953). Due to this, his simulations were very time-consuming , and he studied just a few choices for the values of the interaction parameters ($\frac{J_{AB}}{J_{AA}}=0,1,2,4$ and $\frac{J_{BB}}
{J_{AA}}=4$). He also worked on relatively small $40\times 40$ square lattices with periodic boundary conditions. For each choice of the interaction parameters value he studied three cases for the concentration of the $B$ component $q=0.25$, $q=0.5$ and $q=0.75$ (the cases $q=0$ and $q=1$ are evident). He compared his results with the ones obtained by (Tahir-Kheli [*et al.*]{} 1977), (Thorpe [*et al.*]{} 1978), (Ishikawa [*et al.*]{} 1978) and (Honmura [*et al.*]{} 1982).
We proposed to continue Scholtens work by reconsidering the problem with high-accuracy Monte Carlo simulations, using the more powerful cluster-flip Swendsen and Wang method (Swensen, Wang and Ferrenburg 1992) with an original recursion type algorithm. We considered many choices for the values of the $J_{AB}$ and $J_{BB}$ interaction parameters, the value of $J_{AA}$ being fixed. We proposed to compare our results obtained for the Curie temperature with the ones given by equations (2), (3) and (4).
Our simulations were performed on relatively large lattices, up to $200 \times 200$ lattice sites. The critical temperature was obtained by detecting the maximum in the fluctuation of the absolute value of magnetization. To achieve statistical equillibrium we considered up to $1000$ cluster-flips and then studied the fluctuation for 2000 more iterations. The sensitivity for the determination of the critical temperature was in general of the order of $0.01\cdot T_c(A,A)$. For every chosen set of the interaction parameters we covered the $q\in (0,1)$ concentration interval uniformly with $9$ to $19$ simulation points. The program was written in C and the simulations were performed on a CRAY Y-MP4D/464 computer and IBM R-6000 RISC workstations.
§4. Results {#results .unnumbered}
===========
Our Monte Carlo results concerning the variation of the Curie temperature in function of the B components concentration for the proposed two-dimensional model are plotted with various symbols on figures 1 and 3-7. The curves indicate theoretical results given from equations (2) and (4). In Fig. 1 considering four choices for the $J_{AB}$ interaction parameters ($J_{AA}$ and $J_{BB}$ fixed), we compare our Monte Carlo data with results given by equation (2). In Fig. 2 we show some preliminary results for the three-dimensional (simple-cubic) case, obtained with the same interaction parameters as in Fig. 1, in comparision with the curve given by the (2) molecular-field approximation. As one would expect it, we can also observ that in the real, higher dimensional case the considered molecular-field approximation is workin better. From Fig. 1 we get, that on the square lattice formula (2) predicts much higher results for the Curie temperature than the real values. We checked that equation (3) predict even higher values than (2). In Fig. 3 we show the same Monte Carlo data as in Fig. 1 in comparision with results obtained from equation (4). From Fig. 1 and 2 we conclude that in the considered cases the real critical temperatures are limited by the two curves obtained from equations (2) and (4). In addition to this, in Fig. 4 we show that almost a perfect fit with the realistic Curie temperatures can be obtained, if we use the arithmetic mean of the $T_c(q)$ values obtained from (2) and (4).
In Fig. 5-7 we tried to prove our previous statements. So, we considered other choices for the exchange interaction parameters, and thus for the $T_c(A,A)$, $T_c(B,B)$ and $T_c(A,B)$ critical temperatures. We illustrated with thin dashed lines the results obtained from equation (2) and (4) (dense dashes correspond to the curve calculted from (4)). The continuous darker curve shows the arithmetic mean of the $T_c(q)$ obtained from (2) and (4). We conclude again that in general the values given from equations (2) and (4) limit nicely the realistic simulation data, and their arithmetic mean gives a good estimate for the Curie temperature. This arithmetic mean have stronger differnces with our Monte Carlo data in the case when the $J_{AB}$ exchange interaction parameters does not belong to the interval limited by the $J_{AA}$ and $J_{BB}$ values.
§5. Conclusions {#conclusions .unnumbered}
===============
Our first conclusion is that the Curie temperatures calculated from equations (2), (3) and (4) are not performant approximations of the real values. However as expected, our simulations on simple-cubic lattices reveal that the same (2) molecular-field approach is giving much better results in the real three-dimensional case (Fig. 2).
For the case of the square lattice, generally the curves obtained for the critical temperature from equations (2) and (4) limit rather nicely the real values. Our Monte Carlo simulations indicate, that a few exceptions could be for the small ($q\rightarrow 0$) and big ($q\rightarrow 1$) concentration limit, when the $J_{AB}$ interaction parameter is far from the interval limited by the values of $J_{AA}$ and $J_{BB}$.
Our most important conclusion is that, the theoretical curve constructed as the arithmetic mean of the Curie temperatures obtained from equations (2) and (4) proved to be a good approximation for the critical temperature of a binary Ising ferromagnet on the square lattice.
Similar preliminary results for the three-dimensional case are given in a recent preprint (Neda 1994\[b\]).
Acknowledgements {#acknowledgements .unnumbered}
================
This study was finished during a bursary offered by the Norwegian Research Council. We thank Y. Brechet, A. Coniglio, L. Csernai, and L. Peliti for their continuous help and useful discussions.
References {#references .unnumbered}
==========
[**A**]{}harony, A.H., 1975, [*Phys. Rev. Lett.*]{}, [**34**]{}, 590\
[**B**]{}elokon, V.I., and Semkin, S.V., 1992, [*Sov. Phys. JETP*]{} [**75**]{}, 680\
[**B**]{}inder, K., and Young, A.P., 1986, [*Rev. Mod. Phys.*]{} [**58**]{}, 801\
[**F**]{}oo, E-Ni, and Wu, Der-Hsueh, 1972, [*Phys. Rev. B*]{} [**5**]{}, 98\
[**H**]{}onmura, R., Khater, A.F., Fittipaldi, I.P., and Kaneyoshi, T., 1982, [*Solid State Commun.*]{} [**41**]{}, 385\
[**I**]{}shikawa, T., and Oguchi, T., 1978, [*J. Phys. Soc. Jpn.*]{} [**44**]{}, 1097\
[**K**]{}aneyoshi, T., and Li, Z.Y., 1982, [*Phys. Rev. B*]{} [**35**]{}, 1869\
[**K**]{}aneyoshi, T., 1989, [*Phys. Rev. B*]{} [**39**]{}, 12134\
[**K**]{}aneyoshi, T., Jascur, M., and Tomczak, P., 1992, [*J. Phys. Condens. Matter.*]{} [**4**]{}, L653\
[**K**]{}atsura, S., and Matsubara, F., 1974, [*Can. J. Phys.*]{} [**52**]{}, 120\
[**K**]{}ouvel, J.S., 1969, in [*Magnetism and Metallurgy vol. II.*]{}, eds. A.E. Berkowitz and E. Kneller (Academic Press)\
[**L**]{}uborsky, F.E., 1980, [*J. Appl. Phys.*]{} [**51**]{}, 2808\
[**M**]{}etropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., and Teller, E., 1953, [*J. Chem. Phys.*]{} [**21**]{}, 1087\
[**N**]{}éda, Z., 1994, [*J. Phys. I. (Paris)*]{} (to be published in february issue )\
[**N**]{}éda, Z., 1994, [*Scientific Report* ]{} [**02/1994**]{} (Univ. of Bergen, Norway)\
[**S**]{}cholten, P.D., 1985, [*Phys. Rev. B*]{} [**32**]{}, 345\
[**S**]{}cholten, P.D., 1989, [*Phys. Rev. B*]{} [**40**]{}, 4981\
[**S**]{}ilva, N.R. da, and Salinas, S.R., 1991, [*Phys. Rev. B*]{} [**44**]{}, 852\
[**S**]{}wendsen, R.H., Wang, J.S., and Ferrenburg, A.M., 1992, in [*The Monte-Carlo Method in Condensed Matter Physics*]{}, ed. K. Binder (Springer-Verlag)\
[**T**]{}ahir-Kheli, A., and Kawasaki, T., 1977, [*J. Phys. C*]{} [**10**]{}, 2207\
[**T**]{}atsumi, T., 1977, [*Prog. Theor. Phys.*]{} [**57**]{}, 1799\
[**T**]{}atsumi, T., 1978, [*Prog. Theor. Phys.*]{} [**59**]{}, 1428; [**59**]{} 1437\
[**T**]{}horpe, M.F., and McGurn, A.R., 1978, [*Phys. Rev. B*]{} [**20**]{}, 2142
$\:$\
[**V**]{}onsovskii, S.V., 1940, [*Dokl. Akad. Nauk. SSSR*]{} [**26**]{}, 364\
[**V**]{}onsovskii, S.V., 1948, [*Zhurn. Tekh. Fiz.*]{} [**18**]{}, 131\
[**V**]{}onsovskii, S.V., 1974, [*Magnetism II.*]{}, (John Willey) pp. 776-822\
[**W**]{}u, F.Y., 1982, [*Rev. Mod. Phys.*]{} [**54**]{}, 235\
[**Z**]{}hang, G.M., and Yang, C.Z., 1993, [*Z. Phys. B*]{} [**91**]{}, 145\
Figure Captions {#figure-captions .unnumbered}
===============
$\:$\
[**Fig. 1**]{} Monte Carlo results for the variation of the Curie temperature as a function of the $B$ components concentration for four choices of the $T_c(A,B)$ critical temperature. Solid curve is given by equation (2).
[**Fig. 2**]{} The same plot as in Fig. 1 for simulations done on the simple-cubic lattice.
[**Fig. 3**]{} The Monte Carlo results from Fig.1 in comparision with the Curie temperatures obtained from (4).
[**Fig. 4**]{} The Monte Carlo results from Fig. 1 in comparision with the arithmetic mean of the Curie temperatures obtained from (2) and (4).
[**Fig. 5**]{} The dots and triangles represents Monte Carlo simulations for the given $T_c(A,B)$ critical temperatures. The thin dashed lines indicate the results obtained from formulas (2) and (4) (dense dashes correspond to (4)). The dark continuous line indicate the arithmetic mean obtained from (2) and (4).
[**Fig. 6**]{} The case when we have no exchange interactions between the atoms of the $B$ component ($J_{BB}=0$) and $T_c(A,A)=T_c(A,B)=100$. Dots are Monte-Carlo results and the curves have the same meaning as in Fig. 4.
[**Fig. 7**]{} Monte Carlo results (dots) for $T_c(A,A)=T_c(B,B)=100$ and\
$T_c(A,B)=500$. The curves represents the same as in Fig. 4.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The expected running time of the classical [(1+1) EA]{}on the [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}benchmark function has recently been determined by Hwang et al. (2018) up to additive errors of $O((\log n)/n)$. The same approach proposed there also leads to a full asymptotic expansion with errors of the form $O(n^{-K}\log n)$ for any $K>0$. This precise result is obtained by matched asymptotics with rigorous error analysis (or by solving asymptotically the underlying recurrences via inductive approximation arguments), ideas radically different from well-established techniques for the running time analysis of evolutionary computation such as drift analysis. This paper revisits drift analysis for the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}and obtains that the expected running time ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$, starting from $\cl{n/2}$ one-bits, is determined by the sum of inverse drifts up to logarithmic error terms, more precisely $$\sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} - c_1\log n \le {\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}} \le \sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} - c_2\log n$$ where $\Delta(k)$ is the drift (expected increase of the number of one-bits from the state of $n-k$ ones) and $c_1,c_2 \, >0$ are explicitly computed constants. This improves the previous asymptotic error known for the sum of inverse drifts from $\tilde{O}(n^{2/3})$ to a logarithmic error and gives for the first time a non-asymptotic error bound. Using standard asymptotic techniques, the difference between ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$ and the sum of inverse drifts is found to be $(e/2)\log n+O(1)$.'
author:
- |
Hsien-Kuei Hwang\
Institute of Statistical Science\
Academia Sinica\
Taipei 115\
Taiwan
- |
Carsten Witt\
Technical University of Denmark\
Kgs. Lyngby\
Denmark
title: 'Sharp Bounds on the Runtime of the (1+1) EA via Drift Analysis and Analytic Combinatorial Tools'
---
Introduction
============
The runtime analysis of randomized search heuristics on simple, well-structured benchmark problems has triggered the development of analytical tools for understanding the complexity and considerably contributed to their theoretical foundations. This paper is concerned with the objective function ${{\textsc{OneMax}\xspace}}(x_1,\dots,x_n)=x_1+\dots+x_n$, the arguably most fundamental theoretical benchmark problem in discrete search spaces and the [(1+1) EA]{}, probably the most fundamental search heuristic in the theoretical runtime analysis (see Algorithm \[alg:oneoneea\]).
Already the earliest analysis of the [(1+1) EA]{}[@Muhlenbein92] showed that the [(1+1) EA]{}optimizes [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}in an expected time of $O(n\log n)$, where time corresponds to the number of iterations. The early interest and attempts in obtaining more precise description of the runtime complexity were summarized in Garnier et al.’s fine paper [@Garnier1999] with very strong approximation results claimed. On the other hand, it follows from the analyses in [@djwea02] that the expected time is bounded from above by $en H_n \le en(\log n+1)$, where $H_n=\sum_{j=1}^n 1/j$ denotes the $n$-th harmonic number and $\log n$ the *natural* logarithm. Lower bounds of the kind $\Omega(n\log n)$ that hold for the much larger class of functions with a unique optimum [@djwea02] showed that the results were at least asymptotically tight.
From the beginning of this decade, finer analyses of the expected runtimes have gained increasing attention. Precise expressions for the runtime, dependent not only on the search space dimension $n$ but also on parameters such as the mutation rate, are vital to optimize parameter settings [@WittCPC13] and to compare different algorithms whose runtime only differs in lower-order terms [@DoerrDYGECCO16].
With respect to [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, the first lower bound that explicitly states the leading coefficient $e$ in an expression of the type ${(1-o(1))en\log n}$ was independently derived by Doerr, Fouz and Witt [@DFW10] and Sudholt [@SudholtLowerFitnessLevel] (in the finer form $en\log n-2n\log\log n$) using the techniques of drift analysis and fitness levels, respectively. The lower-order term was sharpened to a linear term $\Omega(n)$ in [@DFW11], and an explicit bound for the coefficient of this linear term, was given by Lehre and Witt [@LehreWittISAAC14], who proved the lower bound $en\log n-7.81791n-O(\log n)$. The main tool to derive these results relies on increasingly refined drift theorems, most notably on variable drift analysis. At roughly the same time, Hwang et al. [@HwangArxiv14] presented a drastically refined analysis, which determines the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}up to terms of order $O((\log n)/n)$: the exact expression given is $$en \log n - C_1 n + (e/2)\log n + C_2 +O((\log n)/n),
\label{eq:exact-hwang}$$ where $C_1=1.89254\dots$ and $C_2=0.59789875\dots$ are explicitly computable constants; see also [@HwangEVCO18] for the journal version and <http://140.109.74.92/hk/?p=840> for the web version with a full asymptotic expansion. To obtain these precise results, techniques fundamentally different from drift analysis and other established methods for the runtime analysis were used, namely matched asymptotics with rigorous error analysis. In addition to the expected runtime, the asymptotic variance as well as the limiting distribution are also worked out there by similar approaches.
While the expression for the asymptotic expected runtime in represents the best of its kind, it also raises important open questions. First, from a more didactical and methodological point of view, one may look for a more elementary derivation of the formula , at least with respect to the linear term $-C_1 n$. Note that one can analyze the related search heuristic RLS, which flips exactly one bit per iteration, on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{} exactly and without any asymptotic terms (see [@DoerrDoerrAlgo16]), at least if it is initialized deterministically with $\cl{n/2}$ one-bits. The expression of the expected runtime equals $nH_{\tr{n/2}}$ then and is accompanied by an intuitive proof appealing to the coupon collector theorem. For a uniform initialization, the analysis become more involved but still an extremely precise result (coming with an asymptotic term, though) exists: $nH_{\tr{n/2}}-1/2+o(1)$. This proof takes only a few pages and uses well-known intuitive concepts such as the binomial distribution. The $o(1)$-term comes without an explicit error bound, though, and it is not discussed how to refine it.
Second, it would be helpful to confirm that the constant in the $O((\log n)/n)$-term is small so that one may call it negligible even for small problem sizes. This question may be approached along two different directions: one via an explicit error bound for all $n$, and the other by combining exact numerical calculations and asymptotic expansions. The former will be realized by the drift analysis presented in Sections \[sec:lower\]–\[sec:upper\] of this paper; we briefly describe here the latter, which depends on the sample size $n$. If $n$ is large enough, say $n\ge50$, then we can use a longer expansion of the form $$en \log n - C_1 n + \sum_{0\le k\le K}
\frac{d_k \log n + e_k}{n^k}$$ where $K$ is chosen large enough depending on the required tolerance error. In particular, by refining the analysis in [@HwangEVCO18], one has $d_0=\frac12$, $d_1=\frac98$ and $d_2=\frac{31}{16}$ (expressions for $e_k$ being more complex). On the other hand, if $n$ is small, one can always compute the exact quantity by the underlying recurrence relations without introducing any error. Such an exact calculation can be made efficient even in portable computing devices such as laptops and for $n$ in the hundreds; it is equally helpful in measuring the error introduced when using $K$ terms of the asymptotic expansion.
While different approaches have their own strengths and weaknesses, it is possible to combine them in many cases in discrete probabilities and algorithmics, and obtain results that are often stronger than a single approach can achieve. The fine approximation we work out in this paper represents another testimony to this statement.
#### Our contribution.
In this paper, we revisit the method of drift analysis and obtain that the expected runtime ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$ of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, started from $\cl{n/2}$ ones, is approximated by the sum of inverse drifts up to logarithmic error terms, more precisely $$\sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} - c_1\log n \le {\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}} \le \sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} - c_2\log n,$$ where $\Delta(k)$ is the drift (expected increase of the number of one-bits from the state of $n-k$ ones) and $c_1,c_2\,>0$ are explicitly computed constants. This gives not only an intuitive approximation of the expected runtime via inverse drifts but for the first time explicit error bounds. Closest to our results, Gie[ß]{}en and Witt [@GiessenWittAlgo18] used new variants of variable drift analysis and showed for the more general class of (1+$\lambda$) EAs that the expected runtime is characterized by the sum of inverse drifts up to an additive error of $\tilde{O}(n^{2/3})$ — we improve further this error term to $c\log n$ for an explicit constant $c>0$. To prove our results, we use elementary techniques and additive drift analysis as the only tool for the treatment of stochastic processes. At the same time, we obtain new drift theorems dealing with error bounds in variable drift analysis that may be of independent interest. The assumption of a fixed starting point for the [(1+1) EA]{} only introduces a difference in $O(1)$ compared to the expected runtime with a uniform initialization [@DoerrDoerrAlgo16].
Finally, from the sum expression of $\Delta(k)$, we prove, by standard asymptotic methods (generating functions and the Euler-Maclaurin formula), that the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}equals $$\sum_{k=1}^{\tr{n/2}} \frac{1}{\Delta(k)} - \frac{e}{2}\log n + O(1),$$ [i.e.]{}, the sum of inverse drifts overestimates the exact expected time only by an additive term of $(e/2)\log n+O(1)$.
This paper is structured as follows. In Section \[sec:preliminaries\], we introduce the concrete problem setting and well-known variable drift theorems. We also revisit the well-known result that the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}is bounded by the sum of inverse drifts over the interval $\{1,\dots,X_0\}$, where $X_0$ is the initial number of zero-bits. Section \[sec:lower\] is concerned with the lower bound $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k) - c_1 \log n$ for a constant $c_1>0$, which we prove using a new, self-contained variable drift theorem. Section \[sec:upper\] complements this result by bounding the expected runtime from above by $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k) - c_2 \log n$, for another constant $c_2>0$, again using a novel variable drift theorem. The following Section \[sec:exact\] then briefly illustrates that drift analysis in principle allows an alternative proof of an exact expression of the expected runtime, before we in Section \[sec:asymptotic-expansion\] apply asymptotic techniques to show that the expression $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k) - (e/2) \log n$ gives the exact time up to additive errors of $O(1)$.
Preliminaries {#sec:preliminaries}
=============
We consider the classical randomized search heuristic [(1+1) EA]{}; see Algorithm \[alg:oneoneea\], which is intensively studied in the theory of randomized search heuristics [@AugerDoerrBook; @JansenBook]. It creates a new search point by flipping each bit of the current search point independently with probability $1/n$ and accepts it if it is not inferior to the previous search point. The algorithm is formulated for pseudo-boolean maximization problems but can straightforwardly be applied to minimization as well. The analysis of the [(1+1) EA]{}is a stepping stone towards the analysis of more advanced search heuristics, but already this simple framework leads to challenging analyses even on very simple problems. In this paper, we focus exclusively on the simple [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}problem, which can be regarded as a simple hillclimbing task.
$t:=0$. Choose uniformly at random $x_0 \in \{0,1\}^n$. Create $x'$ by flipping each bit in $x_t$ independently with probability $1/n$. $x_{t+1}:=x'$ if $f(x') \ge f(x_t)$, and $x_{t+1}:=x_t$ otherwise. $t:=t+1$.
Since the [(1+1) EA]{}is unbiased, [i.e.]{}, it treats one-bits and zero-bits in the same way [@LehreWittAlgorithmica12], all results in this paper hold also for the more general Hamming distance minimization problem $f_z(x)=n-H(x,z)$, where $z\in\{0,1\}^n$ is arbitrary and $H(x,z)$ denotes the Hamming distance of the search points $x$ and $z$. We also remark that our forthcoming analyses can be generalized to different mutation rates, [i.e.]{}, a [[(1+1) EA]{}]{}that flips each bit independently with probability $c/n$ for a constant $c>0$; however, this will not yield new interesting insights. We emphasize that we only consider a static mutation probability here – dynamic schemes, including self-adjusting and self-adaptive mutation rates ([e.g.]{}, [@DoerrDYGECCO16; @DoerrGWYAlgo19]) must usually be analyzed via different techniques.
The *runtime* (synonymously, *optimization time*) is the smallest $t$ such that $x_t$ is optimal, [i.e.]{}, the random number of iterations until sampling an optimum. It corresponds to the number of fitness evaluations (plus $1$ for the initialization) until the optimum is found. In this paper, we are exclusively concerned with the expected runtime; bounds on the tail of the runtime of [(1+1) EA]{}can be found, [e.g.]{}, in [@LehreWittISAAC14].
Additive Drift
--------------
Our main tool for the runtime analysis of the [(1+1) EA]{}is *drift analysis*, which is in fact one of the most versatile and wide-spread techniques for this purpose [@LenglerDriftSurveyArxiv]. Roughly speaking, drift analysis translates information about the expected local change of the process (the so-called drift) into a global statement about the first hitting time of a target state. Drift analysis, which is well known in the theory of stochastic processes [@Hajek1982], was introduced to the field of runtime analysis of evolutionary computation by He and Yao [@heyao2001] in the form of an additive drift theorem. This theorem was continuously refined and given in different formulations. We present it in a very general style, allowing continuous state spaces and non-Markovian processes. As noticed by Lengler [@LenglerDriftSurveyArxiv] and Krejca and K[ö]{}tzing [@KoetzingKrejcaPPSN2018-drift], the process may live on a one-sided unbounded state space if upper bounds on the expected first hitting time are to be derived. We also integrate both variants for upper and lower bounds on expected hitting times in one theorem, sacrificing some generality in the second case [@KoetzingKrejcaPPSN2018-drift].
\[theo:additive-drift\] Let $(X_t)_{t\ge 0}$ be a stochastic process, adapted to a filtration ${\mathcal{F}_t}$, over some state space $S\subseteq {{\mathds{R}}}^{\ge 0}$, where $0\in S$. Let $T:=\min\{t\mid X_t=0\}$ be the first hitting time of state $0$.
1. If there is some $\delta>0$ such that conditioned on $t<T$ it holds that $${\mathord{E}\hspace{0.3ex}\mathord{\left(X_t-X_{t+1}\mid {\mathcal{F}_t}\right)}} \ge \delta,$$ then $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid {\mathcal{F}_0}\right)}} \le \frac{X_0}{\delta}.$$
2. If there is some $\delta>0$ such that conditioned on $t<T$ it holds that both $${\mathord{E}\hspace{0.3ex}\mathord{\left(X_t-X_{t+1}\mid {\mathcal{F}_t}\right)}} \le \delta,$$ and $X_t\le b$ for some constant $b> 0$ then $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid {\mathcal{F}_0}\right)}} \ge \frac{X_0}{\delta}.$$
In a nutshell, Theorem \[theo:additive-drift\] estimates the first hitting time of the target $0$ by the initial distance divided by the average process towards the target. Clearly, if the worst-possible $\delta$ over the state space is very small, then the resulting bound on the expected hitting time (in part 1) may overestimate the truth considerably. To obtain more precise results, one may transform the actual state space $X_t$ to a new state space $g(X_t)$ via a so-called potential (Lyapunov) function $g\colon S\to {{\mathds{R}}}^{\ge 0}$. If the drift of the process $g(X_t)$ is similar all over the search space then more precise bounds are obtained. This idea of smoothing out the drift over the state space underlies most advanced drift theorems such as multiplicative drift [@DJWMultiplicativeAlgorithmica] and variable drift [@Johannsen10]. Since multiplicative drift is a special case of variable drift, we will focus exclusively on additive and variable drift in the remainder of this paper.
Variable Drift
--------------
The first theorems stating upper bounds on the hitting time using variable drift go back to [@Johannsen10] and [@MitavskiyVariable]. These theorems were subsequently generalized in [@RoweSudholtChoiceJournal] and [@LehreWittISAAC14]. Similarly to Theorem \[theo:additive-drift\], we present a general version allowing non-Markovian processes and unbounded state spaces. We also give a self-contained proof.
\[theo:variable-upper\] Let $(X_t)_{t\ge 0}$ be a stochastic process, adapted to a filtration ${\mathcal{F}_t}$, over some state space $S\subseteq \{0\}\cup {{\mathds{R}}}^{\ge {x_{\mathrm{min}}}}$, where ${x_{\mathrm{min}}}>0$. Assume $0\in S$ and define $T:=\min\{t\mid X_t=0\}$.
Let $h\colon {{\mathds{R}}}^{\ge {x_{\mathrm{min}}}}\to{{\mathds{R}}}^+$ be a monotone increasing function and suppose that $E(X_t-X_{t+1} \mid {\mathcal{F}_t}) \ge h(X_t)$ conditioned on $t<T$. Then it holds that $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid {\mathcal{F}_0}\right)}} \le
\frac{{x_{\mathrm{min}}}}{h({x_{\mathrm{min}}})} + \int_{{x_{\mathrm{min}}}}^{X_0} \frac{1}{h(x)} \,\mathrm{d}x\enspace.$$
We will apply Theorem \[theo:additive-drift\] (part 1) with respect to the process $g(X_t)$, where the potential function $g(x)$ be defined by $$g(x) \coloneqq \frac{{x_{\mathrm{min}}}}{h({x_{\mathrm{min}}})} + \int_{{x_{\mathrm{min}}}}^x \frac{1}{h(z)} \,\mathrm{d}z.$$
We note that $g$ is concave since $1/h$ is monotone decreasing by assumption. Considering the drift of $g$, we have $${\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_t)-g(X_{t+1})\mid {\mathcal{F}_t}\right)}} =
\int_{{x_{\mathrm{min}}}}^{X_t} \frac{1}{h(z)} \,\mathrm{d}z
- {\mathord{E}\hspace{0.3ex}\mathord{\left(\int_{{x_{\mathrm{min}}}}^{X_{t+1}} \frac{1}{h(z)} \,\mathrm{d}z\mid {\mathcal{F}_t}\right)}}.$$ By Jensen’s inequality, we obtain for $t<T$ $${\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_t)-g(X_{t+1})\mid {\mathcal{F}_t}\right)}} =
\int_{{x_{\mathrm{min}}}}^{X_t} \frac{1}{h(z)} \,\mathrm{d}z
- \int_{{x_{\mathrm{min}}}}^{{\mathord{E}\hspace{0.3ex}\mathord{\left(X_{t+1}\mid {\mathcal{F}_t}\right)}}} \frac{1}{h(z)} \,\mathrm{d}z,$$ which, since ${\mathord{E}\hspace{0.3ex}\mathord{\left(X_{t+1}\mid {\mathcal{F}_t}\right)}} \le X_t-h(X_t)$, is at least $$ \int_{X_t-h(X_t)}^{X_t} \frac{1}{h(z)} \,\mathrm{d}z \ge
\int_{X_t-h(X_t)}^{X_t} \frac{1}{h(X_t)} \,\mathrm{d}z = 1,$$ where the inequality used that $h(z)$ in non-decreasing. The theorem now follows by Theorem \[theo:additive-drift\], part 1.
We remark that we can avoid applying Jensen’s inequality in the above proof by splitting $$\begin{aligned}
{\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_{t+1})\mid{\mathcal{F}_t}\right)}} & = {\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_{t+1}){\mathds{1}_{X_{t+1}\le X_t}}\mid{\mathcal{F}_t}\right)}} \\
& \quad +
{\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_{t+1}){\mathds{1}_{X_{t+1}>X_t}}\mid{\mathcal{F}_t}\right)}}\end{aligned}$$ and estimating $1/h(z)$ from above by $h(X_t)$ if $X_{t+1}>X_t$ by taking a change of sign into account [@LehreWitt2013arXivPreprint]. However, we find that this leads to a less easily readable proof. In any case, the variable drift theorem upper bounds the expected time to reach state $0$ because $h(x)$ is non-decreasing by assumption. If $h(x)$ was non-increasing, we could conduct an analogous proof to bound ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$ from below; however, usually the drift of a process increases with the distance from its target.
For discrete search spaces, the variable drift theorem can be simplified (see also [@RoweSudholtChoiceJournal]). We present the following version for Markov processes on the integers.
\[cor:discrete-variable\] Let $(X_t)_{t\ge 0}$ be a Markov process on the state space $\{0,\dots,N\}$ for some integer $N$. Let $\Delta\colon \{1,\dots,N\} \to {{\mathds{R}}}^+$ be a monotone increasing function such that $E(X_t-X_{t+1} \mid X_t=k) \ge \Delta(k)$. Then it holds for the first hitting time $T:=\min\{t\mid X_t=0\}$ that $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0\right)}} \le
\sum_{k=1}^{X_0} \frac{1}{\Delta(k)}.$$
First Upper Bound for [[<span style="font-variant:small-caps;">OneMax</span>]{}]{} {#sec:triv-upper}
----------------------------------------------------------------------------------
Corollary \[cor:discrete-variable\] is ready to use for our scenario of the analysis of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}. We identify state $k$ with all search points having $k$ zero-bits ([i.e.]{}, $n-k$ one-bits), think of the [(1+1) EA]{}minimizing the number of zero-bits and note that state $0$ is the optimal state. If we instantiate the corollary with $$\Delta(k) \coloneqq \sum_{\ell=1}^k \sum_{j=0}^{\ell} (\ell-j)\binom{k}{\ell}\binom{n-k}{j} \left(\frac{1}{n}\right)^{\ell+j}\left(1-\frac{1}{n}\right)^{n-\ell-j} ,
\label{eq:delta-i}$$ where as usual $\binom{a}{b}=0$ if $b<0$ or $b>a$, which is the exact expression for the expected decrease in the number of zero-bits from $k$ such bits, then we obtain an upper bound on the runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, started with $X_0$ zero-bits. This result is well known and it can easily be shown that $$\sum_{k=1}^{X_0} \frac{1}{\Delta(k)} \le enH_{X_0}$$ since $\Delta(k)\ge \frac{k}{n}(1-1/n)^{n-1}\ge e^{-1}k/n$ by considering all steps flipping exactly one bit out of the $k$ zeros and no other bits. However, no exact closed-form expression for $\sum_{k=1}^{X_0} \frac{1}{\Delta(k)} $ is known in general.
Lower Bounds {#sec:lower}
============
Variable Drift with Error Bound
-------------------------------
In light of the simple upper bound presented above in Section \[sec:triv-upper\], it is interesting to study how tight this bound is. Previous research addressed this question usually by
- Proving an analytical upper bound on the expected value of $Q_k\coloneqq \sum_{j=1}^{k} \frac{1}{\Delta(j)}$ (for a random starting state $k$)
- Bounding ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$ from below by using specific variable drift theorems for lower bounds. The sum $Q_k$ did not explicitly show up in these bounds.
As a result, this approach estimates the error made by bounding ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=k\right)}}\le Q_k$ only indirectly. One notable example is the work by Gie[ß]{}en and Witt [@GiessenWittAlgo18], who prove the nesting $$(1-O(n^{-1/3}\log n)) Q_k \le {\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=k\right)}} \le Q_k,$$ which shows that the sum of inverse drifts $Q_k$ represents the expected optimization time of (1+$\lambda$) EAs on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}from state $k$ up to polynomial lower order terms (which would be in the order of $O(n^{2/3}\log^2 n)$ for those starting points from which it takes expected time $\Omega(n\log n)$). Interestingly, this result was obtained by a new variable drift theorem for lower bounds that can be instantiated with the concrete setting of optimizing [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}. In this setting, one can identify the sum of inverse drifts $Q_k$ up to lower order terms.
In this section, we follow an even more direct approach to relate ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=k\right)}}$ to $Q_k$. As already mentioned, several variants of variable drift theorems for proving lower bounds on hitting times have been proposed; see again [@GiessenWittAlgo18] for a recent discussion. The main challenge proving such lower bounds is that the potential function $g(x)$ proposed in the proof of Theorem \[theo:variable-upper\] is concave, so Jensen’s inequality cannot be used to bound the drift of the potential function from above. However, if one can estimate the exact drift of the potential function and bound it uniformly from below for all non-optimal states, we get a lower bound for the expected first hitting time. We make this explicit for discrete search spaces in the following; however, the approach would easily generalize to continuous spaces. We restrict ourselves to non-increasing processes for notational convenience but note that we could allow $X_{t+1}$ to be greater than $X_t$ by adjusting the definition of $\eta(k)$ in the following theorem slightly.
\[theo:var-lower-error\]\[Variable drift, lower bound, with error bound\] Let $(X_t)_{t\ge 0}$ be a non-increasing Markov process on the state space $\{0,\dots,N\}$ for some integer $N$. Let $\Delta\colon \{1,\dots,N\} \to {{\mathds{R}}}^+$ be a function satisfying $E(X_t-X_{t+1} \mid X_t=k) \le \Delta(k)$ for $k\in\{1,\dots,N\}$. Let $$\eta(k) \coloneqq {\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=X_{t+1}+1}^{k} \frac{1}{\Delta(j)} \bigm| X_t=k\right)}}$$ and $$\eta^* \coloneqq \max_{k=1,\dots,N} \eta(k).$$ Then it holds for the first hitting time $T:=\min\{t\mid X_t=0\}$ that $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0\right)}} \ge
\sum_{k=1}^{X_0} \frac{1}{ \eta^*\Delta(k)} .$$
Hence, $\eta(k)$ is an error bound quantifying the relative error incurred by using the sum of inverse drifts as an estimate for the expected first hitting time from state $k$, and $\eta^*$ is the worst case of the $\eta(k)$ over all non-target states.
We consider the same potential function $g(k)=\sum_{j=1}^k 1/\Delta(j)$ as in the proof of Theorem \[theo:variable-upper\] and note that its drift at point $k$ equals $$\begin{aligned}
{\mathord{E}\hspace{0.3ex}\mathord{\left(g(k)-g(X_{t+1})\mid X_t=k\right)}} &
= \sum_{j=1}^{k} \frac{1}{\Delta(k)}- {\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=1}^{X_{t+1}} \frac{1}{\Delta(j)} \bigm| X_t=k\right)}} \\
& = \eta(k).\end{aligned}$$ By the additive drift theorem (Theorem \[theo:additive-drift\], part 2) with potential function $g(x)$ and upper bound $\eta^*$ on the drift, the theorem follows.
We will use the previous variable drift theorem to obtain the following lower bound.
\[theo:lower\] Let ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}}$ denote the expected optimization time of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, started with $\cl{n/2}$ one-bits and let $\Delta(k)$ be the drift of the number of zeros as defined in Definition . Then $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\right)}} \ge \sum_{k=1}^{\tr{n/2}} \frac{1}{\Delta(k)} - c_1 \log n$$ for some constant $c_1>0$.
The proof is dealt with in the following subsection. As already mentioned in the introduction, the assumption of a fixed starting point of $\cl{n/2}$ one-bits ([i.e.]{}, $\tr{n/2}$ zero-bits) allows us to concentrate on the essentials; if a uniform at random starting point was chosen, then the expected time would at most change by a constant [@DoerrDoerrAlgo16].
Bounding the Error {#sec:bounding-error}
------------------
This subsection is concerned with the proof of Theorem \[theo:lower\]. In particular, most effort is spent on establishing the claim $$\eta^* - 1 \le c/n$$ for some explicit constant $c>0$, [i.e.]{}, we bound the additive error of the drift of the potential function $g(k)-g(X_{t+1})$, where $X_t=k$ is the current state, compared to the lower bound $1=\Delta(k)/\Delta(k)$ established for the drift of the potential function at $X_t=k$ in Theorem \[theo:variable-upper\]. Here the notions of state (number of zero-bits), drift $\Delta(k)$ and transition probabilities are taken over from the preceding section.
Looking back into , we have already defined the drift (in terms of the number of zero-bits) at point $k$ and observe that $\Delta(k)$ is monotone increasing in $k$, which we will use later. Using the notation $p(k,\ell)$ for the transition probability from the state of $k$ to $\ell$ zero-bits, we note that by definition $$\eta(k) = \sum_{\ell=0}^{k-1} p(k,\ell) \sum_{j=\ell+1}^k \frac{1}{\Delta(j)}$$ and also that $$\sum_{\ell=0}^{k-1} p(k,\ell) \frac{k-\ell}{\Delta(k)} = \frac{\Delta(k)}{\Delta(k)} = 1 ,$$ which is why we pay attention to bounding the terms $$\label{eq:linearize-pot-diff}
p(k,\ell) \left(\sum_{j=\ell+1}^{k} \frac{1}{\Delta(j)} - \frac{k-\ell}{\Delta(k)}\right)$$ with the final aim of showing that $$\eta(k)-1 = \sum_{\ell=0}^{k-1} \left(p(k,\ell) \left(\sum_{j=\ell+1}^{k} \frac{1}{\Delta(j)} - \frac{k-\ell}{\Delta(k)}\right)\right) \le \frac{c}{n}
\label{eq:linearize-pot-full-diff}$$ for some sufficiently large constant $c>0$.
We shall define, as in [@HwangEVCO18], a kind of normalized drift that is easier to handle. Here it becomes relevant to manipulate the number $n$, so that we write more formally $$\begin{aligned}
\Delta_n(k) & \coloneqq \Delta(k) \\
& =
\sum_{\ell=1}^k \sum_{j=0}^{\ell} (\ell-j)\binom{k}{\ell}\binom{n-k}{j} \left(\frac{1}{n}\right)^{\ell+j}\left(1-\frac{1}{n}\right)^{n-\ell-j}\end{aligned}$$
The definition of the normalized drift $\Delta^*$ is then as follows.
\[def:delta-start\] Define, for $k\in\{1,\dots,n+1\}$, $$\begin{aligned}
\Delta_n^*(k)
& \coloneqq \Delta_{n+1}(k)
\Bigl(1-\frac1{n+1}\Bigr)^{-n-1} \\
& = \sum_{\ell=1}^{k}
\binom{k}{\ell}\sum_{j=0}^{\ell-1} (\ell-j)
\binom{n+1-k}{j} \left(\frac{1}{n}\right)^{j+\ell},\end{aligned}$$ Define, for convenience, $\Delta_n^*(0)\coloneqq 0$.
From we are brought to the task of bounding $\frac{1}{\Delta(k)} - \frac{1}{\Delta(k+1)}$, leading to Lemma \[lem:diff-1-over-Delta-k\] below. To this end, it is crucial to bound ${\Delta(k+1)-\Delta(k)}$. While this can be achieved in a tedious analysis comparing terms in the above-given representation of $\Delta(k)$ as a double sum, we follow a more elegant approach involving generating functions here. To this end, let $[z^n]f(z)$ denote the coefficient of $z^n$ in the Taylor expansion of $f(z)$.
\[lmm\] For $k\in\{0,\dots,n+1\}$ and $n\ge1$, the relation $$\begin{aligned}
\Delta_n^*(k)
= [z^{-1}]\frac1{(1-z)^2}\Bigl(1+\frac1{nz}\Bigr)^k
\Bigl(1+\frac zn\Bigr)^{n+1-k}\end{aligned}$$ holds.
Rewrite the sum definition of $\Delta_n^*(k)$ as the Cauchy product of three series: $$\begin{aligned}
\Delta_n^*(k)
&= \sum_{\ell = 1}^k
\binom{k}{\ell}n^{-\ell}\sum_{j=0}^{\ell-1} (\ell-j)
\cdot \binom{n+1-k}{j} n^{-j}\\
&= \sum_{\ell=0}^{k-1}
\binom{k}{\ell}n^{-k+\ell}\sum_{j=0}^{k-\ell-1} (k-\ell-j)
\cdot \binom{n+1-k}{j} n^{-j}\\
&= \sum_{\substack{h,j,\ell\\ h+j+\ell=k-1}}\binom{k}{\ell}
n^{-k+\ell}\cdot (h+1)
\cdot \binom{n+1-k}{j} n^{-j},\end{aligned}$$ implying that $$\begin{aligned}
\Delta_n^*(k) = [z^{k-1}]
\Bigl(z+\frac1{n}\Bigr)^k\frac1{(1-z)^2}
\Bigl(1+\frac zn\Bigr)^{n+1-k}.\end{aligned}$$ The lemma then follows from the relation $$\begin{aligned}
\label{coeff-shift}
[z^{k-1}]f(z) = [z^{-1}]z^{-k}f(z)
\qquad(k\ge1).\end{aligned}$$
We shall prove bounds on the difference $\Delta(k+1)-\Delta(k)$ via bounding the corresponding difference of the $\Delta^*$-values.
\[lem:diff-delta\] For $k\in\{0,\dots,n\}$ it holds that $$\begin{aligned}
\label{df-Delta}
\frac1n\le \Delta_n^*(k+1)-\Delta_n^*(k)
\le \frac{2e}{n}. \end{aligned}$$ and for $k\in\{0,\dots,n-1\}$ that $$\frac{1}{en} \le \Delta_n(k+1) - \Delta_n(k) \le \frac{2}{n-1}.$$
We prove first . From Lemma \[lmm\], we have $$\begin{aligned}
\Delta_n^*(k+1)-\Delta_n^*(k)
&= \frac{1}{n} [z^{-1}]
\frac1{(1-z)^2}\Bigl(1+\frac1{nz}\Bigr)^{k}
\Bigl(1+\frac zn\Bigr)^{n-k}
\Lpa{\frac1z-z}\\
&= \frac{1}{n}[z^{-1}]
\frac{1+z}{z(1-z)}\Bigl(1+\frac1{nz}\Bigr)^{k}
\Bigl(1+\frac zn\Bigr)^{n-k}.\end{aligned}$$
Thus, by taking the coefficients of the Cauchy product, we obtain $$\begin{aligned}
\label{Delta-diff}
n&\lpa{\Delta_n^*(k+1)-\Delta_n^*(k)} \nonumber\\
&= \sum_{\ell=0}^k \binom{k}{\ell}n^{-\ell}
\sum_{j=0}^{n-k}\binom{n-k}{j}n^{-j}
[z^{\ell-j}]\frac{1+z}{1-z} \\
&\le 2 \sum_{\ell=0}^k\binom{k}{\ell}n^{-\ell}
\sum_{j=0}^{\ell-1}\binom{n-k}{j}n^{-j}\nonumber \\
&< 2\sum_{\ell\ge0}\frac{(k/n)^\ell}{\ell!}
\sum_{j\ge0}\frac{(1-k/n)^j}{j!}\nonumber \\
&= 2e^{k/n+1-k/n}=2e.\nonumber\end{aligned}$$ On the other hand, by , $$\begin{aligned}
\label{Delta-mono}
n& \lpa{\Delta_n^*(k+1)-\Delta_n^*(k)}\nonumber\\
& \ge \sum_{\ell=0}^k \binom{k}{\ell}n^{-\ell}
\sum_{j=0}^\ell \binom{n-k}{j}n^{-j}
\ge\Lpa{1+\frac1n}^k > 1.\end{aligned}$$ This proves .
Recalling the definition $$\Delta_n(k) = \Delta_{n-1}^*(k) \left(1-\frac{1}{n}\right)^n,$$ we finally obtain $$\Delta_n(k+1) - \Delta_n(k) \le \left(1-\frac{1}{n}\right)^n \frac{2e}{n-1} \le \frac{2}{n-1}$$ and $$\Delta_n(k+1) - \Delta_n(k) \ge \frac{(1-1/n)^n}{n-1} = \frac{(1-1/n)^{n-1}}{n} \ge \frac{1}{en}$$ as claimed.
Recall that we want to investigate the difference $$\frac{1}{\Delta(k-1)} - \frac{1}{\Delta(k)} = \frac{\Delta(k) - \Delta(k-1)}{\Delta(k-1)\Delta(k)}$$ (and later $\frac{1}{\Delta(k-\ell)} - \frac{1}{\Delta(k)} $ for $\ell\ge 1$); thus we need bounds on $\Delta(k)$ itself. The following lemma gives such bounds along with estimations of the transition probabilities. We will use the notation $p(k,\le j) = \sum_{\ell=0}^{j} p(k,\ell)$ for the probability to change from state $k$ to state at most $j$.
\[lem:delta-bounds\] For $k\ge 1$, $e^{-1}\frac{k}{n}\le \Delta(k) \le \frac{k}{n}$. Moreover, for $\ell\ge 1$ it holds that $p(k,k-\ell)\le p(k,\le k-\ell) \le \binom{k}{\ell}\lpa{\frac{1}{n}}^\ell \le \left(\frac{k}{n}\right)^\ell / \ell!$.
This proof uses well-known standard arguments. The upper bound on the drift follows from considering the expected number of flipping bits among $k$ one-bits and the lower bound from looking into steps flipping one bit only. The bound on the transition probability considers all mutations flipping at least $\ell$ bits.
Intuitively, the parenthesized term in estimates the error incurred by estimating the potential function using the slope at $k$ for a step of size $\ell+1$. This error will below in Lemma \[lem:bound-eta-i\] be weighted by the probability of making a step of such size, more precisely by the probability of jumping from $k$ to $j=k-\ell-1$. Assembling the previous lemmas, we now give a bound for the difference of $1/\Delta(\cdot)$.
\[lem:diff-1-over-Delta-k\] For $k\ge 1$ and $\ell\in\{1,\dots,k\}$ it holds that $$\frac{1}{\Delta(k-\ell)} - \frac{1}{\Delta(k)} \le
\frac{2e^2 \ell n^2}{k(k-\ell)(n-1)}.
$$
Using Lemma \[lem:diff-delta\] and Lemma \[lem:delta-bounds\], $$\begin{aligned}
\frac{1}{\Delta(k-\ell)} - \frac{1}{\Delta(k)} & =
\frac{\Delta(k)-\Delta(k-\ell)}{\Delta(k)\Delta(k-\ell)} \\
& \le
\frac{2\ell/(n-1)}{e^{-2} k(k-\ell)/n^2} = \frac{2e^2 \ell n^2}{k(k-\ell)(n-1)} .
\qedhere\end{aligned}$$
If we jump from $k\ge 2$ to $k-\ell-1$ then the parenthesized term in (intuitively incurred by linearizing the potential function using the slope at $k$) equals $$\left(\frac{1}{\Delta(k-\ell)} - \frac{1}{\Delta(k)}\right)
+ \dots + \left(\frac{1}{\Delta(k-1)} - \frac{1}{\Delta(k)}\right) \le
\frac{\ell}{\Delta(k-\ell)}-\frac{\ell}{\Delta(k) }.$$
Finally, we weigh these differences with the respective probabilities and put everything together to bound the whole expression .
\[lem:bound-eta-i\] For $k\in\{1,\dots,\tr{n/2}\}$ it holds that $\eta(k) \le 1+\frac{2e^{5/2}}{n-1}$.
Using Lemma \[lem:delta-bounds\] and Lemma \[lem:diff-1-over-Delta-k\], $$\begin{aligned}
\sum_{\ell=0}^{k-1}\;& p(k, k-\ell-1)
\left(\frac{\ell}{\Delta(k-\ell)}
- \frac{\ell}{\Delta(k)}\right)\\
&=\sum_{\ell=1}^k p(k, k-\ell)
\left(\frac{\ell-1}{\Delta(k-\ell+1)}
- \frac{\ell-1}{\Delta(k)}\right)\\
&\le \frac{2e^2n^2}{k(n-1)}
\sum_{\ell=1}^k \binom{k}{\ell}n^{-\ell}
\frac{(\ell-1)^2}{k-\ell+1}\\
&\le \frac{2e^2n^2}{k(n-1)}
\sum_{\ell=0}^k\binom{k}{\ell}n^{-\ell}
\frac{\ell(\ell-1)}{k-\ell+1}.\end{aligned}$$ Applying the integral representation $\frac1a=\int_0^1 t^{a-1}\dd t$ for $a>0$, we obtain $$\begin{aligned}
\sum_{\ell=0}^k\binom{k}{\ell}n^{-\ell}
\frac{\ell(\ell-1)}{k-\ell+1}
&= \int_0^1 \sum_{\ell=0}^k \binom{k}{\ell}\ell(\ell-1)n^{-\ell}
t^{k-\ell}\dd t\\
&= \frac{k(k-1)}{n^2}\int_0^1\Lpa{t+\frac1n}^{k-2}\dd t\\
&\le \frac{k}{n^2}\Lpa{1+\frac1n}^{k-1}.\end{aligned}$$ Thus $$\begin{aligned}
&\sum_{\ell=0}^{k-1} p(k, k-\ell-1)
\left(\frac{\ell}{\Delta(k-\ell)}
- \frac{\ell}{\Delta(k)}\right)\\
&\qquad\le \frac{2e^2n^2}{k(n-1)}
\cdot \frac{k}{n^2}\Lpa{1+\frac1n}^{k-1}\\
&\qquad= \frac{2e^2}{n-1}\Lpa{1+\frac1n}^{k}
\le \frac{2e^{5/2}}{n-1}.\end{aligned}$$
We are now ready to complete the proof of Theorem \[theo:lower\].
[Theorem \[theo:lower\]]{} According to Theorem \[theo:var-lower-error\], we have $$\begin{aligned}
{\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=n/2\right)}} & \ge \sum_{k=1}^{\tr{n/2}} \frac{1}{ \eta^*\Delta(k)}
\ge \sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} \left(1-\frac{\eta^*-1}{\eta^*}\right),\end{aligned}$$ and by Lemma \[lem:bound-eta-i\] $$\eta^* \le 1+\frac{2e^{5/2}}{n-1}.$$ Since $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k) \le enH_{\tr{n/2}} \le en\log n$ as observed in Section \[sec:triv-upper\], we altogether obtain $$\begin{aligned}
{\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=n/2\right)}} & \ge \sum_{k=1}^{\tr{n/2}} \frac{1}{ \Delta(k)} - \frac{(en\log n)(2e^{5/2})}{n-1} \\ & \ge
\sum_{k=1}^{\tr{n/2}} \frac{1}{ \Delta(k)} - (4e^{7/2})\log n,\end{aligned}$$ where the last inequality used $n\ge 2$. Altogether, the theorem has been established with $c_1=4e^{7/2}\approx 132.56$.
In conjunction with Section \[sec:triv-upper\], we have determined the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}(starting in state $\tr{n/2}$, [i.e.]{}, with $\cl{n/2}$ one-bits) up to an additive term bounded by $c_1\log n$. As already mentioned in the introduction, terms of even lower order down to $O((\log n)/n)$ have been determined in [@HwangEVCO18] by a more technical analysis. Our result features a non-asymptotic error bound.
Improving the $\bf \sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)}$ Bound {#sec:upper}
==================================================================
The upper bound $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k)$ derived Section \[sec:triv-upper\] precisely characterizes the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}, but is a slight overestimation resulting from the inequality $1/h(X_{t+1})\ge 1/h(X_t)$ in the proof of Theorem \[theo:variable-upper\]; intuitively this corresponds to estimating the progress from state $X_t$ via a linearized potential function of slope $1/h(X_t)$, which is the derivative of $g$ at $X_t$.
We can improve the bound on the expected runtime by estimating the error stemming from this inequality and will gain a logarithmic term. To this end, we study the following simple analogue of Theorem \[theo:var-lower-error\].
\[theo:var-upper-error\]\[Variable drift, upper bound, with error bound\] Let $(X_t)_{t\ge 0}$ be a non-increasing Markov process on the state space $\{0,\dots,N\}$ for some integer $N$. Let $\Delta\colon \{1,\dots,N\} \to {{\mathds{R}}}^+$ be a function satisfying $E(X_t-X_{t+1} \mid X_t=k) \ge \Delta(k)$ for $k\in\{1,\dots,N\}$. Let $$\eta(k) \coloneqq {\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=X_{t+1}+1}^{k} \frac{1}{\Delta(j)} \bigm| X_t=k\right)}}$$ and $$\eta^* \coloneqq \min_{k=1,\dots,N} \eta(k).$$ Then it holds for the first hitting time $T:=\min\{t\mid X_t=0\}$ that $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0\right)}} \le
\sum_{k=1}^{X_0} \frac{1}{ \eta^*\Delta(k)} .$$
We proceed analogously to the proof of Theorem \[theo:var-upper-error\], use the potential function $g(k)=\sum_{j=1}^k 1/\Delta(j)$ and apply additive drift analysis (Theorem \[theo:additive-drift\], part 1) with the lower bound $\eta^*$ on its drift.
We state our improved result, carrying over notation from previous sections such as the definition of the drift $\Delta(k)$ with respect to [(1+1) EA]{}and [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}.
\[theo:improved-upper\] Let $n\ge 4$. Then the expected optimization time of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}(starting at $\cl{n/2}$ ones) is at most $\sum_{k=1}^{\tr{n/2}} 1/\Delta(k) - c_2\log n$ for some constant $c_2>0$.
To prove this result, we need to invert a statement from Section \[sec:bounding-error\].
\[lem:diff-1-over-delta-lower\] For $ k\in\{2,\dots,n/2\}$, $$\frac{1}{\Delta(k-1)} - \frac{1}{\Delta(k)} \ge \frac{n}{ek^2}.$$
We proceed similarly to the proof of Lemma \[lem:diff-1-over-Delta-k\] but aim at lower bounds. First, we recall from Lemma \[df-Delta\] that $$\Delta(k)-\Delta(k-1) \ge \frac{1}{en}.$$ Now, using the upper bound $\Delta(k)\le k/n$ from Lemma \[lem:delta-bounds\], we obtain $$\frac{1}{\Delta(k-1)} - \frac{1}{\Delta(k)} = \frac{\Delta(k)-\Delta(k-1)}{\Delta(k)\Delta(k-1)}
\ge \frac{1}{en (k/n)^2} = \frac{n}{ek^2},$$ which concludes the proof.
We can now present the proof of the improved upper bound.
[Theorem \[theo:improved-upper\]]{} The aim is to apply Theorem \[theo:var-upper-error\] for some $\eta^*=1+c/n$, where $c>0$ is constant. Since state $1$ is special in that it only has one possible successor, we consider $T_1:=\min\{t\mid X_t\le 1\}$ instead and the following straightforward generalization of the theorem: $${\mathord{E}\hspace{0.3ex}\mathord{\left(T_1\mid X_0\right)}} \le
\sum_{k=2}^{X_0} \frac{1}{ \eta^*\Delta(k)},$$ where $\eta^* \coloneqq \min_{k=2,\dots,n} \eta(k)$. This implies $${\mathord{E}\hspace{0.3ex}\mathord{\left(T_0\mid X_0=n/2\right)}} \le
\frac{1}{\Delta(1)} + \sum_{k=2}^{\tr{n/2}} \frac{1}{ \eta^*\Delta(k)}$$ since the expected transition time from state $1$ to $0$ is exactly $1/\Delta(1)$.
We now show that $\eta(k)\ge 1+c_1/n$ for some constant $c_1>0$ and $k\in\{2,\dots,\tr{n/2}\}$. Note that (conditioning on $X_t=k$ everywhere) $$\begin{aligned}
\eta(k) & = {\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=X_{t+1}+1}^{k} \frac{1}{\Delta(j)} \right)}} \\
& = {\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=X_{t+1}+1}^{k} \frac{1}{\Delta(j)} \mid X_{t+1}<k-1 \right)}} \,\, {\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}<k-1\right)}} \\
& \qquad +
\frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-X_{t+1})\mid X_{t+1}\ge k-1\right)}}}{\Delta(k)} \,{\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}\ge k-1\right)}}
\\
& =
{\mathord{E}\hspace{0.3ex}\mathord{\left(\sum_{j=X_{t+1}+1}^{k} \frac{1}{\Delta(j)} \mid X_{t+1}<k-1 \right)}} \, {\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}<k-1\right)}} \\
& \qquad + \frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-X_{t+1}){\mathds{1}_{X_{t+1}\ge k-1}}\right)}}}{\Delta(k)}.\end{aligned}$$
The first term on the right-hand side can be bounded from below by $$\left(\frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-1-X_{t+1}) \mid X_{t+1}<k-1\right)}}}{\Delta(k-1)} + \frac{1}{\Delta(k)}\right){\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}<k-1\right)}}$$ since $\Delta(k)$ is non-decreasing. Using Lemma \[lem:diff-1-over-delta-lower\], the last expression is further bounded from below by $$\begin{aligned}
& \left({\mathord{E}\hspace{0.3ex}\mathord{\left((k-1-X_{t+1}) \mid X_{t+1}<k-1\right)}} \left(\frac{1}{\Delta(k) }
+ \frac{n}{ek^2}\right)+\frac{1}{\Delta(k)}\right) \\
& \hspace*{6cm} \mbox{}\cdot {\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}<k-1\right)}},\end{aligned}$$ which, using $${\mathord{\operatorname{Pr}}\mathord{\left(X_{t+1}<k-1\right)}} \ge \frac{e^{-1}k(k-1)}{2n^2} \ge e^{-1} \frac{k^2}{4n^2}$$ and $${\mathord{E}\hspace{0.3ex}\mathord{\left((k-1-X_{t+1}) \mid X_{t+1}<k-1 \right)}}\ge 1,$$ is at least $$\left(\frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-X_{t+1}) {\mathds{1}_{X_{t+1}<k-1}} \right)}}}{\Delta(k) }\right) + \frac{e^{-2}}{4n}.$$ Putting everything together, we have $$\begin{aligned}
\eta(k) & \ge
\frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-X_{t+1}){\mathds{1}_{X_{t+1}\ge k-1}}\right)}}}{\Delta(k)} \\
& \quad +
\left(\frac{{\mathord{E}\hspace{0.3ex}\mathord{\left((k-X_{t+1}) {\mathds{1}_{X_{t+1}<k-1}} \right)}}}{\Delta(k) }\right) + \frac{e^{-2}}{4n} \\
& = \frac{\Delta(k)}{\Delta(k)} + \frac{e^{-2}}{4n} = 1+\frac{e^{-2}}{4n},\end{aligned}$$ so $\eta^*\ge 1+e^{-2}/(4n)$. We conclude the proof by noting that $$\sum_{k=2}^{\tr{n/2}} \frac{1}{\Delta(k)(1+e^{-2}/(4n))} \le \sum_{k=2}^{\tr{n/2}} \frac{1}{\Delta(k)} -
\sum_{k=2}^{\tr{n/2}} \frac{e^{-2}}{4n (1+e^{-2}/4) \Delta(k)},$$ which, using $\sum_{k=2}^{\tr{n/2}} \frac{1}{\Delta(k)} \ge n (H_{\tr{n/2}}-1) \ge n(\log n)/3$ for $n\ge 4$, amounts to $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=n/2\right)}} \le \sum_{k=1}^{\tr{n/2}} \frac{1}{\Delta(k)} - \frac{e^{-2}}{12 (1+e^{-2}/4)}
\log n.$$ Hence, we can set $c_2=\frac{e^{-2}}{12 (1+e^{-2}/4)}\approx 1/91.69$.
Formulas for The Exact Optimization Time {#sec:exact}
========================================
In light of the Theorems \[theo:var-lower-error\] and \[theo:var-upper-error\] one might wonder whether one should try to choose a potential function that makes the “error” $\eta^*$ vanish and leads to a drift of exactly $1$. It is well known [@Lehre12DriftTutorial; @LenglerDriftSurveyArxiv] that letting $g(k)$ be the expected remaining optimization time from state $k$ actually achieves this.
In this section, we briefly investigate how to choose $g(k)$ with respect to our setting of [(1+1) EA]{}and [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}. We will obtain formulas that can also be derived manually, so the result is by no means new. However, it is still interesting to see that it can be derived via drift analysis. This will turn out in the proof of the following theorem.
\[theo:times-transition\]Let $(X_t)_{t\ge 0}$ be a non-increasing Markov process on the state space $\{0,\dots,N\}$ for some integer $N$ and denote by $p(k,j)$ the transition probability from state $k$ to state $j$. Let the function $g(k)$ be recursively defined by $g(0)\coloneqq 0$ and for $k\ge 1$: $$g(k)\coloneqq \frac{1+\sum_{j=1}^{k-1}p(k,j) g(j)}{\sum_{j=0}^{k-1}p(k,j)}$$ Then it holds for the first hitting time $T:=\min\{t\mid X_t=0\}$ that $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0\right)}} =
g(X_0).$$
We shall use additive drift analysis (Theorem \[theo:additive-drift\]), which gives the exact expected hitting time if ${\mathord{E}\hspace{0.3ex}\mathord{\left(g(X_t)-g(X_{t+1})\mid {\mathcal{F}_t}\right)}} = \delta$, [i.e.]{}, if both the first and the second cases of the theorem hold.
We compute $$\begin{aligned}
& {\mathord{E}\hspace{0.3ex}\mathord{\left(g(k)-g(X_{t+1})\mid X_t=k\right)}} = \sum_{j=0}^{k-1} p(k,j) (g(k)-g(j)) \\
& \qquad\qquad = (1-p(k,k)) g(k) - \sum_{j=0}^{k-1} p(k,j)g(j) \\
& \qquad\qquad= (1-p(k,k)) \frac{1+\sum_{j=0}^{k-1}p(k,j) g(j)}{\sum_{j=0}^{k-1}p(k,j)}
- \sum_{j=0}^{k-1} p(k,j)g(j)\\
& \qquad\qquad= \left(1+\sum_{j=0}^{k-1}p(k,j) g(j)\right) -
\sum_{j=0}^{k-1} p(k,j)g(j) = 1,\end{aligned}$$ with the the definition of $g(k)$ plugged in the third equality. Hence, by Theorem \[theo:additive-drift\] the expected hitting time of state $0$ from state $X_0$ equals $g(X_0)/1$.
That $g(k)$ equals the expected first hitting time from state $k$ to state $0$ can also be proved in an elementary induction. By writing $$g(k) = \frac{1}{\sum_{j=0}^{k-1}p(k,j)} + \sum_{j=1}^{k-1}\frac{p(k,j)}{\sum_{j=0}^{k-1}p(k,j)}\, g(j)$$ we realize that the first term is the expected time to leave state $k$ and the second term is a weighted sum of the remaining optimization times from smaller state, weighted by the respective transition probabilities conditional on leaving state $k$. Such formulas can also be derived by inverting matrices obtained from the transition probabilities of the underlying Markov chain [@ChicanoEC15].
We note that estimations of hitting times in finite search spaces based on the transition probabilities were recently presented in K[ö]{}tzing and Krejca [@KoetzingKrejcaPPSN2018-finite]. These estimations are not recursively defined and easy to evaluate. However, as the underlying scenario does not allow big jumps towards the optimum when estimating the hitting time from below, tight formulas for the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}cannot be proved with this approach.
We exemplarily apply Theorem \[theo:times-transition\] to our scenario of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}. Using the transition probabilities $$p(k,j)= \sum_{\ell=0}^{\min\{j,n-k\}} \binom{k}{k-j+\ell}\binom{n-k}{\ell} \left(\frac{1}{n}\right)^{k-j+2\ell}\left(1-\frac{1}{n}\right)^{n-(k-j)-2\ell}$$ we obtain $g(0)=0$, $g(1) = n(1-1/n)^{1-n}$, and
$$\begin{aligned}
g(2)&=\frac{(3n^3-8n^2+6n-1)(1-1/n)^{1-n}}{2n^2-2n-1}\\[2ex]
g(3)&=\frac{
(22n^7-114n^6+203n^5-117n^4-38n^3+49n^2-7n+2)(1-1/n)^{1-n}
}
{12n^6-36n^5+4n^4+60n^3-23n^2-21n-2}
\\
&\vdots\end{aligned}$$
While these expansions obviously reflect the well-known estimate $g(k)=(1\pm o(1))en H_k$, they do not seem readily useful in expressing the expected runtime of the [(1+1) EA]{}on [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}in a closed-form formula depending on $n$.
The Asymptotics of the Partial Sum $\bf \sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)}$ {#sec:asymptotic-expansion}
=================================================================================
The purpose of this section is to analyze more precisely how far the sum of inverse drifts $\sum_{k=1}^{\tr{n/2}}1/\Delta(k)$ differs from the expected optimization time $${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=\tr{n/2}\right)}}
= en \log n - C_1 n + (e/2)\log n + O(1)$$ derived in [@HwangEVCO18]. We know from the preceding analysis that the sum of inverse drifts overestimates ${\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=\tr{n/2}\right)}}$ by a $\Theta(\log n)$-term. We will prove the following asymptotic approximation for the sum of inverse drifts, which, when compared with , shows their logarithmic difference.
\[thm:S1D\] For large $n$, $$\begin{aligned}
\label{S1D}
\sum_{k=1}^{\tr{n/2}}\frac1{\Delta_n(k)}
= en\log n -C_1n + e\log n +O(1),\end{aligned}$$ where $$\begin{aligned}
\label{C1}
C_1 := -e\left(\gamma-\log 2+\int_0^{1/2}
\left(\frac1{S_1(t)}-\frac1t\right)\dd t\right)
\approx 1.89254\dots\end{aligned}$$ is the same linear constant appearing in . Here $$\begin{aligned}
\label{Srz}
S_r(z) := \sum_{\ell\ge0}\frac{z^\ell}{\ell!}
\sum_{j=0}^{\ell-1}(\ell-j)^r \frac{(1-z)^j}{j!}
\qquad(r\ge0; z\in\mathbb{C}).\end{aligned}$$
Note that if we multiply the left-hand side of by $e^{-1/(2n)}$, then the difference with is bounded, namely, $$e^{-1/(2n)}\sum_{k=1}^{\tr{n/2}}\frac1{\Delta_n(k)}
= en\log n -C_1\,n + \frac{e}2\log n +O(1).$$
To prove Theorem \[thm:S1D\], we use the techniques of generating functions and Euler-Maclaurin summation formula, which are conceptually and methodologically simpler than the asymptotic resolution of the recurrences used in [@HwangEVCO18]. The following lemma can be obtained in style similar to Lemma \[lem:delta-bounds\]. Since it is with respect to the normalized $\Delta^*$, we give a self-contained proof.
For $k\in\{0,\dots,n+1\}$, $$\begin{aligned}
\label{ineq-Delta}
\Lpa{1+\frac1n}^{k-1} \frac kn
\le \Delta_n^*(k)\le \Lpa{1+\frac1n}^{n} \frac kn.\end{aligned}$$
By definition $$\begin{aligned}
\Delta_n^*(k)
&= \frac kn\sum_{\ell=0}^{k-1}
\binom{k-1}{\ell}n^{-\ell}\sum_{j=0}^{\ell}
\frac{\ell+1-j}{\ell+1}
\binom{n+1-k}{j} n^{-j}\\
&\le \frac kn\sum_{\ell=0}^{k-1}
\binom{k-1}{\ell}n^{-\ell}\sum_{j=0}^{\ell}
\binom{n+1-k}{j} n^{-j}\\
&\le \frac kn \sum_{\ell=0}^{k-1}
\binom{k-1}{\ell}n^{-\ell}
\sum_{j=0}^{n+1-k}
\binom{n+1-k}{j} n^{-j}\\
&= \Lpa{1+\frac1n}^{n} \frac kn < e\,\frac kn.\end{aligned}$$ On the other hand, $$\begin{aligned}
\Delta_n^*(k)
&\ge \frac kn\sum_{\ell=0}^{k-1}
\binom{k-1}{\ell}n^{-\ell}
= \Lpa{1+\frac1n}^{k-1} \frac kn.\end{aligned}$$
Note that becomes an identity when $k=0$ and $k=n+1$.
The crucial lemma we need to prove is given as follows.
Let $\ve>0$. Then for $1\le k\le (1-\ve)n$, $$\begin{aligned}
\label{ST}
\Delta_n^*(k)
= S_1(\alpha) + \frac{T_1(\alpha)}{n}
+ O\lpa{n^{-2}},\end{aligned}$$ where $\alpha=k/n$ and $$\label{T1}
\begin{split}
T_1(\alpha) &= \tfrac12S_1(\alpha)-2\alpha S_0(\alpha)
-\alpha \,I_0\lpa{2\sqrt{\alpha(1-\alpha)}}\\
&\qquad -\sqrt{\alpha(1-\alpha)}\,
I_1\lpa{2\sqrt{\alpha(1-\alpha)}}.
\end{split}$$ Here the $I_j$’s represent the modified Bessel functions.
It is possible to extend further the range in $k$, but we do not need it here.
First for small $k$, we have, by Definition \[def:delta-start\] and direct expansion, $$\begin{aligned}
\label{Dnk-small}
\Delta_n^*(k)
= \frac{k}{n}+\frac{3k(k-1)}{2n^2}
+O\lpa{k^3n^{-3}},\end{aligned}$$ which holds uniformly for $1\le k=o(n)$. A simple, readily codable procedure to derive this is as follows. Assuming $k$ to be fixed and expanding $$\Bigl(1+\frac1{nz}\Bigr)^k
\Bigl(1+\frac zn\Bigr)^{n+1-k}
= e^z + \frac{e^z}{2n} \left(\frac{2k}z-2(k-1)z-z^2\right)
+\cdots,
$$ for large $n$. Then multiplying both sides by $(1-z)^{-2}$ and computing the coefficient of $z^{-1}$ term by term (corresponding to the residue of the integrand in the Cauchy integral), giving $$\begin{aligned}
[z^{-1}]\frac{e^z}{(1-z)^2} &= 0,\\
[z^{-1}]\frac{e^z}{2n(1-z)^2}
\left(\frac{2k}z-2(k-1)z-z^2\right)
&= \frac kn,\\
[z^{-1}]\frac{e^z}{2n(1-z)^2}
\left(\frac{k(k-1)}{z^2}-2k(k-1)z\right)
&= \frac {3k(k-1)}{2n^2},\\
&\cdots\end{aligned}$$ On the other hand, by the Taylor expansions $$\label{ST-z}
S_1(z) = z +\tfrac32z^2+\tfrac5{12}z^3
+\cdots
\text{ and }
T_1(z) = -\tfrac32z-\tfrac74z^2-\tfrac18z^3
+\cdots,$$ we see that $$S_1(\alpha) + \frac{T_1(\alpha)}{n}
= \alpha +\frac32\,\alpha^2-\frac{3\alpha}{2n}
+O\lpa{\alpha^3+\alpha^2\,n^{-1}},$$ consistent with . This proves when $k=o(n)$.
Now consider larger values of $k$ and write $k=\alpha n$, where $\alpha\in[\ve,1-\ve]$. Then $$\begin{aligned}
&\alpha n \log\Lpa{1+\frac1{nz}}
+(1-\alpha) n\log\Lpa{1+\frac zn} \\
&\qquad\eqqcolon \frac\alpha z+(1-\alpha)z
-\frac{\alpha+(1-\alpha)z^4}{2nz^2}
+E_0(z),\end{aligned}$$ where $$\begin{aligned}
E_0(z) &= \sum_{\ell\ge2}\frac{(-1)^{\ell}}{n^\ell}
\left(\alpha\,\frac{z^{-\ell-1}}{\ell+1}
+(1-\alpha)\frac{z^{\ell+1}}{\ell+1} \right) \\
&= O\left( \frac{\alpha |z|^{-3}+
(1-\alpha)|z|^3}{n^2}\right).\end{aligned}$$ By the inequality $$|e^z-1| = \left|z\int_0^1 e^{tz}\dd t\right|
\le |z|e^{|z|}\qquad(z\in\mathbb{C}),$$ we have $$\begin{aligned}
&\left|\Bigl(1+\frac1{nz}\Bigr)^k
\Bigl(1+\frac zn\Bigr)^{n-k}
-e^{\frac\alpha z+(1-\alpha)z
-\frac{\alpha z^{-2}+(1-\alpha)z^2}{2n}}\right|\\
&\qquad \le |E_0(z)| e^{|E_0(z)|}
\Bigl|e^{\frac\alpha z+(1-\alpha)z
-\frac{\alpha z^{-2}+(1-\alpha)z^2}{2n}}\Bigr|.\end{aligned}$$ The error is then estimated by using the Cauchy integral representation $$\begin{aligned}
&[z^{-1}]\frac{1}{(1-z)^2}
\Lpa{1+\frac1{nz}}^k\Lpa{1+\frac zn}^{n+1-k}\\
&\qquad= \frac{1}{2\pi i}\oint_{|z|=r}
\frac1{(1-z)^2}\Bigl(1+\frac1{nz}\Bigr)^k
\Bigl(1+\frac zn\Bigr)^{n+1-k}\dd z,\end{aligned}$$ so that ($0<r<1$) $$\begin{aligned}
&\left|\oint_{|z|=r}
\frac{|E_0(z)| e^{|E_0(z)|}}{|1-z|^2}
\Bigl|e^{\frac\alpha z+(1-\alpha)z
-\frac{\alpha z^{-2}+(1-\alpha)z^2}{2n}}
\Bigl(1+\frac zn\Bigr)\Bigr| \dd z\right|\\
&=O\left(n^{-2}\int_{-\pi}^{\pi}
\frac{\alpha r^{-2} + (1-\alpha)r^4}{(1-r)^2}\,
e^{\frac \alpha r\cos t+(1-\alpha)r\cos t}\right)\dd t\\
&=O\lpa{n^{-2}},\end{aligned}$$ since $r$ is away from $1$. Thus $$\begin{aligned}
\Delta_n^*(k)
&= [z^{-1}]\frac{1+\frac zn}{(1-z)^2}
\,e^{\frac\alpha z+(1-\alpha)z
-\frac{\alpha z^{-2}+(1-\alpha)z^2}{2n}}
+O\lpa{n^{-2}}.\end{aligned}$$ By the same argument, we have $$\begin{aligned}
\Delta_n^*(k)
&= [z^{-1}]\frac{e^{\frac\alpha z+(1-\alpha)z}}{(1-z)^2}
\left(1-\frac{\alpha-2z^3+(1-\alpha)z^4}{2nz^2}
\right)+O\lpa{n^{-2}}.\end{aligned}$$ The lemma will then follow from the relations $$\begin{aligned}
\label{S1a}
S_1(\alpha) = [z^{-1}]
\frac{e^{\frac\alpha z+(1-\alpha)z}}{(1-z)^2},\end{aligned}$$ and $$\begin{aligned}
\label{T1a}
T_1(\alpha) &= [z^{-1}]
\frac{e^{\frac\alpha z+(1-\alpha)z}}
{(1-z)^2}\cdot\frac{-\alpha+2z^3-(1-\alpha)z^4}{2z^2}.\end{aligned}$$ To prove , we expand the factor $e^{\frac\alpha z}$ and take the coefficient term by term, yielding $$\begin{aligned}
[z^{-1}]\frac{e^{\frac\alpha z+(1-\alpha)z}}{(1-z)^2}
&= \sum_{\ell\ge0}\frac{\alpha^\ell}{\ell!}
[z^{\ell-1}]\frac{e^{(1-\alpha)z}}{(1-z)^2}\\
&= \sum_{\ell\ge0}\frac{\alpha^\ell}{\ell!}
\sum_{j=0}^{\ell-1}(\ell-j)\frac{(1-\alpha)^j}{j!}
= S_1(\alpha).\end{aligned}$$ Similarly, $$\begin{aligned}
\label{S0a}
S_0(\alpha) = [z^{-1}]
\frac{e^{\frac\alpha z+(1-\alpha)z}}{1-z},\end{aligned}$$ and by the decomposition, $$\frac{-\alpha+2z^3-(1-\alpha)z^4}{z^2(1-z)^2}
= \frac1{(1-z)^2}-\frac{4\alpha}{1-z}
-(1-\alpha)-\frac{2\alpha}z-\frac{\alpha}{z^2},$$ we obtain $$\begin{aligned}
&[z^{-1}]
\frac{e^{\frac\alpha z+(1-\alpha)z}}
{(1-z)^2}\cdot\frac{-\alpha+2z^3-(1-\alpha)z^4}{2z^2} \\
&\quad= \frac12
\sum_{\ell\ge0}\frac{\alpha^\ell}{\ell!}
\sum_{j=0}^{\ell-1}\frac{(1-\alpha)^j}{j!}
(\ell-j -4\alpha)\\
&\qquad-(1-\alpha)\sum_{\ell\ge1}\frac{\alpha^\ell}{\ell!}
\cdot\frac{(1-\alpha)^{\ell-1}}{(\ell-1)!}
-2\alpha \sum_{\ell\ge0}\frac{\alpha^\ell}{\ell!}
\cdot\frac{(1-\alpha)^{\ell}}{\ell!}\\
&\qquad-\alpha \sum_{\ell\ge0}\frac{\alpha^\ell}{\ell!}
\cdot\frac{(1-\alpha)^{\ell+1}}{(\ell+1)!},\end{aligned}$$ which equals $T_1(\alpha)$ by properly grouping the terms. This proves the lemma.
As we will see below, finer calculations give $$\begin{aligned}
\label{ST2}
\Delta_n^*(k)
= S_1(\alpha) + \frac{T_1(\alpha)}{n}
+ \frac{T_2(\alpha)}{n^2}
+ O\lpa{n^{-3}},\end{aligned}$$ where $$\label{T2a}
\begin{split}
T_2(\alpha)
&:= -\frac{S_1(\alpha)}{24}+\alpha S_0(\alpha)
+\frac{1+6\alpha}{12} \,I_0\lpa{2\sqrt{\alpha(1-\alpha)}}\\
&\qquad-\frac{1-10\alpha+4\alpha^2}{12\sqrt{\alpha(1-\alpha)}}
\,I_1\lpa{2\sqrt{\alpha(1-\alpha)}}.
\end{split}$$ In particular, when $\alpha\to0$, we have $T_2(\alpha)
=\frac43\alpha+\frac{215}{144}\alpha^2+\frac{13}{192}\alpha^3+\cdots$.
To obtain formula for $T_2(\alpha)$ we begin with the expression $$T_2(\alpha) = [z^{-1}]
e^{\frac\alpha z+(1-\alpha)z}\cdot
\frac{W_\alpha(z)}{(1-z)^2},$$ where
$$W_\alpha(z)
:= \frac{3\alpha^2+8\alpha z-12\alpha z^3
+6\alpha(1-\alpha)z^4-4(1-\alpha)z^7+3(1-2\alpha)z^8}
{24 z^4}.$$
By the decomposition
$$\begin{aligned}
\frac{W_\alpha(z)}{(1-z)^2}
&= \left(\begin{array}{l}
-\frac1{24}{(1-z)^2}\\
+\frac{\alpha}{1-z}
\end{array}\right)
+\left(
\begin{array}{l}
\frac{\alpha^2}8z^{-4} \\
+\frac{\alpha(4+3\alpha)}{12}z^{-3}\\
+\frac{\alpha(16+9\alpha)}{24}z^{-2}\\
+\frac{\alpha(1-\alpha)}2z^{-1}\\
\end{array}\right)
+\left(\begin{array}{l}
\frac{(1-\alpha)(1-9\alpha)}{24}\\
+\frac{(1-\alpha)(1-3\alpha)}{12}z\\
+\frac{(1-\alpha)^2}8z^2
\end{array}\right) ,
\end{aligned}$$
we then derive by a term-by-term translation using the relations , and $$\begin{aligned}
[z^{-1}] z^m e^{\frac\alpha z+(1-\alpha)z}
&= \sum_{\ell\ge \max\{0,-m+1\}}
\frac{\alpha^\ell(1-\alpha)^{m+\ell-1}}
{\ell!(m+\ell-1)!} \\
&= \left(\frac{1-\alpha}\alpha\right)^{(m-1)/2}
I_{m-1}\lpa{2\sqrt{\alpha(1-\alpha)}},\end{aligned}$$ for $m\in\mathbb{Z}$.
#### Proof of Theorem \[thm:S1D\].
Substituting the expansion into the partial sum $$Q_{\tr{n/2}} = \sum_{k=1}^{\tr{n/2}}
\frac1{\Delta_n^*(k)},$$ and using the expansion $$\begin{aligned}
&\frac1{S_1(\alpha)+\frac{T_1(\alpha)}{n}
+ \frac{T_2(\alpha)}{n^2}
+ O\lpa{n^{-3}}} \\
&\qquad= \frac1{S_1(\alpha)}
-\frac{T_1(\alpha)}{nS_1(\alpha)^2}
-\frac{S_1(\alpha)T_2(\alpha)-T_1(\alpha)^2}
{n^2S_1(\alpha)^3}+O\lpa{\alpha^{-2}n^{-3}}\end{aligned}$$ we obtain $$\begin{aligned}
Q_{\tr{n/2}} &= \sum_{k=1}^{\tr{n/2}}
\frac1{S_1\lpa{\frac kn}}
-\frac1{n}\sum_{k=1}^{\tr{n/2}}
\frac{T_1\lpa{\frac kn}}{S_1\lpa{\frac kn}^2}
+ E_1(n),\end{aligned}$$ where $$\begin{aligned}
E_1(n) &= -\frac1{n^2} \sum_{k=1}^{\tr{n/2}}
\frac{S_1\lpa{\frac kn}T_2\lpa{\frac kn}
-T_1\lpa{\frac kn}^2}{S_1\lpa{\frac kn}^3}\\
&\qquad\qquad\qquad +O\left(n^{-3}\sum_{k=1}^{\tr{n/2}}\frac{n^2}{k^2}\right).\end{aligned}$$ By the local expansion $$\frac{S_1(\alpha)T_2(\alpha)-T_1(\alpha)^2}
{S_1(\alpha)^3} = -\frac{11}{12}\, \alpha^{-1}
+\frac{341}{144}+\cdots,$$ we deduce that $$E_1(n) = O\left(n^{-1}\sum_{k=1}^{\tr{n/2}}k^{-1}+
n^{-1}\right)
= O\lpa{n^{-1}\log n}.$$ On the other hand, since most contribution to the sums come from terms with small $k$, we deduce, by using the expansion $$\frac{T_1(\alpha)}{S_1(\alpha)^2}
= -\frac{3}{2\alpha} +\frac{11}4 -\frac{15}4\,\alpha
+\cdots,$$ and the boundedness of $\frac{\alpha T_1(\alpha)}{S_1(\alpha)^2}$ on the unit interval, that $$-\frac1{n}\sum_{k=1}^{\tr{n/2}}
\frac{T_1\lpa{\frac kn}}{S_1\lpa{\frac kn}^2}
= \frac32\,H_{\tr{n/2}} +O(1)
= \frac32\,\log n +O(1).$$ Define $$R(z) := \frac1{S_1(z)}-\frac1z,$$ which is bounded in the unit interval. We have, by , $$Q_{\tr{n/2}} = nH_{\tr{n/2}}+\sum_{k=1}^{\tr{n/2}}R\Lpa{\frac kn}
+\frac32\,H_{\tr{n/2}}+O(1).$$ In view of the bounded derivative of $R$ in the unit interval, we then deduce, by a standard application of the Euler-Maclaurin summation formula (approximating the sum by an integral), that $$Q_{\tr{n/2}} = n\log n + C_0\, n + \frac32\log n + O(1),$$ where $$C_0 := \gamma-\log 2 +\int_0^{\frac12}
\left(\frac1{S_1(t)}-\frac1t\right)\dd t
\approx -0.69622\,72155\dots$$ By the relation $\Delta_n(k) = \Delta_{n-1}^*(k)\lpa{1-\frac1n}^n$, we then deduce , proving the theorem.
See Figure \[fig:differences\] for the graphical rendering of the various approximations derived.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\Delta_n^*(k)-S_1(\alpha)$ $\Delta_n^*(k)-S_1(\alpha)-\frac{T_1(\alpha)}{n}$ $\Delta_n^*(k)-S_1(\alpha)-\frac{T_1(\alpha)}{n}
-\frac{T_2(\alpha)}{n^2}$
{height="2.2cm"} {height="2.2cm"} {height="2.2cm"}
$\frac1{\Delta_n^*(k)}-\frac1{S_1(\alpha)}$ $\frac1{\Delta_n^*(k)}-\frac1{S_1(\alpha)}+\frac{T_1(\alpha)} $\begin{array}{l}
{nS_1(\alpha)^2}$ \frac1{\Delta_n^*(k)}-\frac1{S_1(\alpha)}
+\frac{T_1(\alpha)}{nS_1(\alpha)^2}\\+
\frac{T_1(\alpha)^2-T_2(\alpha)S_1(\alpha)}{n^2S_1(\alpha)^2}
\end{array}$
{height="2.2cm"} {height="2.2cm"} {height="2.2cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$\sum_{1\le k\le \tr{\frac n2}}\frac1{\Delta_n(k)}-\mathbf{E}(X_n)$ $\sum_{1\le k\le \tr{\frac n2}}\frac1{\Delta_n(k)}-\mathbf{E}(X_n)
-\frac{e}2\log n$
![Differences between the exact expected runtime and $\sum_{1\le k\le \tr{\frac n2}}\frac1{\Delta_n(k)}$ without (left) and with (right) the correction term $\frac e2\log n$.[]{data-label="fig:differences"}](SD-1 "fig:"){height="2.6cm"} ![Differences between the exact expected runtime and $\sum_{1\le k\le \tr{\frac n2}}\frac1{\Delta_n(k)}$ without (left) and with (right) the correction term $\frac e2\log n$.[]{data-label="fig:differences"}](SD-2 "fig:"){height="2.6cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Conclusions {#conclusions .unnumbered}
===========
We have revisited drift analysis for the fundamental problem of bounding the expected runtime of the [(1+1) EA]{}on the [[<span style="font-variant:small-caps;">OneMax</span>]{}]{}problem. Using novel drift theorems involving error bounds, we have bounded the expected runtime when starting from $\tr{n/2}$ ones, up to additive terms of logarithmic order; more precisely we have $$\sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)} - {\mathord{E}\hspace{0.3ex}\mathord{\left(T\mid X_0=\tr{n/2}\right)}} \in [c_1\log n, c_2\log n]$$ for explicitly computed constants $c_1,c_2>0$. This for the first time gives an absolute error bound for the expected runtime. Then by standard asymptotic methods, we have found that $\sum_{k=1}^{\tr{n/2}}\frac{1}{\Delta(k)}$ overestimates the exact expected runtime by a term ${(e/2)\log n+O(1)}$.
Acknowledgements {#acknowledgements .unnumbered}
================
Partially supported by an Investigator Award from Academia Sinica under the Grant AS-IA-104-M03.
[10]{}
Anne Auger and Benjamin Doerr. . World Scientific Publishing, 2011.
Francisco Chicano, Andrew M. Sutton, L. Darrell Whitley, and Enrique Alba. Fitness probability distribution of bit-flip mutation. , 23(2):217–248, 2015.
Benjamin Doerr and Carola Doerr. The impact of random initialization on the runtime of randomized search heuristics. , 75(3):529–553, 2016.
Benjamin Doerr, Carola Doerr, and Jing Yang. Optimal parameter choices via precise black-box analysis. In [*Proc. of GECCO ’16*]{}, pages 1123–1130. ACM Press, 2016.
Benjamin Doerr, Mahmoud Fouz, and Carsten Witt. Quasirandom evolutionary algorithms. In [*Proc. of GECCO ’10*]{}, pages 1457–1464. ACM Press, 2010.
Benjamin Doerr, Mahmoud Fouz, and Carsten Witt. Sharp bounds by probability-generating functions and variable drift. In [*Proc. of GECCO ’11*]{}, pages 2083–2090. ACM Press, 2011.
Benjamin Doerr, Christian Gie[ß]{}en, Carsten Witt, and Jing Yang. The (1+$\lambda$) evolutionary algorithm with self-adjusting mutation rate. , 81(2):593–631, 2019.
Benjamin Doerr, Daniel Johannsen, and Carola Winzen. Multiplicative drift analysis. , 64(4):673–697, 2012.
Stefan Droste, Thomas Jansen, and Ingo Wegener. On the analysis of the (1+1) evolutionary algorithm. , 276:51–81, 2002.
Josselin Garnier, Leila Kallel, and Marc Schoenauer. Rigorous hitting times for binary mutations. , 7(2):173–203, 1999.
Christian Gie[ß]{}en and Carsten Witt. Optimal mutation rates for the (1+$\lambda$) [EA]{} on [OneMax]{} through asymptotically tight drift analysis. , 80(5):1710–1731, 2018.
Bruce Hajek. Hitting-time and occupation-time bounds implied by drift analysis with applications. , 13(3):502–525, 1982.
Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary algorithms. , 127:57–85, 2001.
Hsien[-]{}Kuei Hwang, Alois Panholzer, Nicolas Rolin, Tsung[-]{}Hsi Tsai, and Wei[-]{}Mei Chen. Probabilistic analysis of the (1+1)-evolutionary algorithm. , abs/1409.4955, 2014. <http://arxiv.org/abs/1409.4955>.
Hsien-Kuei Hwang, Alois Panholzer, Nicolas Rolin, Tsung-Hsi Tsai, and Wei-Mei Chen. Probabilistic analysis of the (1+1)-evolutionary algorithm. , 26:299–345, 2018.
Thomas Jansen. . Natural Computing Series. Springer, 2013.
Daniel Johannsen. . PhD thesis, Universit[ä]{}t des Saarlandes, Germany, 2010.
Timo K[ö]{}tzing and Martin S. Krejca. First-hitting times for finite state spaces. In [*Proc. of PPSN ’18*]{}, pages 79–91. Springer, 2018.
Timo K[ö]{}tzing and Martin S. Krejca. First-hitting times under additive drift. In [*Proc. of PPSN ’18*]{}, pages 92–104. Springer, 2018.
Per Kristian Lehre. Drift analysis (tutorial). In [*Companion to GECCO 2012*]{}, pages 1239–1258. ACM Press, 2012.
Per Kristian Lehre and Carsten Witt. Black-box search by unbiased variation. , 64(4):623–642, 2012.
Per Kristian Lehre and Carsten Witt. General drift analysis with tail bounds. Technical report, 2013. <http://arxiv.org/abs/1307.2559>.
Per Kristian Lehre and Carsten Witt. Concentrated hitting times of randomized search heuristics with variable drift. In [*Proc. of ISAAC ’14*]{}, pages 686–697. Springer, 2014.
Johannes Lengler. Drift analysis. , abs/1712.00964, 2018. To appear as a book chapter in Theory of Evolutionary Algorithms in Discrete Search Spaces (eds. B. Doerr and F. Neumann), Springer.
Boris Mitavskiy, Jonathan E. Rowe, and Chris Cannings. Theoretical analysis of local search strategies to optimize network communication subject to preserving the total number of links. , 2(2):243–284, 2009.
Heinz M[ü]{}hlenbein. How genetic algorithms really work: I. [M]{}utation and hillclimbing. In [*Proc. of PPSN ’92*]{}, pages 15–26. Elsevier, 1992.
Jonathan E. Rowe and Dirk Sudholt. The choice of the offspring population size in the (1, [$\lambda$]{}) evolutionary algorithm. , 545:20–38, 2014.
Dirk Sudholt. General lower bounds for the running time of evolutionary algorithms. In [*Proc. of PPSN ’10*]{}, pages 124–133. Springer, 2010.
Carsten Witt. Tight bounds on the optimization time of a randomized search heuristic on linear functions. , 22(2):294–318, 2013.
| {
"pile_set_name": "ArXiv"
} |
[**On a realistic interpretation of quantum mechanics**]{}
[**Arnold Neumaier**]{}\
\
Institut für Mathematik\
Universität Wien\
Strudlhofgasse 4\
A-1090 Wien, Austria\
email: [email protected]\
WWW: http://solon.cma.univie.ac.at/neum/\
\
[**Abstract.**]{} The best mathematical arguments against a realistic interpretation of quantum mechanics – that gives definite but partially unknown values to all observables – are analysed and shown to be based on reasoning that is not compelling.
This opens the door for an interpretation that, while respecting the indeterministic nature of quantum mechanics, allows to speak of definite values for all observables at any time that are, however, only partially measurable.
The analysis also suggests new ways to test the foundations of quantum theory.
[**Keywords**]{}: realistic interpretation, quantum inseparability, probability, projector, measurement, spin, classical model, particle paths, double slit experiment\
[**1991 MSC Classification**]{}: primary 81P10\
[**1990 PACS Classification**]{}: 03.65\
Introduction {#intro}
============
Quantum physics is a very successful theory for predicting nature. However, in spite of many attempts, a mathematically and philosophically convincing basis for the interpretation of quantum phenomena has not yet been found.
As stated recently by [Zeilinger]{} [@Zei], there are at least two levels of interpreting quantum mechanics: the statistical interpretation in the narrower sense introduced by [Born]{} [@Bor], on which there is almost complete consensus between physicists, and the interpretation of the meaning of the quantum mechanical concepts, where no agreement has been reached and all existing interpretations have been found wanting.
The ‘orthodox’ Copenhagen interpretation in terms of state reduction by external measurement loses credit because it becomes meaningless for the universe as a whole. There have been a number of proposals to change the structure of quantum mechanics, e.g., through pilot waves ([Bohm]{} [@Bohm]), consistent histories (e.g., [Omn[è]{}s]{} [@Omn]), event enhancement ([Blanchard & Jadczyk]{} [@BlaJ]), or gravitational objective reduction ([Penrose]{} [@Pen]).
Perhaps the main reason why there is so little progress on the meaning of quantum mechanical concepts is that it is usually seen as thoroughly enmeshed with measurement problems. I believe that this is a mistake, and that more clarity can be obtained by separating the analysis of meaning from that of observability. The discussion thereby becomes more concise and clear.
Here I refer to [*meaning*]{} as a conceptual, logically consistent mathematical framework that allows one to speak unambiguously about [*all*]{} the terms used in the theory, in a way intuitively related to corresponding concepts of external reality. This part is often elegant and concise.
On the other hand, I refer to [*observability*]{} questions (or [*measurement*]{} problems) as the operational explanation of how to obtain quantitative values for the observables of a system, preferably within the framework of a well-defined mathematical theory. This part is usually messy, and it is the part that I propose to avoid until the other part is satisfactory. (This also saves me from entering a discussion about the meaning of words like ‘operational’ or ‘obtain’.)
For example, in classical real analysis, infinitesimals are an ill-defined, meaningless concept, though they can be approximately realized, while random sequences are a well-defined, meaningful concept, though there is no constructive way of finding one, except approximately.
A historical case supporting the power of this view of separating the analysis of meaning from that of observability is the uncertainty principle. It was argued by [Heisenberg]{} [@Hei] in terms of observability, and later by [Robertson]{} [@Rob] in mathematical terms as a simple consequence of the canonical commutation relation $[p,q]=-i\hbar$. The mathematical argument takes only a few lines and is simple and compelling; the discussion of observability is complex but shows that the results of the mathematics cannot be in conflict with what can in principle be measured by experiment.
Further support for the positive effect of the separation of meaning from observability is given by Bell’s inequality ([Bell]{} [@Bel], [Clauser]{} et al. [@ClaHS]), a purely algebraic statement whose clarity is compelling, and the subsequent verification of its violation through experiments by [Aspect]{} [@Asp].
Finally, it seems that measurement problems can be adequately analysed by generalized observables defined as positive operator valued (POV) measures; see, e.g., [Davies]{} [@Dav], [Busch]{} et al. [@BusGL]. The gradual and approximate state reduction can be explained thermodynamically through dissipative interaction by the coupling to a macroscopic apparatus or heat bath; see, e.g., [Zurek]{} [@Zur], [Joos & Zeh]{} [@JooZ], [Ghirardi, Rimini & Weber]{} [@GhiRW].
That obtaining information about a system and state reduction are not equivalent things is a consequence of the possibility of nondemolition measurements ([Braginsky]{} et al. [@BraVT], [Kwiat]{} et al. [@KwiWH]) that produce knowledge about a system but (almost) avoid state reduction. And since the measuring apparatus necessarily involves a huge number of particles, any satisfactory solution of measurement problems should not be a part of the foundational concepts and these but be obtained as their [*consequence*]{}.
In an attempt to separate the discussion about the meaning of quantum mechanics from that of experimental consistency, I searched the literature for concise statements of the basic difficulties in terms unrelated to measurement, but given in a clear mathematical context that allows one to analyse the logic of the problems without losing oneself in philosophical speculations.
The outcome was somewhat surprising: The best presented mathematical arguments turned out to have subtle flaws. This becomes apparent when one reduces them to arguments residing completely within the universally accepted statistical interpretation in the narrower sense.
Probably the weaknesses in the arguments escaped previous attention since the arguments were only used for philosophical discussion and justification of foundations, and never subjected to experiment. I therefore suggest below some specific things that one should try to test with suitable experiments.
In the following, I shall present my analysis of some arguments related to a realistic interpretation of quantum mechanics; [*realistic*]{} in the sense that it allows to speak of definite values for [*all*]{} observables at any time that are, however, only partially measurable.
The findings give rise to the hope that a consistent interpretation is possible that respects the stochastic nature of quantum mechanics, but is also realistic, with all the accompanying advantages for our intuition. In a companion paper ([Neumaier]{} [@Neu1]), a proposal for such a realistic interpretation will be presented.
Feynman’s argument {#feynman}
==================
In his 1948 paper “Space-time approach to non-relativistic quantum mechanics” [@Fey], Richard [Feynman]{} gives on pp.368-369 the following argument.
[*“We define $P_{ab}$ as the probability that if the measurement $A$ gave the result $a$, then measurement $B$ will give the result $b$. \[...\] denote by $P_{abc}$ the probability \[...\] if $A$ gives $a$, then $B$ gives $b$, and $C$ gives $c$. \[...\] If the events between $a$ and $b$ are independent of those between $b$ and $c$, then P\_[abc]{}=P\_[ab]{}P\_[bc]{}. This is true according to quantum mechanics when the statement that $B$ is $b$ is a complete specification of the state. In any event, we expect the relation P\_[ac]{}=\_b P\_[abc]{}. Now, the essential difference between classical and quantum physics lies in Eq. . In classical mechanics it is always true. In quantum mechanics it is often false. \[...\] The classical law, obtained by combining and , P\_[ac]{}=\_b P\_[ab]{}P\_[bc]{} is replaced by \_[ac]{}=\_b \_[ab]{}\_[bc]{}. If is correct, ordinarily is incorrect. \[...\] Looking at probability from a frequency point of view simply results from the statement that in each experiment giving $a$ and $c$, $B$ had some value. \[...\] Noting that gzit[e5]{} replaces only under the circumstance that we make no attempt to measure $B$, we are lead to say that the statement, ‘$B$ had some value,’ may be meaningless whenever we make no attempt to measure $B$.”* ]{}
Feynman’s argument sounds convincing but it has a flaw in the assumption of , the conditional independence of $c$ from $a$, given $b$, for all states summed over in . This assumption is inconsistent with quantum mechanics.
Indeed, one of the characteristic features of quantum mechanics is its inseparability, the impossibility of splitting a quantum mechanical system into two that are independent. According to the Copenhagen interpretation ([Bohr]{} [@Boh]), and verified abundantly by experiment, the [*whole*]{} space-time set-up must be specified to determine the correct outcome of an experiment. The contradiction obtained by Feynman in comparing and therefore need not have the consequence that $B$ sometimes has no value, but serves more naturally as [*a strong argument for quantum inseparability*]{}.
The Markov property does typically not even hold for consecutive events; due to the inherent noncommutativities, states “remember” something about their whole history and “anticipate” something about their whole future. To see this, pick $a,b,c$ as observations of the same observable at increasing times (for memory) and at decreasing times (for anticipation). It should be an interesting project to find the conditions under which (and the accuracy to which) the Markov property for a completely observed time series of a sufficiently isolated system, lost in the quantum domain, is recovered in a thermodynamic limit of a large number of particles.
We now show in detail that is inconsistent with quantum mechanics. According to the standard interpretation of probabilities as expectations of projection operators we have $P_{ab}=\<a|\Pi_b|a\>$ and $P_{bc}=\<b|\Pi_c|b\>$, where $\Pi_b=|b\>\<b|$ is the projector to state $b$ (and similarly for $c$). In order that $P_{abc}$ makes sense in the standard interpretation we need $P_{abc}=\<a|\Pi_{bc}|a\>$ with a projector $\Pi_{bc}$ expressing the proposition $\Pi_b \& \Pi_c$ that $B$ is in state $b$ and $C$ is in state $c$.
In orthodox quantum logics ([Birkhoff & von Neumann]{} [@BirN], [Svozil]{} [@Svo]), one considers a projector as representing the subspace of eigenstates to the eigenvalue 1 of the projector and takes the logical operation $\&$ as subspace intersection. One finds that, whenever $\Pi_b \ne \Pi_c$, we have $\Pi_b \& \Pi_c = 0$ and thus $P_{abc}=0$. This is different from , proving that has nothing to do with quantum mechanics. And when $\Pi_b = \Pi_c$, we have $\Pi_b \& \Pi_c = \Pi_c$, hence $P_{abc}=\<a|\Pi_c|a\>=P_{ac}$, and turns out to be [*always*]{} true in the quantum logic calculus, in contrast to what Feynman claimed.
On the other hand, the statement $P_{abc}=0$ if $\Pi_b \ne \Pi_c$ seems somewhat counterintuitive, since if $\Pi_b\approx \Pi_c$ one might want to have $P_{abc}\approx P_{ac}$. In view of the fact that, unless $\Pi_b$ and $\Pi_c$ commute, there is no simple [*algebraic*]{} expression relating $\Pi_{bc}$ to $\Pi_b$ and $\Pi_c$, one is not compelled to adhere to the rigid assignment $\Pi_b \& \Pi_c = 0$ and might prefer to consider $\Pi_b \& \Pi_c$ to be undefined if $\Pi_b$ and $\Pi_c$ do not commute, interpreting this as an indication that $P_{abc}$ cannot be predicted (in principle) by the formalism of quantum theory (and thus may have any value).
In view of the fact that the counterintuitive abundance of zero projectors has further undesirable consequences to be discussed in Section \[path\], this would seem a more satisfactory solution of Feynman’s argument. A new quantum logic consistent with this interpretation is discussed in [Neumaier]{} [@Neu1].
Wigner’s claim {#wigner}
==============
Similar discussions as those given by Feynman usually suffer from the same problem that the claimed probability cannot be written as the expectation of a quantum proposition, i.e., a projector. For example, in his 1976 lecture notes on the ‘interpretation of quantum mechanics’, [Wigner]{} [@Wig p.288] writes,
[*\[...\] it remains essentially correct to say that the basic statement of quantum mechanics can be given in a formula as simple as .* ]{}
In adapted notation, for a finite sequence $a,b,c,d\dots$ of results of measurements of $A,B,C,D,\dots$, it reads
[ll]{} P\_[abcd...]{}&=\_a\_b\_c\_d \_d\_c\_b / \_a\
&=a|\_b\_c\_d \_d\_c\_b|a
Now there is a problem with . It implies that repeating the measurement of $A$ immediately after a first measurement of $A$ gives the result of the first measurement with certainty. This is possible only if we assume that the measurements performed are ideal.
But (cf. Wigner’s discussion in pp.283-284 of [@Wig]; see also [@JooZ]) one cannot make an ideal measurement of an observable with a continuous spectrum; and the quantum mechanical analysis of the measurement process shows that an ideal measurement of any quantity takes, strictly speaking, an infinite time. And even allowing for that, the only measurable observables would be the functions of scattering invariants (p.298). Hence formula can apply to real situations only approximately.
Therefore, unlike Schroedinger’s equation, cannot be considered a ‘basic statement’ and must be banned from the foundations. It should rather be a consequence of more elementary, exact features of the theory.
Indeed, Wigner ‘derives’ from first principles. The argument, given on pp.286-287, is identical to Feynman’s, assuming (without comment) the Markov property that provides the independence that allows to multiply probabilities:
[*If the first observation \[...\] This probability is $|(a_\kappa,b_\lambda)|^2$. The probability that the next measurement \[...\] is then $|(b_\lambda,c_\mu)|^2$ and the probability of both outcomes \[...\] is $|(a_\kappa,b_\lambda)(b_\lambda,c_\mu)|^2$ \[...\].* ]{}
Thus the formula, based on invalid reasoning, is suspect.
This suspicion is confirmed by realizing that the operator whose expectation is taken is generally not a projector, hence does not correspond to a ‘proposition’ in the traditional quantum logic calculus. But what sense should it make to talk about the probability of a statement that is not even logically well-formed?
It seems to me that the only formula for probabilities verified (abundantly) by experiments is the formula (see, e.g., equation (1.8) in Chapter 3 of [Davies]{} [@Dav]) P(B E|)=B(E) for the probability that, in a mixed state with the density matrix $\rho$, the generalized observable (POV measure) $B$ has a value from a set $E$. For ordinary observables $B$ and singleton sets $E=\{b\}$, the ‘effect’ $B(E)$ reduces to a projector $\Pi_b$. Equation contains a special case of only,
[ll]{} P\_[ab]{}&=\_a\_b / \_a\
&= a|\_b|a= |a|b|\^2 \_a=|aa|, a|a=1.
It is interesting to note that Wigner’s formula reappears as the basic formula in the theories of consistent histories (e.g., equation (4.3) in [Omn[è]{}s]{} [@Omn]). However, as mentioned on p.143 of [@Omn], only the case of two reference times can be proved, and this is equivalent to (with $\rho$ replaced by $E_1\rho E_1$, the mixed state obtained after the first state reduction).
However, we can give the probability an alternative, correct meaning: If $a$ is a pure state then implies that $P_{abcd}$ is the probability of measuring $\Pi_a\Pi_b\Pi_cD\Pi_c\Pi_b\Pi_a$ in that state! In general, therefore, Wigner’s formula gives the probability of measuring in state $a$ the last element in a finite sequence $A$, $\Pi_aB\Pi_a$, $\Pi_a\Pi_bC\Pi_b\Pi_a$, $\Pi_a\Pi_b\Pi_cD\Pi_c\Pi_b\Pi_a$, $\dots$, assuming that the measurements resulted in reductions of the intermediate states to the states represented by the projectors $\Pi_a, \Pi_b, \Pi_c, \dots$.
To see how we can possibly interpret this in terms of the uncontroversial part of quantum mechanics, we consider a quantized relative of Laplace’s demon. Suppose there were a quantum demon with the unusual capacity to ‘see’ every detail of a closed system, without interacting with the system. (The demon doesn’t need to be physically realizable; this is just a fictional argument to make a vivid point.) The demon leaves the whole system (consisting of the measured system together with the measuring device) undisturbed and only interprets our claims on measurements and compares it with the dynamics it sees.
The demon notes the time intervals $t,u,v,\dots$ between the successive measurements $A,B,C,D,\dots$. What does it see us measure? In the Heisenberg picture, with unlabelled operators at the time of the first measurement, the state of the complete system remains unchanged, and the measurements are those of $A,B(t),C(t+u),D(t+u+v),\dots$. If we use the Heisenberg dynamics and introduce the commuting operators $\Phi_a=\exp(itH/\hbar)$, $\Phi_b=\exp(iuH/\hbar)$, $\Phi_c=\exp(ivH/\hbar)$, we see that the demon sees us measure $A, \Phi_a^*B\Phi_a, \Phi_a^*\Phi_b^*C\Phi_b\Phi_a,
\Phi_a^*\Phi_b^*\Phi_c^*D\Phi_c\Phi_b\Phi_a, \dots$. (This is quite different from what Wigner asserts in his equation (43), where he assumes an [*additional*]{} Heisenberg dynamics of the operators between the state-reducing measurements.)
Comparing with the above interpretation of , and noting that projectors are Hermitian, we see that the difference between the unproved formula and the demon’s objective description of our sequence of measurements is the replacement of the commuting unitary $\Phi$’s in the complete system by the noncommuting idempotent Hermitian $\Pi$’s in the measured system alone.
Thus is valid precisely to the extent that the time-dependent operators (in the Heisenberg picture), as calculated by the Heisenberg equation for the [*complete*]{} experimental setup at the time the next measurement is taken, can be approximated in the assumed (time-independent) state of the investigated [*subsystem*]{} by a two-sided multiplication with these projectors.
This shows clearly [*the aim of a correct measurement theory:*]{} It must exhibit broad sufficient conditions for a system that guarantee, for all observables to be measured, the validity of this approximation, and hence a complete reduction of the wave packet.
I’d like to challenge the adherents of to devise and perform experiments testing validity, accuracy and limits of when more than two (noncommuting) projectors are involved and the measurements are not ideal. The findings should coincide with the interpretation in the Heisenberg picture just given. (For the realization of arbitrary discrete projectors, note that these can always be brought into the form $UDU^*$ with diagonal projectors $D$ and unitary $U$; hence [Reck]{} et al. [@RecZB] applies.)
On the basis of such experiments it might also be possible to decide whether it is at all possible to perform experiments for measurements of probabilities that cannot be expressed as the expectation of a projector.
Schr[ö]{}dinger’s argument {#schroe}
==========================
In his 1936 paper “Die gegenw[ä]{}rtige Situation in der Quantenmechanik” [@Sch] (where the famous cat paradox appears), Erwin [Schr[ö]{}dinger]{} gives the following argument (pp. 156-157 of the English translation):
*At first thought one might well attempt likewise to refer back the always uncertain statements of Q.M. to an ideal ensemble of states, of which a quite specific one applies in any concrete instance – but one does not know which one. That this won’t work is shown by the one example of angular momentum, as one of many.*
Imagine \[...\] the point $M$ to be situated at various positions relative to $O$ and fitted with various momentum vectors, and all these possibilities combined into an ideal ensemble. Then one can indeed choose these positions and vectors that in every case the product of vector length by length of normal $OF$ \[where $F$ is the point closest to $O$ on the line through $M$ along the momentum vector\] yields one or the other of the acceptable values – relative to the particular point $O$. But for an arbitrary different point $O'$, of course, unacceptable values occur. \[...\]
One could go on indefinitely with more examples. \[...\] Already for the single instant things go wrong. At no moment does there exist an ensemble of classical states of the model that squares with the totality of quantum mechanical statements of this moment. \[...\] we saw that it is not possible smoothly to take over models and to ascribe, to the momentarily unknown or not exactly known variables, nonetheless determinate values, that we simply don’t know.
The problem is indeed unsolvable if one insists on the existence of a [*single*]{} vector $J$ of values determining all linear combinations $u\cdot J$ of the angular momentum. For example, in measuring the inner product $u\cdot J$ of the angular momentum $J$ with four or more triplewise linearly independent unit vectors $u$, one obtains in each case values from the discrete spectrum of $u\cdot J$, and these spectra are inconsistent with a precise value of $J$. This is explained by Schr[ö]{}dinger in detail on p. 164 in the context of a simpler one-parameter example involving the observables $p^2+a^2q^2$ where $a$ is a parameter. Commenting on it, Schr[ö]{}dinger writes on p. 165:
[*Should one now think that because we are so ignorant about the relations among the variable-values held ready in one system, that none exists, that far-ranging arbitrary combination can occur? That would mean that such a system of “one degree of freedom” would need not merely two numbers for adequately describing it, as in classical mechanics, but rather many more, perhaps infinitely many.* ]{}
Schr[ö]{}dinger dismisses this as not viable, but as I’ll show now, there is a serious possibility that precisely this is the case, thus invalidating Schr[ö]{}dinger’s conclusion that it is not possible to ascribe to the momentarily unknown or not exactly known variables determinate values. His conclusion is only valid if one ascribes to each observable vector $K$ a determinate value that is to be used in the calculations for [*all*]{} $u\cdot K$.
However, as was observed already by [von Neumann]{} [@vNeu IV.1.E], one cannot, in general, combine the measuring recipes for two noncommuting observables to one for their sum, so that the sum of two observables is only implicitly characterized through the axioms. This should forbid the naive use of values for the components of $K$, say, to calculate the values of $u\cdot K$.
As I shall show now, one may consider each $u\cdot K$ as a (not necessarily classical) random variable in its own right, and determine their relationship for different $u$ not by ordinary algebra but by the statistics derived from the standard quantum mechanical recipes.
For simplicity, I’ll take for $K$ in place of the angular momentum $J$ the Pauli spin vector $\sigma$ which shows precisely the same problems as $J$. For each unit vector $u\in \Rz^3$, the operator $u\cdot \sigma$ has the simple eigenvalues $1$ and $-1$; thus these are the possible values of $u\cdot \sigma$. The projector to an eigenstate of $u\cdot \sigma$ corresponding to the eigenvalue $1$ is $\Pi_u:= \half(1+u\cdot \sigma)$, with $\tr \Pi_u=1$, and the corresponding projector for the eigenvalue $-1$ is simply $\Pi_{-u}$. For any unit vector $v\in \Rz^3$, the probability for $v\cdot \sigma$ having the value $1$ when $u\cdot \sigma$ has the value $1$ is then, according to ,
[ll]{} P\_[uv]{}&=\_u\_v = (1+u)(1+v)/4\
&=(1+(u+v)+uv+i(u v))/4\
&=(1+uv),
and the probability for $v\cdot \sigma$ having the value $-1$ when $u\cdot\sigma$ has the value $1$ is therefore $P_{u,-v}= \half(1-u\cdot v)$, adding up to $1$, as it should be. Moreover, $P_{u,-u}=0$, as it should be. The probabilities depend continuously on $u$ and $v$, in a very natural way.
One would get precisely the same probabilities if one had extremely fragile, classical spheres, painted white on one hemisphere and black on the other one, and each sphere would be destroyed after observing the color at a single point. One could still calculate probabilities for the colors of the other points, and any rotationally symmetric classical probability model produces precisely the same probabilities as we just calculated for the spin.
More details on such a classical model for electron spin can be found in [Kochen & Specker]{} [@KocS]. They show that, under natural assumptions, such a classical model is restricted to two-state systems. (In a companion paper [@Neu1], I’ll show how one can reinterpret quantum mechanics in such a way that the assumptions of Kochen and Specker become irrelevant, thus removing this very restictive conclusion.) In any case, one such model is enough to invalidate the cogency of Schr[ö]{}dinger’s conclusion.
To summarize, Schr[ö]{}dinger’s argument demonstrates that, in quantum mechanics, one cannot calculate the values for linear combinations of noncommuting observables from the values of the observables themselves; but this is already obvious from the properties of the spectrum of operators. Only the values of functions of [*commuting*]{} observables (and, in particular, of functions of a single observable) can be predicted with certainty from the values of the observables themselves.
However, as was shown for the case of linear combinations of the Pauli spin matrices, his argument does not demonstrate that one cannot assign natural values to all observables that lead to natural and consistent classical probabilities.
The nature of quantum observations may put severe limits on what is observable through experiment and how these are combined to estimate the values of other observables, but it does not seem to put restrictive limits on the joint values of noncommuting observables. Indeed, [von Neumann]{} [@vNeu] shows in the same section mentioned above how to get estimates of joint probabilities of noncommuting quantities if sufficiently large ensembles are available.
Particle paths {#path}
==============
It is widely believed that it is impossible to ascribe definite paths to moving quantum particles, except at special points where the position has been measured. The particle paths seen in a cloud chamber, say, can still be explained as an illusion created by correlation patterns among the atoms ionized by the particles; see [Mott]{} [@Mot]. No such analysis seems to be available, however, that would explain why we see macroscopic bound states (such as you or me) move along fairly definite paths.
The traditional context for discussing the inconsistency of definite particle paths is the double slit experiment ([Bohr]{} [@Boh]), where it is claimed that electrons (or photons) passing a diaphragm containing the two slits cannot be said to have passed through one or the other slit, or – even worse – are said to have passed through both, in a way (superposition) that defies intuition.
Unfortunately, the discussion seems always to be connected to measurement questions, so that it is difficult to discern the mathematical content of the arguments that claim [*nonexistence*]{} of the path (in contrast to its [*limited measurability*]{} only, [Wootters & Zurek]{} [@WooZ]). However, I gathered some indirect evidence about particle paths by reading between the lines of some papers.
[Feynman]{} [@Fey], in laying ground work for his (nonrelativistic) path integral formalism, discusses on p. 371 the probability of particle paths as his first postulate:
[*If an ideal measurement is performed to determine whether a particle has a path lying in a region of space-time, then the probability that the result will be affirmative is the absolute square of a sum of complex contributions, one from each path in the region.* ]{}
This postulate is used to show (among other things) that the most likely paths are those close to the path determined by the least action principle.
No comment is given on the precise mathematical meaning of the terms involved (except for the term “each”); thus we need to see whether we can interpret this in an orthodox way. In the spirit of the standard interpretation of quantum mechanics, we need to assign to each (let us say open and bounded) region $\Omega$ of space-time a projector $\Pi_\Omega$ to the subspace $\Hz_\Omega$ of all wave functions belonging to states where the particle is with certainty in $\Omega$.
Unfortunately, particle positions at different times are represented by operators $x(t)=\exp(iHt)x(0)\exp(-iHt)$ that generally do not commute with each other, with the consequence that it is unlikely that they have common eigenvectors, which would be the elements of $\Hz_\Omega$. Thus, it is very likely that, for most $\Omega$, this subspace consists of zero only; and $\Pi_\Omega$ would vanish. This casts serious doubt on the applicability of the quantum logic recipe to defining the projector $\Pi_\Omega$.
Perhaps it is even impossible to make consistent assignments of all projectors $\Pi_\Omega$, in view of the fact that alternative attempts in consistent histories interpretations run into contradictions ([Kent]{} [@Ken]). (But possibly these contradictions are due to the non-projector nature of these histories; cf. [Isham]{} [@Ish].)
However, we can be more modest and only look at special sets of paths for which we may define proper projectors. For example, we may define the distance of two paths $\xi^k:[\alpha,\omega]\to \Rz^3$ ($k=1,2$) by d(\^1,\^2)=. Then, for each specific path $\xi$, we can use the position operators $x(t)=\exp(iHt)x(0)\exp(-iHt)$ to define the observable $\Delta_\xi:=d(x,\xi)$ that measures the distance of the particle path from $\xi$. Suitable projectors are now the projectors $\Pi_{\xi,\eps}$ projecting to the subspace spanned by the wave functions corresponding to states in which $\Delta_\xi$ has with certainty a value not exceeding $\eps$. Then, for a system in an arbitrary pure state $a$, P(d(x,))=a|\_[,]{}|a is, according to the most orthodox interpretation of quantum mechanics, a well-defined quantum mechanical probability for the event that the particle path is, in the root mean square sense, within $\eps$ of the path $\xi$. Note that, in principle, this formula can be tested experimentally; and by varying $\eps$, the complete distribution of $d(x,\xi)$ can be found. The natural path to be used for $\xi$ is the expectation (t)=a|x(t)|a or a computable approximation to it (such as the classical path of the particle).
With formulas such as – available, it would be completely unreasonable to assume that a particle has no definite path. If it hadn’t, what should be the meaning of its expectation and of the probability of it being close to a given path!?
It remains to discuss how this finding can be reconciled with the double slit experiment. In an implementation of von Neumann’s method mentioned in Section \[schroe\] to measure probabilities for two noncommuting observables, [Wootters & Zurek]{} [@WooZ] measure in a modified double-slit experiment both position and momentum of particles passing a slit:
*If the measured momentum is positive, then we will guess that the photon passed through slit $A$; if it is negative, then we will guess that the photon passed through slit $B$. Clearly some of our guesses will be wrong – there are photons that have positive values of measured momentum even though their actual momentum was negative and they went through slit $B$.*
\[...\] these two measurements cannot both be performed for the same photons, and so \[the figures\] cannot refer to the same experiment. Hence, in accord with the Copenhagen interpretation of quantum mechanics, there is no paradox. The complementarity principle does not prevent photons from behaving once as waves and once as particles. It only states that the same photon should not reveal this “split personality” in the same experiment. \[...\]
Despite the fact that we know with $99\%$ certainty the paths of the photons, they still have strong wavelike properties. \[...\] we have presented a result which, although not paradoxical, was nevertheless surprising (that is, that one can make a fairly precise determination of the slit through which each photon comes with only a slight disturbance of the interference pattern).
The finding loses its surprise in the light of the above considerations that suggest existence of the path but with intrinsic limitations on its observability.
For the original double-slit experiment, the arguments generally agree on the observation that there is no way to predict through which slit a particle [*will pass*]{}. However, it is possible to compute with high confidence through which a particle [*has passed*]{} when it has been observed at the photographic plate. To do so, one only needs to compute the probabilities for two typical paths $\xi$ connecting each slit with the position where the particle was absorbed by the photographic plate. By comparing the likelihood ratios at a given significance level one will find an exceedingly high likelihood for most particles having passed through the slit closer to their recorded position on the photographic plate.
Thus, after the event, enough information is available to decide reasonably reliably on the particle path. One could also calculate the values of the actions of the paths through the two slits ending at the observed position, and use the stationary phase approximation of the Feynman path integral to get (approximations to) the required probabilities. In any case, the statistical analysis is essentially similar to that used for all experiments where an event must be reconstructed from indirect measurable information and from the way the experiment was prepared.
Conclusion {#conc}
==========
The strongest arguments against realism available in the literature dissolved under a scrutinized analysis, using only that part of quantum mechanics for which there is almost universal agreement about its validity.
This opens the door for an interpretation that, while respecting the indeterministic nature of quantum mechanics, allows to speak of definite but only partially measurable values for all observables at any time. This kind of realism is consistent with the intrinsic indeterminism required by Heisenberg’s uncertainty relation if we distinguish carefully between [*what is*]{} and [*what is measurable*]{}.
Though not proved by the present investigations, it appears that, independent of the detailed description, a cautious realistic interpretation of quantum mechanics is in full accord with the generally accepted quantum mechanical formalism.
Here [*realistic*]{} means that in a completely specified state (not to be confused with the ‘pure states’ of quantum mechanics) [*all*]{} observables have definite values at all times that are, however, only partially measurable, according to the stochastic predictions of quantum mechanics.
And [*cautious*]{} means, that one has to take into account the following four restrictions:
1\. Only probabilities defined by orthogonal projectors (for sharp observables) or POV measures (for smeared observables) via equation are measurable and predictable.
2\. For noncommuting projectors, the logical operation ‘and’ is not defined; the corresponding questions may be asked (and may have definite, i.e., logically consistent answers) but a precise answer cannot be found by experiments.
3\. From the values of commuting observables one can deduce values of functions of these observables, but from the values of noncommuting observables one cannot even deduce values of their linear combinations. Instead, functions of noncommuting observables must be considered as random variables on their own whose expectation values (and probability distributions) must be calculated from the general quantum mechanical formalism and not from a classical inference of measured linear combinations.
4\. Due to quantum inseparability, the Markov property for a completely observed time series is lost.
A constructive proposal for such an interpretation will be given in a companion paper [@Neu1].
Appendix: Some final speculations {#spec .unnumbered}
=================================
In this section I want to be more speculative, at the risk of being less precise, less cogent, and more vulnerable about some issues discussed. I shall mention some ideas and ‘conclusions’ that are related to the preceding analysis though far from being consequences of the rigorous observations of the preceding sections.
A realistic interpretation of quantum mechanics is independent of the measurement problem and in better accord with classical (say, 19th century) intuition. As we have seen, there seem to be no longer strong arguments against a cautious realistic interpretation of quantum mechanics.
A problem reamining, and one that obscured for a long time the underlying simplicity, is the lack of observable information on the quantum level. As [Schr[ö]{}dinger]{} [@Sch] puts it (p.159 of the English translation), referring to the complementarity between position and momentum,
[*\[...\] the momentary statement content of the $\psi$-function is far from complete; it comprises only about $50$ percent of a complete description.* ]{}
However, unlike Schr[ö]{}dinger, I conclude that, while the remaining 50 percent will probably always remain unobservable, they may have a reality just as objective as the (ideally) observable 50 percent. The fact that in many cases we can choose which 50 percent of a complete description we want to observe underlines this conclusion.
The fact that both past and future boundary conditions are needed (and sufficient) to locate a particle path is consistent with the observation in Section \[feynman\] that states anticipate something about their whole future. It is also reminiscent of the action-at-a-distance formulation of classical electrodynamics by [Wheeler & Feynman]{} [@WheF; @WheF2], where electromagnetic radiation is explained through interaction with particles in the future.
By observing the right $50\%$, namely past [*and*]{} future position boundary conditions about [*one*]{} of a pair of conjugate variables at both boundaries, we might be able, in principle, to reconstruct the complete intermediate picture as reliably as in the classical case, where one has the same option. But in the classical case one has the additional and, for us as subjects acting based on past information only, more useful option to predict the particle path from initial conditions only (position and momentum in the past).
This state of affairs can be summarized in the statement:
[ *Physics essentially describes nature as if everything had already happened, and then expresses its laws as information about observed correlations. Since some of the correlations involve time, it is possible to partially predict the future from the past, or the past from the future, or an intermediate situation from past and future observations.* ]{}
This summary also explains neatly why questions such as the flow of time or free will cannot be discussed within the framework of physics. Whether or not time flows, whether or not our will is free, the four-dimensional picture resulting from the course of nature, whether or not influenced by us, can (in a gedankenexperiment) be replayed, after everything has happened, like a movie. In the replay, everything is determined, and there is only the illusion of free will, just as we are used from the cinema. But the physics, expressed in the correlations between the parts of the movie, is identical to that in the original version.
The principle of physics, that it restricts attention to that which remains the same after everything happened, makes physics very powerful in that it allows us to investigate a past of billions of years and to anticipate a future of billions of years. But the same principle also generates its intrinsic limitations, that physics must be silent about everything that cannot be captured by this static, four-dimensional view of nature. This includes both questions such as “how does reality happen?”, and many of the subjects of most interest to people: freedom, purpose, and consciousness.
[99]{}
A. Aspect, Proposed experiment to test the nonseparability of quantum mechanics, Phys. Rev. D 14 (1976), 1944-1951. (Reprinted in [@WheZ].)
J.S. Bell, On the Einstein Podolsky Rosen paradox, Physics 1 (1964), 195-200. (Reprinted in [@WheZ].)
G. Birkhoff and J. von Neumann, The logics of quantum mechanics, Ann. Math. 37 (1936), 823-843.
Ph. Blanchard and A. Jadczyk, Time of events in quantum theory, Helv. Phys. Acta 69 (1996), 613-635.
D. Bohm, A suggested interpretation of the quantum theory in terms of ‘hidden’ variables, I and II, Phys. Rev. 85 (1952), 166-179. (Reprinted in [@WheZ].)
N. Bohr, Discussion with Einstein on epistemological problems in atomic physics, pp. 200-241 in: P.A. Schilpp (ed.), Albert Einstein: Philosopher-Scientist, The Library of Living Philosophers, Evanston 1949. (Reprinted in [@WheZ].)
M. Born, Zur Quantenmechanik der Sto[ß]{}vorg[ä]{}nge, Zeitschrift f. Physik 37 (1926), 863-867. (Engl. translation: Section I.2 in [@WheZ].)
V.B. Braginsky, Yu. I. Vorontsov and K.S. Thorne, Quantum nondemolition measurements, Science 209 (1980), 547-557. (Reprinted in [@WheZ].)
P. Busch, M. Grabowski and P.J. Lahti, Operational Quantum Physics, Springer, Berlin 1995.
J.F. Clauser and M.A. Horne, A. Shimony and R.A. Holt, Proposed experiment to test local hidden-variable theories, Phys. Rev. Lett. 23 (1969), 880-884. (Reprinted in [@WheZ].)
E.B. Davies, Quantum Theory of Open Systems, Academic Press, London 1976.
R.P. Feynman, Space-time approach to non-relativistic quantum mechanics, Rev. Mod. Phys. 20 (1948), 367-387. (Reprinted in J. Schwinger, ed., Quantum Electrodynamics, Dover, New York 1958.)
G.C. Ghirardi, A. Rimini and T. Weber, Unified dynamics for microscopic and macroscopic systems, Phys. Rev. D 34 (1986), 470-491.
W. Heisenberg, [Ü]{}ber den anschaulichen Inhalt der quantentheoretischen Kinematik und Mechanik, Zeitschrift f. Physik 43 (1927), 172-198. (Engl. translation: Section I.3 in [@WheZ].)
C.J. Isham, Quantum logic and the histories approach to quantum theory, J. Math. Phys. 35 (1994), 2157-2185. gr-qc/9308006
E. Joos and H.D. Zeh, The emergence of classical properties through interaction with the environment, Z. Phys. B 59 (1985), 223-243.
A. Kent, Consistent sets contradict, Phys. Rev. Lett. 78 (1997), 2874-2877. gr-qc/9604012
S. Kochen and E.P. Specker, The problem of hidden variables in quantum mechanics, J. Math. Mech. 17 (1967), 59-67. (Reprinted in C.A. Hooker, ed., The Logico-Algebraic Approach to Quantum mechanics, Vol. I: Historical Evolution, Reidel, Dordrecht 1975.)
P.G. Kwiat, H. Weinfurter, T. Herzog, A. Zeilinger, and M. A. Kasevich, Interaction-free measurements, Phys. Rev. Lett. 74 (1995), 4763-4766.
N.F. Mott, The wave mechanics of $\alpha$-ray tracks, Proc. Roy. Soc. London A 126 (1929), 79-84. (Reprinted in [@WheZ].)
A. Neumaier, Noncommutative analysis and quantum physics I. Algebra and logic. In preparation.
J. von Neumann, Mathematische Grundlagen der Quantenmechanik. Springer, Berlin 1932.
R. Omn[è]{}s, The Interpretation of Quantum Mechanics. Princeton Univ. Press, Princeton 1994.
R. Penrose, Shadows of the mind. Oxford Univ. Press, Oxford 1994.
M. Reck, A. Zeilinger, H.J. Bernstein and P. Bertani, Experimental realization of any discrete unitary operator, Phys. Rev. Lett. 73 (1994), 58-61.
H.P. Robertson, The uncertainty principle, Phys. Rev. 34 (1929), 163-164. (Reprinted in [@WheZ].)
E. Schr[ö]{}dinger Die gegenw[ä]{}rtige Situation in der Quantenmechanik, Naturwissenschaften 23 (1935), 807-812; 823-828; 844-849. (Engl. translation: Section I.11 in [@WheZ].)
K. Svozil, Quantum Logic. Springer, Berlin 1998.
J.A. Wheeler and R.P. Feynman, Interaction with the absorber as the mechanism of radiation, Rev. Mod. Phys. 17 (1945), 157-181.
J.A. Wheeler and R.P. Feynman, Classical electrodynamics in terms of direct interparticle action, Rev. Mod. Phys. 21 (1949), 425-433.
J.A. Wheeler and W. H. Zurek, Quantum theory and measurement. Princeton Univ. Press, Princeton 1983.
E.P. Wigner, Interpretation of quantum mechanics, Section II.2 in [@WheZ].
W.K. Wootters and W.H. Zurek, Complementarity in the double-slit experiment: Quantum nonseparability and a quantitative statement of Bohr’s principle, Phys. Rev. D 19 (1979), 473-484. (Reprinted in [@WheZ].)
A. Zeilinger, On the interpretation and philosophical foundation of quantum mechanics, in: “Vastakohtien todellisuus”, Festschrift for K.V. Laurikainen, U. Ketvel et al. (Eds.), Helsinki University Press, 1996. (http://info.uibk.ac.at/c/c7/c704/qo/philosop.html)
W.H. Zurek, Decoherence and the transition from quantum to classical, Physics Today 44 (1991), 36-44.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Recent work has indicated that ghost imaging may have applications in standoff sensing. However, most theoretical work has addressed transmission-based ghost imaging. To be a viable remote-sensing system, the ghost imager needs to image rough-surfaced targets in reflection through long, turbulent optical paths. We develop, within a Gaussian-state framework, expressions for the spatial resolution, image contrast, and signal-to-noise ratio of such a system. We consider rough-surfaced targets that create fully developed speckle in their returns, and Kolmogorov-spectrum turbulence that is uniformly distributed along all propagation paths. We address both classical and nonclassical optical sources, as well as a computational ghost imager.'
author:
- 'Nicholas D. Hardy'
- 'Jeffrey H. Shapiro'
title: Reflective Ghost Imaging through Turbulence
---
Introduction
============
Ghost imaging is a transverse active-imaging technique that exploits the correlation between two light beams to image a target without spatially-resolving measurements of the light beam that has undergone target interaction. The two beams are the signal, which interacts with the target and then is measured by a single-pixel bucket detector, and the reference, which is directly measured by a high spatial-resolution detector. An intensity cross-correlation between the signal beam that encounters the target and the reference beam that impinges on the high spatial-resolution detector imparts target information—the ghost image—to the cross correlation between the photocurrents obtained from the two detectors, information that is unavailable from either photocurrent alone.
Most ghost-imaging experiments and related theory deal with the transmissive case, in which the bucket detector is placed behind the target, and we image the target’s intensity-transmission profile. In the first ghost imaging experiment, the signal and reference beams were the quantum-mechanically entangled outputs from a spontaneous parametric downconverter (SPDC) [@Pittman1995], which afforded them a phase-sensitive cross correlation stronger than permitted by classical physics [@Erkmen2008Unified]. The system was run at low flux, in which the SPDC’s post-selected output state, within the photodetectors’ response time, could be taken to be a biphoton; the ghost image was then formed by counting coincidences between the bucket and high spatial-resolution detections instead of a photocurrent cross correlation. Subsequently, ghost imaging was done with a high-flux pseudothermal source, i.e., a classical-state source realized by sending a laser beam through a rotating ground-glass diffuser and a 50-50 beam splitter to create identical, spatially-incoherent signal and reference beams with a phase-insensitive cross correlation. The ghost image was then obtained by continuous-time cross correlation of the bucket and high spatial-resolution photocurrents [@Valencia2005; @Ferri2005]. Later, the theories of SPDC and pseudothermal ghost imaging were united in a Gaussian-state treatment that showed, in both cases, that ghost-image formation by photocurrent correlation arises from classical coherence propagation, but the nonclassical SPDC source offers much higher image contrast and a modest spatial-resolution advantage in near-field operation [@Erkmen2008Unified; @Erkmen2009SNR; @Erkmen2010AOP].
Recently, a computational ghost-imaging method was introduced, in which the need for a reference beam is removed [@Shapiro2008Comp; @Bromberg2009]. Conventional ghost imaging relies on the intensity cross-correlation between the signal beam on the target and the reference beam on the high spatial-resolution detector. In SPDC and pseudothermal ghost imaging, these beam patterns are random. If, however, a spatial light modulator is used to impart a sequence of known spatial patterns to a laser beam prior to propagation to the target, then diffraction theory can be used to calculate noiseless versions of the associated reference-beam intensity patterns. These computed reference-beam results can be employed, for cross-correlation ghost-image formation, as if they came from photodetection of a physical reference beam. More importantly, computational ghost imaging opens the door for more advanced computational image reconstruction techniques, notably compressive sensing [@Katz2009].
All of the preceding discussion has been framed for ghost imaging of transmissive targets. Recent experiments have shown the feasibility of an alternate ghost-imaging configuration, in which the bucket detector views the target in reflection, rather than in transmission [@Meyers2008]. In this reflective setup, the source and detector can, in principle, be moved arbitrarily far away from the target, suggesting that ghost imaging could be a viable remote-sensing technique. Atmospheric turbulence will surely be a relevant concern in any such standoff-sensing application, just as it is for astronomical imaging and laser radar. To date, there has been theoretical study of the impact of turbulence on ghost imaging in transmission [@Cheng2009], but there has only been an initial turbulence-free theory development for ghost imaging in reflection [@Hardy2010], and that work was limited to pseudothermal sources.
In this paper we will extend the analysis from [@Hardy2010] to include SPDC ghost imaging in reflection and, for both SPDC and pseudothermal operation, the presence of atmospheric turbulence in the propagation paths. The setup we shall consider is the lensless ghost-imaging configuration shown schematically in Fig. \[fig:setup\] [@footnote1]. Pseudothermal ghost imaging can be described quantitatively using semiclassical photodetection theory—in which the fields are described classically and photodetection incurs the shot noise associated with the discreteness of the electron charge—whereas quantum photodetection theory is required for the SPDC case, owing to the nonclassical nature of its output state [@Erkmen2008Unified]. Nevertheless, to provide a unified development, we will employ quantum photodetection to characterize both pseudothermal and SPDC ghost imaging. We begin, in Sec. \[sec:Setup\], with the theoretical framework for our analysis. Here we will describe our models for the pseudothermal and SPDC sources, reflection from the target, atmospheric propagation, quantum photodetection, and image formation. In Sec. \[sec:ResCon\] we will derive expressions for the spatial resolution and image contrast for pseudothermal and SPDC ghost images, and in Sec. \[sec:SNR\] we will derive their respective signal-to-noise ratios (SNRs). Then, in Sec. \[sec:Comp\], we shall consider a computational ghost imager, in which the reference beam is calculated instead of measured [@Shapiro2008Comp]. Because the reference beam is not an optical field, we find it convenient in Sec. \[sec:Comp\] to employ the semiclassical theory. This section will also include a quantitative performance comparison between pseudothermal, SDPC, and computational ghost images formed in reflection. We conclude, in Sec. \[sec:Discussion\], with a discussion of our findings, and an assessment of the situations in which reflective ghost-imaging might have practical value.
![(Color online) Light from a spatially-incoherent optical source—either a type-II phase matched SPDC or a laser rendered spatially incoherent by passage through a rotating ground-glass—is separated into signal and reference beams by a beam splitter. For the pseudothermal (laser) case, this is a 50-50 beam splitter. For the SPDC, this is a polarizing beam splitter. The reference beam travels $L\,$m to the CCD camera, while the signal beam travels $L\,$m to the target, and the reflected light travels $L\,$m to the bucket detector. The image is then formed by cross correlation between the photocurrents obtained from the bucket detector and the high spatial-resolution (CCD camera) detector.[]{data-label="fig:setup"}](Setup.pdf){width="3.5in"}
Setup {#sec:Setup}
=====
Optical Sources
---------------
We are interested in ghost imaging with classical pseudothermal light, and the nonclassical output of an SPDC. In both cases our signal and reference beams will be represented quantum-mechanically as $\sqrt{\mbox{photons/m$^2$s}}$-units, $+z$-going, positive-frequency, field operators—${\hat{E}}_S({\bm\rho},t)e^{-i\omega_0t}$ and ${\hat{E}}_R({\bm\rho},t)e^{-i\omega_ot}$, as functions of transverse-coordinate vector ${\bm\rho}$ and time—at the output of the Fig. \[fig:setup\] beam splitter. Their baseband field operators have the canonical commutation relations $$\begin{aligned}
\label{eq:Comm1}&[{\hat{E}}_J({\bm\rho}_1,t_1),{\hat{E}}_K({\bm\rho}_2,t_2)] = 0, \\[.05in]
\label{eq:Comm2}&[{\hat{E}}_J({\bm\rho}_1,t_1),{\hat{E}}^\dagger_K({\bm\rho}_2,t_2)] = \delta_{JK}\delta({\bm\rho}_1-{\bm\rho}_2)\delta(t_1-t_2),\end{aligned}$$ for $J = S,R$ and $K = S,R$, where $\delta_{JK}$ is the Kronecker delta, and $\delta(\cdot)$ is the unit impulse. Following [@Erkmen2008Unified], we will take both the pseudothermal and SPDC sources to produce signal and reference beams that are in zero-mean, jointly Gaussian states, which are then completely characterized by their nonzero autocorrelation and cross-correlation functions. All the autocorrelations will be assumed to have the same Gaussian-Schell model form, $$\begin{aligned}
\label{eq:AutoCorr}
{\langle {\hat{E}}^\dagger_K&({\bm\rho}_1,t_1){\hat{E}}_K({\bm\rho}_2,t_2) \rangle} \nonumber\\[.05in]
& =\frac{2P}{\pi a_0^2}e^{-\frac{|{\bm\rho}_1|^2+|{\bm\rho}_2|^2}{a_0^2}} e^{-\frac{|{\bm\rho}_1-{\bm\rho}_2|^2}{2\rho_0^2}}e^{-\frac{(t_1-t_2)^2}{2 T_0^2}},\end{aligned}$$ for $K = S,R$, where $P$ is the photon flux, $a_0$ is the source’s $e^{-2}$ intensity radius, $\rho_0$ is its coherence length, and $T_0$ is its coherence time. The distinction between the Gaussian-state models for the pseudothermal and SPDC sources lies in their signal-reference cross correlations, as we will now explain.
The pseudothermal signal and reference beams start as continuous-wave (cw) laser light, that is passed through a rotating ground-glass diffuser to make it spatially incoherent. Signal and reference beams are then obtained by 50-50 beam splitting. The signal and reference thus have no phase-sensitive cross correlation, and the maximum phase-insensitive cross correlation allowed by classical physics given their autocorrelation functions, i.e., $$\begin{aligned}
\label{eq:CCorrPS}
{\langle {\hat{E}}_S&({\bm\rho}_1,t_1){\hat{E}}_R({\bm\rho}_2,t_2) \rangle}_C=0, \\[.05in]
\begin{split}\label{eq:CCorrPIS}
{\langle {\hat{E}}^\dagger_S&({\bm\rho}_1,t_1){\hat{E}}_R({\bm\rho}_2,t_2) \rangle}_C \\[.05in]
& =\frac{2P}{\pi a_0^2}e^{-\frac{|{\bm\rho}_1|^2+|{\bm\rho}_2|^2}{a_0^2}} e^{-\frac{|{\bm\rho}_1-{\bm\rho}_2|^2}{2\rho_0^2}}e^{-\frac{(t_1-t_2)^2}{2 T_0^2}},
\end{split}\end{aligned}$$ where the subscript $C$ indicates that these are classical-state cross correlations. The coherence time is directly related to how fast the ground-glass rotates, the coherence length is typically on the order of a few wavelengths, and the intensity radius is set by the radius of the initial laser beam.
A type-II phase-matched, collinear, cw SPDC source emits orthogonally-polarized, co-propagating signal and idler beams, which are separated by a polarizing beam splitter to become the signal and reference beams for ghost imaging. They are in a maximally-entangled, jointly-Gaussian state with no phase-insensitive cross correlation, and the maximum phase-sensitive cross correlation permitted by quantum theory given their autocorrelation functions, namely [@Erkmen2009SNR]
$$\begin{aligned}
\label{eq:QCorrPIS}
{\langle {\hat{E}}^\dagger_S({\bm\rho}_1,t_1)&{\hat{E}}_R({\bm\rho}_2,t_2) \rangle}_Q = 0, \\[.05in]
\begin{split}\label{eq:QCorrPS}
{\langle {\hat{E}}_S({\bm\rho}_1,t_1)&{\hat{E}}_R({\bm\rho}_2,t_2) \rangle}_Q = \frac{2P}{\pi a_0^2} e^{-\frac{|{\bm\rho}_1|^2+|{\bm\rho}_2|^2}{a_0^2}} \!\!
\left[e^{-\frac{|{\bm\rho}_1-{\bm\rho}_2|^2}{2\rho_0^2}}e^{-\frac{(t_1-t_2)^2}{2T_0^2}} + i \left(\frac{2}{\pi}\right)^\frac{1}{4} \!\sqrt{\frac{a_0^2}{P T_0 \rho_0^2}}e^{-\frac{|{\bm\rho}_1-{\bm\rho}_2|^2}{\rho_0^2}}e^{-\frac{(t_1-t_2)^2}{T_0^2}} \right],
\end{split}\end{aligned}$$
where the subscript $Q$ indicates that these are quantum (nonclassical) cross correlations.
Propagation through Turbulence
------------------------------
The signal and reference beams in Fig. \[fig:setup\] first propagate $L\,$m over spatially-separated paths from their shared source location [@footnote1]. The reference beam is then measured by a high-spatial resolution CCD camera, while the signal field reflects off the target and travels $L\,$m to the bucket detector. We will assume statistically-independent turbulence on all three paths: the reference path ($R$) from the source to the CCD; the signal path ($S$) from the source to the target; and the target-return path ($T$) from the target to the bucket detector.
To account for the Kolmogorov-spectrum turbulence that is uniformly distributed along the three propagation paths, we make use of the quantum version of the extended Huygens-Fresnel principle [@Strohbehn1978; @footnote2] $$\begin{aligned}
\label{eq:Prop}
{\hat{E}}'_m({\bm\rho}',t) = \int\! d{\bm\rho}\, {\hat{E}}_m({\bm\rho},t) \frac{k_0 e^{ik_0(L+|{\bm\rho}'-{\bm\rho}|^2/2L)}}{i2\pi L}\,e^{\psi_m({\bm\rho}', {\bm\rho})},\end{aligned}$$ where $k_0 = \omega_0/c = 2\pi/\lambda_0$, and $\psi_m({\bm\rho}', {\bm\rho})$ is a complex-valued random process that encapsulates the effects of turbulence on path $m$ from ${\bm\rho}$ to ${\bm\rho}'$. The real and imaginary parts of $\psi_m({\bm\rho}',{\bm\rho})$, which we will denote $\chi_m({\bm\rho}',{\bm\rho})$ and $\phi_m({\bm\rho}',{\bm\rho})$, are the logamplitude and phase fluctuations imposed on the field received at transverse coordinate ${\bm\rho}'$ over path $m$ from a point source at transverse coordinate ${\bm\rho}$. Because all the path lengths in Fig. \[fig:setup\] are equal, we have suppressed the $L/c$ line-of-sight propagation delay in Eq. (\[eq:Prop\]). We have also suppressed any time-dependence of the turbulence. Physically, this amounts to assuming that the integration time, $T_I$, employed in the Fig. \[fig:setup\] correlator is shorter than the $\sim$ms coherence time of the turbulence. However, that assumption is *not necessary for deriving the spatial resolution and image contrast achieved in the presence of turbulence, although it *will be relevant to our signal-to-noise ratio analysis.**
The mutual coherence function of $e^{\psi_m({\bm\rho}',{\bm\rho})}$ is taken to be [@Ishimaru1978; @Osche2002] $$\begin{aligned}
\label{eq:TurbCorr}
{\langle &e^{\psi^*_m({\bm\rho}'_1, {\bm\rho}_1)}e^{\psi_m({\bm\rho}'_2, {\bm\rho}_2)} \rangle}\nonumber \\[.05in] &=e^{-(|{\bm\rho}'_1-{\bm\rho}'_2|^2+({\bm\rho}'_1-{\bm\rho}'_2)\cdot({\bm\rho}_1-{\bm\rho}_2) +|{\bm\rho}_1-{\bm\rho}_2|^2)/2\rho_m^2},\end{aligned}$$ which is the square-law approximation to the rigorous $5/3$-law behavior for uniformly-distributed turbulence [@footnote3]. In Eq. (\[eq:TurbCorr\]), $\rho_m$ is the turbulence coherence length on path $m$, given by $$\begin{aligned}
\label{eq:TurbCL}
\rho_m = (1.09 k_0^2 C_{n,m}^2 L)^{-3/5},\end{aligned}$$ in terms of the constant turbulence strength-parameter, $C^2_{n,m}$, along path $m$ [@Ishimaru1978].
We will assume that the path length $L$ satisfies the far-field coherence-propagation conditions for pseudothermal and SPDC light discussed in [@Erkmen2008Unified; @Erkmen2010AOP], viz., $k_0 a_0 \rho_0/2L \ll 1$, and $k_0 a_0^2/2L \ll 1$, respectively. For convenience, we also define $\rho_L = 2L/k_0 a_0$ and $a_L = 2L/k_0\rho_0$. These are the far-field, turbulence-free, on-target coherence length and intensity radius produced by the Gaussian-Schell autocorrelation function from Eq. (\[eq:AutoCorr\]).
Target Reflection
-----------------
Most real-world targets have surfaces that are sufficiently rough—on the scale of an optical wavelength—that reflections from them present fully-developed speckle patterns at appreciable distances from the surface [@Goodman2007]. We will model such a target as a planar 2-D object with random, microscopic height variations—having depths on the order of several wavelengths, and transverse variations on the order a wavelength—and an average intensity-reflection coefficient ${\mathcal{T}}({\bm\rho})$. Thus, following laser-radar theory [@Shapiro1981], we shall assume that this target has a random field-reflection coefficient, $T({\bm\rho})$, that is a zero-mean, complex-valued Gaussian random process characterized by the autocorrelation function $$\begin{aligned}
\label{eq:Scatter}
{\langle T^*({\bm\rho}_1)T({\bm\rho}_2) \rangle} = \lambda_0^2 {\mathcal{T}}({\bm\rho}_1) \delta({\bm\rho}_1-{\bm\rho}_2),\end{aligned}$$ where the deterministic pattern ${\mathcal{T}}({\bm\rho})$ is what we are tying to image.
Target reflection is a passive process, so we require $|T({\bm\rho})| \le 1$, which is in conflict with the Gaussian statistics and the delta-function term in Eq. (\[eq:Scatter\]). That delta function, however, leads to quasi-Lambertian reflection, implying that at standoff distances a realistic bucket detector will only capture a very small fraction of the reflected light. Hence using the reflection statistics from the preceding paragraph is a reasonable approximation for ghost imaging of rough-surfaced targets. However, because some light may be absorbed by or transmitted through the target, the quantum model for target reflection needs to be $$\begin{aligned}
\label{eq:TargetLoss}
{\hat{E}}_T({\bm\rho},t) = T({\bm\rho}){\hat{E}}'_S({\bm\rho},t) +\sqrt{1-|T({\bm\rho})|^2}\,{\hat{E}}_{vac,S}({\bm\rho},t)\end{aligned}$$ Here ${\hat{E}}_{vac,S}({\bm\rho},t)$ is a vacuum-state field operator, needed to preserve the free-field commutator relations given, for the source fields, in Eqs. and .
In what follows we will assume that the target’s intensity-reflection coefficient, ${\mathcal{T}}({\bm\rho})$, is space-limited to the region illuminated by the signal beam, i.e., $|{\mathcal{T}}({\bm\rho})| = 0$ for $|{\bm\rho}| > a_L$. This condition will give the ensemble-average pseudothermal and SDPC ghost images convolution-integral relationships with ${\mathcal{T}}({\bm\rho})$ from which to determine their spatial resolutions. See [@Erkmen2008Unified; @Erkmen2010AOP] for information about ghost imaging field-of-view when the preceding restriction on ${\mathcal{T}}({\bm\rho})$ is not satisfied.
Photodetection and Image Formation
----------------------------------
Loss also occurs in the photodetection process, because the detectors in Fig. \[fig:setup\] will, in general, have sub-unity quantum efficiencies. Therefore, the field operator ${\hat{E}}_p({\bm\rho},t)$ driving the photocurrent, $\hat{i}_p(t)$, from CCD pixel $p$, and the field operator ${\hat{E}}_b({\bm\rho},t)$ driving the photocurrent, $\hat{i}_b(t)$, from the bucket detector satisfy [@footnote4; @Shapiro2009] $$\begin{aligned}
\label{eq:Detector}
{\hat{E}}_m({\bm\rho},t) = \sqrt{\eta}\,{\hat{E}}'_\ell({\bm\rho},t) +\sqrt{1-\eta}\,{\hat{E}}_{vac,\ell}({\bm\rho},t),\end{aligned}$$ for $(m,\ell)=(p,R)$ or $(m,\ell)=(b,T)$. Here, $\eta$ is the quantum efficiency, which is taken to be the same for both the bucket detector and the CCD camera, and ${\hat{E}}_{vac,\ell}({\bm\rho},t)$ is a vacuum-state field operator whose inclusion models the loss incurred by having $\eta < 1$, cf. Eq. (\[eq:TargetLoss\]).
We will assume ghost-image formation via continuous-time, pixel-wise correlation of the photocurrents from the CCD and the bucket detector, formalized as $$\begin{aligned}
\label{eq:Corr}
\hat{C}({\bm\rho}_p) = \frac{1}{T_I}\int_{-T_I/2}^{T_I/2}\!dt\, \hat{i}_p(t) \hat{i}_b(t),\end{aligned}$$ where ${\bm\rho}_p$ is the center coordinate of the CCD’s $p$th pixel, and $T_I$ is the correlation integration time. From quantum photodetection theory [@Shapiro2009], we have that $$\begin{aligned}
\label{eq:Current}
\hat{i}_m(t) = q\int d\tau\, \hat{P}_m(\tau) h(t-\tau), \quad m=p,b\end{aligned}$$ where $$\begin{aligned}
\label{eq:Flux}
\hat{P}_m(t) = \int_{{\cal A}_m} d{\bm\rho}\, {\hat{E}}^\dagger_m({\bm\rho},t) {\hat{E}}_m({\bm\rho},t)\end{aligned}$$ is the effective (sub-unity quantum efficiency) photon flux on the active region ${\cal A}_m$ for detector $m$, and $h(t)$ is the impulse response of the detectors’ output circuit, which we have taken to be the same for both the bucket detector and the CCD camera.
To enable the performance evaluations in the sections to follow, we make the following assumptions. First, each pixel is small enough that the the average photon flux impinging on it obeys $\langle \hat{P}_p(t)\rangle = A_p\langle {\hat{E}}^\dagger_p({\bm\rho}_p,t){\hat{E}}_p({\bm\rho}_p,t)\rangle$, where $A_p$ is the area of the ${\cal A}_p$. Second, we assume that the detector has a Gaussian impulse response, $$\begin{aligned}
h(t) = \frac{\Omega_B}{\sqrt{8\pi}}e^{-\Omega^2_B t^2/8},\end{aligned}$$ whose bandwidth, $\Omega_B$, greatly exceeds that of the incident light for the pseudothermal source ($\Omega_B T_0 \gg 1$), and is much smaller than that of the incident light for the SPDC source ($\Omega_B T_0 \ll 1$). Finally, we assume that the correlation integration time is sufficiently long to capture a ghost image, so $T_I \gg T_0$ for the pseudothermal source, and $T_I \gg \Omega_B^{-1}$ for the SPDC source.
Resolution and Contrast {#sec:ResCon}
=======================
The spatial resolution and contrast of the ghost image are properties of the average cross correlation, $\langle C({\bm\rho}_p)\rangle$. To determine this ensemble-average cross correlation we: (1) write $\hat{i}_b(t)$ and $\hat{i}_p(t)$ in terms of the field operators that drive them, ${\hat{E}}_b({\bm\rho},t)$ and ${\hat{E}}_p({\bm\rho},t)$; (2) apply the canonical commutator relations to normally-order the resulting photocurrent-product expressions; and (3) use Eq. to write the detected fields in terms of ${\hat{E}}'_R({\bm\rho},t)$ and ${\hat{E}}'_T({\bm\rho},t)$. From the linearity of expectation, we can bring the ensemble-average operation inside the time-domain integration to get $$\begin{aligned}
\label{eq:AvgCorr}
{\langle \hat{C}(&{\bm\rho}_p) \rangle} = q^2\eta^2 A_p \int\!d\tau_1\int\!d\tau_2 \int_{{\cal A}_b} \! d{\bm\rho}\, h(t-\tau_1)h(t-\tau_2) \nonumber \\[.05in]
&\times {\langle {\hat{E}}'^\dagger_R({\bm\rho}_p,\tau_1) {\hat{E}}'^\dagger_T({\bm\rho},\tau_2) {\hat{E}}'_R({\bm\rho}_p,\tau_1) {\hat{E}}'_T({\bm\rho},\tau_2) \rangle}.\end{aligned}$$ We then use Eqs. and to back-propagate the target-return field operator to the target, and find $$\begin{aligned}
\lefteqn{
{\langle {\hat{E}}'^\dagger_R({\bm\rho}_p,\tau_1) {\hat{E}}'^\dagger_T({\bm\rho},\tau_2) {\hat{E}}'_R({\bm\rho}_p,\tau_1) {\hat{E}}'_T({\bm\rho},\tau_2) \rangle}}
\nonumber \\[.05in]
&=&\int\! d{\bm\rho}_2 \!\int\! d{\bm\rho}_3\, {\langle T^*({\bm\rho}_2)T({\bm\rho}_3) \rangle}{\langle e^{\psi^*_T({\bm\rho},{\bm\rho}_2)}e^{\psi_T({\bm\rho},{\bm\rho}_3)} \rangle}\nonumber \\[.05in]
&\times& {\langle {\hat{E}}'^\dagger_R({\bm\rho}_p,\tau_1) {\hat{E}}'^\dagger_S({\bm\rho}_2,\tau_2) {\hat{E}}'_R({\bm\rho}_p,\tau_1) {\hat{E}}'_S({\bm\rho}_3,\tau_2) \rangle} \nonumber \\[.05in]
&\times& \frac{e^{-ik_0(|{\bm\rho}_2|^2-|{\bm\rho}_3|^2)/2L + ik_0{\bm\rho}\cdot({\bm\rho}_2-{\bm\rho}_3)/L}}{(\lambda_0L)^2},\end{aligned}$$ where we have used the statistical independence of the target’s reflection coefficient, the signal and reference fields at the source, and the turbulence on all three propagation paths.
Substituting for $ {\langle T^*({\bm\rho}_2)T({\bm\rho}_3) \rangle}$ from Eq. and performing the ${\bm\rho}_2$ integral eliminates the turbulence term from the preceding $\langle C({\bm\rho}_p)\rangle$ expression, which shows that turbulence on the the target-to-bucket path has no impact on the ghost image’s spatial resolution or image contrast. Now, by back-propagating the field operators to the source—by means of Eq. with $k_0a_0^2/2L \ll 1$—we have that $$\begin{aligned}
\lefteqn{{\langle {\hat{E}}'^\dagger_R({\bm\rho}_p,\tau_1) {\hat{E}}'^\dagger_T({\bm\rho},\tau_2) {\hat{E}}'_R({\bm\rho}_p,\tau_1) {\hat{E}}'_T({\bm\rho},\tau_2) \rangle} }\nonumber \\[.05in]
&=&\frac{1}{\lambda_0^4L^6}\int\! d{\bm\rho}_2 \, {\mathcal{T}}({\bm\rho}_2) \int\! d{\bm\rho}'_1 \int\!d{\bm\rho}'_2 \int\!d{\bm\rho}''_1 \int \! d{\bm\rho}''_2 \nonumber \\[.05in]
&\times& {\langle {\hat{E}}^\dagger_R({\bm\rho}'_1,\tau_1) {\hat{E}}^\dagger_S({\bm\rho}'_2,\tau_2) {\hat{E}}_R({\bm\rho}''_1,\tau_1) {\hat{E}}_S({\bm\rho}''_2,\tau_2) \rangle} \nonumber \\[.05in]
&\times& {\langle e^{\psi^*_R({\bm\rho}_p,{\bm\rho}'_1)}e^{\psi_R({\bm\rho}_p,{\bm\rho}''_1)} \rangle}{\langle e^{\psi^*_S({\bm\rho}_2,{\bm\rho}'_2)}e^{\psi_S({\bm\rho}_2,{\bm\rho}''_2)} \rangle} \nonumber \\[.05in]
&\times& e^{ik_0[{\bm\rho}_p\cdot({\bm\rho}'_1-{\bm\rho}''_1) + {\bm\rho}_2\cdot({\bm\rho}'_2 - {\bm\rho}''_2)]/L},\end{aligned}$$ where we have again invoked the statistical independence of the signal and reference fields at the source and the turbulence on the source-to-CCD, and source-to-target paths.
We can now directly evaluate all remaining moments. The turbulence moments are available from Eq. , and the fourth-order field moment can be expressed in terms of second-order moments with the Gaussian-state moment-factoring theorem [@Shapiro1994]: $$\begin{aligned}
\lefteqn{\hspace*{-.1in}{\langle {\hat{E}}^\dagger_R({\bm\rho}'_1,\tau_1) {\hat{E}}^\dagger_S({\bm\rho}'_2,\tau_2) {\hat{E}}_R({\bm\rho}''_1,\tau_1) {\hat{E}}_S({\bm\rho}''_2,\tau_2) \rangle} = } \nonumber \\
&& \hspace{-.1in}{\langle {\hat{E}}^\dagger_R({\bm\rho}'_1,\tau_1) {\hat{E}}^\dagger_S({\bm\rho}'_2,\tau_2) \rangle} {\langle {\hat{E}}_R({\bm\rho}''_1,\tau_1) {\hat{E}}_S({\bm\rho}''_2,\tau_2) \rangle}+ \nonumber \\
&& \hspace*{-.1in}{\langle {\hat{E}}^\dagger_R({\bm\rho}'_1,\tau_1){\hat{E}}_R({\bm\rho}''_1,\tau_1) \rangle} {\langle {\hat{E}}^\dagger_S({\bm\rho}'_2,\tau_2){\hat{E}}_S({\bm\rho}''_2,\tau_2) \rangle} + \nonumber \\
&& \hspace*{-.1in} {\langle {\hat{E}}^\dagger_R({\bm\rho}'_1,\tau_1){\hat{E}}_S({\bm\rho}''_2,\tau_2) \rangle} {\langle {\hat{E}}^\dagger_S({\bm\rho}'_2,\tau_2){\hat{E}}_R({\bm\rho}''_1,\tau_1) \rangle}.
\label{eq:MomentFactor}\end{aligned}$$
Until this point, our analysis applies equally well to the pseudothermal and SPDC ghost imagers. That equivalence disappears, however, when we evaluate the second-order moments in Eq. using Eqs. – for the pseudothermal ghost imager, and using Eqs. , , and for the SPDC ghost imager. The results we obtain are $$\begin{aligned}
\lefteqn{{\langle \hat{C}({\bm\rho}_p) \rangle}_C = \frac{q^2\eta^2 A_p A_b}{L^2}\left(\frac{2P}{\pi a_L^2}\right)^2} \nonumber \\
&\times& \int\!d{\bm\rho}\, {\mathcal{T}}({\bm\rho})\left[1+\frac{ e^{-|{\bm\rho}-{\bm\rho}_p|^2/\alpha\rho_L^2}}{\alpha}\right],
\label{eq:CAvgMean}\end{aligned}$$ for the pseudothermal (classical) imager, and $$\begin{aligned}
\lefteqn{{\langle \hat{C}({\bm\rho}_p) \rangle}_Q= \frac{q^2\eta^2A_p A_b}{L^2} \left(\frac{2P}{\pi a_L^2}\right)^2\int \! d{\bm\rho}\, {\mathcal{T}}(-{\bm\rho})} \nonumber \\[.05in]
&\times& \left[ 1+ \frac{\Omega_B T_0}{4}\frac{e^{-|{\bm\rho}-{\bm\rho}_p|^2/\alpha\rho_L^2}}{\alpha}\left(1+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right)\right].
\label{eq:QAvgMean}\end{aligned}$$ for the SPDC (quantum) ghost imager, where $A_b$ is the area of $\mathcal{A}_b$, $\mathcal{I}\equiv P T_0 \rho_0^2/a_0^2 = PT_0\rho_L^2/a_L^2$ is the source brightness in photons per spatiotemporal mode, and $$\begin{aligned}
\alpha \equiv \frac{2\rho_R^2\rho_S^2+a_0^2(\rho_R^2 +\rho_S^2)}{2\rho_R^2\rho_S^2}\end{aligned}$$ is a turbulence-induced performance-degradation factor. We see that both average correlation functions are the sum of the same featureless background term $$\begin{aligned}
\label{eq:C0}
C_0=\frac{q^2\eta^2 A_p A_b}{L^2}\left(\frac{2P}{\pi a_L^2}\right)^2 \int\!d{\bm\rho}\, {\mathcal{T}}({\bm\rho}),\end{aligned}$$ and different image-bearing terms: $$\begin{aligned}
C_{1,C}&({\bm\rho}_p)=\frac{q^2\eta^2 A_p A_b}{\alpha L^2}\left(\frac{2P}{\pi a_L^2}\right)^2 \int\!d{\bm\rho}\, {\mathcal{T}}({\bm\rho}) e^{-\frac{|{\bm\rho}-{\bm\rho}_p|^2}{\alpha\rho_L^2}},\end{aligned}$$ for the classical-state (pseudothermal) source, and $$\begin{aligned}
C_{1,Q}({\bm\rho}_p)&=&\frac{q^2\eta^2 A_p A_b}{\alpha L^2}\left(\frac{2P}{\pi a_L^2}\right)^2\! \left(1+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right) \nonumber\\[.05in]
&\times& \frac{\Omega_B T_0}{4}\int\!d{\bm\rho}\, {\mathcal{T}}(-{\bm\rho}) e^{-\frac{|{\bm\rho}-{\bm\rho}_p|^2}{\alpha\rho_L^2}},\end{aligned}$$ for the nonclassical-state (quantum) source.
The average pseudothermal ghost image is erect, and consists of a scaled version of the target’s intensity-reflection coefficient ${\mathcal{T}}({\bm\rho})$ convolved with a Gaussian point-spread function whose $e^{-1}$ radius is $\rho_L\sqrt{\alpha}$. The average SPDC ghost image is inverted, with a different scaling but the same point-spread function. Thus, both the pseudothermal and SPDC ghost images have the same spatial resolution, $$\begin{aligned}
\lefteqn{\rho_L\sqrt{\alpha} = \frac{\lambda_0 L}{\pi a_0}\sqrt{1+ \frac{a_0^2}{2}\left[\rho_S^{-2} + \rho_R^{-2}\right]} = \label{eq:Res}}
\\[.05in]
&&\hspace*{-.1in}
\frac{\lambda_0L} {\pi a_0}\sqrt{1+\frac{a_0^2}{2} (1.09 k_0^2 L)^\frac{6}{5}\left[(C_{n,S}^2)^\frac{6}{5} + (C_{n,R}^2)^\frac{6}{5}\right]},\end{aligned}$$ where we have utilized Eq. to write the turbulence coherence lengths on each path in terms of their respective strength parameters. Equation shows that turbulence-limited resolution prevails for the far-field lensless ghost image when either the source-to-target coherence length ($\rho_S$) or the source-to-reference coherence length ($\rho_R$) becomes smaller than the source size ($a_0$). When $\rho_S \gg a_0$ and $\rho_R \gg a_0$, far-field lensless ghost imaging in reflection yields the same $\lambda_0L/\pi a_0$ spatial resolution previously reported for far-field lensless ghost imaging in transmission [@Erkmen2008Unified; @Hardy2010]. However, even when turbulence does not degrade ghost-image spatial resolution, there is still a significant difference between the transmissive and reflective cases, namely the $A_b/L^2$ factor that is present in our results but is absent from those in [@Erkmen2008Unified]. This factor, which will obey $A_b/L^2 \ll 1$ in a standoff-sensing scenario, is the bucket detector’s angular subtense at the target, and represents the fraction of the quasi-Lambertian reflected light that this detector collects. No such factor appeared in [@Erkmen2008Unified], because that work assumed the bucket detector collected all the light that was transmitted through the target.
The featureless background terms in our pseudothermal and SPDC ghost images affect the images’ contrast. To assess these contrasts, we will adopt the definition and approach presented in [@Erkmen2008Unified]. In particular, the image contrast $\mathcal{C}$ is the ratio of the difference between the brightest and darkest pixels in the image to the featureless background: $$\begin{aligned}
\label{eq:ContrastDef}
\mathcal{C} = \frac{\max_{\mathcal{R}}[\hat{C}({\bm\rho}_p)]-\min_{\mathcal{R}}[\hat{C}({\bm\rho}_p)]}{C_0},\end{aligned}$$ where $\mathcal{R}$ is the region within which ${\mathcal{T}}({\bm\rho})$ is nonzero. We will assume that the ghost imager completely resolves the target, so that we can use the approximation $$\begin{aligned}
\label{eq:Resolve}
\int\!d{\bm\rho}\,{\mathcal{T}}(\pm{\bm\rho})e^{-\frac{|{\bm\rho}-{\bm\rho}_p|^2}{\alpha\rho_L^2}} \approx \pi \alpha \rho_L^2 {\mathcal{T}}(\pm{\bm\rho}) .\end{aligned}$$ Defining $A_T=\int\!d{\bm\rho}\, {\mathcal{T}}({\bm\rho})$ as the target’s reflective cross-section, and taking $\max_{\mathcal{R}}[{\mathcal{T}}({\bm\rho})] = 1$, we get the following image contrast expressions for our two ghost imagers, $$\begin{aligned}
\label{eq:CContrast}
&\mathcal{C}_C = \frac{\pi \rho_L^2}{A_T},\end{aligned}$$ and $$\begin{aligned}
\label{eq:QContrast}
&\mathcal{C}_Q = \frac{\pi \rho_L^2}{A_T}\frac{\Omega_BT_0}{4}\left(1+\frac{1}{4\sqrt{\pi}\mathcal{I}} \right).\end{aligned}$$
If $|{\mathcal{T}}({\bm\rho})| \sim 1$ where it is nonzero, then $A_T$ is the area of the target. Moreover, $\rho_L^2$ is the spatial-resolution area in the absence of turbulence. It follows that the pseudothermal ghost-image contrast is approximately $1/N$, where $N$ is number of no-turbulence resolution cells in the image. Note that our analysis to this point has assumed a [dc]{}-coupled correlator in Fig. \[fig:setup\]. The contrast of a pseudothermal ghost image can be improved by suppressing the background term through [ac]{}-coupling, i.e., measuring the cross covariance, rather than the cross correlation, between the bucket and high spatial-resolution photodetector’s outputs, as was done in the pseudothermal experiments of Scarcelli *et al. [@Scarcelli2006].*
The pseudothermal imager’s contrast does not depend on its source brightness, $\mathcal{I}$, but that for the SPDC does. At high source brightness, $\mathcal{I} \gg 1$, the SPDC image contrast reduces to the result we found for the pseudothermal case. The normal operating regime for a cw SPDC, however, is one of low brightness, $\mathcal{I} \ll 1$, for which Eq. reduces to $$\mathcal{C}_Q \approx \frac{\sqrt{\pi}}{16}\,\frac{\Omega_B}{P}\frac{a_L^2}{A_T}.$$ The number of photon pairs emitted by an SPDC in one detector time-constant is low, so $\Omega_B /P \gg 1$, and our assumption that the target is contained within the region illuminated by the signal beam implies that $a_L^2/A_T > 1$ if $|{\mathcal{T}}({\bm\rho})| \sim 1$ where it is nonzero. Consequently, unlike the pseudothermal case, $\mathcal{C}_Q \gg 1$ is the norm for SDPC ghost imaging in [dc]{}-coupled operation.
Signal-to-Noise Ratio {#sec:SNR}
=====================
In keeping with the analysis done in [@Erkmen2009SNR], we will be evaluating the signal-to-noise ratio (SNR) pixel-wise as the ratio of the square of the expected value to the variance for the photocurrent cross-correlation function, $$\begin{aligned}
\label{eq:SNRi}
\text{SNR} = \frac{{\langle \hat{C}({\bm\rho}_p) \rangle}^2}{{\rm Var}[\hat{C}({\bm\rho}_p)]} = \frac{{\langle \hat{C}({\bm\rho}_p) \rangle}^2}{{\langle \hat{C}^2({\bm\rho}_p) \rangle}-{\langle \hat{C}({\bm\rho}_p) \rangle}^2}.\end{aligned}$$ The main complication in evaluating Eq. is the second moment, $$\begin{aligned}
\label{eq:Cm2}
{\langle \hat{C}^2({\bm\rho}_p) \rangle} = \frac{1}{T_I^2}\!\int\limits_{-T_I/2}^{T_I/2}\!\!\!dt\!\!\int\limits_{-T_I/2}^{T_I/2}\!\!\!du\, {\langle \hat{i}_p(t)\hat{i}_b(t)\hat{i}_p(u)\hat{i}_b(u) \rangle}.\end{aligned}$$ After using Eq. to write the currents in terms of the fields, Eqs. and to normally order the fields, and Eqs. and to propagate the fields back to the source, we are left needing to evaluate an eighth-order moment of the fields, two sixth-order moments of the fields, a fourth-order moment of the fields, a fourth-order moment of the target, three fourth-order moments of the turbulence, and propagation of these terms over 12 unique paths. To make the analysis more tractable, we use several simplifying assumptions. First, we [ac]{}-couple both of the detector outputs by inserting a narrowband notch centered at the origin in frequency space, resulting in the frequency response $$\begin{aligned}
\label{eq:DCblock}
H_B(\Omega) = e^{-2\Omega^2/\Omega^2_B}-e^{-2\Omega^2/\Omega^2_N},\end{aligned}$$ where $\Omega_N \ll \Omega_B$. We also assume that the ghost imager can resolve all structure on the target, allowing us to use the approximation in Eq. .
To evaluate the fourth-order turbulence moments, we can no longer use the square-law approximation in Eq. that models the turbulence’s mutual-coherence function as Gaussian, because this approximation is not valid for fourth-order moments [@Wandzura1980]. To evaluate these terms we assume that the turbulence coherence length on each path is much larger than the on-target coherence length $\rho_L$. We also assume the turbulence coherence lengths are all much larger than $a_0$, implying that the turbulence functions are independent of coordinates at the source and detection planes. Although these assumptions are somewhat limiting, they still allow weak-to-medium strength turbulence for a typical situation. For instance, if the target is standing off $1\,$km, the source aperture has radius $a_0=3\,$cm, and we are at a wavelength $\lambda_0=1.5\,\mu$m, this assumption is satisfied for $C_n^2 \le 10^{-14}\,{\rm m}^{-2/3}$. Moreover, this situation has no spatial-averaging of the turbulence at the target, and thus represents a worst-case scenario for the SNR.
For ${\bm\rho}'_\ell$ and ${\bm\rho}''_\ell$ being coordinates at the CCD or bucket detectors, and ${\bm\rho}_\ell$ being coordinates at the target, the preceding coherence-length assumptions imply that $$\begin{aligned}
{\langle e^{\psi^*_m({\bm\rho}'_1, {\bm\rho}_1)}&e^{\psi_m({\bm\rho}''_1, {\bm\rho}_1)}e^{\psi^*_m({\bm\rho}'_2, {\bm\rho}_2)}e^{\psi_m({\bm\rho}''_2, {\bm\rho}_2)} \rangle}\nonumber \\
&\approx {\langle e^{\psi^*_m(\bm 0, {\bm\rho}_1)}e^{\psi_m(\bm 0, {\bm\rho}_1)}e^{\psi^*_m(\bm 0, {\bm\rho}_2)}e^{\psi_m(\bm 0, {\bm\rho}_2)} \rangle}\nonumber \\
&= {\langle e^{2{\rm Re}[\psi_m(\bm 0, {\bm\rho}_1)]+2{\rm Re}[\psi_m(\bm 0, {\bm\rho}_2)]} \rangle}.\end{aligned}$$ Decomposing the turbulence function into its log-amplitude and phase components as $\psi_m({\bm\rho}_1,{\bm\rho}_2)=\chi_m({\bm\rho}_1,{\bm\rho}_2)+i\phi_m({\bm\rho}_1,{\bm\rho}_2)$, and taking the logamplitude to be Gaussian distributed with mean equal to minus its variance [@Strohbehn1978], then gives $$\begin{aligned}
{\langle e^{\psi^*_m({\bm\rho}'_1, {\bm\rho}_1)}e^{\psi_m({\bm\rho}''_1, {\bm\rho}_1)}e^{\psi^*_m({\bm\rho}'_2, {\bm\rho}_2)}e^{\psi_m({\bm\rho}''_2, {\bm\rho}_2)} \rangle} \nonumber \\[.05in]
\approx e^{4 K_{\chi_m} ({\bm\rho}_1-{\bm\rho}_2)},\end{aligned}$$ where $$K_{\chi_m} ({\bm\rho}_1-{\bm\rho}_2) = {\langle \Delta\chi_m (\bm 0, {\bm\rho}_1)\Delta\chi_m (\bm 0, {\bm\rho}_2) \rangle},$$ for $\Delta\chi_m \equiv \chi_m - {\langle \chi_m \rangle}$, is the log-amplitude covariance function. The logamplitude variance will be taken to be [@Osche2002] $$\begin{aligned}
\sigma_m^2=K_{\chi_m} (\bm 0)=0.124 \,C_{n,m}^2 k_0^{7/6}L^{11/6},\end{aligned}$$ where we have used the Rytov-approximation expression, whose validity is ensured by our assumption of weak-to-medium strength turbulence.
Even with these simplifications, the SNR evaluation is quite tedious, if somewhat straightforward. The higher-order field moments and the fourth-order target moment can be reduced to a sum of the products of second-order moments by application of the Gaussian moment-factoring theorem. The [ac]{}-coupling removes several of these terms, and we are left with integrals and Fourier transforms of the product of Gaussian terms, the turbulence correlation functions, and the target’s intensity-reflection coefficient, ${\mathcal{T}}({\bm\rho})$. To simplify our final SNR expressions, we define two new terms, $$\begin{aligned}
\label{eq:ATp}A'_T & = \int\! d{\bm\rho}\, {\mathcal{T}}^2({\bm\rho}), \\
\label{eq:Gamma}\Gamma &= \frac{1}{(4\pi\beta)^2}\int \! d\bm\nu\, \ e^{-|\bm\nu|^2/2}O(\bm\nu,4\sqrt{\beta}),\end{aligned}$$ where $O(\bm\zeta, D)$ is the two-circle overlap function for circles of diameter $D$, $$\begin{aligned}
\lefteqn{\hspace*{-.1in}O(\bm\zeta,D) = } \nonumber \\[.05in]
&&\hspace*{-.2in}\left\{\begin{array}{ll}
\frac{D^2}{2}\left[ \cos^{-1} \left( \frac{|\bm\zeta|}{D}\right) -\frac{|\bm\zeta|}{D}\sqrt{1-\frac{|\bm\zeta|^2}{D^2}} \right], & \mbox{for $|\bm\zeta|\le D$} \\[.05in]
0, & \text{elsewhere,}
\end{array}\right.\end{aligned}$$ $\beta = A_b/\pi a_0^2$, and we have utilized the dimensionless difference coordinates $\bm{\nu} = \rho_Lk_0(\bm{\rho}'-\bm{\rho}'')/L$, where ${\bm\rho}'$ and ${\bm\rho}''$ are coordinates at the bucket detector.
Both $A'_T$ and $\Gamma$ have significant physical interpretations. Similar to $A_T$, $A'_T$ is another measure of the reflective area of the target, and thus directly related to the number of on-target resolution cells, and the subsequent amount of time it takes to form an image. $\Gamma$ is a metric for the spatial averaging of the target-induced speckle over the bucket detector. For very small detectors $\Gamma \approx 1$, while for large detectors it is inversely proportional to the area of the bucket detector, viz. $\Gamma \approx 1/2\beta$. To further simplify our SNR expressions, we make two final assumptions: first that $A'_T/\rho_L^2\gg 30$, which is equivalent to saying that the ghost image consists of at least $10 \times 10$ resolution cells; and second, that the bucket detector is at least as large as the area of the source beam, or $\beta \ge 1$. Applying these conditions we arrive at the following SNR expressions for pseudothermal and SPDC reflective ghost imaging through atmospheric turbulence,
$$\begin{gathered}
\label{eq:CSNR}
\text{SNR}_C = \frac{\displaystyle {\mathcal{T}}^2({\bm\rho}_p)}{\displaystyle\frac{T_0}{T_I}\Delta^2\text{Source} + \Delta^2\text{Path} + \frac{\Omega_BT_0^2}{T_I}\Delta^2\text{Detect}+\frac{T_0}{T_I}\Delta^2\text{Mix}} \\
\label{eq:QSNR}
\text{SNR}_Q = \frac{\displaystyle {\mathcal{T}}^2({\bm\rho}_p) \left[1+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right]^2}{\displaystyle \frac{4}{\Omega_BT_I}\Delta^2\text{Source} + \Delta^2\text{Path}\left[1+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right]^2 + \frac{T_0 4\sqrt{2}}{T_I}\Delta^2\text{Detect}\left[1+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right]+ \frac{T_0}{T_I}\Delta^2\text{Mix}\left[\frac{2}{\sqrt{3}}+\frac{1}{4\sqrt{\pi}\mathcal{I}}\right]}.\end{gathered}$$
where, for ease of comparison with the pseudothermal imager, we have inverted the image coordinates in the SPDC case.
The terms that appear in the noise denominators of Eqs. (\[eq:CSNR\]) and (\[eq:QSNR\]) are as follows: $$\begin{aligned}
&\Delta^2\text{Source} = \frac{A'_T (1+\beta^{-1})e^{4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}}{\sqrt{2\pi}\rho_L^2} , \\
&\Delta^2\text{Path} = {\mathcal{T}}^2({\bm\rho}_p)\left[e^{4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}\left(\Gamma+1\right)-1\right], \\
&\Delta^2\text{Detect} =\frac{{\mathcal{T}}({\bm\rho}_p) \rho_L^2 \sqrt{\pi}}{16 \sqrt{2} A_p \eta^2 \mathcal{I}^2}\frac{L^2}{A_b}, \\
&\Delta^2\text{Mix} = \frac{{\mathcal{T}}({\bm\rho}_p)}{\eta\mathcal{I}}\frac{ L^2}{A_b}e^{4\sigma_R^2}\nonumber\\
&\quad\quad+\frac{\pi\rho_L^2{\mathcal{T}}^2({\bm\rho}_p)}{A_p\eta\mathcal{I}}\left[\frac{4}{3}+\beta^{-1}\right]e^{4(\sigma_S^2+\sigma_T^2)},\end{aligned}$$ where $\sigma^2_\ell$ for $\ell = R, S, T$ denotes the turbulence-induced logamplitude variance on the reference ($R$), signal ($S$), and target ($T$) paths. Each of these noise terms has an important physical interpretation, as we explain below.
The first noise term, $\Delta^2\text{Source}$, captures the fluctuations associated with decorrelation of transverse locations in the image arising from the source’s time evolution. Thus, it is inversely proportional to the number of on-target resolution cells, $\rho_L^2/A'_T$, and the number of different source patterns measured. For the pseudothermal imager, for which $\Omega_BT_0 \gg 1$ holds, this is the number of field-coherence times in the integration time, so we find $\Delta^2\text{Source}$ scaled by $T_0/T_I$. For the SPDC imager we have $\Omega_BT_0 \ll 1$, so it is the detectors’ response time that limits the measurements’ correlation behavior, hence we find $\Delta^2\text{Source}$ scaled by $4/\Omega_BT_I$ in that imager’s SNR denominator.
The $\Delta^2\text{Path}$ term is due to the optical-path interactions that degrade the measurement, i.e., the turbulent atmosphere, and the scattering off the rough target. The noise contribution from target scattering can be mitigated by aperture averaging, which, as noted above, is quantified by $\Gamma$. The effect of the turbulence is somewhat more complicated, as it contributes a scaling factor $e^{4\sigma_\ell^2}$ to all field-variation terms on path $\ell$. In addition, the logamplitude fluctuations also cause a deviation that results in the squared-mean term in the denominator of Eq. no longer being canceled. These effects are particularly devastating because $\Delta^2\text{Path}$ is the only noise term whose strength is independent of the detector integration time. This is because we have assumed time-independent target behavior and $T_I$ less than the $\sim$ms turbulence coherence time. Consequently, regardless of the source brightness and integration time, $\Delta^2\text{Path}$ sets an ultimate upper limit on the SNR.
The noise term $\Delta^2\text{Detect}$ results from the statistics of photodetection at both the reference and bucket detectors, and as such is inversely proportional to the photon fluxes impinging on their active regions. Thus, $\Delta^2\text{Detect}$ is inversely proportional to the size of each pixel, $A_p$, and the solid-angle subtense factor $A_b/L^2$. Detection noise is mitigated by temporal averaging, but exacerbated by having a fast detector, so the pseudothermal $\Delta^2\text{Detect}$ is scaled by $\Omega_BT_0^2/T_I$, while for the SPDC imager it is scaled only by $T_0/T_I$. The SPDC case is also scaled by by $1+1/4\sqrt{\pi}\mathcal{I}$, a result of that source’s stronger-than classical correlation.
The $\Delta^2\text{Mix}$ term contains noise contributed by mixing between source plus path noise on the reference arm with detection noise on the signal arm, and vice versa. For both imagers this term has a temporal scaling of $T_0/T_I$, and for the SPDC imager the quantum correlation imparts an additional factor of $2/\sqrt{3} + 1/4\sqrt{\pi}\mathcal{I}$.
To get a better understanding of the SNR behavior for pseudothermal and SPDC ghost imagers, we will consider three limiting cases. The first is the long integration-time limit in which $\Delta^2\text{Path}$ dominates the noise denominators for both the pseudothermal and SPDC ghost imagers, leading to the *identical maximum (saturation) SNR value $$\begin{aligned}
\label{eq:SNRsat}
\text{SNR}_{\rm sat} =\frac{e^{-4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}}{\left[\Gamma+1\right]-e^{-4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}},\end{aligned}$$ regardless of the strength of the illumination.*
For our next two cases, we will assume the integration time is sufficiently short that SNR saturation is not approached. Then, in the high-brightness ($\mathcal{I}\gg 1$) case, for which detection noise can be neglected, performance is dominated by source fluctuations, $\Delta^2\text{Source}$, yielding $$\begin{aligned}
\label{eq:SNRC_H}
\text{SNR}_{H,C} =\frac{T_I}{T_0}\frac{ \sqrt{2 \pi} \rho_L^2 e^{-4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}}{A'_T (1+\beta^{-1})} {\mathcal{T}}^2(\bm{\rho}_p),\\
\label{eq:SNRQ_H}
\text{SNR}_{H,Q} = \Omega_BT_I \sqrt{\frac{\pi}{8}}\frac{\rho_L^2 e^{-4(\sigma_R^2+\sigma_S^2+\sigma_T^2)}}{A'_T (1+\beta^{-1})} {\mathcal{T}}^2(\bm{\rho}_p),\end{aligned}$$ for pseudothermal ($C$) and SPDC ($Q$) imagers, respectively. With all parameters other than source coherence-time being equal, the SPDC high-brightness SNR exceeds that of the pseudothermal source by a factor $\Omega_BT_0 \gg 1$, where $T_0$ is the pseudothermal source’s coherence time. This is because the much faster fluctuations in the SPDC output fields lead to faster decorrelation, and hence higher SNR. That said, however, the reader is reminded that cw SPDC operation is ordinarily in the low-brightness regime, hence we turn our attention now to that special case.
In the low-brightness, $\mathcal{I}\ll 1$, regime, SNR behavior is dominated by the beat between the detection noises, $\Delta^2\text{Detect}$. Here we find that $$\begin{aligned}
\label{eq:SNRC_L}
&\text{SNR}_{L,C}=\frac{T_I}{T_0}\frac{16\sqrt{2}}{\sqrt{\pi}}\frac{A_p\eta^2\mathcal{I}^2}{\Omega_B T_0\rho_L^2}{\mathcal{T}}(\bm{\rho}_p)\frac{A_b}{L^2}, \\
\label{eq:SNRQ_L}
&\text{SNR}_{L,Q}=\frac{T_I}{T_0}\frac{A_p\eta^2\mathcal{I}}{\pi \rho_L^2}{\mathcal{T}}(\bm{\rho}_p)\frac{A_b}{L^2},\end{aligned}$$ for the pseudothermal and SPDC cases. The SPDC SNR is linear in source brightness, whereas the pseudothermal SNR is quadratic in that parameter. This behavior is to be expected, because the SPDC low-brightness output approaches the biphoton state, which is known to have lower detection noise than classical-state light. Furthermore, $T_0$ for the pseudothermal source will greatly exceed $T_0$ for the SPDC, and, as noted in the preceding paragraph, $\Omega_BT_0 \gg 1$ holds for the pseudothermal source. It follows that the low-brightness SPDC SNR is much higher than that of the pseudothermal imager. The reader is cautioned to remember, however, that pseudothermal ghost imaging is typically performed in the high-brightness regime.
It is also instructive to compare our SNR results—which apply to reflective ghost imaging through atmospheric turbulence—to the transmissive ghost imaging, no-turbulence results from [@Erkmen2009SNR]. The principal differences between the SNR behaviors in these two cases is as follows. Our SNR expressions have the average intensity reflection coefficient ${\mathcal{T}}({\bm\rho}_p)$ appearing in lieu of the magnitude-squared of the field reflection coefficient $|T({\bm\rho}_p)|^2$ that is seen in the transmissive case. Moreover, the transmissive-case SNR from [@Erkmen2009SNR] has neither target-induced speckle nor turbulence contributions, so it has neither $\Delta^2\text{Path}$ term nor any turbulence-induced noise-magnification factors. The reflective case’s $\Delta^2\text{Source}$ term has a factor of $(1+\beta^{-1})$ and its $\Delta^2\text{Mix}$ term has a factor of $(4/3+\beta^{-1})$, but both $\beta^{-1}$ contributions are absent from the transmissive case. These additional reflective-case contributions are direct results of the rough-surface scattering. In both transmissive and reflective operation there is averaging of the source randomness at the target, but in the reflective case the fields scatter incoherently off the target, and their Fourier components are averaged at the bucket detector. This second averaging depends on the ratio of the source size to the bucket size, i.e., $\beta$. As $\beta$ increases, the bucket detector captures more light, this term becomes less significant, and we approach the transmissive behavior for these terms. Finally, the reflective-SNR terms associated with the bucket detector measurement in $\Delta^2\text{Detect}$ and $\Delta^2\text{Mix}$ are scaled by $L^2/A_b$, relative to the corresponding terms in the transmissive case. This scale factor is the inverse of the solid angle subtended at the target by the bucket detector. Its presence in the reflective case is a consequence of the quasi-Lambertian scattering from the rough surface
It follows from the SNR differences highlighted above that the low-brightness SNR asymptotes for transmissive and reflective operation only differ by the latter’s including the solid-angle subtense factor $A_b/L^2$. Their high-brightness SNR asymptotes are also similar, with the only difference being the reflective case’s factor of $(1+\beta^{-1})$. The most profound difference, however, is due to the $\Delta^2\text{Path}$ term that is only present for reflective operation. This term implies that the SNR of reflective ghost imaging has a maximum value, SNR$_{\rm sat}$, whereas no such saturation occurs in transmissive ghost imaging, i.e., its SNR can grow without bound as $T_I/T_0$ increases. Furthermore, in reflective imaging, this limiting effect can be quite severe: even with no turbulence, for $\beta=1$ we find ${\rm SNR}_{\rm sat} =3.26$, and for $\beta = 2$ we have ${\rm SNR}_{\rm sat}=5.54$. So, for realistic standoff sensing, the SNR will be limited to single-digit values if no further measures are taken to suppress the fluctuations arising from target-induced speckle and atmospheric turbulence.
Computational Ghost Imager {#sec:Comp}
==========================
We have developed the performance of reflective ghost imaging from classical and nonclassical sources in a quantum Gaussian-state framework. Next we consider a computational ghost imager, a variant of the pseudothermal case in which the “chaotic” signal field is created by deterministically modulating the wavefront of a laser with a spatial light modulator (SLM), and no reference-arm light beam is required. Instead, knowledge of the SLM phase pattern is used to calculate the signal-beam intensity at the target for use in lieu of a reference-beam measurement [@Shapiro2008Comp]. If an independent pseudorandom phase process is applied to each SLM pixel, then far-field propagation can be argued to yield Gaussian field statistics per the Central Limit Theorem. Thus it is appropriate for us to model the computational ghost imager’s far-field statistics with a Gaussian-Schell model, taking the source’s coherence length to be approximately the width of an SLM pixel, and source’s intensity diameter to be the diameter of the SLM.
The laser that illuminates the SLM will be taken to emit coherent-state light, hence the SLM’s output beam is in classical state and semiclassical theory will give quantitatively identical results for its far-field photodetection statistics as the full quantum treatment [@Erkmen2008Unified; @Erkmen2010AOP; @Shapiro2008Comp]. Also, because the reference field at the target is computed, rather than measured, it is most naturally modeled as a complex-valued Gaussian random process. Consequently, and without loss of generality, we choose to use semiclassical theory for our analysis of computational ghost imaging because it lets us deal with only classical random processes, rather than a combination of quantum field operators and classical random processes. We thus use $E_S'({\bm\rho},t)$ to represent the complex envelope of the target-plane signal field, and $\tilde{E}_R'({\bm\rho},t)$ to represent the computed complex-envelope of the field that would have been present on the CCD camera. These two complex-valued, zero-mean, Gaussian random processes are completely characterized by their nonzero autocorrelation and cross-correlation functions, which we take to be as given in Eqs. – for their field-operator pseudothermal counterparts. That these correlation functions are consistent with classical random-process theory follows from the pseudothermal source’s producing classical-state light.
The principal distinction between computational and pseudothermal ghost imaging—which is also its main advantage—is that there is neither detection noise nor turbulence on its computed reference-arm intensity, which is found via $$\begin{aligned}
\tilde{i}_p(t) = q\eta\int d\tau\, \tilde{P}_p(\tau) h(t-\tau), \end{aligned}$$ with $\tilde{P}_p(t) = A_p|\tilde{E}'_R({\bm\rho}_p,t)|^2$. The bucket detector photocurrent, in semiclassical theory, is given by $$\begin{aligned}
i_b(t) = \int d\tau\, [q\eta P_b(\tau) + \Delta i_b(\tau)] h(t-\tau), \end{aligned}$$ where $P_b(t)=\int_{{\cal A}_b} \!d{\bm\rho}\,|E_T'({\bm\rho},t)|^2$ is the impinging photon flux, and conditioned on that photon flux the shot noise $\Delta i_b(t)$ is a zero-mean random process with autocorrelation function ${\langle \Delta i_b(t_1) \Delta i_b(t_2) \rangle} = q^2 \eta P_b(t_1)\delta(t_1-t_2)$. The computational ghost image is then obtained from the correlation function $$\begin{aligned}
\label{CGI}
\tilde{C}({\bm\rho}_p) = \frac{1}{T_I}\int_{-T_I/2}^{T_I/2}\!dt\, \tilde{i}_p(t) i_b(t).\end{aligned}$$
Because the shot-noise is zero mean, given $P_b(t)$, it does not contribute to $\langle \tilde{C}({\bm\rho}_p)\rangle$, hence the computational ghost imager’s spatial resolution and image contrast analysis is identical to the Sec. \[sec:ResCon\] treatment of the pseudothermal ghost imager once we substitute the computational system’s classical random processes for the pseudothermal imager’s field operators, and then let $\rho_R\rightarrow \infty$, and set $\sigma_R^2=0$ to account for the absence of turbulence on the computed reference. Using this approach we immediately find the spatial resolution to be $\rho_L\sqrt{\tilde{\alpha}}$ where $$\begin{aligned}
\tilde{\alpha} = \frac{2\rho_S^2+a_0^2}{2 \rho_S^2},\end{aligned}$$ and the image contrast is identical to that given for the pseudothermal imager in Eq. . In the latter regard we note that we could replace $\tilde{i}_p(t)$ in Eq. (\[CGI\]) with the zero-mean process, $\Delta\tilde{i}_p(t) \equiv \tilde{i}_p(t) -\langle \tilde{i}_p(t)\rangle$, thereby achieving the same high-contrast operation as is realized with [ac]{}-coupling in pseudothermal ghost imaging.
Turning now to the computational ghost imager’s signal-to-noise ratio, we find more significant differences between its analysis and that of the pseudothermal ghost imager. In our Sec. \[sec:SNR\] SNR analysis, the second moment of the photocurrent cross-correlation function involved an eighth-order field moment, two sixth-order field moments, a fourth-order field-moment, three fourth-order moments of the turbulence, and a fourth-order moment of the target’s field-reflection coefficient. For the computational ghost imager’s SNR analysis we only have to evaluate an eighth-order field moment, a sixth-order field moment, two fourth-order moments of the turbulence, and the fourth-order moment of the target’s field-reflection coefficient. This simplification is because the other terms involved the path and detection noises on the reference arm, which are absent from the computational configuration. The computational SNR is then given by
$$\begin{aligned}
\label{eq:CompSNR}
\text{SNR}_{\tilde{C}}=
\frac{\displaystyle {\mathcal{T}}^2({\bm\rho}_p)\frac{T_I}{T_0}e^{-4(\sigma_S^2+\sigma_T^2)}}
{\displaystyle\frac{A'_T(1+\beta^{-1})}{\sqrt{2\pi}\rho_L^2}
+{\mathcal{T}}^2({\bm\rho}_p)\frac{T_I}{T_0}(\left[\Gamma+1\right]-e^{-4(\sigma_S^2+\sigma_T^2)})
+\frac{{\mathcal{T}}({\bm\rho}_p)}{\eta\mathcal{I}}\frac{ L^2}{A_b}}.\end{aligned}$$
It follows that computational ghost imaging shares the same saturation SNR as pseudothermal and SPDC ghost imaging, from Eq. (\[eq:SNRsat\]), with $\sigma_R^2$ set to zero because there is no turbulence in the computational imager’s reference arm. Likewise, computational ghost imaging has the same high-brightness SNR asymptote as pseudothermal ghost imaging, from Eq. (\[eq:SNRC\_H\]), with $\sigma_R^2 = 0$. On the other hand, computational ghost imaging has a significantly improved low-brightness SNR asymptote, given by $$\begin{aligned}
\label{eq:SNRComp_L}
\text{SNR}_{\tilde{C},L} = \frac{T_I}{T_0}\eta\mathcal{I}{\mathcal{T}}({\bm\rho}_p)\frac{A_b}{L^2},\end{aligned}$$ in comparison with *both pseudothermal and SPDC ghost imaging.*
For quantitative comparison between the SNRs of the pseudothermal, SPDC, and computational systems we will first have to put them on equal footing. The dimensionless brightness quantity $\mathcal{I}$ that appears in our SNR formulas is the photon flux per spatiotemporal mode of the source. However, because of the dramatically different coherence times of the SPDC ($\Omega_BT_0 \ll 1$) and the pseudothermal and SLM sources ($\Omega_BT_0 \gg 1$), comparisons based on equal $\mathcal{I}$ values will be for sources with dramatically different photon fluxes, whereas we would prefer to compare SNRs at equal photon fluxes. To do so we introduce the dimensionless quantity $\mathcal{I}_\Omega = \mathcal{I}\Omega_B T_0 = P \rho_0^2/a_0^2\Omega_B$, which is the photon flux per source spatial mode and detector temporal mode. Because all three systems are taken to have detectors with identical characteristics, identical values for $\mathcal{I}_\Omega$ will yield identical photon fluxes. Making this substitution we can easily compare the low-brightness SNRs from Eqs. , , and . In the extreme low-brightness regime, the computational imager has the best SNR, followed by the SPDC and pseudothermal imagers. We can also compare the high-brightness limits Eqs. and , where the shorter coherence time of the SPDC source gives it a clear advantage.
Because the different sources exit their low- and high-brightness SNR regimes and enter into SNR saturation for different photon fluxes, assessing that behavior requires us to explicitly choose some operating parameters. First, we assume the operating wavelength is $\lambda = 1.5\, \mu$m, the source coherence length is $\rho_0= 0.15/\pi\,$mm, the source intensity radius is $a_0= 3\,$cm, and that the bucket detector matches the source size, i.e., $A_b = \pi a_0^2$. These give us the on-target parameters of $\rho_L = 0.05/\pi\,$m, and $a_L = 10 \,$m. The photodetectors are taken to have $\eta=0.9$ quantum efficiency, with the CCD pixel area being $A_p = 0.1 \rho_L^2$. We assume the target is at $L= 1\,$km range, that it has an effective area of $A'_T = 50\, \text{m}^2$, and ${\mathcal{T}}({\bm\rho}_p) = 1$ at the point for which we will perform our SNR evaluations. All paths are assumed to have moderate-strength ($C_n^2 = 10^{-14}\,{\rm m}^{2/3}$) turbulence. Finally, we take the pseudothermal and computational sources to have temporal coherence times satisfying $T_0 = 10^3/\Omega_B \,$Hz, while for the SPDC source $T_0 =1/ 10^3\Omega_B \,$Hz.
We consider two scenarios: first, when the source is weak enough ($\mathcal{I}_\Omega = 1$) that the SPDC output is still a stream of biphotons, but we are not deep into the low-brightness regime; and second, when the source is sufficiently strong ($\mathcal{I}_\Omega = 10^4$) to make the SPDC output appear classical [@footnote5]. The resulting SNR curves for these two cases are shown in Figs. \[fig:SNR\_L\] and \[fig:SNR\_H\], respectively. From Fig. \[fig:SNR\_L\] we see that the SPDC imager is the worst performer when it is in its biphoton regime, while the computational ghost imager is to the top performer there. From Fig. \[fig:SNR\_H\], however, we see that when the SPDC output appears to be classical, its much shorter coherence time allows it to reach the saturation-SNR limit well before the other two systems, whose SNR curves are nearly identical.
![(Color online) SNRs for pseudothermal, SPDC, and computational ghost imaging vs. normalized integration time $\Omega_B T_I$. The source brightness is $\mathcal{I}_\Omega = 1$, so that the SPDC output is a stream of biphotons and all three systems are well below their respective high-brightness SNR asymptotes. Other parameters used in computing these curves are given in the text.[]{data-label="fig:SNR_L"}](SNRL.pdf){width="3.5in"}
![(Color online) SNRs for pseudothermal, SPDC, and computational ghost imaging vs. normalized integration time $\Omega_B T_I$. The source brightness is $\mathcal{I}_\Omega = 10^4$, so that the SPDC output appears to be classical and the pseudothermal and computational ghost imagers are in their high-brightness limits wherein they only differ by a factor of $e^{4\sigma_R^2}$. Other parameters used in computing these curves are given in the text.[]{data-label="fig:SNR_H"}](SNRH.pdf){width="3.5in"}
Discussion {#sec:Discussion}
==========
We have derived the key performance characteristics of reflective ghost imaging in the context of realistic standoff sensing applications. This was done within a Gaussian-state framework that allowed for analysis of classical pseudothermal sources and the nonclassical output of an SPDC. Our results for the spatial resolution, image contrast, and SNR were analogous to those calculated for transmissive ghost imaging in the absence of turbulence [@Erkmen2008Unified; @Erkmen2009SNR], with a few significant differences that result from the roughness of the target surface, turbulence on the optical paths, and the standoff nature of the measurement. First we shall discuss the impact of rough-surface scattering in the absence of turbulence, then we shall discuss the additional effects incurred when turbulence is present.
Rough-surfaced targets scatter their impinging illumination in a random manner, casting random speckle patterns in their far fields that, on average, correspond to a quasi-Lambertian distribution. Consequently, the ghost-imager’s bucket detector will, on average, collect only a fraction, $A_b/L^2$, of the target-scattered light. This angular subtense factor—which is absent in transmissive ghost imaging—appears in the average photocurrent cross-correlation functions in Eqs. and , and the noise terms associated with the bucket detector, i.e., $\Delta^2\text{Detect}$ and $\Delta^2\text{Mix}$. The surface roughness has another major effect on the SNR: its time-independent speckle pattern implies the existence of a finite saturation-SNR value in the limit of long integration time.
The target-speckle contribution that leads to a saturation SNR appears in $\Delta^2\text{Path}$, and is a function of $\beta$, the ratio of the source size to the bucket detector size. Here we may incur a trade-off between spatial resolution and SNR. Specifically, by enlarging the bucket detector we can increase the averaging of the target speckle, and thus the saturation SNR, without affecting the imager’s spatial resolution. Ultimately, however, there will be a practical limit beyond which the receiving aperture size can no longer be increased. Beyond that point $\beta$, and thus the saturation SNR, can only be increased by decreasing the source size $a_0$. However, as seen in Eq. , spatial resolution is inversely proportional to $a_0$, so increasing the saturation SNR in this manner will degrade spatial resolution.
The effects of turbulence on the image contrast and spatial resolution of reflective ghost imaging are similar those previously found for transmissive ghost imaging [@Cheng2009]. In particular: turbulence does not change the image contrast; turbulence in the target-to-bucket path has no effect on spatial resolution; and turbulence on the signal and reference paths degrades spatial resolution in the same manner, i.e., degradation occurs when they become smaller than the source size. In both SPDC and pseudothermal ghost imaging it is therefore advantageous to propagate the reference field through a controlled (turbulence-free) environment. Turbulence also degrades ghost-image SNR in two ways: first, its logamplitude fluctuations magnify several existing noise terms; and second, these fluctuations also contribute to the time-independent noise term $\Delta^2\text{Path}$, decreasing the saturation SNR from its speckle-only value.
In computational ghost imaging, the reference field is computed, removing the turbulence on that path, and thus improving both the spatial resolution and the SNR. The computational case also has significantly reduced noise for low-brightness illumination, as there is no longer detection noise on the reference arm, making its low-brightness SNR behavior superior to that of both pseudothermal and SPDC ghost imaging. More generally, Figs. \[fig:SNR\_L\] and \[fig:SNR\_H\] show that the computational ghost imager has a higher SNR than the pseudothermal system in both low-brightness and high-brightness operation, and than the SPDC system in its low-brightness (biphoton) regime. However, in high-brightness operation, the SPDC ghost imager’s significantly shorter coherence time enables it to reach its saturation SNR much quicker than either the computational or the pseudothermal ghost imagers. Although a cw SPDC will not be able to reach that high-brightness regime, pulsed systems are capable of doing so [@Gouet2009].
Insofar as standoff sensing is concerned, the essential conclusion to be drawn from the preceding summary of our work is simple: the computational approach is the ghost-imaging configuration that shows the most promise for this application. It has the best spatial resolution, the highest saturation-SNR, and general SNR performance second only to high-brightness SPDC operation. In addition, the computational approach obviates the need for a physical reference path, which makes it naturally amenable to obtaining 3D ghost images. In particular, pseudothermal and SPDC ghost imaging require that reference-arm measurements be made on an intensity pattern corresponding to the one that is projected onto the target. As shown theoretically in [@Shapiro2008Comp] and experimentally in [@Bromberg2009], computational ghost imaging allows reference intensity patterns to be computed at a variety of target ranges so that ghost images can be formed for these target ranges from the same bucket-detector data. This range sectioning is something that *cannot be done, without separate measurements for each possible target range, in pseudothermal or SPDC ghost imaging.*
We have reserved our final comment for our choice of image reconstruction via photocurrent cross-correlation. It has already been demonstrated [@Bromberg2009; @Katz2009] that ghost imaging reconstruction can be performed in transmission through more advanced methods, such as compressive sensing. A preliminary no-turbulence performance comparison between a simple cw-laser radar and ghost imaging in reflection via cross correlation has shown their similarity in spatial resolution and SNR [@Hardy2010]. If advanced image-reconstruction techniques can be successfully applied to computational ghost imaging, the resulting system could have notable advantages over current laser radars.
This work was supported by the U.S. Army Research Office Multidisciplinary University Research Initiative Grant No. W911NF-05-1-0197, the W. M. Keck Foundation Center for Extreme Quantum Information Theory, the DARPA Information in a Photon Program under U.S. Army Research Office Grant No. W911NF-10-1-0404, and the DARPA Quantum Sensors Program under AFRL Contract No. FA8750-09-C-0194,
[99]{}
T. B. Pittman, Y. H. Shih, D. V. Strekalov, and A. V. Sergienko, Phys. Rev. A [**52,**]{} 3429 (1995).
B. I. Erkmen and J. H. Shapiro, Phys. Rev. A [**77,**]{} 043809 (2008).
A. Valencia, G. Scarcelli, M. D’Angelo, and Y. Shih, Phys. Rev. Lett. [**94,**]{} 063601 (2005).
F. Ferri, D. Magatti, A. Gatti, M. Bache, E. Brambilla, and L. A. Lugiato, Phys. Rev. Lett. [**94,**]{} 183602 (2005).
B. I. Erkmen and J. H. Shapiro, Phys. Rev. A [**79,**]{} 023833 (2009).
B. I. Erkmen and J. H. Shapiro, Adv. Opt. Photon. [**2,**]{} 405 (2010).
J. H. Shapiro, Phys. Rev. A [**78,**]{} 061802 (2008).
Y. Bromberg, O. Katz, and Y. Silberberg, Phys. Rev. A [**79,**]{} 053840 (2009).
O. Katz, Y. Bromberg, and Y. Silberberg, Appl. Phys. Lett. [**95,**]{} 131110 (2009).
R. E. Meyers, K. S. Deacon, and Y. Shih, Phys. Rev. A [**77,**]{} 041801 (2008).
J. Cheng, Opt. Express [**17,**]{} 7916 (2009).
N. D. Hardy and J. H. Shapiro, Proc. SPIE [**7815,**]{} 78150L (2010),
For the sake of generality, we have allowed there to be turbulence in the reference path, although a lens could be used to cast a minified reference-beam pattern onto a CCD camera within a protected environment near the source, hence avoiding turbulence on that path.
J. H. Shapiro, “Imaging and optical communication through atmospheric turbulence,” in *Laser Beam Propagation in the Atmosphere, edited by J. W. Strohbehn (Springer-Verlag, Berlin 1978) Chap. 6.*
That no additional quantum-noise term is needed to ensure that this expression for ${\hat{E}}'_m({\bm\rho}',t)$ preserves free-field commutator brackets follows from the normal-mode decomposition for propagation through turbulence \[J. H. Shapiro, Appl. Opt. [**13,**]{} 2614 (1974)\].
A. Ishimaru, *Wave Propagation and Scattering in Random Media. Vol. 2: Multiple Scattering, Turbulence, Rough Surfaces, and Remote-Sensing (Academic Press, New York, 1978).*
G. R. Osche, *Optical Detection Theory for Laser Applications (Wiley-Interscience, Hoboken 2002).*
A similar square-law approximation is available for cases in which the turbulence strength is not uniformly distributed along each propagation path, but we have opted to limit our attention to uniform distributions.
J. W. Goodman, *Speckle Phenomena in Optics: Theory and Applications (Roberts & Co., Englewood, 2007).*
J. H. Shapiro, B. A. Capron, and R. C. Harney, Appl. Opt. [**20,**]{} 3292 (1981).
In quantum photodetection, the currents at the outputs of the bucket and CCD detectors are classical stochastic processes, $i_b(t)$ and $i_p(t)$, whose statistics are equivalent to those of the quantum measurements $\hat{i}_b(t)$ and $\hat{i}_p(t)$ given in the next paragraph.
J. H. Shapiro, IEEE J. Sel. Top. Quantum Electron. [**15,**]{}1547 (2009).
J. H. Shapiro and K.-X. Sun, J. Opt. Soc. Am. B [**11,**]{} 1130 (1994).
G. Scarcelli, V. Berardi, and Y. Shih, Phys. Rev. Lett. [**96,**]{} 063602 (2006).
S. M. Wandzura, J. Opt. Soc. Am. [**70,**]{} 745 (1980).
We say “appear classical” because the phase-sensitive cross-correlation function from Eq. (\[eq:QCorrPS\]) *always exceeds the classical limit set by the autocorrelation functions from Eq. (\[eq:AutoCorr\]). However, for $\mathcal{I} \gg 1$, that cross-correlation function is so close in value to the classical limit, given in Eq. (\[eq:CCorrPIS\]), to render this measure of nonclassicality exceedingly small.*
J. Le Gouët, D. Venkatraman, F. N. C. Wong, and J. H. Shapiro, Opt. Express [**17,**]{} 17874 (2009).
| {
"pile_set_name": "ArXiv"
} |
---
bibliography:
- './IDS-NF\_IDR.bib'
---
00a-Title-page/00a-Title-page 00b-AuthorList/00b-AuthorList 00c-Foreword/00c-Foreword 00d-Executive-summary/00d-Executive-summary 01-PPEG/01-0-PPEG 02-AccWG/02-0-AccWG 03-DetWG/03-0-DetWG 04-RandD-plan/04-RandD-plan
Acknowledgements {#acknowledgements .unnumbered}
================
During the course of the IDS-NF to date, we have been welcomed at a number of laboratories across the world and therefore thank the CERN, FNAL, and RAL laboratories and the Tata Institute of Fundamental Research for hosting the IDS-NF plenary meetings. The authors acknowledge the support of the European Community under the European Commission Framework Programme 7 Design Study: EUROnu, Project Number 212372. The authors acknowledge the support of Grants-in-Aid for Scientific Research, Japan Society for the Promotion of Science and the World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan. The work was supported by the Science and Technology Facilities Council under grant numbers PP/E003192/1, ST/H001735/1, ST/H003142/1 and through SLAs with STFC supported laboratories. This research was partially supported by the Director, Office of Science, Office of High Energy Physics, of the U.S. Department of Energy, under contract numbers DE-AC02-07CH11359, DE-AC02-05CH11231, DE-AC02-98CH10886, DE-AC05-06OR23177, and DE-AC05-00OR22725.
Appendices {#appendices .unnumbered}
==========
06-Appendix/06-Appendix
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'D. Yadykin, Y.Q. Liu, R. Paccagnella'
title: '[**Effect of kinetic resonances on the stability of Resistive Wall Mode in Reversed Field Pinch**]{}'
---
Introduction
============
Understanding physics and stabilization of the Resistive Wall Mode (RWM) is an important task for the successful operation of the present day and the future fusion devices. Converted from the ideal MHD mode in the presence of the wall with finite conductivity, and growing on the time scale of the magnetic field penetration time through the wall, RWM sets pressure limit in the advanced scenario of the tokamak device [@garofallo_prl_99_1; @gryaznev_baps_03_1; @shilov_pop_04_1]. It also limits the discharge duration of the Reversed Field Pinch (RFP) device [@alper_ppcf_89_1; @brunsell_pop_03_1; @bolzonella_32eps_1]. Control of RWM growth is necessary for the discharge times longer than the wall time.
Two general methods are known for the RWM stabilization: active control and stabilization via the mode resonance with continuum spectra or particle motions. For the active control the perturbed magnetic field measured by the set of sensors facing the plasma surface is used to generate the control signal by the set of active control coils. Such mechanism is extensively studied theoretically [@bishop_ppcf_89_1; @fitzpatrick_pop_96_1; @okabayashi_nf_98_1; @fitzpatrick_pop_99_1; @liu_pop_00_1; @pustovitov_ppr_01_1] and is successfully applied in the present day fusion devices, allowing operation with the plasma pressures above no-wall limit for tokamaks [@okabayashi_pop_01_1; @strait_pop_04_1; @sabbagh_prl_04_1] and resulting in substantial increase of the plasma discharge duration in RFPs[@brunsell_prl_04_1; @pacaggnella_prl_06_1] .
RWM suppression via the mode resonance with continuum spectra or particle motions is another possible control mechanism (also called ’rotational stabilization’). It was experimentally observed for the tokamak configuration [@reimerdes_phys_plasams_06_1] that RWM is stabilized when the plasma rotation frequency is sufficiently high. Several mechanisms were proposed that could explain the stabilization. In ideal MHD description the dissipation caused by the RWM resonance with Alfvén [@gregoratto_ppcf_01_1; @zheng_prl_05_1] or sound [@bondeson_prl_94_1; @betti_prl_95_1] continuum spectra could lead to the RWM stabilization. The dissipation strength depends on the plasma rotation frequency. The plasma rotation frequency $\Omega$ of the order of sound frequency $\omega_s$ is needed for the RWM stabilization by the resonance with continuum spectrum. Another mechanism that could lead to the RWM suppression in the region $\Omega\sim\omega_s$ is ion Landau damping. Accurate description of this mechanism requires kinetic treatment of the ion motion parallel to the magnetic field [@bondeson_ph_fluids_89_1]. In ideal MHD description this mechanism is modelled by adding dissipation term [@chu_phys_plasmas_95_1]. More physically consistent approach is used in semi-kinetic model [@bondeson_phys_plasmas_96_1] where damping term is calculated using drift kinetic energy principle for the large aspect ration approximation. Mode resonance with circulating and trapped ions are included in the model. Calculated damping term are then included in the MHD description. In order to explain observed RWM stabilization for the very low plasma rotation frequency [@reimerdes_prl_07_1] ($\Omega \ll \omega_s$) other model was proposed [@hu_prl_04_1] where the main dissipation channel is due to the mode resonance with the precession drift motion of the trapped particles. This model in fact predicts the RWM stabilization even without plasma rotation. The accurate prediction of the RWM stabilization for the low plasma rotation frequencies is an important issue for the future reactor experiment ITER, where the highest plasma rotation in steady-state scenario is predicted to be in the sound range [@polevoi_02_1].
Previous studies of the rotational stabilization of the RWM were performed mainly for the tokamak configuration. Available results for the RFP [@jiang_phys_plasmas_95_1; @guo_phys_plasmas_99_1] show that the plasma rotation frequency in the range of the Alfvén frequency ($\Omega \sim \omega_a$) is needed for the RWM stabilization by the mode resonance with continuum spectrum. Such high plasma rotation is not observed in the present day RFP devices, therefore rotational stabilization mechanism was not considered as realistic for RFP. On the other hand mode resonance with particle motions occurs for the low plasma rotation frequency and could be considered as a possible RWM stabilization mechanism also for RFP configuration. In this work numerical studies of the effect of kinetic resonances on the RWM stability are performed for reversed-field pinch (RFP) configuration. Particularly, parameters corresponding to the RFX device [@sonato_fid_03_1] is used. In RFP the values of poloidal and toroidal components are of the same order of magnitude. Toroidal component of the equilibrium magnetic field changes sign at the plasma edge. Several Fourier harmonics with different poloidal and toroidal mode numbers are seen in the RFP mode spectrum. Resonant harmonics with poloidal mode numbers $m=0,1$ are usually linearly stable. Several non-resonant harmonics with different toroidal mode numbers (positive and negative according to the sign of the safety factor $q$ of the resonant magnetic surface) are unstable growing on the times comparable with the equilibrium magnetic field penetration through the resistive wall and therefore are classified as RWMs. Experimental studies of RWM control in RFP have shown possibility of simultaneous suppression of several unstable harmonics with different toroidal mode numbers using active feedback [@brunsell_prl_04_1; @pacaggnella_prl_06_1].
Model {#sec2}
=====
In this section physical model is briefly described that is used in the present work. More complete derivation can be found in [@liu_phys_plasmas_08_1]. Stability of RWM for these studies is determined by solving numerically the system of MHD equations with toroidal plasma rotation: $$\begin{aligned}
(-i\omega+in\Omega)\mathbf{\xi}&=\mathbf{v}+(\mathbf{\xi} \cdot \nabla\Omega)R^2\nabla \phi \label{eq1}\\
\rho(-i\omega+in\Omega)\mathbf{v}&=-\nabla \cdot \mathbf{p} +\mathbf{j}\times \mathbf{B}+\mathbf{J}\times \mathbf{Q}- \rho[2\Omega\mathbf{\hat{Z}}\times \mathbf{v} +(\mathbf{v} \cdot\nabla\Omega)R^2 \nabla\phi] \label{eq2}\\
%+ \framebox{$\nabla \cdot \mathbf{\Pi}$} \\
(-i\omega+in\Omega)\mathbf{Q}&=\nabla\times(\mathbf{v}\times \mathbf{B})+(\mathbf{Q}\cdot\nabla\Omega)R^2\nabla\phi \label{eq3}\\
(-i\omega+in\Omega)p&=-\mathbf{v}\cdot\nabla P \label{eq4}\\
\mathbf{j}&=\nabla\times \mathbf{Q} \label{eq5}\end{aligned}$$ where $\omega$ is the complex mode frequency ($\omega=i\gamma-\omega_r$ where $\gamma$ - mode growth or damping rate and $\omega_r$ - mode rotation frequency), $\mathbf{B}$,$\mathbf{J}$,$P$- equilibrium magnetic field, current density and pressure respectively, $\mathbf{\hat{Z}}$ - unit vector in the vertical direction, $\rho$ - plasma density, $\Omega$ - plasma rotation frequency in toroidal direction $\phi$, $\mathbf{\xi}$,$\mathbf{v}$,$\mathbf{j}$,$\mathbf{Q}$ - plasma displacement,perturbed velocity,perturbed current,perturbed magnetic field respectively, $\textbf{p}$ - pressure tensor.
Kinetic terms are included into the MHD equations via the pressure tensor components. Pressure tensor is defined as $$\mathbf{p}=\textbf{I}p+p_\parallel\mathbf{\hat{b}\hat{b}}+p_\bot (\mathbf{I}-\mathbf{\hat{b}\hat{b}})
\label{pres_tensor}$$ where $p$ is the scalar fluid pressure perturbation, $p_\parallel,p_\bot$ are the particle parallel and perpendicular components of the kinetic pressure perturbations, $\mathbf{\hat{b}}=\textbf{B}/B$, $B=\mid \textbf{B} \mid$, $\textbf{I}$ is the unit tensor. For the particular perturbation with the mode number $n$, the parallel and perpendicular components of the pressure tensor are written as: $$\begin{aligned}
p_\parallel e^{-i\omega t+in\phi}& = \sum_{e,i}\int\int d\epsilon d\Lambda M v_\parallel^2 f_L^1 \\
p_\bot e^{-i\omega t+in\phi}& = \sum_{e,i}\int\int d\epsilon d\Lambda \frac{1}{2} M v_\bot^2 f_L^1
\end{aligned}$$ Here the summation is over the electron and ion plasma components, integration is carried out over the particle energy $\epsilon$ and pitch angle $\Lambda$, M is the particle mass, $v_\parallel, v_\bot$ are the parallel and perpendicular velocity components respectively with respect to the equilibrium magnetic field, $f_L^1$ is the perturbed particle distribution function. It is derived as: $$f_L^1=-f_\epsilon^0\epsilon_k e^{-i\omega t+in\phi}\sum_{m,l} X_m H_{ml}\lambda_{ml}e^{-in\tilde{\phi}(t)+im<\dot{\chi}>+il\omega_b t}
\label{pert_distr}$$ where subscripts $n$,$m$,$l$ mark Fourier components along toroidal and polidal angles and along bounce particle orbit, $f_\epsilon^0$ - derivative of the equilibrium distribution function (taken to be Maxwellian for thermal particles) with respect to the particle energy $\epsilon$, $\epsilon_k$ - kinetic energy of the particles, $H_{ml}$ - perturbed Lagrangian component in Fourier space, $\lambda_{ml}$ mode-particle resonance operator, $\tilde{\phi}(t)=\phi(t)-<\dot{\phi}> t $ - the periodic component of the particle velocity in toroidal direction, $<..>$ means average over the particle bounce period. The mode-particle resonant operator is calculated as: $$\lambda_{ml}=\frac{n(\omega_{*N}+(\epsilon_k-3/2)\omega_{*T}+\Omega)-\omega}{n\omega_d+ n\Omega+[\alpha(m+nq)+l]\omega_b-i\nu_{eff} -\omega}
\label{res_oper}$$ where $\omega_{*N},\omega_{*T}$ - diamagnetic frequencies associated with the density and temperature gradients respectively, $\omega_b$ - bounce frequency, $\omega_d$ - bounce averaged magnetic drift frequency, $\nu_{eff}$ - effective collision frequency, $\alpha=$0 for the trapped particles, $\alpha=$1 for the passing particles. The energy transfer between the mode and the particles is described by the imaginary part of the resonant operator. It can be seen that in case of finite mode growth (damping) rate and collision frequency energy is transferred between the mode and the particles for all plasma rotation frequency values. Mode-particle interaction is enhanced in the resonant regions, where the real part of the resonant operator is equal to zero. The collisionallity effect on the RWM stability (via collision frequency term $i\nu_{eff}$ of the resonant operator) is not considered in this work. Using and to calculate the pressure tensor components and substituting in , the self-consistent formulation is obtained, that allows to study RWM stability including the effect from kinetic resonances. System of ideal MHD equations where the ion Landau damping is modelled by the parallel viscosity can be obtained from the equations - by using only scalar pressure part of the pressure tensor , adding the viscous stress tensor $\nabla \cdot \Pi=\kappa \vert k_\Vert v_{thi}\vert\rho(\mathbf{v \cdot \hat{b}\hat{b}})$ in eq. , where $\kappa$ is the coefficient determining the damping strength, $k_\Vert=(n-\frac{m}{q})/R$, $v_{thi}$ is the ion thermal velocity and adding the term $-5/3P \nabla \cdot \mathbf{v}$ in the eq. .
Numerical details
=================
System of equations - is an eigenvalue problem that can be written in the matrix form as $$\gamma_0 B X=A X
\label{eigen}$$ where $\gamma_0$ is the complex eigenvalue, $A,B$ - MHD operators, $X$ - eigenvector. In this work the eigenvalue problem is solved numerically using MHD stability code MARS-K [@liu_phys_plasmas_08_1] in toroidal coordinates ($s$,$\chi$,$\phi$), where $s=\sqrt{1-\frac{\psi}{\psi_0}}$ is the normalized poloidal flux ($\psi_0$ is the poloidal flux value on the axis), $\chi$ is the generalized poloidal angle and $\phi$ is the geometrical toroidal angle. Equilibrium quantities are obtained using equilibrium solver CHEASE [@lutjens_cpc_96_1]. Note that eigenvalue problem is non-linear as resonant operator includes eigenvalue $\gamma_0$ and therefore $A=A(\gamma_0)$. Outer iterative loop is required in order to obtain the solution of the eigenvalue problem.
Results
=======
Equilibrium {#eqsec}
-----------
Equilibria closely modelling the RFP device RFX [@sonato_fid_03_1] is used. Circular plasma cross-section is taken with major radius $R_0=2.0$ m and aspect ratio $\varepsilon=0.23$. The RFP equilibrium parameters are $F\equiv\frac{B_\phi (a)}{<B_\phi>}$=-0.06, $\Theta\equiv\frac{B_\theta (a)}{<B_\phi>} $=1.41, where $B_\theta (a)$,$B_\phi (a)$ are the values of toroidal and poloidal components of the equilibrium magnetic field at the plasma edge and $<..>$ means averaging over the plasma column. The plasma current value is $I_p$=1.6 MA, on-axis toroidal field $B_0$=1.53 T. The poloidal beta value is $\beta_{pol}$=4$\%$, the on-axis values of electron and ion temperatures and densities are: $T_{e0}$=1 kEv, $T_{i0}$=400 eV, $n_{e0}=n_{i0}$=2.0$\cdot$ 10$^{19}$ m$^{-3}$. The pressure profile $p(r)$ is given by the expression $p(r)=n(r)*T(r)$, where $n(r)=n_0(1-(r/a)^6)$, $T(r)=T_0(1-(r/a)^3)$. The on-axis and edge safety factor values are: $q_0$=0.161, $q_a$=-0.01. In Fig. \[equil\_profs\] the poloidal and toroidal components of the magnetic field, pressure profile, and safety factor profile are shown for the described equilibrium as a functions of the normalized poloidal flux.
![[*Equilibrium quantities as a functions of the normalized poloidal flux. a) normalized equilibrium magnetic field components (solid -toroidal component, dashed - poloidal component), b) normalized pressure, c) safety factor* ]{}[]{data-label="equil_profs"}](equil_profs_new.eps)
RWM stability without rotation
------------------------------
The RWM spectrum in RFP is usually characterized by the presence of several unstable RWMs with different toroidal mode numbers $n$. Fourier harmonics with different helicities (positive and negative toroidal mode numbers) are visible in the spectrum due to the edge reversal of the equilibrium toroidal field. In Fig. \[n\_spec\] the spectrum of unstable RWMs is shown calculated for the equilibrium described in the previous section .
![[*Spectrum of unstable RWMs. Black - $n<0$, grey - $n>0$.*]{}[]{data-label="n_spec"}](n_spec_new.eps)
The poloidal structure (different poloidal Fourier harmonics) of the normal displacement $\xi_n$ of the most unstable toroidal harmonic (n=-6) are shown in Fig. \[n6\_struc\]. It is observed that the harmonic with m=1 is dominant. This result points to the weak poloidal coupling in the RFP (differently from the tokamak configuration). In the further discussion the stability of the Fourier harmonic with n=-6 will be discussed if not stated otherwise.
![[*Poloidal harmonics ($m$=-5..5) of the RWM with n=-6. Dominant m=1 poloidal harmonic is shown in magenta solid line. All other harmonics are not visible.*]{}[]{data-label="n6_struc"}](n6_struc_new1.eps)
Ideal MHD stability with plasma rotation. Continuum resonances {#cont_res}
--------------------------------------------------------------
Plasma rotation opens the possibility for the interaction of the static non-resonant RWM with the stable waves in the plasma. In ideal MHD such interaction is due to the mode resonance with continuum spectrum. Two continuum spectra are known to be resonant with RWM: Alfvén continuum and sound continuum. The condition for the resonance appearance can be written in general as $n\Omega+\omega_r=\omega_c$ where $\Omega$ is the plasma rotation frequency, $\omega_r$ is the RWM rotation frequency and $\omega_c$ is the continuum frequency. For the static RWM ($\omega_r=0$) this condition becomes $n\Omega= \omega_c$. RWM resonance with the Alfvén continuum spectrum appears when the condition $\vert n\Omega\vert =\vert \omega_{ca} \vert \equiv \vert k_\| v_a \vert$ is satisfied. Here $v_a\equiv B/(\mu_0\rho)^{1/2}$ is the Alfvén velocity and $k_\|\equiv (m/q-n)$ is the parallel component of the RWM wave vector. In Fig. \[al\_res\] the mode growth rate $\gamma$ and mode rotation frequency $\omega_r$ (both normalized by the wall time $\tau_w$) dependence is shown on the plasma rotation frequency $\Omega$ (normalized by the Alfvén frequency on the plasma axis $\omega_a^0$). The pressureless equilibrium with the equilibrium parameters $F,\Theta$ given in sec \[eqsec\] is used.
![[*Mode resonance with the Alfvén continuum. Complex eigenvalue (mode growth rate - $\gamma \tau_w$ and mode frequency $\omega_r \tau_w$) dependence on the plasma rotation $\Omega/\omega_a^^0$ normalized to the on-axis value of Alfvén frequency.*]{}[]{data-label="al_res"}](al_res.eps)
Because of the non-resonance nature of the RWM in RFP, no resonance between the mode and the Alfvén waves is possible at vanishing or slow plasma rotation. The obtained value of the plasma rotation frequency that starts to give the resonance with Alfvén wave (seen at peak of the growth rate $\gamma\tau_w$) is $\Omega/\omega_a^0 \approx $0.06. This critical value agrees very well with the analytical prediction that is calculated using the resonance condition given above. The RWM growth rate $\gamma\tau_w$ starts to decrease for plasma rotation frequencies $\Omega>\omega_{ca}$ accompanied by the finite mode frequency appearance. This points to the damping effect appearance from the mode-continuum resonance, although no compete RWM stabilization is seen for the studied $\Omega$ range. The obtained value of the plasma rotation frequency for the mode resonance with the Alfvén continuum is much less that the one obtained in the previous studies [@guo_phys_plasmas_99_1] due to the fact that the internal RWM with $n=-6$ is used in the present studies with the resonant surface closer to the plasma (and therefore smaller $k_\Vert$). It should be noted that the experimentally observed plasma rotation frequency in RFX is still much smaller than the value obtained here for the mode resonance with Alfvén continuum ($\Omega$ value is of order of $10^{-3}\omega_a$ [@guazotto_ppcf_09_1]).
The resonance of the static RWM with sound continuum appears when $n\Omega=\omega_{cs}\equiv k_\|v_s$ where $v_s=(p/\rho)^{1/2}$ is the sound velocity. The behaviour of the complex eigenvalue in the plasma rotation frequency range corresponding to the resonance condition is shown in Fig. \[so\_res\] (line with $\times$) where an equilibrium with finite pressure is chosen. The obtained plasma frequency value that corresponds to the resonance is $\Omega/\omega_a^0 \approx$ 0.004 and also agrees well with the value given by the analytical predictions. Note also that $\omega_{cs}\ll \omega_{ca}$.
According to the studies performed in [@bondeson_ph_fluids_89_1] the resonant behaviour seen at $\omega_{cs}$ is not physical due to the fact that the ideal MHD model is unable to describe accurately the particle motion along the field lines. A reasonable damping model can be introduced [@chu_phys_plasmas_95_1] to remove the unphysical resonant behaviour observed in Fig. \[so\_res\] by damping the parallel sound wave in the ideal MHD model. This parallel wave is damped physically by ion Landau damping mechanisms. In this work viscous damping model is used (see Sec. \[sec2\]) where a numerical coefficient $\kappa$ is introduced to measure the strength of the parallel sound wave damping. In Fig. \[so\_res\] the complex eigenvalue behaviour is shown for the different values of the sound wave damping coefficient.
![[*Complex eigenvalue (mode growth rate - $\gamma \tau_w$ and mode frequency $\omega_r \tau_w$) dependence on the plasma rotation $\Omega/\omega_a$ normalized to the Alfvén frequency. Blue (and $\times$)- $\kappa$=0.0, green (and $o$) - $\kappa$=0.5, red (and $\square$) - $\kappa=1.0$, cyan (and $\lozenge$) - $\kappa=1.5$*]{}[]{data-label="so_res"}](so_res_new.eps)
It is seen that the the large $\kappa$ values (strong damping) removes the resonant behaviour of the RWM growth rate. At the same time viscous damping results in RWM suppression, with the larger suppression rate for the larger $\kappa$ values.
Kinetic resonances {#kin_res}
------------------
A more physically consistent description of RWM stability for the low values of the plasma frequency is obtained considering the mode interaction with plasma particle drift motions. Mode particle interaction is introduced in the MHD model through the pressure tensor and is included into the resonant operator : $$\lambda=\sum_{e,i}\sum_{m,l}\sum_\alpha\frac{n(\omega_{*N}+(\epsilon_k-3/2)\omega_{*T}+\Omega)-\omega}{n\omega_d+ n\Omega+[\alpha(m+nq)+l]\omega_b-i\nu_{eff} -\omega}$$ where sum is over the poloidal Fourier harmonics $m$, bounce harmonics $l$, particle fraction $\alpha$, and particle species ($e,i$). It is seen that resonance with several particle motions is possible depending on the value of the plasma rotation frequency. In these studies the mode resonance with precession drift and bounce motions is considered. Full picture of the mode-particle resonance is complex due to the fact that contribution from the both trapped and passing fractions of the different plasma species (electrons and ions) should be taken into account. Moreover, particular particle motion (characterized by the motion’s frequency value) have in general complex dependence on the pitch angle $\Lambda$ and particle kinetic energy $\epsilon_k$ on each flux surface. Total effect of the RWM-particle resonances on the mode stability is obtained numerically considering mentioned dependencies. Kinetic frequencies averaged both over $\epsilon$ and $\Lambda$ are discussed in this section. Such simplified approach allows to make some qualitative predictions for the mode-particle resonances. The average kinetic frequencies for the ion plasma component are shown in Fig. \[dfreq\] as functions of the normalized poloidal flux.
![[*Ion kinetic frequencies averaged over the pitch angle $\Lambda$ and particle kinetic energy $\epsilon_k$. Solid line -$\omega_b$\*0.1, dotted line - $\omega_p$\*0.01, dashed line -$\omega_d$, dashed-dotted line - $\omega_*$. All frequencies are normalized to the central Alfvén frequency value$\omega_a^0$.*]{}[]{data-label="dfreq"}](kin_freq_new2.eps)
The frequency for the passing particles $\omega_p$ is equal to $\omega_b$ when $\alpha=1$ in . Note that $\omega_p$ and $\omega_b$ are scaled (by factors 0.01 and 0.1 respectively) to be able to compare profiles of different kinetic frequencies on one figure. The value of $\omega_*$ is the sum of the diamagnetic frequencies due to the density and temperature gradients. The plasma rotation frequency is not shown on the figure above, but comparing the experimental value ($\Omega\sim 10^{-3}\omega_a$ [@guazotto_ppcf_09_1]) with the calculated results it is seen that $\Omega \sim\omega_d$. The following approximate frequency ordering is obtained for the studied case: $$\omega_r\ll\Omega\sim\omega_d\sim\omega_*\ll\omega_b \sim\omega_{ca}\sim\omega_s^{i0}<\omega_p<\omega_a^0
\label{freq_order}$$ where $\omega_s^{i0}\simeq 0.1 \omega_a^0$ is the central ion sound frequency. It is seen that the ion bounce frequency value is in the range both with the sound frequency and also with the frequency of the mode resonance with Alfvén continuum.
In the present work the plasma frequency range $\Omega \lesssim \omega_s^{i0}$ is studied. Two specific frequency subregions can be deduced following the frequency ordering obtained above: $\Omega\sim\omega_d$ and $\omega\sim\omega_b$. In each of this frequency region the plasma rotation frequency is close to the certain kinetic frequency (precession drift of bounce frequencies) and therefore the mode-particle resonance is expected to affect the RWM stability. It is useful for the further studies to have qualitative estimation of particle fractions and particle species that possibly could be involved in the mode-particle resonance in the mentioned frequency regions. The following qualitative analysis aims the estimation of the dominant terms of the resonant operator $\lambda$ for the two frequency subregions for different values of $\alpha$ (particle trapping) and $l$ (bounce harmonic number). In the subregion $\Omega \sim \omega_d$ the following conditions could lead to the mode-particle resonance:
- $\alpha=0$. In this case $\lambda \simeq \frac{\omega_*+\Omega}{\omega_d+\Omega+(l/n)\omega_b}$. For $l=0$ mode resonance with precession drift motion has dominant effect. Both electrons and ions contribute. For $l \neq 0$ bounce motion could contribute to the mode-particle resonance for the slow rotation when $\omega_d\sim(l/n)\omega_b$. As $\omega_d/\omega_b\approx 0.02$ across most of the plasma column (see Fig. \[dfreq\]) condition $l\ll n$ should be satisfied that is not the case in the present studies ($min\vert (l/n) \vert=1/6$). Note that $\omega_d\sim\omega_b$ is seen locally in the plasma core or the plasma edge regions and therefore bounce motion can affect the RWM stability in these regions.
- $\alpha=1$. In this case resonant operator has general form . Passing particles can contribute to the mode-particle resonance for the slow plasma rotation when $(m+nq+l)/n \sim \omega_d/\omega_p$. As $\vert \omega_d/\omega_p \vert<10^{-2}$ in the present studies the contribution from the passing particles can be expected when $m=-l, \vert q \vert \ll 1.0$ that is satisfied locally near the resonant surface. The main effect on the RWM stability in the low rotation frequency subregion is expected to come from the mode resonance with the precession drift frequencies. The effect from the bounce motion or passing particles is local.
Similar analysis for $\Omega \sim \omega_b$ gives:
- $\alpha=0$ As from the frequency ordering $\omega_d \ll \omega_b$ mode resonance only with bounce motion is possible, i.e. when $l \neq 0$. Then $\lambda \simeq \frac{\omega_*+\Omega}{(l/n)\omega_b+\Omega}$. The mode-particle resonance condition is satisfied when $(l/n)\omega_b\simeq -\Omega$. For present studies ($n=-6$) bounce harmonic with $l=6$ should have the dominant contribution. Note also that this condition could give the contribution to the RWM stability in the intermediate values of the plasma rotation frequency (i.e $\omega_d\lesssim\Omega<\omega_b$). This happens when $l/n<1$. Only ion component is involved as the electron bounce frequency is much higher.
- $\alpha=1$ In this case $\lambda\simeq \frac{\omega_*+\Omega}{[(m+nq+l)/n]\omega_p+\Omega}$. Passing particles can contribute to the mode resonance with the bounce motion when $(m+nq+l)/n \sim \omega_b/\omega_p$. As $\vert \omega_b/\omega_p \vert \sim 10^{-1}$ in this studies the contribution of the passing particles can be expected for $l=-m, \vert q \vert \lesssim 0.1$ that points to the possibility of the global effect from the passing particles (note that $\vert q \vert <0.16$ in the present studies).
Both trapped and passing particles could contribute to the mode-particle resonance for the plasma rotation in the order of ion sound frequency (note that $\omega_s\sim \omega_b$ from the frequency ordering ). Trapped particle contribution comes from the bounce motion, passing particles effect could be global (note that this is different from the case of the low rotation frequencies $\Omega\sim\omega_d$).
Comparison of the viscous and kinetic damping models
----------------------------------------------------
It was mentioned above that the reasonable damping model can be used in ideal MHD description to model physical ion Landau damping. Viscous damping model is used here with the damping coefficient $\kappa$=1.0 (strong damping) for the comparison with the MHD description including kinetic effects. The results of comparison are shown in Fig. \[comp\_fl\_kin\]. It is seen that kinetic damping model gives slightly lower growth rate for the low plasma rotation frequencies ($\Omega\lesssim \omega_d$) while the higher suppression rate is obtained for $\Omega \sim \omega_b$ using the viscous damping model. Increase of the RWM growth rate for $\Omega>5\cdot 10^{-2}\omega_a^0$ is explained by the resonant RWM behaviour in vicinity of the mode resonance with Alfvén continuum (see frequency ordering ).
![[*Complex eigenvalue (mode growth rate - $\gamma \tau_w$ and mode frequency $\omega_r \tau_w$) dependence on the plasma rotation $\Omega/\omega_a$ normalized to the Alfvén frequency. Solid line with $\times$- fluid damping (sound wave damping with $\kappa=1.0$), dashed line with $\square$ - model with kinetic resonances included.*]{}[]{data-label="comp_fl_kin"}](comp_fl_kin_new1.eps)
Different kinetic resonances
----------------------------
Several kinetic frequencies are included in the resonant operator and therefore can contribute to the total effect on the RWM stability. In the previous section ’full’ model is used taking into account mode resonance with precession drift and bounce motions for trapped and passing particles. In order to evaluate the importance of the particular kinetic resonance and particle fraction on the total effect the following studies are performed. It can be seen from the definition of the resonant operator that the frequencies of the kinetic resonances are included in additive manner, giving the possibility to exclude all but one resonance from the calculations. Particle fractions (trapped or passing) can be also treated separately. In Fig. \[kin\_res\_comp1\] the ’full’ model (blue curve) is compared to the ’only bounce’ model (red curve) when only bounce frequencies are taken into account and with ’only precession drift’ model (green curve) when only precession drift frequencies are taken into account. Both trapped and passing particles are included.
![[*Complex eigenvalue (mode growth rate - $\gamma \tau_w$ and mode frequency $\omega_r \tau_w$) dependence on the plasma rotation $\Omega/\omega_a$ normalized to the Alfvén frequency. Solid line with $\square$ -full model, dashed line with $\times$ - resonance only with $\omega_d$ is considered, dashed-dotted line with $o$ - resonance only with $\omega_b$ is considered. Both trapped and passing particles are included*]{}[]{data-label="kin_res_comp1"}](kin_res_comp1_new1.eps)
Two plasma rotation frequency regions can be distinguished separated by the plasma rotation frequency value $\Omega/\omega_a^0 \approx 10^{-2}$. For the region $\Omega/\omega_a^0 < 10^{-2}$ eigenvalue calculated for the ’full’ model case is closely follows that calculated for the ’only precession drift’ model. Note that the mode growth rate for these two cases is slightly less than that for the ’only bounce’ model. For the region $\Omega/\omega_a^0 > 10^{-2}$ situation becomes opposite i.e. the eigenvalue trace corresponding to the ’full’ model follows that for the ’only bounce’ model. Resonant behaviour of the RWM growth rate is seen for the ’only precession drift’ model in this frequency range. It is attributed to the mode resonance with the Alfvén continuum spectrum.
In Fig. \[kin\_comp2\] the same resonances are compared where only trapped particles are included. It is seen that the mode growth rate behaviour is different in the second plasma rotation frequency region $\Omega/\omega_a^0 > 10^{-2}$ for the ’full’ and ’only bounce’ models. Now similar resonant behaviour is seen for all three models at $\Omega/\omega_a^0 \approx 0.05$. The growth rate for the ’full’ model is almost identical with that for the ’only precession drift’ model for the whole range of the studied plasma rotation frequencies.
![[*Complex eigenvalue (mode growth rate - $\gamma \tau_w$ and mode frequency $\omega_r \tau_w$) dependence on the plasma rotation $\Omega/\omega_a$ normalized to the Alfvén frequency. Solid line with $\square$ -full model, dashed line with $\times$ - resonance only with $\omega_d$ is considered, dashed-dotted line with $o$ - resonance only with $\omega_b$ is considered. Only trapped particles are included*]{}[]{data-label="kin_comp2"}](kin_res_comp2_new1.eps)
The effect of the mode resonance with different particle fractions for two frequency regions ($\Omega/\omega_a^0 < 10^{-2}$ and $\Omega/\omega_a^0 > 10^{-2}$ ) can be seen. In the region $\Omega/\omega_a^0 < 10^{-2}$ the effect of the trapped and passing particles on the RWM stability is observed. RWM is suppressed due to the resonance with the precession drift motion ($\times$ curve in Fig. \[kin\_res\_comp1\]) and destabilized due to the resonance with the passing particles($o$ curve in Fig. \[kin\_res\_comp1\]). This two effects almost cancel each other as it is seen from the behaviour of the ’full’ model eigenvalue ($\square$ curve in Fig. \[kin\_res\_comp1\]). Note that there is no effect from the bounce motion in this frequency region (compare $o$ curves in Figs.\[kin\_res\_comp1\] and \[kin\_comp2\]). In the region $\Omega/\omega_a^0 > 10^{-2}$ only passing particles have the effect on the RWM stability. Indeed, similar eigenvalue behaviour for the three studied cases seen in Fig. \[kin\_comp2\] points (somewhat surprisingly) that the mode resonance with bounce motion does not affects the RWM stability. The resonant behaviour and subsequent mode suppression seen for $\Omega/\omega_a^0 > 5\cdot10^{-2}$ is attributed to the mode resonance with the Alfvén continuum. On the other hand substantial mode suppression is observed in Fig. \[kin\_res\_comp1\] ($o$ and $\square$ curves) in this frequency range that effectively prevents the mode growth due to the resonance with Alfvén continuum. Note that although the frequency of the passing particles is in general much higher than the frequencies of the bounce and precession drift motions, the possibility of the mode resonance with the passing particles for low frequencies exists and the conditions for such interaction are qualitatively estimated in Sec. \[kin\_res\] (cases $\alpha=1$ for bounce and precession drift frequency ranges).
Equilibrium plasma pressure
---------------------------
In the previous studies of the kinetic effects for tokamak configuration [@liu_phys_plasmas_08_2] it was shown that equilibrium pressure affects the mode interaction with the particle motions. Both plasma pressure profile and absolute value can change the behaviour of the complex eigenvalue. In order to investigate the effect of pressure on the mode resonance with the plasma particles for RFP configuration the studies of the plasma pressure profile pressure absolute value (characterized by the $\beta_p$ parameter) are performed below.
### Pressure profile
Pressure profile shape affects the value of the pressure derivative and therefore changes the values of $\omega_*$ frequencies and also (according to [@liu_phys_plasmas_08_2]) affects the precession drift frequency term of the resonance operator . Two pressure profiles are compared in the present studies: the one used for the studies above (see Sec.\[eqsec\]) and more flat one that is characterized by the following density and temperature profiles: $n^{i,e}(r/a)=n_0^{i,e}(1-(r/a)^12)$,$T^{i,e}(r/a)=T_0^{i,e}(1-(r/a)^6)$. The results of comparison are shown in Fig. \[dif\_pres\]. It is see that by changing pressure profile the diamagnetic frequency profile is changed (dashed-dotted line in Fig. \[dif\_pres\]b). Note also the slight change of $\omega_d$ (dashed line) and $\omega_p$(solid line) profiles. The RWM is suppressed more (Fig. \[dif\_pres\]c) in the region $\Omega\sim\omega_d$ for the case of the flat pressure profile, but for $\Omega\sim\omega_b$ more peaked profile gives better suppression. RWM destabilization is seen for $\Omega\gtrsim 5\cdot 10^{-2}\omega_a^0$ in both cases attributed to the effect from the mode resonance with the Alfvén continuum.
![[*Results of the pressure profile studies. a) normalized pressure radial profiles; b) kinetic frequencies (solid - $\omega_p*0.01$, dashed - $\omega_d$, dashed-dotted - $\omega_*$), $\omega_b$ profiles are not shown as no effect on RWM stability is found from the mode resonance with the bounce motion ; c) RMW growth rate dependence on the plasma rotation frequency; d) RWM rotation frequency dependence on the plasma rotation frequency. Black - $n(r)=n_0(1-(r/a)^6,T(r)=T_0(1-(r/a)^3$; grey - $n(r)=n_0(1-(r/a)^{12},T(r)=T_0(1-(r/a)^6$;*]{}[]{data-label="dif_pres"}](dif_pres_new2.eps)
### Effect of $\beta_p$
The absolute value of the plasma pressure (characterized usually by the beta value) is one of the most important parameters in the fusion research. Operation with the high beta value is necessary for the efficient operation of the fusion reactor. In this work poloidal beta $\beta_p\equiv\frac{8\pi<p>}{I_p^2}V_{tot}$ is used for the pressure value characterization, where $<p>$ is the equilibrium plasma pressure averaged over the plasma volume $V_{tot}$, $I_p$ - plasma current. It is seen from the frequency ordering that for $n=-6$ Fourier harmonic the effects from the kinetic and continuum resonances occurs for the same plasma rotation value $\omega\sim\omega_s^{i0}$. In order to separate these effects another unstable Fourier harmonic is chosen with $n$=-5 for which the mode resonance with the Alfvén continuum is shifted towards higher plasma rotation frequencies. Indeed using analytical expression (see Sec. \[cont\_res\]) the mode resonance with Alfvén continuum appears for this harmonic, when $\Omega\approx 0.2\omega_a^0$. The results of the studies for three $\beta_p$ values ($\beta_p$=5%,15%,25%) are shown in Fig. \[kin\_pr\_val\]
![(Colour online) [*Results of the pressure value studies for the Fourier component with $n=-5$. a) normalized pressure radial profiles; b) kinetic frequencies (solid - $\omega_p*0.01$, dashed - $\omega_d$, dashed-dotted - $\omega_*$); c) RMW growth rate dependence on the plasma rotation frequency; d) RWM rotation frequency dependence on the plasma rotation frequency. Blue - $\beta_p$=5%; red - $\beta_p$=15%, green -$\beta_p$=25%;*]{}[]{data-label="kin_pr_val"}](kin_res_pr_val_new.eps)
It is observed (Fig. \[kin\_pr\_val\]a) that the values of kinetic frequencies depend on poloidal beta. Dependence for different kinetic frequencies can be approximated as: $\omega_p(\beta_p)\sim \sqrt{\beta_p}$, $\omega_d(\beta_p)\sim \beta_p$, $\omega_*(\beta_p) \sim \beta_p$ that is similar as for the large aspect ration scaling for tokamak equilibrium [@liu_phys_plasmas_08_2]. Change of the kinetic frequencies magnitude causes also the change of the plasma rotation frequency of the mode-particle resonances. The consequence of such resonance shift is that the stabilizing effect from the kinetic resonances in the range $\Omega\sim\omega_b$ is seen only for the low value of poloidal beta (blue and red curves In Fig. \[kin\_pr\_val\]c) for the studied range of plasma rotation frequencies. The mode destabilization seen for the $\Omega/\omega_a^0\gtrsim 0.01$ is similar for all cases and can be attributed to the effect of the mode resonance with Alfvén continuum. Note that no such substantial shift of the kinetic frequencies is seen for tokamaks, due to the fact that the beta range where unstable RWM is observed is of the order of one percent that is much smaller than in RFP.
Mode resonance with electron and ion plasma components for $\Omega\sim\omega_d$
-------------------------------------------------------------------------------
It was shown in [@liu_phys_plasmas_08_1] for tokamak that temperature difference between electron and ion components can sufficiently affect the RWM stability for the low plasma rotation frequencies ($\Omega\sim\omega_d$) pointing to the different contribution from the mode resonance with electrons and ions. In general, contribution both from ion and electron components to the total effect is present at each flux surface due to the specific dependence of the precession drift frequency on the pitch angle for each flux surface. In particular, $\omega_d$ changes sign as a function of $\Lambda$ (see Fig. 1 in [@liu_phys_plasmas_08_1]) and therefore mode resonance both with electrons and ions is possible on the same flux surface as $\omega_d$ depends on the sign of the particle charge (note that such $\omega_d(\Lambda)$ dependence makes possible RWM suppression with zero plasma rotation). However for the present studies it is seen (Fig. \[dfreq\]) that the average ion precession drift frequency is positive for all radial points. This means that precession drift frequency for particular plasma component has predominantly one sign as a function of $\Lambda$ on each flux surface and therefore the contributions from electrons and ions could be separated by the sign of the plasma rotation (i.e. mode resonance only with ion or electron component will determine the total effect on the RWM stability for the particular plasma rotation direction). In Fig. \[dif\_rot\] the complex eigenvalue behaviour is shown as a function of the plasma rotation frequency in the low plasma rotation region ($\Omega\sim\omega_d$). The complex eigenvalue is calculated for the Fourier harmonic $n=-6$ taking into account only the mode resonance with the precession drift motion. Two values of the ion- electron temperatures ratio $C_p=\frac{T_0^i}{T_0^i+T_0^i}$ ( $C_p=0.33$ and $C_p=0.5$) are compared for the positive and negative plasma rotation directions. First $C_p$ value corresponds to the ion-electron temperature ratio used in the present work ($T_{e0}$=1 $keV$, $T_{i0}$=0.5 $keV$), second $C_p$ value corresponds to the case of equal central temperatures of electrons and ions. In this case $\omega_*^i=-\omega_*^e, \omega_d^i=-\omega_d^e$.
![[*Complex eigenvalue dependence on the plasma rotation frequency for the Fourier component with $n=-6$. Solid line with $\square$ - $\Omega>0,C_p=0.33$, solid line with $\lhd$ - $\Omega>0,C_p=0.5$, dashed line with $\diamondsuit$ - $\Omega<0, C_p=0.33$, dashed line with $\rhd$ - $\Omega<0,C_p=0.5$. Mode growth rate $\gamma$ for each curve is normalized to the value for $\Omega=10^{-4}$.*]{}[]{data-label="dif_rot"}](cp_effect_new1.eps)
Absolute value of the mode rotation frequency ($\vert \omega_r \vert \tau_w$) is plotted in order to make comparison easier (mode frequency rotation changes sign when the plasma rotation direction is reversed). It is seen that the mode growth rate behaves differently for the positive and negative plasma rotation directions in the case when $C_p=0.33$. For negative $\Omega$ (dashed line with $\diamondsuit$) mode stabilization effect is observed for the lower $\Omega$ values than for the positive rotation direction (solid line with $\square$). This could mean that $\vert \omega_d^e \vert < \vert \omega_d^i \vert$ and the mode resonance with electron component occurs for the lower $\Omega$ values. In the case when $C_p=0.5$ the same effect is observed for the positive and negative $\Omega$ directions (solid line with $\lhd$ and dashed line with $\rhd$ are overlapped). This is expected result as the kinetic frequency values are equal for $C_p=0.5$ and therefore the effect from the different plasma components is independent on the plasma rotation direction.
Discussion and conclusions
==========================
The results of the present investigations show that the kinetic effects do not change the RWM stability substantially for the studied RFP equilibrium. An attempt to explain such result could be made by considering the dispersion relation for the RWM written in terms of the fluid and kinetic energy components (see eq. 8 in [@hu_prl_04_1]). This dispersion relation is used for the studies of the kinetic effects via the perturbative approach. Qualitatively, the stability of RWM is determined by the relation between fluid and kinetic energy components. It was shown [@liu_phys_plasmas_09_1] that for the pressure driven RWM seen in tokamak configuration kinetic and fluid terms are comparable. Therefore changes in kinetic term can substantially affect the overall stability of RWM. The calculations made for the studied here RFP equilibrium show that for the current driven RWM fluid term is much larger that the kinetic term. Therefore qualitatively kinetic term does not play an important role for the RWM stability in RFP. More work should be done in order to make the general conclusion about the role of the kinetic energy on the RWM stability in RFP.
It is also seen during the present studies, that the observed effect on the RWM stability from the kinetic resonances comes mainly from the mode resonance with the [*passing particles*]{} for the plasma rotation of the order of ion sound frequency. Surprisingly no effect from the trapped particles (bounce motion) is seen in this region. It is observed that also for the low rotation (of the order of precession frequency) mode resonance with the passing particles have quantitatively similar effect on RWM stability as compared to the mode resonance with the precession drift motion.
In conclusion the studies of the RWM interaction with particle motions was investigated for RFP configuration. The equilibrium close to the experimental was used. The effect from the kinetic resonances is small, but the eigenvalue behaviour for the low plasma rotation frequency is explained better by the model with kinetic resonances than the ideal MHD model. Two regions of stabilization was found similarly to the previous work\[\] that correspond to the different plasma particles.
[1]{} A. M. Garofalo, A. D. Turnbull, M. E. Austin, J. Bialek, M. S. Chu, K. J. Comer, E. D. Fredrickson,R. J. Groebner, R. J. La Haye, L. L. Lao, E. A. Lazarus, G. A. Navratil, T. H. Osborne, B. W. Rice, S. A. Sabbagh, J. T. Scoville, E. J. Strait, and T. S. Taylor, Phys. Rev. Lett., [**82**]{}, 3811 (1999). M Gryaznevich, C G Gimblett, T C Hender, D F Howell, S Pinches, Y Liu, A Bondeson, Bull. Am. Phys. Soc., [**48**]{}, 307 (2003). M. Shilov, C. Cates, R. James, A. Klein, O. Katsuro-Hopkins, Y. Liu, M. E. Mauel, D. A. Maurer, G. A. Navratil, T. S. Pedersen, and N. Stillits, R. Fitzpatrick, S. F. Paul, Phys. Plasmas, [**11**]{}, 2573 (2004). B Alper, M K Bevir, H A B Bodin, C A Bunting, P G Carolan, J Cunnane, D E Evans, C G Gimblett, R J Hayden, T C Hender, A Lazaros, R W Moses, A A Newton, P G Noonan, R Paccagnella, A Patel, H Y W Tsui and P D Wilcock, Plasma Phys. Controlled Fusion, [**31**]{}, 205 (1989). P. R. Brunsell, J.-A. Malmberg, D. Yadikin, and M. Cecconello, Phys. Plasmas, [**10**]{}, 3823 (2003). T. Bolzonella, E. Martines, D. Terranova, P. Zanca, R. Cavazzana, L. Grando, N. Pomaro, G. Serianni, N. Vianello, M. Zuin, Proceedings of the 32nd EPS Plasma Physics Conference (Tarragona) ECA (European Physical Society, Mulhouse Cedex, France, 2005), Vol. 29C, p. 1.107. C. M. Bishop, Plasma Phys. Controlled Fusion, [**31**]{}, 1179 (1989). R. Fitzpatrick and T. H. Jensen, Phys. Plasmas [**3**]{}, 2641 (1996). M. Okabayashi, N. Pomphrey and R.E. Hatcher, Nucl. Fusion, [**38**]{}, 1607 (1998). R. Fitzpatrick and E. P. Yu, Phys. Plasmas, [**6**]{}, 3536 (1999). Y. Q. Liu, A. Bondeson, C. M. Fransson, B. Lennartson, and C. Breitholtz, Phys. Plasmas, [**7**]{}, 3681 (2000). V. D. Pustovitov, Plasma Phys. Rep., [**27**]{} 195 (2001). M. Okabayashi, J. Bialek, M. S. Chance, M. S. Chu, E. D. Fredrickson, A. M. Garofalo, M. Gryaznevich, R. E. Hatcher, T. H. Jensen, L. C. Johnson, R. J. La Haye, E. A. Lazarus, M. A. Makowski, J. Manickam, G. A. Navratil, J. T. Scoville, E. J. Strait, A. D. Turnbull, and M. L. Walker, Phys. Plasmas, [**8**]{}, 2071 (2001). E. J. Strait, J. M. Bialek, I. N. Bogatu, M. S. Chance, M. S. Chu, D. H. Edgell, A. M. Garofalo, G. L. Jackson, R. J. Jayakumar, T. H. Jensen, O. Katsuro-Hopkins, J. S. Kim, R. J. La Haye, L. L. Lao, M. A. Makowski, G. A. Navratil, M. Okabayashi, H. Reimerdes, J. T. Scoville, A. D. Turnbull, and DIII-D Team, Phys. Plasmas, [**11**]{} 2505 (2004). S. A. Sabbagh, R. E. Bell, J. E. Menard, D. A. Gates, A. C. Sontag, J. M. Bialek, B. P. LeBlanc, F. M. Levinton, K. Tritz, and H. Yuh, Phys. Rev. Lett., [**97**]{}, 045004 (2006). P. R. Brunsell, D. Yadikin, D. Gregoratto, R. Paccagnella, T. Bolzonella, M. Cavinato, M. Cecconello, J. R. Drake, A. Luchetta, G. Manduchi, G. Marchiori, L. Marrelli, P. Martin, A. Masiello, F. Milani, S. Ortolani, G. Spizzo, and P. Zanca, Phys. Rev. Lett., [**93**]{}, 225001 (2004). R. Paccagnella, S. Ortolani, P. Zanca, A. Alfier, T. Bolzonella, L. Marrelli, M. E. Puiatti, G. Serianni, D. Terranova, M. Valisa, M. Agostini, L. Apolloni, F. Auriemma, F. Bonomo, A. Canton, L. Carraro, R. Cavazzana, M. Cavinato, P. Franz, E. Gazza, L. Grando, P. Innocente, R. Lorenzini, A. Luchetta, G. Manduchi, G. Marchiori, S. Martini, R. Pasqualotto, P. Piovesan, N. Pomaro, P. Scarin, G. Spizzo, M. Spolaore, C. Taliercio, N. Vianello, B. Zaniol, L. Zanotto, and M. Zuin, Phys. Rev. Lett., [**97**]{}, 075001 (2006). H. Reimerdes, T. C. Hender, S. A. Sabbagh, J. M. Bialek, M. S. Chu, A. M. Garofalo, M. P. Gryaznevich, D. F. Howell, G. L. Jackson, R. J. La Haye, Y. Q. Liu, J. E. Menard, G. A. Navratil, M. Okabayashi, S. D. Pinches, A. C. Sontag, E. J. Strait, W. Zhu,1 M. Bigi, M. de Baar, P. de Vries, D. A. Gates, P. Gohil,R. J. Groebner, D. Mueller, R. Raman, J. T. Scoville, W. M. Solomon, the DIII-D Team, JET-EFDA Contributors, and the NSTX Team, Phys. Plasmas [**13**]{}, 056107 (2006).
D. Gregoratto, A. Bondeson, M. S. Chu and A. M. Garofalo, Plasma Phys. Controlled Fusion [**43**]{}, 1425 (2001). L. J. Zheng, M. Kotschenreuther, and M. Chu, Phys. Rev. Lett. [**95**]{}, 255003 (2005). A. Bondeson and D. J. Ward, Phys. Rev. Lett. [**72**]{}, 2709 (1994). R. Betti, J. P. Freidberg, Phys. Rev. Lett. [**74**]{}, 2949 (1995). A. Bondeson and R. Iacono, Phys. Fluids B [**1**]{}, 1431 (1989). M. S. Chu, J. M. Greene, T. H. Jensen, R. L. Miller, A. Bondeson, R. W. Johnson, and M. E. Mauel, Phys. Plasmas [**2**]{}, 2236 (1995). A. Bondeson and M. S. Chu, Phys. Plasmas [**3**]{}, 3013 (1996). H. Reimerdes, A. M. Garofalo, G. L. Jackson, M. Okabayashi, E. J. Strait, M. S. Chu, Y. In, R. J. La Haye, M. J. Lanctot, Y. Q. Liu, G. A. Navratil, W. M. Solomon, H. Takahashi, and R. J. Groebner, Phys. Rev. Lett. [**98**]{}, 055001 (2007). Bo Hu and R. Betti, Phys. Rev. Lett. [**93**]{}, 105002 (2004). A.R. Polevoi, S.Yu. Medvedev, V.D. Pustovitov, V.S. Mukhovatov, M. Shimada, A.A. Ivanov, Yu.Yu. Poshekhonov, M.S. Chu Proceedings of the 19th Int. Conf. on Fusion Energy (Lyon) (IAEA, Vienna 2002), CD-ROM file CT/P-08 Z. X. Jiang, A. Bondeson and R. Paccagnella, Phys. Plasmas [**2**]{}, 442 (1995). S. C. Guo, J. P. Freidberg, and R. Nachtrieb, Phys. Plasmas [**6**]{}, 3868 (1999). P. Sonato, G. Chitarin, P. Zaccaria, F. Gnesotto, S. Ortolani, A. Buffa, M. Bagatin, W. R. Baker, S. Dal Bello, P. Fiorentin, L. Grando, G. Marchiori, D. Marcuzzi, A. Masiello, S. Peruzzo, N. Pomaro and G. Serianni, Fusion Eng. Des. [**66**]{}, 161 (2003). Y. Q. Liu, M. S. Chu, I. T. Chapman, and T. C. Hender, Phys. Plasmas [**15**]{}, 112503 (2008) H. Lütjens, A. Bondeson, O. Sauter, Comp. Phys. Commun. [**97**]{}, 219 (1996). L. Guazzotto and R. Paccagnella, Plasma Phys. Controlled Fusion [**51**]{}, 065013 (2009). Y. Q. Liu, M. S. Chu,C. G. Gimblett, and R. J. Hastie, Phys. Plasmas [**15**]{}, 092505 (2008). Y. Q. Liu, I. T. Chapman, M. S. Chu, H. Reimerdes, F. Villone, R. Albanese, G. Ambrosino, A. M. Garofalo, C. G. Gimblett, R. J. Hastie, T. C. Hender, G. L. Jackson, R. J. La Haye, M. Okabayashi, A. Pironti, A. Portone, G. Rubinacci, and E. J. Strait, Phys. Plasmas [**16**]{} 056113 (2009).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
W. R. Gibbs\
New Mexico State University\
Las Cruces NM, 88003
title: 'Low-energy Antiproton Interaction with Helium'
---
=-.17in =1.5cm =2ex =6ex =2ex
=1.2
Introduction
============
Recent work on the trapping of low-energy antiprotons [@holz] presented the measurement of the annihilation rate on helium atoms which remained in the partial vacuum. It was found that the rate increased sharply as the cooling took place and then suddenly dropped to the background level. There has been some speculation concerning the cause of this sudden decrease in the rate. One possibility could be that there is a barrier of a few millivolts which would only be important when the energy of the antiprotons reaches this level.
While these observations formed the initial motivation for the work presented here there are additional reasons why the knowledge of this potential is interesting.
First, the experiment mentioned above points out the possibility of carrying out experiments in which this annihilation cross section could be measured and, in a more advanced version, the x-ray transitions might be measured as well.
Second, there are more general reasons for knowing this potential. For example in the particle remnants of the big bang there could well be some antiprotons left in space. To the extent that these antiprotons remained free they would have thermalized with the ambient background of photons and so would have an average energy corresponding to a temperature of around 2.7 K. Their interactions would be with naturally occurring hydrogen and helium.
As a third reason, it is conceivable that there could exist pockets of potential or perhaps changes in slope in the antiproton-atom potential which could allow metastable states to be formed. These might be associated with configurations of the electrons which are strongly deformed from a spherical distribution. Such pockets of potential are perhaps more probable in larger aggregates of matter but this simple case of helium provides an entry into the subject.
There exist some calculations [@voronin; @hughes] of a nature related to this reaction but the author knows of no calculation of this precise process.
The calculation of the annihilation cross section as a function of energy requires the knowledge of the electrostatic potential between the $\bar{p}$ and the neutral helium atom, the annihilation potential of the nucleus and the solution of the Schrödinger equation with these two potentials.
It is natural to approach the problem of antiproton-helium electrostatic interaction, considering at the same time the proton-helium potential. In this way a check on the calculation is provided. The calculation of the proton-helium ab initio (variational) potential by Kolos and Peek[@kolos2] is considered to be the standard reference.
While one might at first think that the calculations are almost the same (just the charge of the external particle is changed), because of the art of choosing the variational wave function they can be very different. For instance, Kolos and Peek (following earlier work by Kolos[@kolos1] and Wolniewicz[@wol]) chose a trial electron wave function which was elliptical in form, thus naturally tending to surround the two positive charges. This function is quite appropriate for the two nuclei separated by small or intermediate distances. For large distances, where the electrons are nearly spherically distributed around the helium nucleus, this wave function can be expected to less efficient. For the case of the antiproton this form is perhaps less appropriate.
Since it is desirable to do both calculations on the same footing, a single trial function has been use for both cases. It is based on an expansion in spherical coordinates about the helium nucleus. This function is very efficient for small or large distances between the two nuclei. It has more difficulty representing the wave function well at intermediate distances where it may require a large number of terms.
In the following section the expressions for the calculation of the variational ratio are derived. In section \[formal\] the results of the potentials are given and in section \[ann\] the annihilation cross section is computed.
Calculation of the potentials\[formal\]
=======================================
The trial wave function chosen for this work is based on a spherical expansion about the helium nucleus in terms of spherical harmonics and orthogonal polynomials in the (scaled) radial distance of the electrons from the center. It is taken of the same form for the proton and antiproton problem so that a direct comparison can be made. This partial wave expansion goes over into the polarizability expansion naturally.
Variational Integrals
---------------------
A completely general form of the wave function for two electrons can be expressed as
$$\psi({\bf r}_1,{\bf r}_2)=\sum \psi_{\ell_1,m_1,\ell_2,m_2,n_1,n_2}
Y_{\ell_1}^{m_1}(\hat{{\bf r}_1})Y_{\ell_2}^{m_2*}(\hat{{\bf r}_2})
L_{n_1}(y_1)L_{n_2}(y_2)e^{-\frac{1}{2}(y_1+y_2)}$$
where $L_n(r)$ is the Laguerre polynomial of order 2 (usual notation $L_n^{(2)}(r)$) and $y_1=r_1/a$, $y_2=r_2/a$, , “a” being a variational parameter which sets the scale of the system. The quantities $\psi_{\ell_1,m_1,\ell_2,m_2,n_1,n_2}$ are the components of the wave function in this basis and embody (reduced by one index, see below) the other variational parameters of the calculation. It has been assumed, as usual [@friedrich], that the singlet spin state dominates the lowest energy configuration leaving the spatial wave function symmetric.
Since the two-electron wave function can depend only on the relative value of the $\phi$ angles, because of the symmetry around the $^4$He-$\bar{p}$ axis, the most general form reduces to
$$\psi({\bf r}_1,{\bf r}_2)=\sum \psi_{m,\ell_1,\ell_2,n_1,n_2}
Y_{\ell_1}^m(\hat{{\bf r}_1})Y_{\ell_2}^{m*}(\hat{{\bf r}_2})
L_{n_1}(y_1)L_{n_2}(y_2)e^{-\frac{1}{2}(y_1+y_2)}.$$
It is this expression which will be used in the following work. The sums on $\ell$ and $n$ are taken to $\ell_{max}$ and $n_{max}$. The condition that the spatial electron wave function is symmetric requires that the $\psi$’s are symmetric under the interchange $(\ell_1,n_1)\leftrightarrow (\ell_2,n_2)$.
The normalization of the wave function is given by
$$\sum (n_1+1)(n_1+2)(n_2+1)(n_2+2)\psi_{m,\ell_1,\ell_2,n_1,n_2}^2.$$
The hamiltonian for the problem is
$$H=-\frac{\hbar^2}{2m}\nabla^2_1-\frac{\hbar^2}{2m}\nabla^2_2
-\frac{2e^2}{r_1}-\frac{2e^2}{r_2}+\frac{e^2}{|{\bf r}_1-{\bf r}_2|}
\pm\frac{e^2}{|{\bf r}_1-{\bf R}|}\pm\frac{e^2}{|{\bf r}_2-{\bf R}|},$$
where ${\bf R}$ is the vector separating the antiproton (proton) from the helium nucleus.
In order to carry out the variational calculation the computation of the expectation value of the trial wave function of each of these terms is needed.
The radial derivative part of the kinetic energy can be expressed as
$$\frac{-\hbar^2}{2ma^2}K(n_1,n'_1,n_2,n'_2)\delta_{\ell_1,\ell'_1}
\delta_{\ell_2,\ell'_2}\delta_{m,m'}$$
where $K$ is an integer.
Twice the contribution to the kinetic energy for one of the electrons is given by $$KK(n',n)=\delta_{n,n'}\frac{(n+1)(n+2)}{2}-2na_{n',n}+2(n+2)a_{n',n-1}
-2(n+1)b_{n',n}$$ so that the contribution to both will be $$K(n_1,n'_1,n_2,n'_2)=\frac{1}{2}\left[KK(n'_1,n_1)(n_2+1)(n_2+2)\delta_{n'_2,n_2}
\right.$$ $$\left.+KK(n'_2,n_2)(n_1+1)(n_1+2)\delta_{n'_1,n_1}\right] ,$$ with $$a_{n_1,n_2}\equiv \int_0^{\infty} e^{-y}L_{n_1}(y)L_{n_2}(y)dy$$ $$a_{n,n}=\frac{(n+1)(n+2)(2n+3)}{6};\ \ a_{n,n+m}=a_{n,n}+
\frac{m(n+1)(n+2}{2}$$
$$b_{n_1,n_2}\equiv \int_0^{\infty} ye^{-y}L_{n_1}(y)L_{n_2}(y)dy
=\frac{(n+1)(n+2)}{2};\ \ \ \ n={\rm min\ of\ } (n_1,n_2)$$
The angular momentum part contributes
$$K_L(m,m',\ell_1,\ell'_1,\ell_2,\ell'_2,n_1,n'_1,n_2,n'_2)=
\delta_{m,m'}\delta_{\ell_1,\ell'_1}\delta_{\ell_2,\ell'_2}$$
$$\times \left[
\ell_1(\ell_1+1)(n_2+1)(n_2+2)a_{n_1,n'_1}\delta_{n'_2,n_2}+
\ell_2(\ell_2+1)(n_1+1)(n_1+2)a_{n_2,n'_2}\delta_{n'_1,n_1} \right].$$
The contribution to the potential energy of the electron-He interaction is given by
$$\frac{
2e^2\delta_{m,m'}\delta_{\ell_1,\ell'_1}\delta_{\ell_2,\ell'_2}}{a}
\left[(n_1+1)(n_1+2)b_{n_2,n'_2}\delta_{n_1,n'_1}+
(n_2+1)(n_2+2)b_{n_1,n'_1}\delta_{n_2,n'_2}\right] .$$
The expectation value of the electron-electron interaction involves calculating
$$\frac{e^2}{a}\int d{\bf y}_1d{\bf y}_2 Y_{\ell'_1}^{*m'}(\hat{{\bf y}}_1)
Y_{\ell'_2}^{m'}(\hat{{\bf y}}_2)
L_{n'_1}(y_1)L_{n'_2}(y_2)\frac{e^{-(y_1+y_2)}}{|{\bf y}_1-{\bf y}_2|}
Y_{\ell_1}^{m}(\hat{{\bf y}}_1)Y_{\ell_1}^{*m}(\hat{{\bf y}}_2)
L_{n'_1}(y_1)L_{n'_1}(y_1).$$
$$\frac{e^2}{a}\sum \sqrt{\frac{(2\ell_1+1)(2\ell_2+1)}
{(2\ell'_1+1)(2\ell'_2+1)}}C_{\ell_1,L,\ell'_1}^{0,0,0}
C_{\ell_2,L,\ell'_2}^{0,0,0}C_{\ell_1,L,\ell'_1}^{m,M,m'}
C_{\ell_2,L,\ell'_2}^{m,M,m'} I(n_1,n'_1,n_2,n'_2,L),$$
where
$$I(n_1,n'_1,n_2,n'_2,L)\equiv
\int y_1^2 dy_1 y_2^2 dy_2 e^{-(y_1+y_2)}L_{n'_1}(y_1)L_{n'_2}(y_2)
\frac{y_<^L}{y_>^{L+1}}L_{n_1}(y_1)L_{n_2}(y_2)$$
The coefficient for the expansion of a product of Laguerre polynomials
$$L_{n_1}(y)L_{n_2}(y)=\sum A_{n_1,n_2,n_3}L_{n_3}(y)$$
can be found by using the explicit expression of the Laguerre polynomials
$$L_n(y)=\sum_{m=0}^n (-1)^m\left(\begin{array}{c} n+2\\n-m\end{array}\right)
\frac{y^m}{m!}= \sum_{m=0}^n d_{n,m}y^m$$
$$(n_3+1)(n_3+2)A_{n_1,n_2,n_3} =(n_1+2)!(n_2+2)!(n_3+2)!$$
$$\times\sum_{m_1,\ m_2,\ m_3=0}^{n_1,\ n_2,\ n_3}
\frac{(-1)^{m_1+m_2+m_3}(m_1+m_2+m_3+2)!}{
(n_1-m_1)!(m_1+2)!m_1!(n_2-m_2)!(m_2+2)!m_2!(n_3-m_3)!(m_3+2)!m_3!}$$
With the definition
$$I_{k,n}=\int_0^{\infty}y^ke^{-y}dy\int_0^yx^ne^{-x}dx$$
$$I_{k,n}=\left\{\begin{array}{l}
n!\sum_{m=n+1}^{\infty}\frac{(k+m)!}{m!2^{m+k+1}}
\ \ \ \ \ \ \ k\le 0 \\
\\
n!\left[ k!-\sum_{m=0}^n\frac{(k+m)!}{m!2^{m+k+1}}\right]
\ \ \ k\ge 0 \end{array}\right\}$$
we have
$$I(n_1,n'_1,n_2,n'_2,L)=$$ $$\sum A_{n_1',n_1,j_1}A_{n_2',n_2,j_2} d_{j_1,k_1}
d_{j_2,k_2} \left[I_{k_1+L-1,k_2-L+1}+I_{k_2+L-1,k_1-L+1}\right].$$
To evaluate the interaction energy with the antiproton (proton) we need
$$\int_0^{\infty} \frac{y^ne^{-y}}{|{\bf y} -{\bf Y} |}dy=
\sum P_L(\cos\theta)\left[\int_0^Y\frac{y^{n+L}}{Y^{L+1}}e^{-y}dy
+\int_Y^{\infty}y^{n-L-1}Y^Le^{-y}dy \right]$$
$$=Y^n\sum P_L(\cos\theta)\left[\int_0^1t^{n+L}e^{-Yt}dt
+\int_1^{\infty}t^{n-L-1}Y^Le^{-Yt}dt \right]$$
where Y is the distance from the He nucleus to the antiproton in units of $a$ along the z-axis.
If we define
$$\alpha_m(Y)\equiv \int_1^{\infty} t^me^{-Yt}dt;\ \ \ m\ge 0,$$
then
$$\alpha_0(Y)=\frac{e^{-Y}}{Y};\ \ \ Y\alpha_m(Y)=e^{-Y}+m\alpha_{m-1}(Y)
.$$
Note
$$\int_0^1t^me^{-Yt}dt=\frac{m!}{Y^{m+1}}-\alpha_m(Y)\equiv \gamma_m(Y)$$
If $m<0$ then the exponential integral is needed
$$E_m(Y)=\int_1^{\infty} t^{-m}e^{-Yt}dt;\ \ \ E_{m+1}=\frac{[e^{-Y}-YE_m(Y)
]}{m}$$
$E_1(Y)$ must be calculated to high accuracy.
With the definition $$\beta_m(Y)\equiv \left\{\begin{array}{l} \alpha_m(Y)\ \ \ m\ge 0\\
E_{-m}(Y)\ \ \ m<0\end{array}\right\}$$ we can write
$$\int_0^{\infty}\frac{y^ne^{-y}}{|{\bf y}-{\bf y}|}=\sum P_L(\cos\theta)
Y^n[\gamma_{n+L}(Y)+\beta_{n-L-1}(Y)]$$
and
$$\int d{\bf y} \frac{Y_{\ell '}^{*m'}(y)Y_{\ell}^m(y)L_{n'}(y)L_{n}(y)}
{|{\bf y} -{\bf y}|}=$$
$$=\delta_{mm'}\sum A_{n,n',\bar{n}}d_{\bar{n},m}C_{\ell,L,\ell '}^{0,0,0}
C_{\ell,L,\ell'}^{m,0,m}\sqrt{\frac{2\ell +1}{2\ell '+1}}
\left[\gamma_{m+2+l}(Y)+\beta_{m-L+1}(Y)\right]$$
Results for the proton potential\[pot\]
---------------------------------------
The results for the proton potential are shown in Fig. \[ppot\]. Calculations were made at various radii for the pair $(\ell_{max},n_{max})=(2,2),\
(3,3)$ and $(4,4)$. The (4,4) results (solid squares) are shown in the figure. Also shown are the results of ref. [@kolos1; @kolos2] and the parameterization of their potential by Bosanac and Knesaurek [@bosanac]. The (4,4) results give an adequate representation of the previous results except in the minimum where there is a significant cancellation between the electronic potential energy and the direct proton-nucleus interaction.
A \[1,1\] Padé approximate was used with the three determinations mentioned above to estimate the result of the limit of the sequence. This result is shown as the open circles in Fig \[ppot\]. While this last extrapolation is significant the result gives a satisfactory agreement with the previous work.
A specific comparison was made with ref. [@kolos1; @kolos2] at R=3.704 Å. They found –8.32 meV while the present work gives –8.10 meV for the (4,4) search.
Results for the antiproton potential
------------------------------------
The potential energy for the electrons alone is shown in Fig. \[heapee\]. It is seen to vary between the limits $E_0$ at $R=0$ (approximately equal to the binding energy of $H^-$) to $E_1$ (the binding energy of isolated helium) at large distances.
Writing the electron potential as a ratio of two fourth-order polynomials, including the direct He-$\bar{p}$ electrostatic potential, assuming that the potential varies as $1/R^4$ at large values of $R$ and subtracting the asymptotic value, $E_1$, the full antiproton-helium atom electrostatic potential can be expressed as
$$V(R)=-\frac{\beta z(z^3+\frac{d}{E_0-E_1}z^4)}
{1+gz+\frac{\beta}{E_0-E_1}z^4}\label{pfit}$$
where $z\equiv 1/R$.
In the fit $ E_0=-14.3477$ eV and $E_1=-78.9847$ eV were used. The constant “d” ($=e^2$) has the value 28.798 eV $\AA^{-1}$ . By searching on parameters in this form against the (4,4) results, a best fit was found for $g=0.6474$ Å and $\beta =1595 $ meV-Å$^4$. The last value can be compared with 1557 meV-Å$^4$ from Ref [@davison] and 1678 meV-Å$^4$ from Ref [@singh].
The antiproton potential is shown in Figs. \[appot1\] and \[appot2\]. It is seen that there is no apparent change in slope which might aid in the formation of a metastable state. Since any such effect would be expected to be small (perhaps in the few meV range) it would have to be active at moderately large distances to be visable, say beyond 3 Å. We can perhaps understand why it is likely that there is no such effect by looking at the structure of the electron wave function.
The variation in the scale parameter “a” with R is shown if Fig. \[scale\]. It is seen that the system undergoes a rapid growth for a baryonic separation inside of 1 Å. Figure \[dipole\] shows the first dipole component of the wave function which provides a measure of the deformation of the electron cloud from spherical symmetry. It also shows little effect outside of 1 Å. Thus any possible inflection in the potential curve would occur inside of 1 Å when the potential is completely dominated by the simple coulomb attraction of the He-$\bar{p}$ system.
Annihilation cross section\[ann\]
=================================
The energy of the antiproton at the nuclear surface will be given by the incident energy plus the gain due to the acceleration in the coulomb potential. Since (at 1 fm) the coulomb energy is 2.88 MeV, in the range of incident energies considered here the annihilation takes place at constant energy to a good approximation. The nuclear potential was taken to be the product of a purely absorbed single nucleon strength with a Woods-Saxon density with radius 2 fm and diffuseness 0.5 fm for 4 nucleons.
The system was treated from the point of view of a nuclear optical model, i.e. no consideration was given to the possibility of the knock-out of electrons or the electromagnetic transition into atomic bound states. In general such corrections might be expected to increase slightly the annihilation cross section.
The problem could be treated in a perturbation approach. In that case one would solve the problem of scattering from the purely electrostatic potential generated in the previous section and calculate the annihilation rate from the expectation value of the potential just introduced. Since the nuclear potential is very short ranged, the amplitude will be proportional to the square of the wave function at the origin. It was found in calculating this quantity that it had very nearly the same value as the pure coulomb case
$$C^2_0(\eta )=2\pi \eta (e^{2\pi\eta} -1)^{-1}\approx -2\pi \eta
;\ \ \ \eta =-\alpha c/v$$
for energies above 10 $\mu V$. From these considerations it is seen that the cross section should vary as 1/$v^2$. Since the potential is strongly absorbing the wave function will be modified significantly in the region of the nuclear potential so that it is better to solve the full equation as now described.
Since there exists no known analytic form for the solution of the Schrödinger equation in these potentials at short range or in the region of the “surface” of the atom, the wave function was calculated over the entire region from 0 to 80 Å. The step size was changed by an order of magnitude 4 times with 20,000 points calculated in each of the first four (overlapping) regions and 160,000 points in the region of largest distance from the helium nucleus.
The results for the annihilation cross section are presented in Fig. \[sigma\]. It is seen that there is a change in energy dependence around $10 \mu V$. Above that energy the cross section is proportional to $1/v^2$ as can be expected for a coulomb system. Below that energy the cross section varies as $1/v$ as expected for a low-energy neutral system.
One can estimate this transition point using the asymptotic form of the potential. We can define a radius representing the “surface” of a potential as that value where the incident wave number is significantly affected by the potential, say where the potential becomes a fraction, f, of the incident kinetic energy. In the present case the radius $R_0$ will be defined by $$E=\frac{k^2}{2m}=f\beta/R_0^4.$$ The long wave length limit is reached at $kR_0<1$. Combining these conditions, it is seen that the transition should take place around 2/f $\mu$ V or about 20 $\mu V$ for f=0.1.
The s-wave cross section above $~10^{-5}$ eV can be represented by the expression $$\sigma_{abs}=\frac{0.536\pi}{k^2} \label{totabs}$$ Since this is greater than half of the unitarity limit the full cross section cannot be much larger than this. While the overall accuracy of the present calculation does not justify the three significant figures quoted in Eq. \[totabs\], the result is stable to this precision in this energy range. Thus if the absolute cross section can be established at any one value of the energy it will be known over a wide range. The measurement of the annihilation rate on a small, known, amount of helium introduced into a trap could thus provide a direct measurement of the temperature over a certain region of energy.
Conclusion
==========
The authors in Ref. [@holz] estimate that the pressure in the container was of the order of $10^{-11}$ Torr. If we assume that the helium and antiprotons are cooled to 4.2 $^{\circ}$K then the annihilation rate calculated using the cross section from Fig. \[sigma\] is $\sim 1.2\times
10^{-3}$ 1/sec. Since a maximum rate of $\sim 8\times 10^{-3}$ 1/sec was observed, there is a slight discrepancy. Possible explanations of this discrepancy include the existance of a higher pressure than estimated or the presence of heavier material.
It was seen that no barrier or inflection of the potential occurred. However, the possible corrections beyond the Born-Oppenheimer approximation have not yet been considered and could influence this conclusion.
The author gratefully acknowledges discussions with T. Goldman, M. Nieto and M. H. Holzscheiter.
This work was supported by the U. S. Department of Energy.
[1000]{} M. H. Holzscheiter, X. Feng, T. Goldman, N. S. P. King, R. A. Lewis, M. M. Nieto and G. A. Smith, [*Phys. Lett.*]{} [**A214**]{}, 279(1996)
A. Yu. Voronin, Sov. Phys. JETP 75, 416(1992
D. L. Morgan and V. W. Hughes, [*Phys. Rev.*]{} [**A7**]{}, 1811(1973)
Harald Friedrich, “Theoretical Atomic Physics,” Springer-Verlag
W. Kolos, [*International Journal of Quantum Chemistry*]{}, 10, 217(1976)
W. Kolos and J. M. Peek, [*Chem. Phys.*]{} [**12**]{}, 381 (1976) L. Wolniewicz, [*The Journal of Chemical Physics*]{}, 43, 1087(1965) S. Bosanac and K. Knesaurek, [*Phys. Rev.*]{} [**A28**]{}, 2173(1983)
W. D. Davison, [*Proc. Phys. Soc.*]{} [**87**]{}, 133(1966)
T. R. Singh, [*Chem. Phys. Lett.*]{} [**11**]{}, 598(1971)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The question of whether all shared objects with consensus number 2 belong to Common2, the set of objects that can be implemented in a wait-free manner by any type of consensus number 2, was first posed by Herlihy. In the absence of general results, several researchers have obtained implementations for restricted-concurrency versions of FIFO queues. We present the first Common2 algorithm for a queue with two enqueuers and any number of dequeuers.'
author:
- 'David Eisenstat[^1]'
bibliography:
- 'paper.bib'
title: 'Two-enqueuer queue in Common2'
---
Introduction
============
Model
=====
Queue implementations
=====================
Proof of correctness
====================
The two-enqueuer case
=====================
Discussion
==========
[^1]: No current affiliation. 55 Autumn Street, New Haven, CT 06511, USA. `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Multiple matrix sampling is a survey methodology technique that randomly chooses a relatively small subset of items to be presented to survey respondents for the purpose of reducing respondent burden. The data produced are missing completely at random (MCAR), and special missing data techniques should be used in linear regression and other multivariate statistical analysis. We derive asymptotic variances of regression parameter estimates that allow us to conduct power analysis for linear regression models fit to the data obtained via a multiple matrix sampling design. The ideas are demonstrated with a variation of the Big Five Inventory of psychological traits. An exploration of the regression parameter space demonstrates instability of the sample size requirements, and substantial losses of precision with matrix-sampled regressors. A simulation with non-normal data demonstrates the advantages of a semi-parametric multiple imputation scheme.
MCAR, multiple imputation, multiple matrix sampling, power analysis, respondent burden.
author:
- 'Stanislav (Stas) Kolenikov$^*$'
- 'Heather Hammer[^1]'
bibliography:
- 'everything.bib'
title: |
Power analysis for a linear regression model\
when regressors are matrix sampled
---
Introduction and motivation
===========================
This work was conceived and carried out in the context of a task to reduce respondent burden in a mental health study. We were interested in a range of outcome variables, and our analytical goal was fitting regression models to explain the mental health outcomes. The instrument collects demographic explanatory variables, as well as scores from the Big Five Inventory [@john:sriv:1999:big5], a commonly used set of five psychological traits that are often found to be correlated with behaviors and outcomes. We expected that multiple matrix sampling would allow us to reduce the instrument length from over an hour to about 20–25 minutes. A key component of sampling design, sample size determination, will be based on a linear regression power analysis. However, complexities of regression analysis with missing data required custom derivations of power analyses, which is what this technical paper addresses.
Regression setting
==================
Consider a regression analysis problem where an outcome $y$ is predicted by a set of explanatory variables $x_1, \ldots, x_p$:
$$y = \beta_0 + \beta_1 x_1 + \ldots + \beta_p x_p + \varepsilon
\label{eq:regress}$$
In the simplest possible case of no missing data and homoskedastic normal errors ${\mathbb{V}}[\epsilon_i]=\sigma^2 \, \forall i$, the maximum likelihood estimates are the OLS estimates
$$\begin{gathered}
\hat\beta_{\rm OLS} = (X'X)^{-1} X'Y;
\quad {\mathbb{V}}[ \hat\beta_{\rm OLS} ] = \sigma^2 (X'X)^{-1};
\notag \\
v[ \hat\beta_{\rm OLS} ] = s^2 (X'X)^{-1};
\quad s^2 = \frac1n \sum_{i=1}^n (y_i - x_i'\hat\beta)^2
\label{eq:ols}\end{gathered}$$
Inference on regression coefficients is based on normality of coefficient estimates, $\hat\beta \sim N(\beta,\sigma^2(X'X)^{-1})$.
An unbiased estimate of $s^2$ can be obtained by changing the denominator (degrees of freedom) from $n$ to $n-p$. The OLS estimates hold desirable properties in more general settings, e.g., by dropping the normality requirement. When regressors $X$ are stochastic, the OLS estimates and their variance estimates given above only have an asymptotic justification, and require independence of regressors $X$ and errors $\varepsilon$. When the basic assumptions are violated, sandwich-type or resampling variance estimates need to be used.
Power analysis and sample size determination in regression setting {#subsec:reg:power}
------------------------------------------------------------------
Power analysis and sample size determination are statistical tasks of addressing, quantifying and controling type I error. In a typical power analysis problem, a null hypothesis, $H_0: \theta \in \Theta_0$, and an alternative, $H_1: \theta \in \Theta_1$, are formulated; a test statistic $t(X)$ is selected, for which a critical region of level $\alpha$ is specified. E.g., assuming that high values of the test statistic indicate disagreement between the data and the null, as is typical with $\chi^2$ or $F$-statistic tests common in regression models, the rejection region would have the form $T_\alpha = [c,+\infty)$ so that $\rm{Prob} [t(X) > c] \le \alpha$ when the true value of the parameter $\theta \in \Theta_0$. Finally, power analysis addresses the issue of Type II error, i.e., $\rm{Prob} [t(X)\le c]$ under the alternative $\theta \in \Theta_1$. While the null hypothesis typically represents a simple hypothesis $\theta=\theta_0$ or a subset of reasonably small dimension, the alternative is necessarily complex. Hence researchers often formulate a measure of *effect size* $\delta$ and consider power analysis for parameter values under the alternative that are at least $\delta$ away from the specific value $\theta_0$ or the subset $\Theta_0$ in an appropriate metric.
As the power to reject the null typically grows with the sample size $n$, the task of sample size determination is to find the value $n_\beta$ that guarantees a given level of power $1-\beta$. While the size of the test is often taken to be $\alpha=5\%$, the traditional type II error rate is $\beta=20\%$ leading to $1-\beta=80\%$ power.
While testing the location parameter of two populations, for example, the natural hypotheses might be $H_0: \mu_1 = \mu_2$ vs. the (two-sided) alternative $H_1: |\mu_1 - \mu_2| \ge \delta$, with relatively straightforward testing based on the Student $t$ distribution, linear regression models feature a variety of statistics that may be subject to testing and power analysis:
1. Test of overall fit: $H_0: \mathbf{\beta}=0$.
1. A version of this hypothesis can be formulated as $H_0: R^2=0$. Depending on how easy or difficult it is to conduct inference on parameter estimates or regression sums of squares, one or the other may be preferred in applications.
2. An increase of overall fit: $H_0: R^2 \le R_0^2$ vs. $H_1: R^2 \ge R_0^2 + \delta$.
3. Specific regression coefficients: the coefficient of the $j$-th explanatory variable is zero, $H_0: \beta_j = 0$ vs. $H_1: | \beta_j | \ge \delta$.
4. Linear hypothesis $H_0: R\beta = r_0$, which covers cases like:
1. Equality of two regression coefficient for the $j$-th and the $k$-th explanatory variable, $H_0: \beta_j - \beta_k = 0$, so $R = (0, \ldots, 1, \ldots, -1), r=0$
2. No impact of a set of variables $j_1, j_2, \ldots$: $H_0: \beta_{j_1}=0, \beta_{j_2}=0, \ldots$, so that $R$ is a subset of rows of a unit matrix, and $r$ is a zero vector of conforming dimension.
5. Tests on error variance $\sigma^2$, e.g. $H_0: \sigma_2 \le \sigma^2_0$ vs. $H_1: \sigma^2 \ge \sigma_0^2 + \delta$.
Multiple matrix sampling
========================
*Multiple matrix sampling* of a survey questionnaire consists of administering only a specific subset of items to a given respondent, out of all items this respondent is potentially eligible to be asked. The name stems from representation of the data with respondents as rows, and items as columns, so that matrix sampling concerns selecting specific entries in the matrix to be administered, rather than the full row as is typically done. The focus of the technique is on selecting items out of all the relevant ones that the respondent could be asked, with the potential skip patterns already taken into account. Similar or equivalent techniques are also known as *partitioned designs* and *questionnaire splitting*. The method originated in educational testing [@shoemaker:1972], where it was first used to select items from a large pool of available ones. The educational testing companies have identified the need to implement multiple matrix sampling methods to protect the integrity of their data products, so that the students taking a standardized test are not able to get trained on a small subset of items known to be administered on standardized tests, thus biasing the estimates of achievement.
Given the relatively esoteric nature of the method, the existing publications have addressed some specific niche problems in matrix sampling (and the current paper is no exception). @gonzalez:eltinge:2007 provided a review of matrix sampling and applications in Consumer Expenditure Quarterly Survey. @chipperfield:steel:2009 put the problem of matrix sampling into a cost optimization framework, where proper subsets of $K$ items can be administered on up to $2^K-1$ forms at specific cost per item rates. They demonstrated that with two items (or groups of items), the split questionnaire best linear unbiased estimator (BLUE; also related to the GLS estimator) provides modest efficiency gains over a design in which all items are administered at once, and over a two-phase design in which all items are administered to a fraction of respondents, and one subset of items is administered to the remainder of respondents. @merkouris:2010 extended their work to provide simplified composite estimation using the estimates based on the form-specific subsamples, where compositing is based on the second-order probabilities of selection and the way they are utilized in estimating the variance of the Horvitz-Thompson estimator. @eltinge:2013 discussed connections to and relations with multiple frame and multiple systems estimation methods (e.g., integration of survey and administrative data, where administrative data may fill some of the survey items when available). We add to this literature by providing the asymptotic variance-covariance matrix of the coefficient estimates under matrix sampling of regressors, assuming that the outcome is always collected. We also discuss implications for power analysis and sample size determination.
A simple example
----------------
Consider the following matrix sampling design, in which the outcome $y$ is collected on every form, while the explanatory variables differ between forms.
Form $X_1$ $X_2$ $X_3$ $n$
------ ------- ------- ------- -------
1 + $n_1$
2 + $n_2$
3 + $n_3$
: Three questionnaire forms for data collection: Design 1. \[tab:forms3:design1\]
With this design, summaries (means, totals) of all the variables ($x_1,x_2,x_3,y$) can be obtained, and the bivariate relations between each of the regressors and the outcome $y$ can be analyzed. However, estimation of a multiple regression model requires estimability of all of the entries of the $(X'X)$ matrix, which this specific matrix sampling design does not provide.
To conduct regression analysis, we need to observe the cross-entries of the $X'X$ matrix, which necessitates the following matrix sampling design.
Form $X_1$ $X_2$ $X_3$ $n$
------ ------- ------- ------- -------
1 + + $n_1$
2 + + $n_2$
3 + + $n_3$
: Three questionnaire forms for data collection: Design 2. \[tab:forms3:design2\]
Parameter estimation under matrix sampling
------------------------------------------
Since the components of $X'X$ and/or $X'y$ necessary to obtain the OLS regression estimates may not be jointly available, more complex estimation strategies may need to be employed. We study two such strategies.
One possibility is to utilize structural equation modeling (SEM) with missing data, in which the marginal regression model of interest is formulated by using the regressors as exogenous variables, the dependent variable is introduced as the only endogenous variable explained by the model [@bollen:selv:1989], and the existing SEM estimation methods are applied [@yuan:bentler:2000; @savalei:2010].
Alternatively, since the data are missing by design, and can be treated as MCAR, multiple imputation [@rubin:1996; @vanbuuren:2012 MI] can be used to fill in the missing values, with Rubin’s variance formulae used to combine MI estimates and provide inference. Of the several existing flavors of multiple imputation, one of the simplest strategies is imputation under multivariate normality (which we expect to behave in ways similar to the estimation methods for SEM with missing data under multivariate normality). A less model-dependent method is predictive mean matching [@little:1988] in which a regression model is fit for each imputed variable, a linear prediction is obtained for each case with missing variable, and an imputation is made by choosing the value of the dependent variable from one of the nearest neighbors in terms of the linear prediction score.
Set up and notation
===================
All of the derivations in this paper concern the joint matrix of the first and second order moments of the data: $$\Omega =
{\mathbb{E}}\left[
\begin{pmatrix}
1 \\
{\mathbf{x}}\\
y
\end{pmatrix}
\begin{pmatrix}
1 &
{\mathbf{x}}' &
y
\end{pmatrix}
\right]
=
\begin{pmatrix}
1 & \mu_\mathbf{x}' & \mu_y \\
\mu_\mathbf{x} & {\mathbb{E}}[ {\mathbf{x}}{\mathbf{x}}'] & {\mathbb{E}}[{\mathbf{x}}y] \\
\mu_y & {\mathbb{E}}[{\mathbf{x}}' y] & {\mathbb{E}}[y^2]
\end{pmatrix}
\equiv
\begin{pmatrix}
\omega_{00} & \Omega_{0x} & \omega_{0y} \\
\Omega_{0x}' & \Omega_{xx} & \Omega_{xy} \\
\omega_{0y} & \Omega_{xy}' & \omega_{yy}
\end{pmatrix}
\label{eq:Omega}$$
The maximum likelihood estimates of the coefficients in the regression of $y$ on $x$ (obtained, for instance, through SEM modeling using maximum likelihood estimates with multivariate normal missing data method; or approximated through multiple imputation) are obtained as $$\hat\beta_{\rm FIML} =
\begin{pmatrix}
\omega_{00} & \hat\Omega_{0x} \\
\hat\Omega_{0x}' & \hat\Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
\hat\omega_{0y} \\
\hat\Omega_{xy}
\end{pmatrix}
\label{eq:ml:regression}$$ where $\hat\Omega$ is the maximum likelihood estimator of the joint parameter matrix: $$\hat\Omega =
\begin{pmatrix}
\omega_{00} & \hat\Omega_{0x} & \hat\omega_{0y} \\
\hat\Omega_{0x}' & \hat\Omega_{xx} & \hat\Omega_{xy} \\
\hat\omega_{0y} & \hat\Omega_{xy}' & \hat\omega_{yy}
\end{pmatrix}
=
\begin{pmatrix}
\omega_{00} & \hat\omega_{01} & \ldots & \hat\omega_{0p} & \hat\omega_{0y} \\
\hat\omega_{01} & \hat\omega_{11} & \ldots & \hat\omega_{1p} & \hat\omega_{y1} \\
\vdots & \ldots & \vdots & \vdots & \vdots \\
\hat\omega_{0p} & \hat\omega_{1p} & \ldots & \hat\omega_{pp} & \hat\omega_{yp} \\
\hat\omega_{0y} & \hat\omega_{y1} & \ldots & \hat\omega_{yp} & \hat\omega_{yy}
\end{pmatrix}
\label{eq:hat:Sigma}$$ where $x_0=1$ is the regression intercept by convention, so that $\omega_{00}\equiv 1$, $\hat\omega_{0j}=\hat\mu_j$ are the (estimated) means of the $j$-th explanatory variable, and $\hat\omega_{0y}=\hat\mu_y$ is the estimated mean of $y$.
To derive the likelihood, we need the form-specific submatrices obtained by multiplying the overall matrix by selector matrices. For instance, in Design 2 above, for the first form, the relevant covariance matrix is $${\rm Cov}(x_2, x_3, y)' = F_1 \Omega F_1',
\quad
F_1 =
\begin{pmatrix}
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 1
\end{pmatrix}
\label{eq:selector:1}$$ Matrices necessary to form $F_2\Omega F_2'$ and $F_3 \Omega F_3'$ are defined in a similar way.
Define the unit selector vector that picks up the estimates of the means $e_0=(1,0,\ldots,0)$, which is the unit vector with 1 in the “zeroth” position corresponding to the intercepts in the parameter matrix $\Omega$. In addition to $e_0$ selecting the first order moments, define the unit selection vectors $e_y=(0,0,0,0,1)'$ as the unit vector selecting the last row/column of $\Omega$ corresponding to the $y$-parameters, and $e_j=(0,\ldots,0,1,0,\ldots)$ is a unit vector with 1 in the $j$-th position corresponding to the $j$-th variable (with the convention of indexing starting at zero). Then we observe that $$\begin{aligned}
F_1' & = (e_2, e_3, e_y) \notag \\
F_2' & = (e_1, e_3, e_y) \notag \\
F_3' & = (e_1, e_2, e_y)\end{aligned}$$
Likelihood and derivatives
==========================
Likelihood
----------
Indexing the forms by $k$, and observations within forms by $i$, the likelihood can be written as $$\begin{aligned}
\ln L(\omega;X) =
\sum_{k=1}^3 \sum_{i=1}^{n_k} &
\Bigl\{
- \frac12 {\mathop{\rm tr}}(F_k F_k') \ln(2\pi)
- \frac12 \ln \det ( F_k \Omega F_k' )
\notag \\ &
- \frac12 \bigl[ (x_i',y) - e_0' \Omega F_k' \bigr] (F_k \Omega F_k')^{-1}
\bigl[ (x_i', y)' - F_k \Omega e_0 \bigr]
\Bigr\}
\label{eq:log:lkhd}\end{aligned}$$ where $n_k$ is the number of observations on which the $k$-th form is collected, and $F_k$ is the selector matrix corresponding to the $k$-th form.
Derivations of the asymptotic properties of the MLE estimate $\widehat\Omega$ are based on the matrix differential [@magn:neud:1999] $$\begin{aligned}
{{\sf d}}\Omega = &
{{\sf d}}\omega_{00} \, e_0 e_0'
+ \sum_{j=1}^p {{\sf d}}\omega_{0j} \, (e_0 e_j' + e_j e_0')
+ {{\sf d}}\omega_{0y} \, (e_0 e_y' + e_y e_0')
+ \sum_{j=1}^p {{\sf d}}\omega_{jj} \, e_j e_j'
\notag \\ &
+ \sum_{j=1}^p \sum_{i\neq j} {{\sf d}}\omega_{ij} \, (e_i e_j' + e_j e_i')
+ \sum_{j=1}^p {{\sf d}}\omega_{yj} \, (e_y e_j' + e_j e_y'),
\label{eq:omega:as:sum:of:e:crossed}\end{aligned}$$
After some tedious algebra, the following information matrix ${\mathbb{E}}\nabla^2 \ln L(\omega;X)$ results.
$$\begin{aligned}
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{0s}\partial \omega_{0t}} \Bigr]
& =
- \sum_{k=1}^3 n_k \tau^{(k)}_{st},
\label{eq:d2lnl:domega0:domega0}
\\
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{0s}\partial \omega_{uu}} \Bigr]
& = 0,
\label{eq:d2lnl:domega0:domega:uu}
\\
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{0s}\partial \omega_{uv}} \Bigr]
& = 0
\label{eq:d2lnl:domega0:domega:uv}\end{aligned}$$
The zero expected cross-derivatives indicate that the estimates of the multivariate normal means and the variance-covariance parameters are independent. (This may not be the case in general if the missing data mechanism coded by the matrices $F_k$ is not MCAR, and instead related to the data values.)
$$\begin{aligned}
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{ss}\partial \omega_{uu}} \Bigr]
& = - \frac12 \sum_{k=1}^3 n_k \bigr[ \tau^{(k)}_{su} \bigl]^2
\label{eq:d2lnl:domega:ss:domega:uu}
\\
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{ss}\partial \omega_{uv}} \Bigr]
& = - \sum_{k=1}^3 n_k \tau^{(k)}_{su} \tau^{(k)}_{sv}
\label{eq:d2lnl:domega:ss:domega:uv}
\\
{\mathbb{E}}\Bigl[ \frac {\partial^2 \ln L(\omega;X)}{\partial \omega_{st}\partial \omega_{uv}} \Bigr]
& = - \sum_{k=1}^3 n_k \bigl[ \tau^{(k)}_{su} \tau^{(k)}_{tv} + \tau^{(k)}_{sv} \tau^{(k)}_{tu} \bigr]
\label{eq:d2lnl:domega:st:domega:uv}
\\
\tau^{(k)}_{st} & = e_s' F_k' (F_k \Omega F_k')^{-1} F_k e_t
\label{eq:tau:kst}\end{aligned}$$
where $\tau^{(k)}_{st}$ is the $(s,t)$-th entry of the inverse of the form-specific covariance matrix; and indices $s,t,u,v$ can enumerate the explanatory variables $x_j$ and the response $y$. As $x_i$ and $y$ are considered jointly multivariate normal at this point, there is no separation into dependent and explanatory variables.
Putting these entries together into a matrix, and using the standard maximum likelihood estimation theory results, the asymptotic variance of the maximum likelihood estimates of $\mathop{\rm vech}\Omega$ is given by $${\rm As.} {\mathbb{V}}[ \hat\omega ] = - {\mathbb{E}}\bigr[ \nabla^2 \ln L(\omega;X) \bigr]^{-1}
\label{eq:asvar:omega-hat}$$
The delta method derivation of the asymptotic variance of $\hat\beta$
---------------------------------------------------------------------
Let us now return to the task of estimating the coefficients of the regression equation $$y=\beta'x + \epsilon$$ via (\[eq:ml:regression\]). The asymptotic variance-covariance matrix of $\hat\beta_{\rm FIML}$ can be obtained from the asymptotic covariance matrix of $\hat\Omega$ using the delta-method, i.e., linearization of the relation (\[eq:ml:regression\]): $$\begin{aligned}
{{\sf d}}\beta = &
-
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
0 & {{\sf d}}\Omega_{0x} \\
{{\sf d}}\Omega_{0x}' & {{\sf d}}\Omega_{xx}
\end{pmatrix}
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
\omega_{0y} \\
\Omega_{xy}
\end{pmatrix}
+
\notag \\
& \hspace{5cm} +
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
{{\sf d}}\omega_{0y} \\
{{\sf d}}\Omega_{xy}
\end{pmatrix}
\label{eq:diff:beta}\end{aligned}$$ where the individual components of ${{\sf d}}\Omega$ can be obtained from (\[eq:omega:as:sum:of:e:crossed\]). Thus $$\begin{aligned}
\frac{\partial\beta}{\partial\omega_{0j}} = &
-
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
0 & e_j' \\
e_j & \underline{0}
\end{pmatrix}
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
\omega_{0y} \\
\Omega_{xy}
\end{pmatrix}
\notag \\
\frac{\partial\beta}{\partial\omega_{0y}} = &
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
1 \\
\vec{0}
\end{pmatrix}
\notag \\
\frac{\partial\beta}{\partial\omega_{jj}} = &
-
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
0 & \vec{0}' \\
\vec{0} & e_j e_j'
\end{pmatrix}
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
\omega_{0y} \\
\Omega_{xy}
\end{pmatrix}
\notag \\
\frac{\partial\beta}{\partial\omega_{ij}} = &
-
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
0 & \vec{0}' \\
\vec{0} & e_i e_j' + e_j e_i'
\end{pmatrix}
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
\omega_{0y} \\
\Omega_{xy}
\end{pmatrix}
\notag \\
\frac{\partial\beta}{\partial\omega_{yj}} = &
\begin{pmatrix}
\omega_{00} & \Omega_{0x} \\
\Omega_{0x}' & \Omega_{xx}
\end{pmatrix}^{-1}
\begin{pmatrix}
0 \\
e_j
\end{pmatrix}
\notag \\
\nabla_\omega \beta = & \Bigl(
\frac{\partial\beta}{\partial\omega_{01}}, \frac{\partial\beta}{\partial\omega_{02}}, \ldots,
\frac{\partial\beta}{\partial\omega_{y0}},
\frac{\partial\beta}{\partial\omega_{11}}, \frac{\partial\beta}{\partial\omega_{12}}, \ldots, \frac{\partial\beta}{\partial\omega_{y1}}, \frac{\partial\beta}{\partial\omega_{22}}, \ldots,
\frac{\partial\beta}{\partial\omega_{yp}}, 0 \Bigr)
\label{eq:dbeta:domega}\end{aligned}$$ where the derivatives are with respect to the components of the vectorization $\mathop{\rm vech}\Omega$, of which the last term is $\frac{\partial\beta}{\partial\omega_{yy}}=0$. By the standard multivariate delta-method results [@newey:mcfadden:1994; @vandervaart:1998], $${\rm As.} {\mathbb{V}}[ \hat\beta ] = \nabla_\omega \beta \, {\mathbb{V}}[ \hat\omega ] \, \nabla_\omega' \beta
\label{eq:asvar:beta-hat}$$
Example: Big Five Inventory
===========================
In our application, we wanted to analyze the relation between mental health outcomes and the Big Five personal traits:
- Openness to experience (inventive/curious vs. consistent/cautious)
- Conscientiousness (efficient/organized vs. easy-going/careless)
- Extraversion (outgoing/energetic vs. solitary/reserved)
- Agreeableness (friendly/compassionate vs. challenging/detached)
- Neuroticism (sensitive/nervous vs. secure/confident)
These personal traits have been found in numerous studies to be related to academic performance, disorders, general health, and many other behaviors and outcomes. The standard Big Five scale consists of 44 items, some of which are reverse worded and reverse scored to minimize the risk of straightlining, and with items from different subscales mixed throughout the scales. Each item is a 5 point Likert scale with a clear midpoint.
In the population of interest, the Big Five traits are expected to have the following correlations, based on preceding research:
$$\label{eq:bigfive:corr}
{\rm Cov}[x] =
\begin{pmatrix}
1 & 0.26 & 0.47 & 0.20 & -0.16 \\
0.26 & 1 & 0.28 & 0.46 & -0.28 \\
0.47 & 0.28 & 1 & 0.20 & -0.35 \\
0.20 & 0.46 & 0.20 & 1 & -0.37 \\
-0.16 & -0.28 & -0.35 & -0.37 & 1
\end{pmatrix}
\equiv
\Sigma_{\rm Big 5}$$
We thus consider a regression model $$y_i = \beta_0 + \beta_1 x_{i1} + \ldots \beta_5 x_{i5} + \varepsilon_i$$ where $x_{i1},\ldots,x_{i5}$ are subscale scores of the Big Five traits. Measurement error in these scores is ignored, although more accurate methods are available to account for it [@skrondal:laake:2001].
A balanced multiple matrix sampling design would consist of ten forms, each administering the outcome $y$ and two of the Big Five subscales:
Form 1 2 3 4 5 6 7 8 9 10
----------------------- --- --- --- --- --- --- --- --- --- ----
O $\vphantom{\Bigl|}$ + + + +
C $\vphantom{\Bigl|}$ + + + +
E $\vphantom{\Bigl|}$ + + + +
A $\vphantom{\Bigl|}$ + + + +
N $\vphantom{\Bigl|}$ + + + +
$y\vphantom{\Bigl|}$ + + + + + + + + + +
: Multiple matrix sampling design with five explanatory variables. \[tab:bigfive\]
Simulation 1: Parameter space exploration {#sec:explore}
=========================================
In this simulation exercise, we explore the parameter space of regression coefficients to gauge the degree of variability of sample size determination results. Asymptotic variance resulting from (\[eq:asvar:beta-hat\]) is used to obtain the sample sizes for the tasks outlined in Section \[subsec:reg:power\]. Simulation 1 consists of the following steps.
1. Population regression parameters are simulated from $\beta \sim N(\mathbf{0},I_5)$.
2. To provide the scale of the residual variance, the fraction of explained variance is set to $R^2=0.15$, a moderate effect for behavioral and social science data, and the associated residual variance $\sigma_\varepsilon^2$ is calculated based on this value of $R^2$.
3. The complete data variances stemming from (\[eq:ols\]) are recorded. \[enum:explor:full:var\]
4. The multiple-matrix-sampled data variances stemming from (\[eq:asvar:beta-hat\]) are recorded. \[enum:explor:mcar:var\]
5. Sample size to reject the test of overall significance $H_0: \beta_1 = \ldots = \beta_5 = 0$ at 5% level with 80% power is recorded. \[enum:explor:n:overall\]
6. Sample size to detect an increase in $R^2$ by 0.01 (i.e., from 0.15 to 0.16), through a uniform multiplicative increase in the values of the regression parameters, keeping the residual variance $\sigma_\varepsilon^2$ constant, at 5% level with 80% power, is recorded. \[enum:explor:n:infl\]
7. Sample size to detect an increase in $R^2$ by 0.01 (i.e., from 0.15 to 0.16), through an increase in the value of the coefficient $\beta_j, j=1, \ldots, 5$, keeping the residual variance $\sigma_\varepsilon^2$ constant and other regression parameters constant, at 5% level with 80% power, is recorded. \[enum:explor:n:poke\]
8. Fraction of missing information (FMI) is computed as one minus the ratio of the variance of regression parameter estimate with complete data (obtained in step \[enum:explor:full:var\]) to the variance of regression parameter estimate with missing data (obtained in step \[enum:explor:mcar:var\]) \[enum:explor:fmi\]
1,000 Monte Carlo draws of the $\beta$ vector, and subsequent analytical computation of asymptotic variances and power, were done. Results are presented graphically. Figure \[fig:explore:n\] presents the sample sizes obtained in steps \[enum:explor:n:overall\]–\[enum:explor:n:poke\] of the parameter exploration. A striking feature of the plot is wide variability of the sample sizes as a function of the specific configuration of parameters. While the lower limit of the sample size necessary to detect an overall increase in $R^2$ by $0.01$ is about $n=82K$, the median value is $n=110K$, the 95th percentile is $n=220K$, and the maximum (worst case scenario) identified in this simulation is $n=400K$. The patterns of the coefficients of the worst case scenarios typically indicate large coefficients of opposite signs of the positively correlated variables ($x_1$ through $x_4$), or large coefficients of similar size of one of the positively correlated factors ($x_1$ through $x_4$) and a high value of factor $x_5$ that is negatively correlated with all other subscales. This wide range of variability makes it difficult to provide a definite recommendation concerning the sample size for the study to the stakeholders. A conservative value based on a high percentile (80% or 90%) can be recommended, to protect against bad population values of regression parameters at the expense of a potentially unnecessary increase in costs.
Figure \[fig:explore:fmi\] presents the exploration distribution of the fraction of missing information due to the missing data. FMI for the intercept is generally low, below 0.2. FMI for regression slopes are generally high, in the range of about 70% to 80%. Given the structure of the missing data shown by the multiple matrix sampling design in Table \[tab:bigfive\], each of the predictor variables is observed in 40% of the data (informing the diagonal entries of the $X'X$ matrix), and each pairwise combination of the regressors is observed in 10% of the data (informing the off-diagonal entries). This yields an expected information loss for the predictor variables somewhere between 60% and 90%.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ \[fig:explore:n\] Sample size to ensure the necessary detectable effect. (a) Overall test $H_0: R^2=0$; (b) $R^2$ increase due to overall explanatory power increase from $R^2=0.15$ by $0.01$; (c) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_1$; (d) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_5$. ](explor_n_overall.png "fig:") ![ \[fig:explore:n\] Sample size to ensure the necessary detectable effect. (a) Overall test $H_0: R^2=0$; (b) $R^2$ increase due to overall explanatory power increase from $R^2=0.15$ by $0.01$; (c) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_1$; (d) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_5$. ](explor_n_infl01.png "fig:")
(a) (b)
![ \[fig:explore:n\] Sample size to ensure the necessary detectable effect. (a) Overall test $H_0: R^2=0$; (b) $R^2$ increase due to overall explanatory power increase from $R^2=0.15$ by $0.01$; (c) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_1$; (d) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_5$. ](explor_n_poke1.png "fig:") ![ \[fig:explore:n\] Sample size to ensure the necessary detectable effect. (a) Overall test $H_0: R^2=0$; (b) $R^2$ increase due to overall explanatory power increase from $R^2=0.15$ by $0.01$; (c) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_1$; (d) $R^2$ increase due to an increase in explanatory power from $R^2=0.15$ by $0.01$ due to $x_5$. ](explor_n_poke5.png "fig:")
(c) (d)
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ \[fig:explore:fmi\] Fraction of missing information: (a) intercept; (b) slope of $x_1$; (c) slope of $x_5$. ](explor_fmi0.png)
(a)
![ \[fig:explore:fmi\] Fraction of missing information: (a) intercept; (b) slope of $x_1$; (c) slope of $x_5$. ](explor_fmi1.png)
(b)
![ \[fig:explore:fmi\] Fraction of missing information: (a) intercept; (b) slope of $x_1$; (c) slope of $x_5$. ](explor_fmi5.png)
(c)
Simulation 2: Performance in finite samples
===========================================
To study the performance of estimation methods based on SEM estimation with missing data, and on multiple imputation procedures, a simulation with microdata was also performed. For each simulation draw, the following steps were taken.
1. Sample size is set to $n=1,000$ (i.e., $100$ observations per form).
2. Multivariate non-normal factor scores are simulated:
1. The non-normal principal components of $x_1, \ldots, x_5$ are simulated as $$\begin{aligned}
f_1 & = -\ln u_1-1, \quad u_1 \sim U[0,1] \\
f_2 & = (2 b-1) (-\ln u_2-1), \quad b \sim {\rm Bernoulli}(0.5), u_2 \sim U[0,1] \\
f_3, f_4, f_4 & \sim N(0,1)
\end{aligned}$$ so that each principal component has a mean of 0 and variance of 1, with all the underlying random variables being drawn independently of each other. The first component $f_1$ has a marginal exponential distribution with a heavy right tail, ensuring the overall skewness of each factor. The second component has a bimodal distribution with two exponential components and heavy tails. The remaining three components are normal.
2. The factor values are reconstructed as $$\begin{pmatrix}
x_1 \\ x_2 \\ x_3 \\ x_4 \\ x_5
\end{pmatrix}
=
\sum_{j=1}^5 \mathbf{u}_j \sqrt{\lambda_j} f_j
\label{eq:Sigma:big5:eigenproblem}$$ where $\Sigma_{\rm Big 5} = U' \Lambda U$ is the eigenvalue decomposition of the target covariance matrix (\[eq:bigfive:corr\]) of the Big Five factors.
3. The outcome is obtained as $y=0.3x_1 + 0.3x_4 + \varepsilon, \varepsilon \sim N(0,1.248602)$ where the specific value of the residual variance was chosen to ensure that $R^2=0.15$ in the population.
4. The regression model with the complete data is fit to obtain the benchmark for FMI calculation.
5. The values of regressors were deleted in accordance with the multiple matrix sampling design in Table \[tab:bigfive\].
6. The normal theory based SEM model for missing data was fit; regression parameter estimates and their asymptotic standard errors based on the inverse Hessian were recorded.
7. $M=50$ complete data sets were imputed using multivariate normal imputation model.
8. The regression model was estimated using the first $M=5$ data sets, in accordance with the traditional recommendation regarding the number of imputed data sets. Regression parameter estimates and their asymptotic standard errors based on the Rubin’s rules were recorded.
9. The regression model was estimated using all of the $M=50$ data sets. Regression parameter estimates and their asymptotic standard errors based on the Rubin’s rules were recorded.
10. $M=50$ complete data sets were imputed using predictive mean matching imputation model for each of the missing variables.
11. The regression model was estimated using the first $M=5$ data sets, in accordance with the traditional recommendation regarding the number of imputed data sets. Regression parameter estimates and their asymptotic standard errors based on the Rubin’s rules were recorded.
12. The regression model was estimated using all of the $M=50$ data sets. Regression parameter estimates and their asymptotic standard errors based on the Rubin’s rules were recorded.
There were 1,200 Monte Carlo samples drawn.
![Sampling distributions of the parameter estimates $\hat\beta_1$ across different methods.[]{data-label="fig:simul:b1"}](simul_hist_b1_cens.png)
Figure \[fig:simul:b1\] reports the simulated distributions of the estimates of parameter $\beta_1$. The population value of 0.3 is shown as a vertical line on the plot. As expected, the complete data regression model demonstrates higher efficiency. Estimates based on the multivarate normal methods are biased up, while those based on MI with predictive mean matching are biased down. Distributions of the estimates based on the multivariate normal methods are more spread out than the asymptotic variance based on (\[eq:asvar:beta-hat\]), while those based on PMM MI are less spread out, with apparent efficiency gains extracted from higher moments of the data. The plots in Figure \[fig:simul:b1\] are truncated, with about 3% of the Monte Carlo simulations outside the right range of the plot (the value of $\beta_1=0.6$), and about 1% of the Monte Carlo simulations outside the left range of the plot (the value of $\beta_1=0$) for each of the methods based on multivariate normality. Details for $\hat\beta_1$ and other regression coefficient estimates are provided in Table \[tab:simul:b1\].
Method $\hat\beta_1$ $\hat\beta_2$ $\hat\beta_3$ $\hat\beta_4$ $\hat\beta_5$
-------------- ---------------------------- ---------------------------- ---------------------------- ---------------------------- ----------------------------
Complete 0.3002 0.0016 0.0006 0.3002 0.0015
data \[0.298,0.303\] \[-0.001,0.004\] \[-0.002,0.003\] \[0.298,0.303\] \[-0.001,0.004\]
regression $\langle$ 0.0418 $\rangle$ $\langle$ 0.0408 $\rangle$ $\langle$ 0.0440 $\rangle$ $\langle$ 0.0414 $\rangle$ $\langle$ 0.0413 $\rangle$
SEM with 0.3277 -0.0203 -0.0130 0.3324 0.0096
MVN \[0.320,0.336\] \[-0.028,-0.013\] \[-0.022,-0.004\] \[0.324,0.340\] \[0.003,0.017\]
missing data $\langle$ 0.1414 $\rangle$ $\langle$ 0.1356 $\rangle$ $\langle$ 0.1588 $\rangle$ $\langle$ 0.1429 $\rangle$ $\langle$ 0.1249 $\rangle$
MI using 0.3369 -0.0253 -0.0173 0.3393 0.0091
MVN model, \[0.329,0.345\] \[-0.033,-0.017\] \[-0.027,-0.008\] \[0.331,0.347\] \[0.002,0.016\]
$M=5$ $\langle$ 0.1390 $\rangle$ $\langle$ 0.1415 $\rangle$ $\langle$ 0.1645 $\rangle$ $\langle$ 0.1435 $\rangle$ $\langle$ 0.1259 $\rangle$
MI using 0.3430 -0.0314 -0.0208 0.3466 0.0109
MVN model, \[0.334,0.352\] \[-0.040,-0.023\] \[-0.031,-0.011\] \[0.338,0.355\] \[0.003,0.018\]
$M=50$ $\langle$ 0.1556 $\rangle$ $\langle$ 0.1507 $\rangle$ $\langle$ 0.1760 $\rangle$ $\langle$ 0.1531 $\rangle$ $\langle$ 0.1336 $\rangle$
MI using 0.2661 0.0356 0.0261 0.2666 -0.0056
PMM model, \[0.262,0.270\] \[0.032,0.039\] \[0.022,0.030\] \[0.263,0.271\] \[-0.009,-0.002\]
$M=5$ $\langle$ 0.0707 $\rangle$ $\langle$ 0.0679 $\rangle$ $\langle$ 0.0758 $\rangle$ $\langle$ 0.0707 $\rangle$ $\langle$ 0.0631 $\rangle$
MI using 0.2678 0.0361 0.0251 0.2671 -0.0043
PMM model, \[0.264,0.272\] \[0.032,0.040\] \[0.021,0.029\] \[0.263,0.271\] \[-0.008,-0.001\]
$M=50$ $\langle$ 0.0676 $\rangle$ $\langle$ 0.0656 $\rangle$ $\langle$ 0.0719 $\rangle$ $\langle$ 0.0665 $\rangle$ $\langle$ 0.0591 $\rangle$
Population 0.3 0 0 0.3 0
$\langle$ 0.0791 $\rangle$ $\langle$ 0.0856 $\rangle$ $\langle$ 0.0926 $\rangle$ $\langle$ 0.0824 $\rangle$ $\langle$ 0.0832 $\rangle$
: Monte Carlo means, \[95% confidence intervals\] for the means and $\langle$standard deviations$\rangle$ for regression parameter estimates.[]{data-label="tab:simul:b1"}
Figure \[fig:simul:se1\] provides the Monte Carlo distributions of the standard errors reported for the missing data methods. The dotted vertical line is the asymptotic standard error based on (\[eq:asvar:beta-hat\]), 0.0791. The dashed lines are empirical means of the standard errors. All distributions are skewed with heavy right tails. The distributions of the standard errors based on multivariate data contain outliers outside the range of the plot (3% of the SEM with missing data results; 6% of the results for MI using the multivariate normal model with $M=5$; 8% of the results for MI using the multivariate normal model with $M=50$; the range of the plots is from 0 to $3\times$ the asymptotic standard error, 0.0791). Distributions of the standard errors for the multivariate normal methods are significantly higher that this asymptotic standard error, which reflects, to some extent, the greater variability of the estimates observed above in Figure \[fig:simul:b1\] and Table \[tab:simul:b1\]. Distributions of the standard errors for the PMM MI method are significantly lower that the asymptotic standard error, which reflects, to some extent, the lower variability of the estimates based on this method. A higher number of multiple imputations $M=50$ vs. $M=5$ helps to stabilize the variance estimates, particularly in the case of PMM.
![Sampling distributions of the standard errors of $\hat\beta_1$ across different methods.[]{data-label="fig:simul:se1"}](simul_se1_sep.png)
Coverage of the nominal 95% confidence intervals is analyzed in Table \[tab:simul:cover95\]. Despite the shortcomings of both the point estimates and the standard errors noted above, things seem to balance out and provide confidence interval coverage fairly close to the target.
[l|ccccc]{} Method & $\hat\beta_1$ & $\hat\beta_2$ & $\hat\beta_3$ & $\hat\beta_4$ & $\hat\beta_5$\
\
Complete data regression & 95.5% & 95.4% & 95.1% & 95.8% & 97.3%\
SEM with MVN missing data & 97.8% & 98.6% & 97.3% & 98.7% & 98.4%\
MI using MVN model, $M=5$ &93.3% & 93.8% & 93.0% & 93.3% & 96.8%\
MI using MVN model, $M=50$ &92.8% & 93.4% & 93.2% & 93.1% & 97.9%\
MI using PMM model, $M=5$ &94.4% & 95.6% & 94.5% & 95.0% & 94.2%\
MI using PMM model, $M=50$ &96.3% & 96.6% & 95.8% & 96.8% & 97.2%\
Estimated fractions of missing information reported by the software are shown on Figure \[fig:simul:fmi\]. The dotted line is the value based on asymptotic variance, 73.6%. Dashed lines are the empirical FMI, based on the ratios of the Monte Carlo variance of $\hat\beta_1$ based on a given missing data method to the variance of $\hat\beta_1$ based on the complete data. The latter empirical FMI is greater than the theoretical one for the MI methods based on the multivariate normality assumption, and lower than the theoretical one for the PMM MI methods. The methods based on multivariate normality appear to underestimate FMI, as the distributions of the reported empirical FMI appear to the left of the true value (dashed line). The FMI that come out of PMM MI appear to be more accurate. An increase in the number of completed data sets from $M=5$ to $M=50$ helps to improve stability of the FMI estimates, making the distributions of the empirical FMI more concentrated.
![Reported fraction of missing information.[]{data-label="fig:simul:fmi"}](simul_fmi1_sep.png)
Concluding remarks
==================
This paper provides an analytical framework for analysis of regression models (and, more generally, other statistical methods that are based on the covariance matrices of observed items or scales) that allows for quick power analysis avoiding computationally intensive simulations.
Revisiting the initial motivation of burden reduction, the results are underwhelming. Is burden really reduced by multiple matrix sampling in the example considered? Out of five explanatory variables (based on approximately 8 survey items each) and one outcome, only three variables are collected on each of the matrix sampled instrument forms. This translates to about 50% burden reduction per respondent. However, given that the loss of information quantified by the fraction of missing information (FMI) is about 75–-80%, the data collection sample sizes would need to be about 4–5 times larger compared to the traditional data collection of all items at once. Unless the response rate drops sharply by a factor of more than two due to the increase in questionnaire length, the total public burden is increased.
The sample sizes necessary to detect the required effect sizes in increased $R^2$ demonstrate long tails in the exploration of parameter spaces. These long tails make it difficult to plan for the worst-case scenarios associated with “unfortunate” regression parameter configurations. Should a specific decision need to be made based on the parameter explorations akin to those undertaken in Section \[sec:explore\], the trade-off between the survey costs due to large sample sizes and risks of having an underpowered study should the coefficient estimates be found to have an “unfortunate” configuration should be carefully discussed with the survey stakeholders to find the most appropriate course of action.
We conducted a finite sample simulation with non-normal data and several missing data methods, and determined that the methods that assume multivariate normality generally perform poorly, and generate a non-negligible proportion of really bad outliers. In comparison, semiparametric multiple imputation by predictive mean matching with sufficiently large number of imputed data sets seem to work best.
Our work can be extended in a number of additional dimensions. The derivations of asymptotic variances are based on the working assumption of multivariate normality and using the inverse information matrix to estimate variances. With non-normal data, the problem can be formulated in terms of estimating equations, and sandwich variance estimators should be formed. As our simulation demonstrated, asymptotic standard errors based on inverse information matrix are inadequate for the analysis methods that we used, leading to underestimates with misspecified normality-based methods, and overestimates with a more accurate semiparametric method.
The current paper assumed independence of respondents. In practice, complex survey features such as strata, clusters, unequal probabilities of selection, and weight calibration would affect asymptotic properties of the estimates. In particular, the sandwich variance estimation will be required. Many practical survey statistics issues may also interact with multiple matrix sampling in unusual ways. How would differential nonresponse by form affect the results? What should we do when a stratum has fewer than two cases of a given form? These and other questions related to design-based inference would need to be answered when multiple matrix sampling is applied in practice.
Finally, in terms of ensuring adequate measurement properties, we note that psychometric properties are usually established and validated for scales, but not necessarily subscales that respondents are exposed to in multiple matrix sampling instruments. In particular, if the order of the items, or the degree of mixing of items from the different subscales of the Big Five Inventory is important for the validity of the scale and its subscales, these properties may be violated when shorter subscales are administered that require the respondent to answer similar questions more frequently.
[^1]: Abt Associates, 8405 Colesville Rd, Suite 300, Silver Spring, MD 20910. Corresponding author: `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce a novel and efficient algorithm called the stochastic approximate gradient descent (SAGD), as an alternative to the stochastic gradient descent for cases where unbiased stochastic gradients cannot be trivially obtained. Traditional methods for such problems rely on general-purpose sampling techniques such as Markov chain Monte Carlo, which typically requires manual intervention for tuning parameters and does not work efficiently in practice. Instead, SAGD makes use of the Langevin algorithm to construct stochastic gradients that are biased in finite steps but accurate asymptotically, enabling us to theoretically establish the convergence guarantee for SAGD. Inspired by our theoretical analysis, we also provide useful guidelines for its practical implementation. Finally, we show that SAGD performs well experimentally in popular statistical and machine learning problems such as the expectation-maximization algorithm and the variational autoencoders.'
author:
- |
Yixuan Qiu^1^ and Xiao Wang^2^\
^1^Department of Statistics and Data Science, Carnegie Mellon University, [email protected]\
^2^Department of Statistics, Purdue University University, [email protected]
bibliography:
- 'ref.bib'
title: Stochastic Approximate Gradient Descent via the Langevin Algorithm
---
Introduction
============
The stochastic gradient descent method (SGD, [@bottou2010large]; [@bottou2018optimization]) is one of the most popular and widely-used optimization techniques in large-scale machine learning problems. In many cases, the objective function one needs to optimize can be written as an expectation, $F(\theta)=\mathbb{E}[f(\theta;\xi)]$, over some random variable $\xi\in\mathbb{R}^{r}$ whose distribution is independent of the parameter vector $\theta\in\Theta\subset\mathbb{R}^{p}$. Under very mild regularity conditions, the true gradient of $F(\theta)$ is also an expectation, obtained as $g(\theta)\coloneqq\nabla F(\theta)=\mathbb{E}[\nabla f(\theta;\xi)]$. When the computational cost of $g(\theta)$ is massive, SGD makes use of the stochastic gradient, denoted by $\tilde{g}(\theta)$, to update the parameter vector. It has been well studied that by appropriately choosing the step sizes, SGD has good convergence properties [@robbins1951stochastic]. As an important special case, SGD is frequently used in the scenario where $F(\theta)$ is an average over the data points, $F(\theta)=n^{-1}\sum_{i=1}^{n}f(\theta;X_{i})$. If data $X_{1},\ldots,X_{n}$ are assumed to be independent and identically distributed, then an unbiased stochastic gradient can be trivially obtained as $\tilde{g}(\theta)=\nabla f(\theta;X_{I})$, where $I$ follows a uniform distribution on $\{1,2,\ldots,n\}$.
However, there are a much broader class of problems where $\xi$ follows a general probability distribution $\pi(\xi)$. Unlike the previous simple scenario, in many cases an unbiased stochastic gradient cannot be easily obtained due to the complexity of $\pi(\xi)$. If $\pi(\xi)$ is beyond the scope of standard distribution families, then some general-purpose sampling techniques such as Markov chain Monte Carlo (MCMC, [@gilks1995markov]; [@metropolis1953equation]; [@hastings1970monte]; [@geman1984gibbs]; [@brooks2011handbook]) have to be adopted, which can be quite slow in practice.
In this article, we propose a novel and efficient algorithm called the stochastic approximate gradient descent (SAGD), as an alternative to SGD for cases where unbiased stochastic gradients cannot be trivially computed. The key idea of SAGD is to construct the stochastic gradient using the Langevin algorithm [@roberts1996exponential; @roberts2002langevin; @cheng2018underdamped], a sampling method whose statistical error can be rigorously quantified. In addition, we use an adaptive sampling scheme that allows larger errors in the early stage of the optimization, and gradually improves the precision as the procedure goes on.
These heuristics are formalized in the SAGD algorithm, and various theoretical results are developed to guarantee its convergence. Moreover, our analysis gives clear rates of the relevant hyperparameters, which provide useful guidelines for practical implementations of SAGD. The highlights and main contributions of this article are as follows:
- We develop a new computational framework for SGD problems in which a stochastic gradient cannot be trivially obtained. The proposed SAGD algorithm is fully automated with a solid convergence guarantee.
- New theoretical contributions are made to the underdamped Langevin algorithm for sampling from sophisticated distributions, which are of interest by their own.
- We discuss the application of the proposed SAGD framework in some important statistical and machine learning problems, including the expectation-maximization algorithm (EM algorithm), and the variational autoencoders (VAE). We show that SAGD is able to automate the EM algorithm for complex models and effectively remove the bias of VAE.
**Notation:** Throughout this article we adopt the following notation. Let $\mathbb{R}^{r}$ be the $r$-dimensional Euclidean space with the inner product $\langle\cdot,\cdot\rangle$ and norm $\Vert\cdot\Vert$. For matrices and higher-order tensors, $\Vert\cdot\Vert$ denotes the operator norm. Let $C\subset\mathbb{R}^{r}$ be a closed convex set, and then the notation $\mathcal{P}_{C}(x)$ means the projection of $x\in\mathbb{R}^{r}$ onto $C$. A mapping $\phi:\mathbb{R}^{r}\to\mathbb{R}^{s}$ is said to have polynomial growth if there exist a constant $C>0$ and an integer $m\ge0$ such that $\Vert\phi(x)\Vert\le C(1+\Vert x\Vert^{m})$ for all $x\in\mathbb{R}^{r}$. The notation $\nabla^{i}\phi$ is used to denote the $i$-th derivative of a multivariate function $\phi:\mathbb{R}^{r}\to\mathbb{R}$, and in particular $\nabla^{0}\phi\equiv\phi$. We use $\mathscr{C}_{poly}^{m}$ to denote the space of mapping $\phi$ such that $\phi$ is $m$-times differentiable, and $\phi$ and its derivatives have polynomial growth. A function $\phi:\mathbb{R}^{r}\to\mathbb{R}$ is said to be $L$-Lipschitz continuous if $|\phi(x)-\phi(y)|\le L\Vert x-y\Vert$ for all $x,y\in\mathbb{R}^{r}$.
Related Work
============
The two main ingredients of the proposed SAGD framework are SGD and the Langevin algorithm. SGD has been extensively studied in the literature, and recent research mainly focused on its acceleration, for example variance reduction methods [@johnson2013accelerating; @reddi2016stochastic], adaptive step sizes [@duchi2011adaptive; @zeiler2012adadelta], momentum methods [@kingma2014adam; @luo2019adabound], etc. In this article, the proposed SAGD framework is based on the original version of SGD, but it can be easily adapted to those acceleration methods.
The Langevin algorithm has two variants, the *overdamped* and the *underdamped* versions. Most of the analysis in literature was based on the overdamped version [@dalalyan2017theoretical; @durmus2016high; @durmus2017nonasymptotic; @cheng2017convergence], whereas some recent research suggests that the underdamped version has faster convergence for some special classes of distributions [@cheng2018sharp; @ma2019there]. Due to this reason, we use the underdamped Langevin algorithm to develop the SAGD framework. As a byproduct, we have derived new results for the underdamped Langevin algorithm that complement prior art.
The Langevin algorithm can be compared to MCMC, as they are both useful sampling techniques. In fact, there are MCMC algorithms derived from the Langevin algorithm such as the Metropolis-based overdamped Langevin algorithm [@roberts1996exponential] and the underdamped Langevin MCMC [@cheng2018underdamped]. Nevertheless, the Langevin algorithm has several advantages in our problem. First, the Langevin algorithm skips the Metropolis adjustment step existing in most MCMC methods, which saves computational time and avoids duplicated values in the sample. Second, as our theoretical analysis shows, the Langevin algorithm has transparent hyperparameter setting and requires less manual intervention. The downside is the resulting bias of the Langevin algorithm, but the theoretical analysis shows that it does not harm the convergence of SAGD.
The idea to combine SGD with the Langevin algorithm has been seen in articles such as @xie2018cooperative and @han2019divergence, but in these works the Langevin algorithm was merely used as an MCMC-like sampling technique, and its statistical error and impact on the convergence of optimization were ignored. Instead, in SAGD the two ingredients are connected in a coherent way, with a rigorous theoretical analysis. One recent work that is similar to SAGD is @de2019efficient, but the major difference is that they used the overdamped Langevin algorithm for sampling, whose theoretical analysis is very different from SAGD. Another direction of research that combines SGD and the Langevin algorithm is the stochastic gradient Langevin dynamics (SGLD, [@welling2011bayesian]; [@vollmer2016exploration]). However, SGLD utilizes SGD to accelerate the Langevin sampling method, while our work aims at extending SGD by using the Langevin algorithm to construct the approximate gradient.
The Underdamped Langevin Algorithm {#sec:langevin}
==================================
In this section we provide some background knowledge of the underdamped Langevin algorithm, and derive a few important results that are crucial to the convergence of SAGD. At a high level, the underdamped Langevin algorithm is an approach to obtaining approximate samples from a target distribution $\pi(\xi)$. In many cases, we can only compute $\pi(\xi)$ up to some normalizing constant, *i.e.*, we have access to $V(\xi)\coloneqq-\log(\pi(\xi))+C$, where $C$ is free of $\xi$. Then the *underdamped Langevin diffusion* is defined by the following stochastic differential equation (SDE) for $W(t)=(\xi^{\mathrm{T}}(t),\rho^{\mathrm{T}}(t))^{\mathrm{T}}\in\mathbb{R}^{2r}$ with $\xi(t),\rho(t)\in\mathbb{R}^{r}$, $$\begin{aligned}
\mathrm{d}\xi(t) & =\rho(t)\mathrm{d}t,\label{eq:sde_xi}\\
\mathrm{d}\rho(t) & =-\gamma\rho(t)\mathrm{d}t-\nabla V(\xi(t))\mathrm{d}t+\sqrt{2\gamma}\mathrm{d}B(t),\label{eq:sde_rho}\\
\xi(0) & =\xi_{0},\ \rho(0)=\rho_{0},\ t\ge0,\nonumber \end{aligned}$$ where $\gamma>0$ is a fixed constant but can be chosen arbitrarily, and $B(t)$ is an $r$-dimensional Brownian motion. Under mild conditions, Proposition 6.1 of @pavliotis2014stochastic shows that the invariant distribution of $W(t)$ is unique, with the density function $$\pi_{W}(\xi,\rho)\propto\exp\{-V(\xi)-\Vert\rho\Vert^{2}/2\}\propto\pi(\xi)\cdot\exp(-\Vert\rho\Vert^{2}/2),$$ where $\rho$ is an auxiliary variable, and our main interest is in $\xi$. The form of $\pi_{W}(\xi,\rho)$ indicates that $\xi$ and $\rho$ are independent with $\xi\sim\pi(\xi)$ and $\rho\sim N(0,I_{r})$. That is, if we can solve the SDE exactly, then $\xi$ follows the target distribution in the long run.
However, in general the solution to and has no closed form, so some discretization methods have to be adopted. Consider the following discretized chain for $W_{k}=(\xi_{k}^{\mathrm{T}},\rho_{k}^{\mathrm{T}})^{\mathrm{T}}$, $k\ge0$: $$\begin{aligned}
\xi_{k+1} & =\xi_{k}+\delta\rho_{k},\label{eq:langevin_xi}\\
\rho_{k+1} & =(1-\gamma\delta)\rho_{k}-\delta\cdot\nabla V(\xi_{k})+\sqrt{2\gamma\delta}\eta_{k},\label{eq:langevin_rho}\end{aligned}$$ where $\delta$ is the step size, $\{\eta_{k}\}_{k=0}^{\infty}\overset{iid}{\sim}N(0,I_{r})$, and $\eta_{k}$ is independent of $\{W_{k}\}_{i=0}^{k-1}$. The iterations and are typically referred to as the *underdamped Langevin algorithm*.
The importance and usefulness of the $\{W_{k}\}$ sample will be illustrated in Theorem \[thm:control\_bias\_mse\]. Before that we need to first guarantee that $\{W_{k}\}$ is well defined and does not explode as time goes on. Formally, we show that under some mild conditions, $\{W_{k}\}$ is stable in the sense that it has finite moments of any order, uniformly in the step count $k$. The result is summarized in Theorem \[thm:finite\_moments\], along with the assumptions we need to impose.
\[assu:v\_boundedness\](a) $V(x)$ is bounded from below, i.e., $V(x)\ge\nu_{0}$ for some constant $\nu_{0}\in\mathbb{R}$ and all $x\in\mathbb{R}^{r}$. (b) The operator norm of the second derivative of $V$ is bounded, i.e., $\Vert\nabla^{2}V(x)\Vert\le\nu$ for some constant $\nu>0$ and all $x\in\mathbb{R}^{r}$. (c) $V(x)\in\mathscr{C}_{poly}^{\infty}$.
For Assumption \[assu:v\_boundedness\](a), we can assume $\nu_{0}=0$ without loss of generality. This is because we can always work on a scale-transformation of $\xi$, $\xi'=c\xi$, resulting in a transformed $V$, $V'(x)=V(x/c)+\log c$. In what follows we adopt this simplification, so that we have $V(x)\ge0$.
\[assu:v\_dissipative\]There exist constants $\alpha>0$ and $0<\beta<1$ such that for all $x\in\mathbb{R}^{r}$, $$\frac{1}{2}\langle\nabla V(x),x\rangle\ge\beta V(x)+\gamma^{2}C_{\beta}\Vert x\Vert^{2}-\alpha,\quad C_{\beta}=\frac{\beta(2-\beta)}{8(1-\beta)}.$$
Assumption \[assu:v\_dissipative\] is a common and standard regularity condition on $V$ coming from @mattingly2002ergodicity. We then have the following conclusion:
\[thm:finite\_moments\]Suppose Assumptions \[assu:v\_boundedness\] and \[assu:v\_dissipative\] hold, and choose $\delta$ small enough such that $\delta\le\min\{1/\gamma,\gamma/(2\nu),(D+1-\sqrt{D^{2}+1})/\gamma\}$, $D=\gamma^{4}C_{\beta}/\nu^{2}$. Then for any fixed $l>0$ and all $k\ge0$, there exist constants $C=C(l,\delta)>0$, $\lambda=\lambda(l,\delta)>0$, and an integer $m=m(l)>0$ such that $$\mathbb{E}\left(\Vert\xi_{k}\Vert^{2l}+\Vert\rho_{k}\Vert^{2l}\right)\le C\left\{ 1+\left(\Vert\xi_{0}\Vert^{m}+\Vert\rho_{0}\Vert^{m}\right)e^{-\lambda k}\right\} .$$
Next, we present the main result for the underdamped Langevin algorithm. Let $\varphi:\mathbb{R}^{2r}\to\mathbb{R}$ be a multivariate function with the notation $\varphi(w)\equiv\varphi(\xi,\rho)$, where $w=(\xi^{\mathrm{T}},\rho^{\mathrm{T}})^{\mathrm{T}}$. Then define its expectation with respect to $\pi_{W}$ as $\bar{\varphi}=\mathbb{E}_{\pi_{W}}\varphi\coloneqq\int\varphi(\xi,\rho)\pi_{W}(\xi,\rho)\mathrm{d}\xi\mathrm{d}\rho$. It is easy to see that if $\varphi(w)=\nabla f(\theta;\xi)$, then $\bar{\varphi}=\mathbb{E}_{\pi_{W}}\varphi=\nabla F(\theta)$ is exactly the true gradient function we are interested in. Driven by the motivation to approximate $\bar{\varphi}$, Theorem \[thm:control\_bias\_mse\] below shows that we can construct an estimator $\hat{\varphi}$ using the sequence $\{W_{k}\}$, where $\hat{\varphi}=K^{-1}\sum_{k=0}^{K-1}\varphi(W_{k})$.
\[thm:control\_bias\_mse\]Let $\varphi$, $\bar{\varphi}$ , and $\hat{\varphi}$ be defined as above, with $\varphi\in\mathscr{C}_{poly}^{r+5}$. Assume that the conditions in Theorem \[thm:finite\_moments\] hold. Then there exist constants $C_{1}>0$ and $C_{2}>0$ such that for any $\delta>0$ in the range and any integer $K>0$, we have $$\begin{aligned}
\left|\mathbb{E}(\hat{\varphi})-\bar{\varphi}\right| & \le C_{1}\left(\frac{1}{K\delta}+\delta\right),\\
\mathbb{E}\left[\left(\hat{\varphi}-\bar{\varphi}\right)^{2}\right] & \le C_{2}\left(\frac{1}{K\delta}+\delta^{2}\right).\end{aligned}$$
Theorem \[thm:control\_bias\_mse\] shows that in general $\hat{\varphi}$ is a biased estimator for $\bar{\varphi}$, but its bias and mean squared error can be made arbitrarily small by appropriately choosing the algorithm parameters $\delta$ and $K$.
Here we make a few remarks about the results in this section. Theorem \[thm:finite\_moments\] is similar to Proposition 2.7 of @kopec2015weak, but they use the implicit Euler scheme to discretize the Langevin SDE, which is computationally much harder. Therefore, Theorem \[thm:finite\_moments\] is a new result for the explicit Euler scheme given by and . The rates in Theorem \[thm:control\_bias\_mse\] are known results [@chen2015convergence]. However, in most prior art the assumptions to make Theorem \[thm:control\_bias\_mse\] hold are highly non-trivial and very difficult to check for real machine learning models. For example, @chen2015convergence needs to assume that our conclusion in Theorem \[thm:finite\_moments\] holds, along with other technical conditions. In contrast, our assumptions are only made on the log-density function $V(\xi)$, which is the actual model that machine learning practitioners are given. In this sense, the results developed in this article have much broader practical use.
The benefit of our new results is that we can easily verify the assumptions for popular machine learning models. For example, the following corollary justifies the use of Langevin algorithm to sample from deep generative models (e.g. VAE). Consider a single-layer neural network $h(z)=a(Wz+b)$, where $z\in\mathbb{R}^{r}$, $b\in\mathbb{R}^{m}$, $W\in\mathbb{R}^{m\times r}$, and the activation function is $a(x)=\log(1+e^{x})$. Then we have the following result.
\[cor:nn\]Assume that $Z\sim N(0,I_{r})$ and $X|\{Z=z\}\sim N(h(z),\sigma^{2}I)$, where $\sigma^{2}$ is a constant. Let $p(z|x)$ denote the conditional density of $Z$ given $X=x$, and then $V(z)=-\log p(z|x)$ satisfies Assumptions \[assu:v\_boundedness\] and \[assu:v\_dissipative\].
For brevity we omit the multi-layer case, but it can be analyzed similarly. In later part of this article we will discuss the application of SAGD to VAE model in more details.
Stochastic Approximate Gradient Descent
=======================================
With the statistical properties of the underdamped Langevin algorithm studied in Theorem \[thm:control\_bias\_mse\], the SAGD framework can then be readily developed. Recall that our target is to minimize the function $F(\theta)=\mathbb{E}[f(\theta;\xi)]$, whose true gradient $g(\theta)=\mathbb{E}[\nabla f(\theta;\xi)]$ is hard to compute exactly. Using the technique developed in the previous section, we can construct a stochastic gradient, $\tilde{g}(\theta)=K^{-1}\sum_{k=0}^{K-1}\nabla f(\theta;\xi_{k})$, to approximate $g(\theta)$. Unlike most existing SGD settings, $\tilde{g}(\theta)$ is not an unbiased estimator for $g(\theta)$, as suggested by Theorem \[thm:control\_bias\_mse\]. Therefore, we refer to the optimization method based on such a $\tilde{g}(\theta)$ as the stochastic *approximate* gradient descent. The outline of SAGD is given in Algorithm \[alg:sagd\].
Despite the fact that $\tilde{g}(\theta)$ is a biased estimator for the true gradient, we show that by carefully choosing the hyperparameters, we can actually guarantee the overall convergence of SAGD. Interestingly, the convergence rate for a convex objective function, in terms of the number of gradient updates, is the same as the vanilla SGD method with an order of $\mathcal{O}(1/\sqrt{T})$, as is shown in Theorem \[thm:convergence\_convex\].
\[assu:f\_condition\]$f(\theta;\cdot)\in\mathscr{C}_{poly}^{r+5}$ for each $\theta\in\Theta$, and there exist a constant $C>0$ and an integer $m\ge0$ such that $\Vert\nabla^{i}f(\theta;\cdot)\Vert\le C(1+\Vert\cdot\Vert^{m})$ for all $\theta\in\Theta$ and $0\le i\le r+5$.
\[thm:convergence\_convex\]Suppose that $F(\theta)$ is convex and $L$-Lipschitz continuous in $\theta\in\Theta$, and $\Theta$ is a closed convex set with diameter $D<\infty$. Also assume that Assumption \[assu:f\_condition\] and the conditions in Theorem \[thm:finite\_moments\] hold. Then by choosing $\delta_{t}=C_{1}/\sqrt{t}$, $K_{t}=C_{2}t$, and $\alpha_{t}=\alpha_{0}/\sqrt{t}$, where $C_{1},C_{2},\alpha_{0}>0$ are constants, we have $\mathbb{E}[F(\hat{\theta})]-F^{*}\le\mathcal{O}(1/\sqrt{T})$.
The significance of Theorem \[thm:convergence\_convex\] is that it provides clear rates for the hyperparameters $\delta_{t}$ and $K_{t}$ in the sampling algorithm, which are crucial for practical algorithm implementation but are typically missing in other MCMC-based methods. Of course, the preservation of the SGD rate is not without a price. Theorem \[thm:convergence\_convex\] indicates that the number of inner iterations, *i.e.*, $K_{t}$ in Algorithm \[alg:sagd\], needs to increase with $t$. However, the developed error bounds are typically conservative, so for practical use, we advocate the following techniques to speed up SAGD: (1) An educated initial value $\xi_{0}$ can be used to initialize the Langevin algorithm, for example in VAE $\xi_{0}$ is sampled from the trained encoder; (2) A persistent Langevin Markov chain is stored during optimization, motivated by the persistent contrastive divergence [@tieleman2008training]; (3) Some advanced gradient update schemes such as Adam can be used.
More generally, we consider objective functions that are nonconvex but smooth, and assume that $\Theta=\mathbb{R}^{p}$. Theorem \[thm:convergence\_nonconvex\] indicates that with a proper choice of hyperparameters, the algorithm again has a nice convergence property.
\[thm:convergence\_nonconvex\]Suppose that $g(\theta)$ is $G$-Lipschitz continuous in $\theta$, and assume that Assumption \[assu:f\_condition\] and the conditions in Theorem \[thm:finite\_moments\] hold. Let $\delta_{t}=C_{1}t^{-c}$, $K_{t}=C_{2}t^{2c}$, and $\alpha_{t}=\alpha_{0}/t$ for some constants $0<\alpha_{0}<1/(2G)$ and $C_{1},C_{2},c>0$. Then we have $\lim\inf_{t\rightarrow\infty}\mathbb{E}[\Vert g(\theta_{t})\Vert^{2}]=0$.
Theorem \[thm:convergence\_nonconvex\] is an analog to Theorem 4.9 of @bottou2018optimization. It is not meant to be the strongest conclusion, but to provide insights on the convergence property of SAGD for nonconvex objective functions.
Applications: EM Algorithm and VAE {#sec:application}
==================================
Automated EM Algorithm
----------------------
The SAGD framework is very useful for implementing an automated version of the EM algorithm [@dempster1977maximum]. EM algorithm is a powerful and indispensable tool to solve missing data problems and latent variable models. Given the data set $X$ that follows a probability distribution with density function $f(x;\theta)$, we are interested in computing the maximum likelihood estimator $\hat{\theta}=\arg\max_{\theta}\,\ell(\theta;x)$ for the unknown parameter vector $\theta$, where $\ell(\theta;x)=\log[f(x;\theta)]$ is the log-likelihood function.
However, in many cases the computation of the marginal distribution $f(x;\theta)$ is intractable, but with an additional random vector $U$, the complete log-likelihood $L(\theta;x,u)=\log[f(x,u;\theta)]$ is simple. This phenomenon typically happens when $U$ represents missing data or latent variables in the model. The EM algorithm computes $\hat{\theta}$ in an iterative way. Given the current value of $\theta$, denoted by $\theta_{k}$, the EM algorithm proceeds by the following two steps:
- **Expectation Step (E-step)**: Compute the expected value of $L(\theta;x,U)$ with respect to the conditional distribution of $U$ given $X=x$ under the current parameter estimate $\theta_{k}$, and define the function $Q(\theta;\theta_{k})=\mathbb{E}_{U|X=x,\theta_{k}}[L(\theta;x,U)]$.
- **Maximization Step (M-step)**: Update the estimate of $\theta$ by maximizing the $Q$ function: $\theta_{k+1}=\arg\max_{\theta}\,Q(\theta;\theta_{k})$.
The EM algorithm has a remarkable monotonicity property, *i.e.*, the marginal log-likelihood $\ell(\theta;x)$ is always nondecreasing on the $\{\theta_{k}\}$ sequence. Due to such nice properties, the EM algorithm has been the standard optimization technique for Gaussian mixture models and many other missing data models. However, one serious problem of the EM algorithm is that the expectation defining the $Q$ function usually has no simple closed form, so the Monte Carlo EM algorithm (MCEM, [@wei1990monte]; [@levine2001implementations]) proposes to use Monte Carlo methods to approximate the expectation. Using MCMC to approximate the expectation in the E-step is not a new idea, but what really matters is how to properly choose the hyperparameters to guarantee the convergence of the M-step.
In this sense, Theorem \[thm:convergence\_convex\] provides a clear way to make the EM algorithm effectively automated. It is easy to see that the target distribution $\pi(\xi)$ is $p(u|x;\theta_{k})$, the conditional density of $U$ given $X=x$, which is proportional to the joint density of $(X,U)$ under $\theta_{k}$. Therefore, we can define $V(\xi)=-L(\theta_{k};x,\xi)$, and then apply Algorithm \[alg:sagd\] to directly solve the M-step, whose convergence is readily guaranteed. Finally, one only needs to create an outer loop to iteratively update the $\{\theta_{k}\}$ sequence, until some convergence condition is met.
Debiased VAE {#subsec:vae_bias_correction}
------------
The automated EM algorithm can be further used to improve the popular VAE model [@kingma2014stochastic]. VAE has the same goal of seeking the maximum likelihood estimator for $\theta$, but it uses the variational Bayes technique to maximize a lower bound of $\ell(\theta;x)$. Let $q(u|x)$ be any conditional density function, and then VAE maximizes the function $\tilde{\ell}(\theta;x)$, defined by $$\tilde{\ell}(\theta;x)=\mathbb{E}_{u\sim q}[\log p(x|u;\theta)]-\mathrm{KL}[q(u|x)\Vert p(u)],\label{eq:elbo}$$ where $p(u)$ is the marginal density of $u$, and $p(x|u;\theta)$ is the conditional distribution of $X$ given $U=u$. In most VAE settings, $U\sim N(0,I)$, and $q(u|x)$ is taken to be a normal distribution whose mean and variance parameters are represented by a deep neural network. VAE has been successfully applied to many problems, but its most critical weakness is that VAE does not maximize the exact log-likelihood, which induces a bias in the final $\theta$.
Here we show that using the SAGD framework, the bias of VAE can be removed via an additional refining step. First, it is easy to show that $\tilde{\ell}(\theta;x)$ has another representation, $\tilde{\ell}(\theta;x)=\ell(\theta;x)-\mathrm{KL}[q(u|x)\Vert p(u|x;\theta)]$. That is, if the distribution $q(u|x)$ matches the true $p(u|x;\theta)$, then $\tilde{\ell}(\theta;x)$ is the genuine log-likelihood function $\ell(\theta;x)$. In this case, the objective function can be optimized via an EM algorithm with a $Q$ function $Q(\theta;\theta_{k})=\mathbb{E}_{U\sim p(u|x;\theta_{k})}[\log p(x|u;\theta)+\log p(u)]$, and we update the current parameter $\theta_{k}$ by a gradient move $$\theta_{k+1}=\theta_{k}+\alpha_{k}\cdot[\partial Q(\theta;\theta_{k})/\partial\theta]|_{\theta=\theta_{k}},$$ with the true expectation replaced by the Langevin-based approximate gradient. The consequence of this refining step is that we are now optimizing the true log-likelihood function $\ell(\theta;x)$ instead of the lower bound $\tilde{\ell}(\theta;x)$, and hence the bias of VAE is removed.
We emphasize that we do not position SAGD as a replacement for VAE; in fact, VAE is computationally more efficient and has a lower variance. Instead, the major virtue of SAGD is its bias-correction capacity that fixes the intrinsic gap between the evidence lower bound of VAE and the true likelihood. Therefore, we suggest using VAE to pre-train models, and then fine-tuning the generative network using SAGD due to its theoretical guarantee.
Numerical Experiments
=====================
EM Algorithm
------------
In this section we use numerical experiments to demonstrate the applications of SAGD in EM algorithm and VAE as discussed in the previous section. First consider a simple model such that the parameter estimation procedure can be easily visualized. Assume that given latent variables $Z_{1},\ldots,Z_{n}\overset{iid}{\sim}N(0,1)$, the data are independently generated as $X_{i}|\{Z_{1}=z_{1},\ldots,Z_{n}=z_{n}\}\sim\mathsf{Gamma}(10\cdot\sigma(a+bz_{i}))$, where $\sigma(x)=1/(1+\exp(-x))$ is the sigmoid function, and $\mathsf{Gamma}(s)$ stands for a gamma distribution with shape parameter $s$. The target is to estimate the unknown parameters $\theta=(a,b)$ from the observed data $X_{1},\ldots,X_{n}$. In our simulation, the true parameters are set to $a=2$ and $b=0.5$, and a sample size $n=100$ is used to simulate $X_{i}$.
For a single variable pair $(x,z)$, it is easy to show that the complete log-likelihood function is $L(\theta;x,z)=-z^{2}/2+(s-1)\log(x)-\log\{\Gamma(s)\}+C$, where $s=10\cdot\sigma(a+bz)$, $\Gamma(\cdot)$ is the gamma function, and $C$ is a constant. The EM algorithm is then used to solve this problem as follows. In the $k$-th M-step, we fix parameter estimate at $\theta_{k}=(a_{k},b_{k})$, and then optimize the objective function $Q(\theta;\theta_{k})=\mathbb{E}_{Z|X=x,\theta_{k}}[L(\theta;x,Z)]$ using SAGD. The next $\theta$ value is set to the optimum of $Q(\theta;\theta_{k})$.
Since for this model we can evaluate the true derivatives of $Q(\theta;\theta_{k})$ using numerical integration, it is of interest to compare SAGD with the exact gradient descent (GD) method. We set the initial value to be $\theta_{0}=(0,1)$, and run both SAGD and exact GD for $T=100$ iterations in each M-step, with a constant step size $\alpha_{t}=0.2$. For SAGD, Langevin parameters are specified as $\delta_{t}=0.1/\sqrt{t}$ and $K_{t}=t+20$, with the first 100 Langevin iterations discarded as burn-in, similar to that in MCMC. Figure \[fig:path\_loglik\](a) demonstrates the path of $(a,b)$ values on the surface of the true log-likelihood function after three M-steps, and Figure \[fig:path\_loglik\](b) gives the log-likelihood values at each gradient update. Clearly, Figure \[fig:path\_loglik\] shows that the path of SAGD nicely approximates that of exact GD, which further verifies the validity of the SAGD algorithm.
[0.48]{} {width="\textwidth"}
[0.36]{} {width="\textwidth"}
Debiased VAE {#debiased-vae}
------------
In the second experiment, we use synthetic data to show that even in the simplest setting, VAE can lead to biased distribution estimation, but its bias can be effectively corrected by SAGD. The observed data are generated as follows: given independent latent variables $Z_{i}\sim\pi(z)$, we set $X_{i}=Z_{i}+e_{i}$, where $e_{i}\sim N(0,1)$ is independent of $Z_{i}$, $i=1,2,\ldots,n$. The target is to recover the unknown latent distribution $\pi(z)$ from $X_{1},\ldots,X_{n}$. In the VAE framework, we first represent $Z$ by a deep neural network transformation $Z=h_{\theta}(U)$, where $U\sim N(0,1)$, and then we have $p(x|u;\theta)\equiv N(h_{\theta}(u),1)$. Once the neural network function $h_{\theta}$ has been learned, we can simulate random variates of $Z=h_{\theta}(U)$ by generating random $U\sim N(0,1)$, and $\pi(z)$ is approximated by the empirical distribution of a large sample of $Z$. Therefore, by evaluating the quality of the $Z$ sample, we can study the accuracy of the learned $h_{\theta}$ function.
In our experiment, we consider three true latent distributions and generate the corresponding data sets: (a) $\pi=N(1,0.5^{2})$; (b) an exponential distribution of mean 2; (c) a mixture of normal distributions, $\pi=0.4\cdot N(0,0.5^{2})+0.6\cdot N(3,0.5^{2})$. For each case, we first train a VAE model with 5000 iterations, and then fine-tune the neural network parameter $\theta$ by running the following four training algorithms for additional 1000 iterations: (a) VAE; (b) the importance weighted autoencoders (IWAE, [@burda2015importance]) with $k=50$ importance samples; (c) Hamiltonian Monte Carlo (HMC) to approximate the true gradient; (d) SAGD. In HMC we use the same step size and chain length as SAGD, and run $L=5$ leapfrog steps to get each proposal. After training is finished, we simulate random variates of $Z$, and compare its empirical distribution $\hat{\pi}$ with the true latent distribution $\pi$. The Kolmogorov–Smirnov distance and 1-Wasserstein distance between $\hat{\pi}$ and $\pi$ are computed. Figure \[fig:latent\_estimation\] shows the data distribution, true latent distribution $\pi$, and the estimated $\hat{\pi}$ in each setting, based on one simulated data set of sample size $n=1000$.
{width="80.00000%"}
Figure \[fig:latent\_estimation\] reflects the following remarkable results. For the normal case, VAE has little bias, since the true conditional distribution $p_{U|X}(u|x)$ is indeed normal, which is well characterized by the encoder. However, in other two cases, neither the latent distribution $\pi(z)$ nor $p_{U|X}(u|x)$ is normal, and hence VAE gives highly biased estimates for $\pi$. For the three debiasing methods, the refining steps indeed reduce the bias of VAE. However, the debiasing effect of HMC is smaller than that of SAGD, even though theoretically they are similar. HMC also takes more computing time due to the leapfrog steps and the calculation of acceptance probability. For IWAE, it tends to overly truncate the support of the distribution and exaggerate the density of modes. Overall, SAGD provides the most favorable bias reduction results.
[>m[0.11]{}>m[0.01]{}>p[0.15]{}>p[0.15]{}>p[0.15]{}>p[0.15]{}]{} & & VAE & IWAE & HMC & SAGD[\
]{} \[-1.2em\][Normal]{} & \[-0.5em\][$D$]{} & 0.033 (0.0024) & 0.043 (0.0035) & 0.032 (0.0019) & 0.033 (0.0020)[\
]{} & \[-0.5em\][$W$]{} & 0.052 (0.0040) & 0.066 (0.0047) & 0.050 (0.0032) & 0.052 (0.0035)[\
]{} \[-1.2em\][Exp(2)]{} & \[-0.5em\][$D$]{} & 0.095 (0.0018) & 0.059 (0.0039) & 0.085 (0.0016) & 0.064 (0.0029)[\
]{} & \[-0.5em\][$W$]{} & 0.226 (0.0079) & 0.125 (0.0091) & 0.161 (0.0051) & 0.115 (0.0082)[\
]{} \[-1.2em\][Mixture]{} & \[-0.5em\][$D$]{} & 0.127 (0.0027) & 0.098 (0.0061) & 0.104 (0.0015) & 0.085 (0.0019)[\
]{} & \[-0.5em\][$W$]{} & 0.320 (0.0025) & 0.165 (0.0059) & 0.276 (0.0021) & 0.197 (0.0031)[\
]{} \[-1.2em\][High-Dim.]{} & \[-0.5em\][$D$]{} & 0.093 (0.0010) & 0.080 (0.0012) & 0.092 (0.0008) & 0.065 (0.0005)[\
]{} & \[-0.5em\][$W$]{} & 0.222 (0.0027) & 0.158 (0.0029) & 0.212 (0.0030) & 0.093 (0.0015)[\
]{}
To further take into account the randomness in data generation, we simulate 30 replications of the data set in each setting, and compute the mean and standard errors of the distance metrics, shown in Table \[tab:latent\_estimation\_rep\]. Also included in this table is a high-dimensional data set with sample size $n=10000$ and dimension $p=100$: each dimension independently follows an exponential distribution with mean 2. We pre-train this data set using VAE for 10000 epochs, and then fit each method with 1000 more epochs. The mean and standard errors are computed over all dimensions. Both Figure \[fig:latent\_estimation\] and Table \[tab:latent\_estimation\_rep\] indicate that SAGD provides good bias reduction results for VAE in both simple and high-dimensional settings, which highlights the importance of optimizing the correct objective function in model fitting.
Generative Model for MNIST Data
-------------------------------
![\[fig:mnist\]Representative examples from randomly generated digits that show significant improvement after the refining step using SAGD.](mnist-comparison){width="0.95\columnwidth"}
In the last experiment, we consider the MNIST handwritten digits data set, and fit generative models on it. The dimension of the latent space is set to 20, and the generative network is a combination of convolutional filters and fully-connected layers. We first train a VAE model for 500 epochs with a batch size of 200, and then run SAGD for 100 epochs for fine-tuning. In SAGD, twenty independent chains are used to compute the approximate gradient, each with five burn-in’s.
Since SAGD basically refines the generative network of VAE, we can directly compare their output images. We randomly generate 100 digits from the trained VAE model and the debiased model, respectively, and in Figure \[fig:mnist\] we show some representative pairs of generated digits, with VAE-trained ones on the top, and SAGD-refined ones on the bottom. It is clear that the SAGD refining step improves the quality of the generated images. For example, in the first column of Figure \[fig:mnist\], VAE shows an ambiguous digit between “9” and “7”, but the refined one is a definite “7”.
Conclusion
==========
In this article we have developed the SAGD framework for optimizing objective functions that can be expressed as a mathematical expectation with intractable gradients. SAGD uses the Langevin algorithm to construct an approximate gradient in each iteration, whose accuracy is carefully controlled. Theoretical analysis shows that SAGD has the same convergence property as SGD, and more importantly, all the hyperparameters of SAGD are transparent so that the algorithm can be practically implemented. We have successfully applied SAGD to both the automated EM algorithm and the debiased VAE. To summarize, SAGD is an alternative to the ordinary SGD in a broader realm, and it is hoped that SAGD can be used to solve more statistical and machine learning problems both efficiently and reliably.
We mention two future directions for the research on SAGD. First, one might be interested in improving the Langevin algorithm, as the assumptions we have made are mild yet not the weakest. A second direction is to study the convergence of SAGD combined with various acceleration techniques, such as the momentum methods.
Appendix
========
Techincal Lemmas
----------------
In this section we provide a number of inequalities and lemmas that are useful for other theorems. First define the function $$\Gamma_{c}(x,y)=\frac{\gamma^{2}}{4}\Vert x\Vert^{2}+V(x)+\frac{\gamma}{2}\langle x,y\rangle+\frac{c}{2}\Vert y\Vert^{2}+1,$$ where $x,y\in\mathbb{R}^{r}$ and $c>1$. Under Assumption \[assu:v\_boundedness\](a), we immediately obtain $$\begin{aligned}
\Gamma_{c}(x,y) & \ge\frac{1}{6}\left\Vert \gamma x+\frac{3}{2}y\right\Vert ^{2}+\frac{\gamma^{2}}{12}\Vert x\Vert^{2}+\frac{1}{8}\Vert y\Vert^{2}+1\nonumber \\
& \ge\frac{\gamma^{2}}{12}\Vert x\Vert^{2}+\frac{1}{8}\Vert y\Vert^{2}+1,\label{eq:lemma_gamma_c_lower_bound}\end{aligned}$$ and $$\Gamma_{c}(x,y)\le\frac{3\gamma^{2}}{8}\Vert x\Vert^{2}+V(x)+\frac{c+1}{2}\Vert y\Vert^{2}+1.\label{eq:lemma_gamma_c_upper_bound}$$ Next, note that $V(x_{1})-V(x_{2})=\langle\nabla V(x_{2}),x_{1}-x_{2}\rangle+\int_{0}^{1}(1-s)(x_{1}-x_{2})^{\mathrm{T}}\nabla^{2}(x_{2}+s(x_{1}-x_{2}))(x_{1}-x_{2})\mathrm{d}s$. By Assumption \[assu:v\_boundedness\](b), we get $$\begin{aligned}
& \langle\nabla V(x_{2}),x_{1}-x_{2}\rangle-\frac{\nu}{2}\Vert x_{1}-x_{2}\Vert^{2}\le V(x_{1})-V(x_{2})\nonumber \\
\le & \langle\nabla V(x_{2}),x_{1}-x_{2}\rangle+\frac{\nu}{2}\Vert x_{1}-x_{2}\Vert^{2}.\label{eq:lemma_v_diff}\end{aligned}$$ Also due to $\Vert\nabla^{2}V(x)\Vert\le\nu$, we have $$\Vert\nabla V(x)\Vert\le\Vert\nabla V(0)\Vert+\Vert\nabla V(x)-\nabla V(0)\Vert\le\Vert\nabla V(0)\Vert+\nu\Vert x\Vert.\label{eq:lemma_gradv_bound}$$
Let $\tau_{k}=(1-\gamma\delta)\rho_{k}-\delta\cdot\nabla V(\xi_{k})$, and for $t\in[k\delta,(k+1)\delta]$ define $R_{k}(t)=\tau_{k}+\sqrt{2\gamma}(B(t)-B(k\delta))$. Clearly $\rho_{k+1}=\tau_{k}+\sqrt{2\gamma\delta}\eta_{k}=R_{k}((k+1)\delta)$. We then present the following two lemmas.
\[lem:gamma\_c\_contraction\]Let $D=\gamma^{4}C_{\beta}/\nu^{2}$, $\delta\le\min\{1/\gamma,\gamma/(2\nu),(D+1-\sqrt{D^{2}+1})/\gamma\}$, and $c=(1-\gamma\delta/2)/(1-\gamma\delta)$. Then there exist constants $\varepsilon=\varepsilon(\delta)>0$ and $C_{1}=C_{1}(\delta)>0$ such that for all $k\ge0$, $$\Gamma_{c}(\xi_{k+1},\tau_{k})\le(1-\varepsilon\gamma\delta)\Gamma_{c}(\xi_{k},\rho_{k})+C_{1}.$$
\[lem:gamma\_c\_l\]Fix any integer $l>0$. Then there exists a constant $C_{2}=C_{2}(l,\delta)>0$ such that for all $k\ge0$ and $t\in[k\delta,(k+1)\delta]$, $$\begin{aligned}
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},R_{k}(t))\right\} ^{l}\right] \le \\
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l}\right]+C_{2}\sum_{i=0}^{l-1}(t-k\delta)^{l-i}\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{i}\right].\end{aligned}$$
For the Langevin diffusion process $W(t)$, let $\mathcal{A}$ be its generator, defined by $$\begin{aligned}
(\mathcal{A}\phi)(w)= & \ \langle y,\nabla_{x}\phi(w)\rangle-\langle\nabla V(x)+\gamma y,\nabla_{y}\phi(w)\rangle \\
& +\gamma\mathrm{tr}(\nabla_{y}^{2}\phi(w)),\end{aligned}$$ where $\phi:\mathbb{R}^{2r}\to\mathbb{R}$ is any twice differentiable function, $w=(x^{\mathrm{T}},y^{\mathrm{T}})^{\mathrm{T}}$, $\nabla_{x}\phi=\partial\phi(w)/\partial x$, $\nabla_{y}\phi=\partial\phi(w)/\partial y$, and $\nabla_{y}^{2}\phi=\partial^{2}\phi(w)/\partial y\partial y^{\mathrm{T}}$. Consider the functional equation $\mathcal{A}\psi=\varphi-\bar{\varphi}$, which is called the Poisson equation, and we have the following lemma for the solution $\psi$.
\[lem:existence\_poisson\]Assume that $\varphi\in\mathscr{C}_{poly}^{r+5}$, and the conditions in Theorem \[thm:finite\_moments\] hold. Then $\nabla^{i}\psi$, $i=0,1,2,3,4$ exist, and $\nabla^{i}\psi$ have polynomial growth.
Proof of Lemma \[lem:gamma\_c\_contraction\]
--------------------------------------------
Without loss of generality consider $k=0$, and then we expand $\Gamma_{c}(\xi_{1},\tau_{0})$ using $\xi_{1}=\xi_{0}+\delta\rho_{0}$ and $\tau_{0}=(1-\gamma\delta)\rho_{0}-\delta\nabla V(\xi_{0})$. From and , we have $$\begin{aligned}
& V(\xi_{1})=V(\xi_{0}+\delta\rho_{0})\le V(\xi_{0})+\delta\langle\nabla V(\xi_{0}),\rho_{0}\rangle+\frac{\nu\delta^{2}}{2}\Vert\rho_{0}\Vert^{2},\\
& \Vert\nabla V(\xi_{0})\Vert^{2}\le2\Vert\nabla V(0)\Vert^{2}+2\nu^{2}\Vert\xi_{0}\Vert^{2}.\end{aligned}$$ Then with some tedious calculations, it can be shown that $$\begin{aligned}
& \Gamma_{c}(\xi_{1},\tau_{0}) \\
\le\ & \Gamma_{c}(\xi_{0},\rho_{0})+\left(\frac{(2c-1)\gamma\delta(\gamma\delta-2)}{4}+\frac{\nu\delta^{2}}{2}\right)\Vert\rho_{0}\Vert^{2} \\
& +\left(\delta-\frac{\gamma\delta^{2}}{2}-c(1-\gamma\delta)\delta\right)\langle\rho_{0},\nabla V(\xi_{0})\rangle \\
& -\frac{\gamma\delta}{2}\langle\xi_{0},\nabla V(\xi_{0})\rangle+c\delta^{2}\left(\Vert\nabla V(0)\Vert^{2}+\nu^{2}\Vert\xi_{0}\Vert^{2}\right)+1.\end{aligned}$$ Setting $\delta-\gamma\delta^{2}/2-c(1-\gamma\delta)\delta=0$ yields $c=(1-\gamma\delta/2)/(1-\gamma\delta)$. Let $\delta$ be sufficiently small such that $\gamma\delta<1$ and $\nu\delta^{2}<\gamma\delta/2$, *i.e.*, $\delta<\min\{1/\gamma,\gamma/(2\nu)\}$. Then $c>1$ and $$\begin{aligned}
& \frac{(2c-1)\gamma\delta(\gamma\delta-2)}{4}+\frac{\nu\delta^{2}}{2} \\
=\ &-\frac{(2c-1)\gamma\delta}{2}+\frac{(2c-1)\gamma^{2}\delta^{2}}{4}+\frac{\nu\delta^{2}}{2} \\
\le\ & -\frac{(c-1)\gamma\delta}{2}.\end{aligned}$$ Therefore, $\Gamma_{c}(\xi_{1},\tau_{0})\le\Gamma_{c}(\xi_{0},\rho_{0})-T+C$, where $$\begin{aligned}
T & =\frac{\gamma\delta}{2}\langle\xi_{0},\nabla V(\xi_{0})\rangle+\frac{(c-1)\gamma\delta}{2}\Vert\rho_{0}\Vert^{2}-c\delta^{2}\nu^{2}\Vert\xi_{0}\Vert^{2}, \\
C & =c\delta^{2}\Vert\nabla V(0)\Vert^{2}+1.\end{aligned}$$ Moreover, Assumption \[assu:v\_dissipative\] indicates that $\langle\nabla V(\xi_{0}),\xi_{0}\rangle/2\ge\beta V(\xi_{0})+\gamma^{2}C_{\beta}\Vert\xi_{0}\Vert^{2}-\alpha,$ so $$\begin{aligned}
& \frac{\gamma\delta}{2}\langle\xi_{0},\nabla V(\xi_{0})\rangle-c\delta^{2}\nu^{2}\Vert\xi_{0}\Vert^{2} \\
\ge\ & \gamma\delta\beta V(\xi_{0})+(\gamma^{3}\delta C_{\beta}-c\delta^{2}\nu^{2})\Vert\xi_{0}\Vert^{2}-\gamma\delta\alpha.\end{aligned}$$ Since $c$ is decreasing in $\delta$, we further decrease $\delta$ if necessary to guarantee that $\gamma^{3}\delta C_{\beta}-c\delta^{2}\nu^{2}>\gamma^{3}\delta C_{\beta}/2$, $i.e.$, $\delta<(D+1-\sqrt{D^{2}+1})/\gamma$, $D=\gamma^{4}C_{\beta}/\nu^{2}$. Then $$\begin{aligned}
T & \ge\gamma\delta\beta V(\xi_{0})+\frac{\gamma^{3}\delta C_{\beta}}{2}\Vert\xi_{0}\Vert^{2}+\frac{(c-1)\gamma\delta}{2}\Vert\rho_{0}\Vert^{2}-\gamma\delta\alpha\\
& =\gamma\delta\left\{ C_{\beta}\gamma^{2}\Vert\xi_{0}\Vert^{2}/2+\beta V(\xi_{0})+(c-1)\Vert\rho_{0}\Vert^{2}/2\right\} -\gamma\delta\alpha.\end{aligned}$$ Let $\varepsilon=\min\{4C_{\beta}/3,\beta,(c-1)/(c+1)\}$, and then by inequality , we have $T\ge\varepsilon\gamma\delta\Gamma_{c}(\xi_{0},\rho_{0})-(\alpha+\varepsilon)\gamma\delta.$ Consequently, let $C_{1}=C+(\alpha+\varepsilon)\gamma\delta$, and then the claimed result holds.
Proof of Lemma \[lem:gamma\_c\_l\]
----------------------------------
Due to the Markov property, we only need to show that for $t\in[0,\delta]$, $$\begin{aligned}
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{1},R_{0}(t))\right\} ^{l}\right]\nonumber \\
\le & \left\{ \Gamma_{c}(\xi_{1},\tau_{0})\right\} ^{l}+C_{2}\sum_{i=0}^{l-1}t^{l-i}\left\{ \Gamma_{c}(\xi_{1},\tau_{0})\right\} ^{i}.\label{eq:gamma_c_l_k0}\end{aligned}$$
First, for $l=1$, we have $B(t)-B(0)\overset{d}{=}\sqrt{t}\zeta$, where $\zeta\sim N(0,I_{r})$. Then $$\begin{aligned}
& \Gamma_{c}(\xi_{1},R_{0}(t))-\Gamma_{c}(\xi_{1},\tau_{0}) \\
=\ & \frac{\gamma}{2}\langle\xi_{1},\sqrt{2\gamma t}\zeta\rangle+\frac{c}{2}\Vert\tau_{0}+\sqrt{2\gamma t}\zeta\Vert^{2}-\frac{c}{2}\Vert\tau_{0}\Vert^{2}\\
=\ & \frac{\gamma}{2}\langle\xi_{1},\sqrt{2\gamma t}\zeta\rangle+c\sqrt{2\gamma t}\langle\tau_{0},\zeta\rangle+c\gamma t\Vert\zeta\Vert^{2}.\end{aligned}$$ Using the fact that $E(\zeta)=0$ and $E(\Vert\zeta\Vert^{2})=r$, we prove the case of $l=1$.
Then we prove the case of $l>1$ by induction. Assume that holds for all $j\le l-1$. Since $R_{0}(t)$ admits the SDE $\mathrm{d}R_{0}(t)=\sqrt{2\gamma}\mathrm{d}B(t)$, $R_{0}(0)=\tau_{0}$, applying Itô’s formula yields $$\begin{aligned}
\left\{ \Gamma_{c}(\xi_{1},R_{0}(t))\right\} ^{l}= & \left\{ \Gamma_{c}(\xi_{1},\tau_{0})\right\} ^{l}+\int_{0}^{t}\mathscr{L}G(s,R_{0}(s))\mathrm{d}s \\
& +\int_{0}^{t}\langle\nabla_{R_{0}(s)}G(s,R_{0}(s)),\sqrt{2\gamma}\mathrm{d}B(s)\rangle,\end{aligned}$$ where $G(\cdot,y)=\{\Gamma_{c}(\xi_{1},y)\}^{l}$, $\mathscr{L}G(\cdot,y)=\gamma\mathrm{tr}(\nabla_{y}^{2}G(\cdot,y))$, and $\nabla_{y}G(\cdot,y)=l\{\Gamma_{c}(\xi_{1},y)\}^{l-1}(\gamma\xi_{1}/2+cy)$. Let $v=\gamma\xi_{1}/2+cy$, $u=\Gamma_{c}(\xi_{1},y)$, and then it can be verified that $$\begin{aligned}
\nabla_{y}^{2}G(\cdot,y) & =clu^{l-1}I_{r}+l(l-1)u^{l-2}vv^{\mathrm{T}}, \\
\mathscr{L}G(\cdot,y) & =\gamma clru^{l-1}+\gamma l(l-1)u^{l-2}\Vert v \Vert^{2}.\end{aligned}$$ We also have $\Vert v \Vert^{2}\le 2c^2u$, because $$\begin{aligned}
& \left\Vert \frac{\gamma}{2}x+cy\right\Vert ^{2}=2c^{2}\left(\frac{\gamma^{2}}{8c^{2}}\Vert x\Vert^{2}+\frac{\gamma}{2c}\langle x,y\rangle+\frac{1}{2}\Vert y\Vert^{2}\right)\\
\le\ & 2c^{2}\left(\frac{\gamma^{2}}{4}\Vert x\Vert^{2}+\frac{\gamma}{2}\langle x,y\rangle+\frac{c}{2}\Vert y\Vert^{2}\right)\le2c^{2}\Gamma_{c}(x,y),\end{aligned}$$ Finally, $\mathbb{E}\{\int_{0}^{t}\langle\nabla_{R_{0}(s)}G(s,R_{0}(s)),\sqrt{2\gamma}\mathrm{d}B(s)\rangle\}=0$, so $$\begin{aligned}
& \mathbb{E}\left\{ \Gamma_{c}(\xi_{1},R_{0}(t))\right\} ^{l} \\
\le\ & \left\{ \Gamma_{c}(\xi_{1},\tau_{0})\right\} ^{l}+C\int_{0}^{t}\{\Gamma_{c}(\xi_{1},R_{0}(s))\}^{l-1}\mathrm{d}s,\end{aligned}$$ where $C=\gamma\left\{ clr+2l(l-1)c^2\right\}$. Using the induction hypothesis, we prove that also holds for $l$.
Proof of Lemma \[lem:existence\_poisson\]
-----------------------------------------
Define $u(t,x,y)=\mathbb{E}(\varphi(\xi(t),\rho(t))|\xi(0)=x,\rho(0)=y)$, and let the notation $\nabla^{i}u(t,x,y)$ denote the derivative of $u$ with respect to $w=(x^{\mathrm{T}},y^{\mathrm{T}})^{\mathrm{T}}$. Proposition 6.1 of @kopec2015weak shows that there exist constants $C>0$ and $\lambda>0$ and an integer $s>0$ such that $$\Vert\nabla^{i}u(t,x,y)\Vert\le C(1+\Vert x\Vert^{s}+\Vert y\Vert^{s})e^{-\lambda t},\quad i=0,1,2,3,4,$$ for all $t>0$. Moreover, it is known that $\psi$ has the representation $\psi(w)=\int_{0}^{+\infty}u(t,x,y)\mathrm{d}t$, provided that the integral exists. Indeed, since $$\begin{aligned}
& \int_{0}^{+\infty}\Vert\nabla^{i}u(t,x,y)\Vert\mathrm{d}t \\
\le & \int_{0}^{+\infty}C(1+\Vert x\Vert^{s}+\Vert y\Vert^{s})e^{-\lambda t}\mathrm{d}t=\frac{C}{\lambda}(1+\Vert x\Vert^{s}+\Vert y\Vert^{s}),\end{aligned}$$ we obtain the existence of $\psi$ for $i=0$, and by the dominated convergence theorem, we can interchange the integral and differential operators, which shows that $\nabla^{i}\psi$, $i=1,2,3,4$ exist and have polynomial growth.
Note that @kopec2015weak is based on a set of assumptions ***B***-1 to ***B***-4, but only ***B***-1, ***B***-2, and the condition $V(x)\ge0$ are used to prove Proposition 6.1. Those three conditions are implied by Assumptions \[assu:v\_boundedness\] and \[assu:v\_dissipative\] in this article.
Proof of Theorem \[thm:finite\_moments\]
----------------------------------------
In Lemma \[lem:gamma\_c\_l\], take $t=(k+1)\delta$, and then $$\begin{aligned}
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\rho_{k+1}\right\} ^{l}\right] \\
\le\ & \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l}\right]+C_{2}\sum_{i=0}^{l-1}\delta^{l-i}\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{i}\right].\end{aligned}$$ For $i=1,\ldots,l-2$, let $q_{i}=(l-1)/i>1$ and $p_{i}=q_{i}/(q_{i}-1)>1$, and then using Young’s inequality we have $$\delta^{l-i}\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{i}\le\frac{1}{p_i}\delta^{p_{i}(l-i)}+\frac{1}{q_i}\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l-1}.$$ As a result, there exist constants $\varepsilon_{3}=\varepsilon(l,\delta)>0$ and $C_{3}=C_{3}(l,\delta)>0$ such that $$\begin{aligned}
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\rho_{k+1})\right\} ^{l}\right] \\
\le\ & \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l}\right]+\varepsilon_{3}\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l-1}\right]+C_{3}.\end{aligned}$$ Moreover, for any $\varepsilon_{4}>0$, we can pick a constant $C_{4}=C_{4}(\varepsilon_{3},\varepsilon_{4},C_{3})>0$ to guarantee $\varepsilon_{3}x^{l-1}+C_{3}<\varepsilon_{4}x^{l}+C_{4}$ for all $x>0$. Using these constants, we have $$\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\rho_{k+1})\right\} ^{l}\right]\le(1+\varepsilon_{4})\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l}\right]+C_{4}.\label{eq:gamma_c_l_result1}$$ With a similar argument, Lemma \[lem:gamma\_c\_contraction\] indicates that for any $\varepsilon_{5}>0$, there is a constant $C_{5}=C_{5}(\varepsilon_{5},l,\delta)>0$ such that $$\left\{ \Gamma_{c}(\xi_{k+1},\tau_{k})\right\} ^{l}\le(1+\varepsilon_{5})(1-\varepsilon\gamma\delta)^{l}\left\{ \Gamma_{c}(\xi_{k},\rho_{k})\right\} ^{l}+C_{5}.\label{eq:gamma_c_l_result2}$$ Putting and together, we have $$\begin{aligned}
& \mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k+1},\rho_{k+1})\right\} ^{l}\right] \\
\le\ & (1+\varepsilon_{4})(1+\varepsilon_{5})(1-\varepsilon\gamma\delta)^{l}\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k},\rho_{k})\right\} ^{l}\right]+(1+\varepsilon_{4})C_{5}+C_{4}.\end{aligned}$$ Clearly, by choosing $\varepsilon_{4}$ and $\varepsilon_{5}$ such that $r_{l}=(1+\varepsilon_{4})(1+\varepsilon_{5})(1-\varepsilon\gamma\delta)^{l}<1$, we get $$\mathbb{E}\left[\left\{ \Gamma_{c}(\xi_{k},\rho_{k})\right\} ^{l}\right]\le r_{l}^{k}\left\{ \Gamma_{c}(\xi_{0},\rho_{0})\right\} ^{l}+\{(1+\varepsilon_{4})C_{5}+C_{4}\}\sum_{i=0}^{k-1}r_{l}^{i}.$$ Formula shows that $\left\{ \Gamma_{c}(\xi_{k},\rho_{k})\right\} ^{l}$ is lower bounded by by a polynomial of order $2l$, and Assumption \[assu:v\_boundedness\](c) and show that $\left\{ \Gamma_{c}(\xi_{0},\rho_{0})\right\} ^{l}$ is upper bounded by a polynomial. Then the desired result is proved.
Proof of Theorem \[thm:control\_bias\_mse\]
-------------------------------------------
The proof of this theorem is inspired by @mattingly2010convergence and @vollmer2016exploration. First let $U(W_{k})=(\rho_{k}^{\mathrm{T}},-\gamma\rho_{k}^{\mathrm{T}}-\nabla V(\xi_{k})^{\mathrm{T}})^{\mathrm{T}}$ and $\zeta_{k}=(\mathbf{0}^{\mathrm{T}},\eta_{k}^{\mathrm{T}})^{\mathrm{T}}$, and then we have the relation $W_{k+1}=W_{k}+\delta U(W_{k})+\sqrt{2\gamma\delta}\zeta_{k}$. For simplicity denote $\varphi_{k}=\varphi(W_{k})$, $u_{k}=U(W_{k})$, $\psi_{k}=\psi(W_{k})$, and $d_{k}=W_{k+1}-W_{k}$. We also use $\nabla^{i}\psi_{k}$ to denote the $i$-th derivative of $\psi$ at $W_{k}$, and $\nabla^{i}\psi(x)[d,\ldots,d]$ to denote the derivative $\nabla^{i}\psi(x)$ evaluated in the directions $(d,\ldots,d)$. In addition, $\nabla_{y}^{i}\psi_{k}$ stands for the partial derivative of $\psi$ at $W_{k}$ with respect to the second component. Since $\psi$ has a fourth-order derivative, the following Taylor expansion holds, $$\begin{aligned}
\psi_{k+1}= & \psi_{k}+\langle d_{k},\nabla\psi_{k}\rangle+\frac{1}{2}\langle d_{k},(\nabla^{2}\psi_{k})d_{k}\rangle\nonumber \\
& +\frac{1}{6}\nabla^{3}\psi_{k}[d_{k},d_{k},d_{k}]+R_{k+1},\label{eq:taylor_psi}\end{aligned}$$ where $$\begin{aligned}
& R_{k+1} \\
=\ & \frac{1}{6}\left(\int_{0}^{1}s^{3}\nabla^{4}\psi(sW_{k}+(1-s)W_{k+1})\mathrm{d}s\right)[d_{k},d_{k},d_{k},d_{k}]\end{aligned}$$ is the remainder term. By expanding $d_{k}=\delta u_{k}+\sqrt{2\gamma\delta}\cdot\zeta_{k}$ and using the definition of the generator $\mathcal{A}$, we can show that $$\begin{aligned}
\langle d_{k},\nabla\psi_{k}\rangle & =\delta\langle u_{k},\nabla\psi_{k}\rangle+\sqrt{2\gamma\delta}\cdot\langle\zeta_{k},\nabla\psi_{k}\rangle\\
& =\delta\mathcal{A}\psi_{k}-\gamma\delta\mathrm{tr}(\nabla_{y}^{2}\psi_{k})+\sqrt{2\gamma\delta}\cdot\langle\zeta_{k},\nabla\psi_{k}\rangle\\
& =\delta(\varphi_{k}-\bar{\varphi})-\gamma\delta\mathrm{tr}(\nabla_{y}^{2}\psi_{k})+\sqrt{2\gamma\delta}\cdot\langle\zeta_{k},\nabla\psi_{k}\rangle\end{aligned}$$ and $$\begin{aligned}
\frac{1}{2}\langle d_{k},(\nabla^{2}\psi_{k})d_{k}\rangle=\ & \frac{1}{2}\delta^{2}\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle \\
& +\sqrt{2\gamma\delta^{3}}\cdot\langle\zeta_{k},(\nabla^{2}\psi_{k})u_{k}\rangle \\
& +\gamma\delta\langle\xi_{k},(\nabla_{y}^{2}\psi_{k})\xi_{k}\rangle.\end{aligned}$$ Since $\xi_{k}\sim N(0,I_{r})$ and $\xi_{k}$ is independent of $W_{k}$, we have for all $k$, $\mathbb{E}\langle\zeta_{k},\nabla\psi_{k}\rangle=0$, $\mathbb{E}\langle\zeta_{k},(\nabla^{2}\psi_{k})u_{k}\rangle=0$, $\mathbb{E}\langle\xi_{k},(\nabla_{y}^{2}\psi_{k})\xi_{k}\rangle=\mathrm{tr}(\nabla_{y}^{2}\psi_{k})$ etc.. Therefore, taking the expectation on both sides of cancels many terms involving $\xi_{k}$. Let the notation $X\overset{E}{=}Y$ stand for $\mathbb{E}(X)=\mathbb{E}(Y)$, and then after some simplification, we get $$\begin{aligned}
\psi_{k+1}\overset{E}{=}\ & \psi_{k}+\delta(\varphi_{k}-\bar{\varphi})+\frac{\delta^{2}}{2}\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle\nonumber \\
& +\gamma\delta^{2}\nabla^{3}\psi_{k}[u_{k},e,e]+\frac{\delta^{3}}{6}\nabla^{3}\psi_{k}[u_{k},u_{k},u_{k}]\nonumber \\
& +R_{k+1},\label{eq:psi_expectation}\end{aligned}$$ where $e=(\mathbf{0}^{\mathrm{T}},\mathbf{1}^{\mathrm{T}})^{\mathrm{T}}$. Summing over the first $K$ terms, and dividing both sides by $K\delta$, we obtain $$\begin{aligned}
& \frac{1}{K\delta}(\psi_{K}-\psi_{0})\nonumber \\
\overset{E}{=}\ & \hat{\varphi}-\bar{\varphi}+\frac{\delta}{2K}\sum_{k=0}^{K-1}\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle+\frac{\gamma\delta}{K}\sum_{k=0}^{K-1}\nabla^{3}\psi_{k}[u_{k},e,e]\nonumber \\
& +\frac{\delta^{2}}{6K}\sum_{k=0}^{K-1}\nabla^{3}\psi_{k}[u_{k},u_{k},u_{k}]+\frac{1}{K\delta}\sum_{k=0}^{K-1}R_{k+1}.\label{eq:psi_average}\end{aligned}$$
Now we attempt to bound each term in . First, $|\psi_{K}-\psi_{0}|\le|\psi_{0}|+|\psi_{K}|$, so by Lemma \[lem:existence\_poisson\] we know that it is bounded by a polynomial of $W_{K}$. Then Theorem \[thm:finite\_moments\] indicates that the expectation $\mathbb{E}|\psi_{K}-\psi_{0}|$ is bounded by a constant, denoted by $A_{1}$. Next, $\vert\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle\vert\le\Vert\nabla^{2}\psi_{k}\Vert\cdot\Vert v_{k}\Vert^{2}$, which can be bounded by a product of polynomials of $W_{k}$. Using the same argument, we get $\mathbb{E}\vert\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle\vert\le A_{2}$ for some constant $A_{2}$. Similar analysis for higher order terms shows $\mathbb{E}\vert\nabla^{3}\psi_{k}[u_{k},e,e]\vert\le A_{3}$ and $\mathbb{E}\vert\nabla^{3}\psi_{k}[u_{k},u_{k},u_{k}]\vert\le A_{4}$. For the remainder term, since $d_{k}=\sqrt{\delta}(\sqrt{\delta}u_{k}+\sqrt{2\gamma}\zeta_{k})$, we have $\mathbb{E}\vert R_{k+1}\vert\le\delta^{2}A_{5}$ for some constant $A_{5}$. Combining all these terms together, we eventually get the first inequality.
For the second part, let $T=K\delta$, and then from we get $$\begin{aligned}
& \frac{1}{T}(\psi_{K}-\psi_{0}) \\
=\ & \hat{\varphi}-\bar{\varphi}+\frac{\gamma\delta}{T}\sum_{k=0}^{K-1}\left\{ \langle\xi_{k},(\nabla_{y}^{2}\psi_{k})\xi_{k}\rangle-\mathrm{tr}(\nabla_{y}^{2}\psi_{k})\right\} \\
& +\frac{\sqrt{2\gamma\delta}}{T}\sum_{k=0}^{K-1}\langle\zeta_{k},\nabla\psi_{k}\rangle\\
& +\frac{\sqrt{2\gamma\delta^{3}}}{T}\sum_{k=0}^{K-1}\langle\zeta_{k},(\nabla^{2}\psi_{k})u_{k}\rangle+\frac{\delta^{2}}{2T}\sum_{k=0}^{K-1}\langle u_{k},(\nabla^{2}\psi_{k})u_{k}\rangle\\
& +\frac{1}{6T}\sum_{k=0}^{K-1}\nabla^{3}\psi_{k}[d_{k},d_{k},d_{k}]+\frac{1}{T}\sum_{k=0}^{K-1}R_{k+1}\\
\coloneqq\ & \hat{\varphi}-\bar{\varphi}+\frac{\gamma\delta}{T}S_{1}+\frac{\sqrt{2\gamma\delta}}{T}S_{2}+\frac{\sqrt{2\gamma\delta^{3}}}{T}S_{3}+\frac{\delta^{2}}{2T}S_{4} \\
& +\frac{1}{6T}S_{5}+\frac{1}{T}S_{6}.\end{aligned}$$ Therefore, $$\begin{aligned}
\mathbb{E}(\hat{\varphi}-\bar{\varphi})^{2}\le\ & \frac{7}{T^{2}}\mathbb{E}\bigg\{ (\psi_{K}-\psi_{0})^{2}+\gamma^{2}\delta^{2}S_{1}^{2}+2\gamma\delta S_{2}^{2} \\
& \left . +2\gamma\delta^{3}S_{3}^{2}+\frac{\delta^{4}}{4}S_{4}^{2}+\frac{1}{36}S_{5}^{2}+S_{6}^{2}\right\} .\end{aligned}$$ Let $B_{0},\ldots,B_{6}$ represent some positive constants. Similar to the first part, we first show $\mathbb{E}(\psi_{K}-\psi_{0})^{2}\le B_{0}$. Then note that $S_{1}$ is a martingale, so $$\begin{aligned}
\mathbb{E}(S_{1}^{2}) & =\sum_{k=0}^{K-1}\mathbb{E}\left\{ \langle\xi_{k},(\nabla_{y}^{2}\psi_{k})\xi_{k}\rangle-\mathrm{tr}(\nabla_{y}^{2}\psi_{k})\right\} ^{2}\le B_{1}K.\end{aligned}$$ With analogous calculations, we can verify that $\mathbb{E}(S_{2}^{2})\le B_{2}K$, $\mathbb{E}(S_{3}^{2})\le B_{3}K$, $\mathbb{E}(S_{4}^{2})\le B_{4}K^{2}$, $\mathbb{E}(S_{5}^{2})\le B_{5}(\delta^{4}K^{2}+\delta^{3}K)$, and $\mathbb{E}(S_{6}^{2})\le B_{6}\gamma^{4}K^{2}$. Then adding up the terms gives the desired result.
Proof of Corollary \[cor:nn\]
-----------------------------
It is easy to show that $V(z)=-\log p(x,z)+C$, where $p(x,z)$ is the joint density of $(X,Z)$ and $C$ is free of $z$. For simplicity we let $\sigma=1$, since it only affects constant terms or scaling factors. Let $y=Wz+b=(y_{1},\ldots,y_{m})^{\mathrm{T}}$, and then $$V(z)=-\log p(x,z)+C=\frac{1}{2}\sum_{i=1}^{m}\{a(y_{i})-x_{i}\}^{2}+\frac{1}{2}\Vert z\Vert^{2}+C.$$ Since the activation function $a(x)=\log(1+e^{x})$ is smooth, Assumption 1 trivially holds.
Define $l_{1}(y)=\max(y,0)-1$ and $l_{2}(y)=\max(y,0)+1$, then clearly $l_{1}(y)<a(y)<l_{2}(y)$ for all $y\in\mathbb{R}$. Therefore, $\{a(y_{i})-x_{i}\}^{2}$ must be bounded above by a quadratic function of $y_{i}$. As $y$ is a linear transformation of $z$, it is also true that $V(z)$ is bounded above by a quadratic function of $z$: $V(z)\le C_{1}(1+\Vert z\Vert^{2})$ for some $C_{1}\ge0$.
On ther other hand, $\partial V/\partial y_{i}=\{a(y_{i})-x_{i}\}a'(y_{i})$, where $a'(y)=e^{y}/(1+e^{y})$. Let $u=(\partial V/\partial y_{1},\ldots,\partial V/\partial y_{m})^{\mathrm{T}}$, and then $\nabla V(z)=W^{\mathrm{T}}u+z$ and $$\langle\nabla V(z),z\rangle=u^{\mathrm{T}}Wz+z^{\mathrm{T}}z=u^{\mathrm{T}}(y-b)+\Vert z\Vert^{2}.$$ We can show that $0<a'(y)<1$, $l_{1}(y)<a(y)a'(y)<l_{2}(y)$, $l_{1}(y)<a'(y)y<l_{2}(y)$, and $a(y)a'(y)y>-1$. So $$\begin{aligned}
u_{i}(y_{i}-b_{i})=\ & a(y_{i})a'(y_{i})y_{i}-x_{i}a'(y_{i})y_{i} \\
& -b_{i}a(y_{i})a'(y_{i})-b_{i}x_{i}a'(y_{i})\end{aligned}$$ is bounded below by a piecewise linear function of $y$, which is also a piecewise linear function of $z$. Consequently, $\langle\nabla V(z),z\rangle$ is bounded below by a quadratic function of $z$: $\langle\nabla V(z),z\rangle\ge C_{2}\Vert z\Vert^{2}-C_{3}$ for some $C_{2},C_{3}\ge0$. Combining with the upper bound of $V(z)$, we show that Assumption 2 holds with a sufficiently small $\beta$ and a sufficiently large $\alpha$.
Proof of Theorem \[thm:convergence\_convex\]
--------------------------------------------
The Lipschitz continuity of $F$ implies that $\Vert g(\theta)\Vert=\Vert\nabla F(\theta)\Vert\le L$ for all $\theta\in\Theta$. Let $F^{*}=\min_{\theta\in\Theta}\,F(\theta)$, and $\theta^{*}\in\arg\min_{\theta\in\Theta}\,F(\theta)$. Then by the convexity of $F$ we have $F(\theta)-F^{*}\le\langle g(\theta),\theta-\theta^{*}\rangle$ for all $\theta\in\Theta$.
In what follows $\mathbb{E}_{t}[\cdot]$ denotes the expectation with respect to $\{\xi_{t,k}\}_{k=0}^{K_{t}-1}$, and $\mathbb{E}[\cdot]$ is the total expectation. When no confusion is caused, we write $g\equiv g(\theta_{t})$ and $\tilde{g}_{t}\equiv\tilde{g}_{t}(\theta_{t})$ for brevity. Suppose that $\Vert\mathbb{E}_{t}[\tilde{g}_{t}]-g\Vert\le\varepsilon_{t}$ and $\mathbb{E}_{t}\Vert\tilde{g}_{t}-g\Vert^{2}\le\omega_{t}$, and then we have $$\begin{aligned}
\quad\mathbb{E}_{t}\langle\tilde{g}_{t}-g,\theta_{t}-\theta^{*}\rangle & \ge_{(i)}-\Vert\mathbb{E}_{t}[\tilde{g}_{t}]-g\Vert\cdot\Vert\theta_{t}-\theta^{*}\Vert\\
& \ge-\varepsilon_{t}\Vert\theta_{t}-\theta^{*}\Vert\ge-\varepsilon_{t}D\end{aligned}$$ and $$\begin{aligned}
\quad\mathbb{E}_{t}\Vert\tilde{g}_{t}\Vert^{2} & =\mathbb{E}_{t}\left[\Vert\tilde{g}_{t}-g\Vert^{2}+2\langle\tilde{g}_{t}-g,g\rangle+\Vert g\Vert^{2}\right]\\
& \le_{(ii)}\omega_{t}+2\Vert\mathbb{E}_{t}[\tilde{g}_{t}]-g\Vert\cdot\Vert g\Vert+L^{2}\\
& \le\omega_{t}+2\varepsilon_{t}L+L^{2},\end{aligned}$$ where $(i)$ and $(ii)$ use the Cauchy–Schwarz inequality. Therefore, $$\begin{aligned}
\mathbb{E}_{t}\langle\tilde{g}_{t},\theta_{t}-\theta^{*}\rangle & =\mathbb{E}_{t}\langle\tilde{g}_{t}-g,\theta_{t}-\theta^{*}\rangle+\langle g,\theta_{t}-\theta^{*}\rangle\\
& \ge F(\theta_t)-F^{*}-\varepsilon_{t}D,\end{aligned}$$ and the update formula for $\theta_{t+1}$ indicates that $$\begin{aligned}
& \,\,\quad\mathbb{E}_{t}\Vert\theta_{t+1}-\theta^{*}\Vert^{2} \\
& \le_{(iii)}\mathbb{E}_{t}\Vert\theta_{t}-\alpha_{t}\cdot\tilde{g}_{t}-\theta^{*}\Vert^{2}\\
& \le\Vert\theta_{t}-\theta^{*}\Vert^{2}-2\alpha_{t}\mathbb{E}_{t}\langle\tilde{g}_{t},\theta_{t}-\theta^{*}\rangle+\alpha_{t}^{2}\mathbb{E}_{t}\Vert\tilde{g}_{t}\Vert^{2}\\
& \le\Vert\theta_{t}-\theta^{*}\Vert^{2}-2\alpha_{t}(F(\theta_{t})-F^{*})+2\alpha_{t}\varepsilon_{t}D \\
& \quad +\alpha_{t}^{2}(\omega_{t}+2\varepsilon_{t}L+L^{2}),\end{aligned}$$ where $(iii)$ comes from the nonexpansion property of the projection operator. Reorganizing the inequality above yields $$F(\theta_{t})-F^{*}\le\frac{\Vert\theta_{t}-\theta^{*}\Vert^{2}}{2\alpha_{t}}-\frac{\mathbb{E}_{t}\Vert\theta_{t+1}-\theta^{*}\Vert^{2}}{2\alpha_{t}}+\mu_{t},\label{eq:telescope}$$ where $\mu_{t}=\varepsilon_{t}D+\alpha_{t}(\omega_{t}+2\varepsilon_{t}L+L^{2})/2$. Summarizing over $t=1,2,\ldots,T$ and taking the total expectation, we obtain $$\begin{aligned}
& \mathbb{E}\sum_{t=1}^{T}\{F(\theta_{t})-F^{*}\} \\
\le\ & \frac{\mathbb{E}\Vert\theta_{1}-\theta^{*}\Vert^{2}}{2\alpha_{1}}+\sum_{t=2}^{T}\left(\frac{1}{2\alpha_{t}}-\frac{1}{2\alpha_{t-1}}\right)\mathbb{E}\Vert\theta_{t}-\theta^{*}\Vert^{2}+\sum_{t=1}^{T}\mu_{t}.\end{aligned}$$ Take $\alpha_{t}=\alpha_{0}/\sqrt{t}$, so $(2\alpha_{t})^{-1}-(2\alpha_{t-1})^{-1}>0$, and hence $$\begin{aligned}
\mathbb{E}\sum_{t=1}^{T}\{F(\theta_{t})-F^{*}\} & \le\frac{D^{2}}{2\alpha_{1}}+D^{2}\sum_{t=2}^{T}\left(\frac{1}{2\alpha_{t}}-\frac{1}{2\alpha_{t-1}}\right)+\sum_{t=1}^{T}\mu_{t}\\
& =\frac{D^{2}\sqrt{T}}{2\alpha_{0}}+\sum_{t=1}^{T}\mu_{t}.\end{aligned}$$ Choose $\delta_{t}=1/\sqrt{t}$ and $K_{t}=t$, and then $\varepsilon_{t}=2C_{1}/\sqrt{t}$ and $\omega_{t}=C_{2}(1/\sqrt{t}+1/t)$. Consequently, we see that $\mu_{t}=\mathcal{O}(1/\sqrt{t})$. Since $\sum_{t=1}^{T}1/\sqrt{t}\le2\sqrt{T}$, we conclude that $\mathbb{E}\sum_{t=1}^{T}\{F(\theta_{t})-F^{*}\}\le\mathcal{O}(\sqrt{T})$. Finally, by the convexity of $F$ we have $T^{-1}\sum_{t=1}^{T}F(\theta_{t})\ge F(\hat{\theta})$, and then the proof is complete.
Proof of Theorem \[thm:convergence\_nonconvex\]
-----------------------------------------------
Similar to the proof of Theorem \[thm:convergence\_convex\], denote $g\equiv g(\theta_{t})$ and $\tilde{g}_{t}\equiv\tilde{g}_{t}(\theta_{t})$. Suppose that $\Vert\mathbb{E}_{t}[\tilde{g}_{t}]-g\Vert\le\varepsilon_{t}$ and $\mathbb{E}_{t}\Vert\tilde{g}_{t}-g\Vert^{2}\le\omega_{t}$, and then we have $$\begin{aligned}
\mathbb{E}_{t}\langle\tilde{g}_{t},g\rangle & =\mathbb{E}_{t}\langle\tilde{g}_{t}-g,g\rangle+\Vert g\Vert^{2}\\
& \ge-\frac{1}{2}\left\{ \Vert\mathbb{E}_{t}[\tilde{g}_{t}]-g\Vert^{2}+\Vert g\Vert^{2}\right\} +\Vert g\Vert^{2}\\
& \ge\frac{1}{2}\Vert g\Vert^{2}-\frac{1}{2}\varepsilon_{t}^{2},\\
\mathbb{E}_{t}\Vert\tilde{g}_{t}\Vert^{2} & \le2\mathbb{E}_{t}\left[\Vert\tilde{g}_{t}-g\Vert^{2}+\Vert g\Vert^{2}\right]\le2\omega_{t}+2\Vert g\Vert^{2}.\end{aligned}$$ It is well known that if $g(\theta)\coloneqq\nabla F(\theta)$ is $G$-Lipschitz continuous, then for any $\theta'$ and $\theta$, $$F(\theta')-F(\theta)\le\langle g(\theta),\theta'-\theta\rangle+\frac{1}{2}G\Vert\theta'-\theta\Vert^{2}.$$ Therefore, $$\begin{aligned}
& \mathbb{E}_{t}[F(\theta_{t+1})]-F(\theta_{t})\nonumber \\
\le\ & \mathbb{E}_{t}\langle g,\theta_{t+1}-\theta_{t}\rangle+\frac{G}{2}\mathbb{E}_{t}\Vert\theta_{t+1}-\theta_{t}\Vert^{2}\nonumber \\
=\ & -\alpha_{t}\mathbb{E}_{t}\langle g,\tilde{g}\rangle+\frac{1}{2}\alpha_{t}^{2}G\mathbb{E}_{t}\Vert\tilde{g}_{t}\Vert^{2}\nonumber \\
\le\ & \frac{1}{2}\alpha_{t}\varepsilon_{t}^{2}-\frac{\alpha_{t}}{2}\Vert g\Vert^{2}+\alpha_{t}^{2}\omega_{t}G+\alpha_{t}^{2}G\Vert g\Vert^{2}\nonumber \\
=\ & -\frac{1}{2}\alpha_{t}(1-2\alpha_{t}G)\Vert g\Vert^{2}+\frac{1}{2}\alpha_{t}\varepsilon_{t}^{2}+\alpha_{t}^{2}\omega_{t}G.\label{eq:f_decrease}\end{aligned}$$ Let $\beta_{t}=\alpha_{t}(1-2\alpha_{t}G)$, and then take the total expectation on both sides of , yielding $$\mathbb{E}[F(\theta_{t+1})]-\mathbb{E}[F(\theta_{t})] \le-\frac{1}{2}\beta_{t}\mathbb{E}[\Vert g\Vert^{2}]+\frac{1}{2}\alpha_{t}\varepsilon_{t}^{2}+\alpha_{t}^{2}\omega_{t}G$$ and $$\begin{aligned}
\sum_{t=1}^{T}\beta_{t}\mathbb{E}[\Vert g\Vert^{2}] & \le2\mathbb{E}[F(\theta_{1})-F(\theta_{T+1})]+\alpha_{t}\varepsilon_{t}^{2}+2\alpha_{t}^{2}\omega_{t}G\\
& \le2\mathbb{E}[F(\theta_{1})]-2F^{*}+\alpha_{t}\varepsilon_{t}^{2}+2\alpha_{t}^{2}\omega_{t}G,\end{aligned}$$ where $F^{*}$ is the optimal value. The choice of $\{\alpha_{t}\}$, $\{\varepsilon_{t}\}$, and $\{\omega_{t}\}$ in the theorem guarantees that $\sum_{t=1}^{\infty}\beta_{t}=\infty$ and $\sum_{t=1}^{\infty}\beta_{t}\mathbb{E}[\Vert g\Vert^{2}]<\infty$, so the conclusion holds.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Motion of a single electron in a disordered alloy and or interacting electrons systems such as magnetic materials, strongly correlated systems and superconductors is replaced by motion of that in an effective medium which is denoted by self-energy. The study of disordered alloy and interacting electrons systems based on single electron motion is an old challenge and an important problem in condensed matter physics. In this paper we introduce a real space approximation beyond super cell approximation for the study of these systems to capture multi-site effects. Average disordered alloy or interacting system is replaced by a self-energy, $\Sigma(i,j,E)$. We divided self-energy in q-space $\Sigma({\bf q}; E)=\frac{1}{N}\sum_{ij}e^{i{\bf q}.{\bf r}_{ij}}\Sigma(i,j; E)$ into two parts $\Sigma({\bf q}; E)=\frac{1}{N_{c}}\sum_{IJ\in\; \mbox{\tiny same cluster}}e^{i{\bf q}.{\bf r}_{IJ}}\Sigma(I,J; E)+\frac{1}{N}\sum_{ij\notin \:\mbox{\tiny same cluster}}e^{i{\bf q}.{\bf r}_{IJ}}\Sigma(I,J,E)$ where $\{Lc_{1}, Lc_{2},Lc_{3}\}$ are dimensions of the super cell. We show that neglecting the second term of q-space self-energy leads to super cell approximation $e^{iq_{j} Lc_{j}}=1$, hence $ q_{j}$ determined by $ q_{j} Lc_{j}=2\pi n_{j}$. Then we kept this correction in the second step to add self energies of sites in different super cells which leads to fully q-dependent self energy in the first Brillouin zone (FBZ). Our self-energy in FBZ is casual, fully q-dependent, and continuous with respect to ${\bf q}$. It recovers coherent potential approximation in the single site approximation and is exact when the number of sites in the super cell approaches to the total number of lattice sites. We illustrate that this approximation undertakes electrons localization for one and two dimensional alloy systems which isn’t observed by previous multi site approximations.'
author:
- 'Rostam Moradian$^{1,2}$, Sina Moradian$^3$'
title: 'Beyond real space super cell approximation, corrections to the real space cluster approximation . '
---
2.5cm
Introduction
============
The treatment of disordered and interacting electron systems based on single electron motion in an effective medium is an important problem in many fields such as alloys, strongly correlated systems, magnetism and superconductivity in condensed matter physics. Coherent Potential Approximation (CPA) and Dynamical Mean Field Theory (DMFT) are single site approximations for calculating effective medium denoted by self-energy. In these approximations muti-site effects is neglected. Metzner and Vollhardt[@Metzner] and Muler-Hartmann[@Muller-Hartmann] found that in the limit of infinite dimensions both single site approximations coherent potential approximation (CPA) [@Soven67] for disordered system and dynamical mean field theory for interacting systems are exact[@Metzner; @Muller-Hartmann]. This means self-energy for systems with high dimensions is k-independent. However, outside of systems with infinite dimensions especially in one and two dimensional systems self-energy is far from local which means it is k-dependent. To treat effective features of disorder systems, k-dependent relation of self-energy $\Sigma({\bf k}; E)$ must be identified. In lower approximations such as Born approximation, T-matrix approximation, and Coherent Potential Approximation (CPA) [@Soven67], which are single site approximations, self-energy is k-independent $\Sigma({\bf k}; E)=\Sigma( E)$. In these approximations multi-site scattering is neglected which leads to overestimation band splitting and also losing short range effects. To add these effects cluster CPA with k-independent self-energy is used [@Gonis]. Dynamical Cluster Approximation (DCA)[@Hettler98; @Jarrell01; @Jarrell01-2] for systems with weak k-dependent self energies by considering periodic boundary condition for both interacting electron and disorder systems introduced. Also cellular dynamical mean field theory (CDMFT) approximation with open boundary condition[@Kotliar] was introduced. In DCA, first Brillouin zone (FBZ) is divided to $N_{c}$ grain regions where self-energy inside of these grains is k-independent although they could be different. So their self-energy in the FBZ is not continues. The wave vectors at center of these grains called cluster wave vectors and are denoted by $\{{\bf K}_{1}, ...,{\bf K}_{N_{c}}\}$. For disordered systems they claimed these cluster wave vectors, $\{{\bf K}_{n}\}$, corresponds to a $N_{c}$ real cluster sites [@Hettler98]. Although both cluster approximations DCA and CDMFT are successful in importing multi site effects but these methods have two major weakness, first their grain self energies are discontinuous, second at low dimension self energy is strongly k-dependent. In real space effective medium super-cell approximation (EMSCA)[@Moradian02; @Moradian04] are used to approximate self-energy of interacting disordered systems. We show that DCA could be a super cell approximation. Here in real space we first introduce super cell approximation by neglecting k-space self-energy contribution of all sites in different super cells. Then by keeping this contribution we go beyond super cell approximation. In our formalism, self energy is k-dependent and continuously varying in FBZ. Our self-energy is more close to real self-energy. Hence the average Green function calculated with our method used for calculation of physical quantities is more close to real average Green function.
The organization of the paper is as follows. In Sec. II the model Hamiltonian and super cell approximation equations are presented. Beyond super cell approximation equations derived and applied to a two dimensional alloy system in Sec. III. In this section we calculated and compared density of states for single site, super cell and beyond super cell approximations. Also electron localization in these approximations are discussed.
Model Hamiltonian and self-energy in the super cell approximation
=================================================================
The starting point is a tight-binding model for a disorder alloy system which is given by, $$\begin{aligned}
H&=&-\sum_{ij\sigma\sigma}t_{ij}c^{\dagger}_{i\sigma}c_{j\sigma}
\nonumber\\&+&\sum_{i\sigma} (\varepsilon_{i}-\mu)c^{\dagger}_{i\sigma}c_{i\sigma},
\label{eq:Hamiltonian}\end{aligned}$$ where $c^{\dagger}_{i\sigma}$ ($c_{i\sigma}$) is the creation (annihilation) operator of an electron with spin $\sigma$ on lattice site $i$ and $\hat{n}_{i\sigma}=c^{\dagger}_{i\sigma}c_{i\sigma}$ is the number operator. $t^{\sigma\sigma}_{ij}$ are the hopping integrals between $i$ and $j$ lattice sites with spin $\sigma$ respectively. $\varepsilon_{i}$ is the random on-site energy and takes $-\delta$ with probability $1-c$ for the host sites and $\delta$ with probability $c$ for impurity sites and $\mu$ is the chemical potential.
The electron equation of motion for Hamiltonian, Eq.\[eq:Hamiltonian\], is given by, $$\begin{aligned}
\sum_{l} \left(
\begin{array}{c}
(E-\varepsilon_{i}+\mu)\delta_{il}-t_{il}\end{array}\right){ G}(l,j; E)=\delta_{ij}
\label{eq:equation of motion}\end{aligned}$$ where $G(i,j)$ is the random single particle Green function. Relation between electron’s random Green function matrix, ${\bf G}$, and average Green function matrix, ${\bf \bar G}$, is given by $${\bf G}=\bf{\bar{G}}+\bf{\bar{G}}({\mbox{\boldmath$\varepsilon$}}-{\mbox{\boldmath$\Sigma$}}){\bf G}.
\label{eq:random average dyson equation}$$
Note that although \[eq:random average dyson equation\] is exact, due to randomness no exact solutions exists. For calculation of self-energy, different single site approximations such as coherent potential approximation (CPA), T-matrix, and Born approximation are introduced. Although attempts have been made with DCA to include multi-site scattering, its coarse grained self energies are discontinuous therefore in k-space these attempts have been unsuccessful. A real space multi site approximation which preserves continuity of k-space dependence of self-energy in the FBZ is not introduced. Here we implement a real space cluster approximation beyond super cell approximation in which not only includes multi site scattering but in the FBZ k-space dependence of self-energy varying continuously. Consider a lattice with dimensions $\{{\bf L}_{1}=N_{1}{\bf a}_{1}, {\bf L}_{2}=N_{2}{\bf a}_{2}, {\bf L}_{3}=N_{3}{\bf a}_{3}$ and sites number $N=N_{1}N_{2}N_{3}$ where ${\bf a}_{j}$ are lattice primitive vectors. Divide this lattice to super cells with dimensions $\{{\bf Lc}_{1}=N_{c1}{\bf a}_{1},{\bf Lc}_{2}=N_{c2}{\bf a}_{2},{\bf Lc}_{3}=N_{c3}{\bf a}_{3}\}$, original lattice symmetries and super cell lattice sites number $N_{c}=N_{c1}N_{c2}N_{c3}$. Position of $N_{c}$ sites in side of each cell denoted by capital letters $\{I\}$. Number of super cells is $\frac{N}{N_{c}}$. Since for alloy system at the band splitting regime for $c=0.5$ and average band filling ${\bar n}=1$ all sites with onsite energy $\delta$ are empty with sites with onsite energy $-\delta$ are filled by two electrons, just super cell with even sites number are acceptable. Fig.\[figure:real-space-nlcpa\] shows this for a two dimensional square lattice with $N_{c}=16$. Note that impurity configuration of super cells are not same.
![(Color online) Show a two dimensional square lattice which is divided to similar super cells of $N_{c}=16$ with original lattice symmetry. The super cell vectors are ${\bf Lc}_{1}=4 a{\bf e}_{x}$ and ${\bf Lc}_{2}=4 a {\bf e}_{y}$ where $\{{\bf a}_{1}=a{\bf e}_{x}, {\bf a}_{2}=a{\bf e}_{y}\}$ and [*a*]{} is lattice constant. []{data-label="figure:real-space-nlcpa"}](real-space-nlcpa-1){width="6cm"}
Since real space self energies only depend on difference of two lattice sites positions $\Sigma(E;i,j)=\Sigma(i-j; E)$, self energies divided to two categories, first self energies between intra sites of each super cell, $\Sigma(I-J; E)$, second self energies of one site inside of a super cell but another site belongs to another super cell $\Sigma(I-j; E)$ in which $$\begin{aligned}
{\bf r}_{j}=l{\bf Lc}_{1}+m{\bf Lc}_{2}+n{\bf Lc}_{3} +{\bf r}_{J} .
\label{eq:vector}\end{aligned}$$ where $l=\{0,1,...,\frac{N_{1}}{N_{c1}}-1\}$, $m=\{0,1,...,\frac{N_{2}}{N_{c2}}-1\}$, $n=\{0,1,...,\frac{N_{3}}{N_{c3}}-1\}$ are integer numbers. The exact q-space self-energy is $$\begin{aligned}
&\Sigma&({\bf q}; E)=\frac{1}{N_{c}}\sum_{IJ} \Sigma(I, J; E)e^{i{\bf q}.{\bf r}_{IJ}}+\sum_{\bf q'} \Sigma({\bf q'}; E)\frac{1}{N_{c}N}\nonumber\\&\times&\sum_{IJ} e^{i({\bf q}-{\bf q'}).{\bf r}_{IJ}}\Pi^{3}_{j=1} \left(\frac{1-e^{-iN_{j}a_{j}(q_{j}-q'_{j})}}{1-e^{-iN_{cj}a_{j}(q_{j}-q'_{j})}}-1\right).
\label{eq:self energy-k2-4-0}\end{aligned}$$ Our first approximation for exact q-space self-energy Eq.\[eq:self energy-k2-4-0\] is that, contribution of summation over all two lattice sites in different super cells become zero, $$\begin{aligned}
\left(\frac{1-e^{-iN_{j}a_{j}(q_{j}-q'_{j})}}{1-e^{-iN_{cj}a_{j}(q_{j}-q'_{j})}}-1\right)=0
\label{eq:super-cell-self energy-k2-4-1}\end{aligned}$$ The Born von Karman periodic boundary condition[@Ashcroft87] imply that $e^{-iN_{j}a_{j}(q_{j}-q'_{j})}=1$, hence in Eq.\[eq:super-cell-self energy-k2-4-1\] we have $$\begin{aligned}
e^{-iN_{cj}a_{j}(q_{j}-q'_{j})}=1.
\label{eq:super-cell-self energy-k-condi}\end{aligned}$$ From Eq.\[eq:super-cell-self energy-k-condi\] we have $$\begin{aligned}
N_{cj}{\bf a}_{j}.{\bf q}_{j}=2\pi n_{j},\;\;j=1,2,3 .
\label{eq:math3}\end{aligned}$$
where $\{n_{j}\}$ are integer numbers such that one of ${\bf q}$ must be center of FBZ. Wave vectors ${\bf q}={\bf K}_{n}$ that satisfy Eq.\[eq:math3\] are $$\begin{aligned}
{\bf K}_{n}={\bf K}_{m_{1}m_{2}m_{3}}=\sum^{3}_{i=1}\frac{m_{i}}{N_{ci}} {\bf b}_{i} .
\label{eq:math4}\end{aligned}$$ where $\{{\bf b}_{1}, {\bf b}_{2}, {\bf b}_{3}\}$ are reciprocal lattice primitive vectors, $\{m_{1},\;m_{2},\;m_{3}\}$ are integer such that ${\bf K}_{n}$ remains in the FBZ. By substitution Eqs.\[eq:math4\] and \[eq:super-cell-self energy-k-condi\] in to Eq.\[eq:self energy-k2-4-0\] we obtain $$\begin{aligned}
&\Sigma&({\bf K}_{n'}; E)=\frac{1}{N_{c}}\sum_{IJ} \Sigma_{sc}(I, J; E)e^{i{\bf K}_{n'}.{\bf r}_{IJ}}
\label{eq:self energy-K_n'}\end{aligned}$$ By times both sides of Eq.\[eq:self energy-K\_n’\] by $e^{-i{\bf K}_{n'}.{\bf r}_{I'J'}}$ and summation over ${\bf K}_{n'}$ and using this fact that $$\begin{aligned}
\frac{1}{N_{c}}\sum_{J} e^{i({\bf K}_{n}-{\bf K}_{n'}).{\bf r}_{J}}=\delta_{{\bf K}_{n}{\bf K}_{n'}},\;\; \frac{1}{N_{c}}\sum_{{\bf K}_{n}} e^{i{\bf K}_{n}.{\bf r}_{IJ}}=\delta_{IJ}\nonumber\\
\label{eq:self energy-K_n'f}\end{aligned}$$ we have $$\begin{aligned}
\Sigma_{sc}(I, J; E)=\frac{1}{N_{c}}\sum_{{\bf K}_{n'}} \Sigma({\bf K}_{n'}; E)e^{-i{\bf K}_{n'}.{\bf r}_{IJ}}.
\label{eq:self energy-I-J}\end{aligned}$$ Eqs.\[eq:super-cell-self energy-k-condi\] and \[eq:math4\] imply that in this approximation $e^{-ilN_{c1}{\bf a}_{1}.{\bf K}_{n'}}=1$, $e^{-imN_{c2}{\bf a}_{2}.{\bf K}_{n,}}=1$, $e^{-inN_{c3}{\bf a}_{3}.{\bf K}_{n'}}=1$. By use of Eq.\[eq:self energy-I-J\] and considering Eq.\[eq:vector\], real space self energies of two sites $I$ and $j$ in different super cells are periodic with respect to super cells center vector position $$\begin{aligned}
\Sigma_{sc}(I,j; E)=\Sigma_{sc}(I,J; E).
\label{eq:self energy approximation1}\end{aligned}$$ Therefore in lattice sites, space self-energy matrix is constructed from just super cell self-energy matrices. This illustrated in Fig.\[figure:impurity-super-cell\].
![(color online) (a ) shows an $N_{c}=4$ impurity super cell in an effective super cell self-energy medium in the lattice sites space matrix. (b) a $N_{c}=4$ cavity super cell embedded in super cell self energies in real space lattice sites matrix. []{data-label="figure:impurity-super-cell"}](impurity-super-cell){width="7cm"}
Note that for $N_{c}=1$, ${\bf K}_{n}=0$ hence converts to CPA self-energy which is k-independent. For $lim_{N_{c}\rightarrow N;\;I,J\rightarrow i,j}\;\frac{1}{N_{c}} \sum_{J}e^{i{\bf K}_{n}.{\bf r}_{IJ}} \Sigma(I,J;E)= \frac{1}{N} \sum_{ij} e^{i{\bf k}.{\bf r}_{ij}} \Sigma(i,j;E)= \Sigma({\bf k};E)$ it is exact k-space self-energy.
By taking impurity average over all random lattice sites except central super cell sites, Eq.\[eq:random average dyson equation\] reduces to a $N_{c}\times N_{c}$ matrix of super cell impurity embedded is an effective medium of super cell self energies $${\bf G}^{im}_{N_{c}\times N_{c}}={\bf\bar{G}}_{N_{c}\times N_{c}}+{\bf\bar{G}}_{N_{c}\times N_{c}}({\mbox{\boldmath$\varepsilon$}}-{\mbox{\boldmath$\Sigma$}}_{sc})_{N_{c}\times N_{c}} {\bf G}^{im}_{N_{c}\times N_{c}}.
\label{eq:imp-random average dyson equation}$$ as illustrated in Fig.\[figure:impurity-super-cell\] (a). Eq.\[eq:imp-random average dyson equation\] can be written as $${\bar{\bf G}}^{-1}_{sc}-{\mbox{\boldmath$\Sigma$}}_{sc}= {{\bf G}^{im}_{sc} }^{-1}-{\mbox{\boldmath$\varepsilon$}}={\mathcal G} .
\label{eq:cavity-imp-random average dyson equation}$$ where ${\mathcal G}$ is called cavity super cell Green function as shown in \[figure:impurity-super-cell\](b). Eq.\[eq:cavity-imp-random average dyson equation\] separates to two following super cell Dysons like equations $$\begin{aligned}
G^{im}_{sc} (I,J; E)={\mathcal G}(I,J; E)+ \sum_{L}{\mathcal G}(I,L; E) \varepsilon_{L}G^{im}_{sc} (L,J; E).\nonumber\\
\label{eq:math5-1}\end{aligned}$$ and $$\begin{aligned}
&{\bar G}_{sc} &(I,J; E)={\mathcal G}(I,J; E)+\nonumber\\ &\;&\sum_{LL'}{\mathcal G}(I,L; E)\Sigma_{sc} (L,L'; E){\bar G}(L',J; E).
\label{eq:math5-2}\end{aligned}$$ The Fourier transform of real space super cell average Green function and cavity Green function to super cell wave vectors $\{{\bf K}_{n}\}$ and vice versa are $$\begin{aligned}
{\bar G}({\bf K}_{n}; E)&=& \frac{1}{N_{c}}\sum_{IJ}{\bar G}_{sc} (I,J; E)e^{-i{\bf K}_{n}.{\bf r}_{IJ}},\nonumber\\ {\bar G}_{sc} (I,J;E)&=& \frac{1}{N_{c}}\sum_{{\bf K}_{n}}{\bar G}({\bf K}_{n}; E)e^{i{\bf K}_{n}.{\bf r}_{IJ}}
\label{eq:math5-3}\end{aligned}$$ $$\begin{aligned}
{\mathcal G}({\bf K}_{n}; E)&=& \frac{1}{N_{c}}\sum_{IJ}{\mathcal G}(I,J; E)e^{-i{\bf K}_{n}.{\bf r}_{IJ}},\nonumber\\{\mathcal G}(I,J; E)&=& \frac{1}{N_{c}}\sum_{{\bf K}_{n}}{\mathcal G}({\bf K}_{n}; E)e^{i{\bf K}_{n}.{\bf r}_{IJ}}.
\label{eq:math5-4}\end{aligned}$$ Substituting Eqs.\[eq:self energy-I-J\] , \[eq:math5-3\] and \[eq:math5-4\] in Eq.\[eq:math5-2\] and using Eq.\[eq:self energy-K\_n’f\] we have $$\begin{aligned}
{\bar G}({\bf K}_{n}; E)={\mathcal G}({\bf K}_{n}; E)+ {\mathcal G}({\bf K}_{n}; E) \Sigma({\bf K}_{n};E){\bar G}({\bf K}_{n}; E).
\label{eq:math5-7}\end{aligned}$$
To calculate $\Sigma({\bf K}_{n}; E)$ the FBZ is divided into $N_{c}$ regions with FBZ symmetries and $\frac{N}{N_{c}}$ wave vectors where each of ${\bf K}_{n}$ are in the center of one of these grains .Inside each grain self-energy is k-independent therefore, it is grain CPA self-energy. At $lim_{N_{c}\rightarrow N}$ number of wave vectors $\{\bf k\}$ in each grain reduces to just one (${\bf K}_{n}={\bf k}$). The $n$th grain CPA average Green function is defined by $$\begin{aligned}
\bar{G}({\bf K}_{n}; E)=\frac{N_{c}}{N}\sum_{{\bf k}\in nth\; grain}\frac{1}{E-\epsilon_{\bf k}+\mu-\Sigma({\bf K}_{n}; E)}.
\label{eq:nth grain green}\end{aligned}$$ and its real space Fourier transform is $$\begin{aligned}
\bar{G}(I,J; E)=\frac{1}{N_{c}}\sum_{{\bf K}_{n}}\bar{G}({\bf K}_{n}; E)e^{-i{\bf K}_{n}.{\bf r}_{IJ}} .
\label{eq:real space nth grain green}\end{aligned}$$
Note that DCA[@Jarrell01-2] could be real space super cell approximation.
beyond super cell approximation
===============================
To go beyond super cell approximation and add self energies contribution of $i$ and $j$ which are not in the same super cell we use super cell approximation $\Sigma(I,J; E)\approx\Sigma_{sc}(I,J; E)=\frac{1}{N_{c}}\sum_{{\bf K}_{n}}\Sigma({\bf K}_{n}; E)e^{i{\bf K}_{n}.{\bf r}_{IJ}}
$ hence $$\begin{aligned}
\sum_{IJ} \Sigma(I,J; E)e^{i{\bf k}.{\bf r}_{IJ}}\approx \frac{1}{N_{c}}\sum_{IJ}\sum_{{\bf K}_{n}}\Sigma({\bf K}_{n}; E) e^{i({\bf k}-{\bf K}_{n}).{\bf r}_{IJ}}.\nonumber\\
\label{eq:byond-super-cell}\end{aligned}$$ Note that beyond super cell approximation where $1<N_{c}<N$, for $q_{j}\neq K_{nj}$ we have $1-e^{-iN_{cj}a_{j}(q_{j}-q'_{j})}\neq 0$. By inserting Eq.\[eq:byond-super-cell\] in to Eq.\[eq:self energy-k2-4-0\] we have $$\begin{aligned}
\Sigma({\bf k}; E)&=&\frac{1}{N^{2}_{c}}\sum_{IJ}\sum_{{\bf K}_{n}}\Sigma({\bf K}_{n},E) e^{i({\bf k}-{\bf K}_{n}).{\bf r}_{IJ}}-\nonumber\\&\;&\sum_{\bf q'} \Sigma({\bf q'}; E)\frac{1}{N_{c}N}\sum_{IJ} e^{i({\bf k}-{\bf q'}).{\bf r}_{IJ}}
\label{eq:byond-super-cell-self energy-k2}\end{aligned}$$ Eq.\[eq:byond-super-cell-self energy-k2\] is centerpiece of our approximation. By iteration, Eq.\[eq:byond-super-cell-self energy-k2\] up to first order reduces to $$\begin{aligned}
\Sigma({\bf k}; E)=\frac{1}{N^{2}_{c}}\sum_{IJ}\sum_{{\bf K}_{n}}\Sigma({\bf K}_{n}; E) e^{i({\bf k}-{\bf K}_{n}).{\bf r}_{IJ}}(1-\frac{1}{N_{c}})\nonumber\\
\label{eq:byond-super-cell-self energy-k3}\end{aligned}$$ where $1<N_{c}<N$. For calculation of self energy $\Sigma({\bf k}; E)$ in Eq.\[eq:byond-super-cell-self energy-k3\] first we calculate $\Sigma({\bf K}_{n}; E)$. Algorithm for super cell calculation of average Green function is as follows\
1- A guess is made for real space and K-space self energies,$\Sigma(I,J; E)$, and $\Sigma({\bf K}_{ n})$. The starting values are usually zero.\
2- By inserting $\Sigma_{sc}({\bf K}_{ n}; E)$ in Eq.\[eq:nth grain green\] ,$\bar{G}({\bf K}_{ n};E)=\frac{N_{c}}{N}\sum_{{\bf k}\in\;nth\;grain}(G^{-1}_{0}({\bf k}; E)-\Sigma({\bf K}_{ n};E))^{-1}$, calculate the grain average k-space Green functions, $\bar{G}({\bf K}_{n}; E)$ .\
3-From Eq.\[eq:math5-7\] calculate K-space cavity Green function $\mathcal{G}({\bf K}_{ n}; E)=({\bar G}^{-1}({\bf K}_{ n};E)+\Sigma({\bf K}_{n}; E))^{-1}$.\
4- Obtain real space cavity Green function ${\mathcal G}(I,J; E)=\frac{1}{N_{c}}\sum_{{\bf K}_{n}}e^{i{\bf K}_{n}.{\bf r}_{IJ}}\mathcal{G}({\bf K}_{ n}; E)$ by Fourier transform of k-space $\mathcal{G}({\bf K}_{ n}; E)$.\
5- Calculate real space super cell impurity Green function matrix $G^{imp}=({\mathcal G}^{-1}-\mathbf{\varepsilon})^{-1}$ .\
6- Calculate super cell impurity average Green function matrix $\bar{ G}(I,J; E)=<({\mathcal G}^{-1}-\mathbf{\varepsilon})^{-1}>_{IJ}$ by taking average over all possible impurity configurations.\
7-Calculate real space new super cell self-energy matrix from ${\Sigma}_{sc}={\mathcal G}^{-1}-{\bf \bar G}^{-1}$.\
8-Inverse Fourier transform of new average super cell self-energy to calculate ${\Sigma}({\bf K}_{n}; E)=\frac{1}{N_{c}}\sum_{IJ}e^{-i{\bf K}_{n}.{\bf r}_{IJ}}{\Sigma}_{sc}(I,J; E)$.\
9- Return to 2 and repeat until convergence.\
10- Calculate self-energy beyond super cell approximation by substitution ${\Sigma}({\bf k}; E)$ in Eq.\[eq:byond-super-cell-self energy-k3\].\
11-Calculate average green function from ${\bar G}({\bf k};E)=G^{-1}_{0}({\bf k}; E)-\Sigma({\bf k}; E)$.\
Now we apply this method to a two dimensional square alloy system in which $\delta=3t$, $c=0.5$ and $\mu=0$. For this system we calculate, self-energy and density of states in the super cell and beyond super cell approximations and compared them. Fig.\[figure:Re-Im-selfenergy-nc4-d3t-n1-mu0\] (a) and (b) shows real and imaginary part of self energy$\Sigma({\bf K}_{n};0)$ in terms of $k_{x}$ and $k_{y}$ in super cell approximation for $N_{c}=4$. self-energy at the borders of grains have discontinuity and inside of each grain is k-independent. (c) and (d) shows real and imaginary parts of self-energy $\Sigma({\bf k};0)$ in the beyond $N_{c}=4$ super cell approximation which are fully k-dependent and causal.
![(color online) (a) and (b) show real and imaginary parts of super cell self energies $\frac{1}{4t}\Sigma({\bf K}_{n},0+i\eta)$ for the $N_{c}=4$ of a two dimensional square alloy. (c) and (d) shows real and imaginary parts of self energy $\frac{1}{4t}\Sigma({\bf k},0+i\eta)$ of a two dimensional square alloy system in beyond $N_{c}=4$ super cell approximation for $\delta=3t$, $c=0.5$ and $\mu=0$. In the super cell approximation k-space self energy in FBZ is discontinuous and k-independent in each grain, but in our beyond super cell approximation it is continuous and fully k-dependent. []{data-label="figure:Re-Im-selfenergy-nc4-d3t-n1-mu0"}](Re-Sigma-Kn-0-nc4-d3t-mu0-n1 "fig:"){width="7cm"} ![(color online) (a) and (b) show real and imaginary parts of super cell self energies $\frac{1}{4t}\Sigma({\bf K}_{n},0+i\eta)$ for the $N_{c}=4$ of a two dimensional square alloy. (c) and (d) shows real and imaginary parts of self energy $\frac{1}{4t}\Sigma({\bf k},0+i\eta)$ of a two dimensional square alloy system in beyond $N_{c}=4$ super cell approximation for $\delta=3t$, $c=0.5$ and $\mu=0$. In the super cell approximation k-space self energy in FBZ is discontinuous and k-independent in each grain, but in our beyond super cell approximation it is continuous and fully k-dependent. []{data-label="figure:Re-Im-selfenergy-nc4-d3t-n1-mu0"}](Im-Sigma-Kn-0-nc4-d3t-mu0-n1 "fig:"){width="7cm"} ![(color online) (a) and (b) show real and imaginary parts of super cell self energies $\frac{1}{4t}\Sigma({\bf K}_{n},0+i\eta)$ for the $N_{c}=4$ of a two dimensional square alloy. (c) and (d) shows real and imaginary parts of self energy $\frac{1}{4t}\Sigma({\bf k},0+i\eta)$ of a two dimensional square alloy system in beyond $N_{c}=4$ super cell approximation for $\delta=3t$, $c=0.5$ and $\mu=0$. In the super cell approximation k-space self energy in FBZ is discontinuous and k-independent in each grain, but in our beyond super cell approximation it is continuous and fully k-dependent. []{data-label="figure:Re-Im-selfenergy-nc4-d3t-n1-mu0"}](Re-sigma-k-0-nc4-d3t-mu0-n1 "fig:"){width="7cm"} ![(color online) (a) and (b) show real and imaginary parts of super cell self energies $\frac{1}{4t}\Sigma({\bf K}_{n},0+i\eta)$ for the $N_{c}=4$ of a two dimensional square alloy. (c) and (d) shows real and imaginary parts of self energy $\frac{1}{4t}\Sigma({\bf k},0+i\eta)$ of a two dimensional square alloy system in beyond $N_{c}=4$ super cell approximation for $\delta=3t$, $c=0.5$ and $\mu=0$. In the super cell approximation k-space self energy in FBZ is discontinuous and k-independent in each grain, but in our beyond super cell approximation it is continuous and fully k-dependent. []{data-label="figure:Re-Im-selfenergy-nc4-d3t-n1-mu0"}](Im-sigma-k-0-nc4-d3t-mu0-n1 "fig:"){width="7cm"}
Fig.\[figure:nc-nc4-d3t-n1-mu0\] (a) shows calculated average density of states for super cells $N_{c}=1$, $N_{c}=4$, and $N_{c}=16$ for $\delta=3t$, $c=0.5$ and band filling ${\bar n}=1$. (b) shows average density of states calculated by our beyond super cell approximation for $N_{c}=1$, $N_{c}=4$, and $N_{c}=16$. However bands of this system in this regime splitted in CPA and super cell approximation but in our approximation beyond super cell it is at beginning of splitting.
![(color online) Show comparison of average density of states of a two dimensional square alloy system for (a) CPA $N_{c}=1$, super cell approximation $N_{c}=4$, $N_{c}=16$ and (b) beyond $N_{c}=4$ and $N_{c}=16$ super cell approximation. The strength length, $\delta=3t$, impurity concentration is $c=0.5$ and $\mu=0$. The difference between density of states are due to nonlocal corrections.[]{data-label="figure:nc-nc4-d3t-n1-mu0"}](nc1-nc4-nc16-d3t-n1-mu0 "fig:"){width="5cm"} ![(color online) Show comparison of average density of states of a two dimensional square alloy system for (a) CPA $N_{c}=1$, super cell approximation $N_{c}=4$, $N_{c}=16$ and (b) beyond $N_{c}=4$ and $N_{c}=16$ super cell approximation. The strength length, $\delta=3t$, impurity concentration is $c=0.5$ and $\mu=0$. The difference between density of states are due to nonlocal corrections.[]{data-label="figure:nc-nc4-d3t-n1-mu0"}](beyod-nc1-nc4-nc16-d3t-n1-mu0 "fig:"){width="5cm"}
One of advantage of super cell approximation is to take in to account electron localization in one and two dimensional disordered alloys which calculates by[@T; @McK] $$\begin{aligned}
P(\infty)&=&lim_{t\;\rightarrow\:\infty} <|G_{ll}(E)|^{2}>\nonumber\\&=&lim_{\eta\;\rightarrow\;0}\int d\epsilon <|G_{ll}(\epsilon+i\eta)|^{2}>
\label{eq:local-super-cell}\end{aligned}$$ Fig.\[figure:localization-nc1-nc16\] shows probability of remaining electron at site $l$ for (a) a one dimensional lattice in the CPA and $N_{c}=16$ super cell approximations for $\delta=3t$, $c=0.5$ and $\mu=0$. CPA $P(\frac{\eta}{t})$ extrapolated to zero while for $N_{c}=16$ it is fitted by $P(\frac{\eta}{t})=0.007613+14.21 (\frac{\eta}{t})-300.7 (\frac{\eta}{t})^{2}+28.28(\frac{\eta}{t})^{3}$ hence $P(0)=0.007613$. (b) shows it for a square two dimensional alloy in the CPA and $N_{c}=16$ super cell approximation. In the CPA it is extrapolates to zero but for $N_{c}=16$ it is extrapolating to non zero value $P(\frac{\eta}{t}=0)=2.45\times 10^{-4}$ .
![(color online) Shows electron localization probability at site $l$ for: (a) one dimensional alloy for $\delta=3t$, $c=0.5$ and average band filling ${\bar n}=1$. For $N_{c}=1$, $P(\frac{\eta}{t})$ extrapolated to zero while for $N_{c}=16$ is fitted to $P(\frac{\eta}{t})=0.007613+14.21 (\frac{\eta}{t})-300.7 (\frac{\eta}{t})^{2}+28.28(\frac{\eta}{t})^{3}$. Hence probability of localization is $P(0)=0.007613$. (b) a two dimensional square alloy system for CPA $N_{c}=1$ and super cell approximation $N_{c}=16$. The strength length, $\delta=3t$, impurity concentration is $c=0.5$ and $\mu=0$. CPA doesn’t shows localization but for $N_{c}=16$ super cell, $P(\frac{\eta}{t}=0)=2.45\times 10^{-5}$ which is due to electron back scattering in the super cell.[]{data-label="figure:localization-nc1-nc16"}](localization-1d-nc16-nc1 "fig:"){width="6cm"} ![(color online) Shows electron localization probability at site $l$ for: (a) one dimensional alloy for $\delta=3t$, $c=0.5$ and average band filling ${\bar n}=1$. For $N_{c}=1$, $P(\frac{\eta}{t})$ extrapolated to zero while for $N_{c}=16$ is fitted to $P(\frac{\eta}{t})=0.007613+14.21 (\frac{\eta}{t})-300.7 (\frac{\eta}{t})^{2}+28.28(\frac{\eta}{t})^{3}$. Hence probability of localization is $P(0)=0.007613$. (b) a two dimensional square alloy system for CPA $N_{c}=1$ and super cell approximation $N_{c}=16$. The strength length, $\delta=3t$, impurity concentration is $c=0.5$ and $\mu=0$. CPA doesn’t shows localization but for $N_{c}=16$ super cell, $P(\frac{\eta}{t}=0)=2.45\times 10^{-5}$ which is due to electron back scattering in the super cell.[]{data-label="figure:localization-nc1-nc16"}](localization-nc1-nc16 "fig:"){width="6cm"}
Conclusion
==========
A successful approximation beyond super cell approximation is introduced. In this approximation self-energy is casual and full k-dependent in the first Brillouin zone. For derivation of the approximation, the entire lattice is divided to super cells with $N_{c}$ sites and no overlap. We proved that self-energy of one site in a definite super cell but another in other super cells are periodic with respect to super cell lengths. Correction to k-space super cell self-energy comes from sites in different super cells. We added this part to the k-space super cell self-energy. Our approximation recovers CPA in the single site cell limit and as the number of super cell sites approaches the number of lattice sites, $N_{c}\longrightarrow N$, becomes exact. This approximation opens a new channel for observing multi sites scattering effects such as localization that are not observed by other approximations such as DCA and CMDFT. It is overcomes discontinuity and weakly k-dependent of DCA and CMDFT especially for low dimensional systems that k-space self energy is k-dependent significantly. Also it is applicable to both disordered and interacting systems.
[0]{} W. Metzner, D. Vollhardt, Phys. Rev. Lett. [**62**]{}, 324 (1989). E. Muller-Hartmann, Z.Phys. B: Condens. Matter [**74**]{}, 507 (1989). P. Soven, Phys. Rev. B , [**156**]{}, 809 (1967).
A. Gonis, Green Functions for Ordered and Disordered Systems, in the series Studies in Mathematical Physics, edited by E. van Groesen and E. M. De- Jager North Holland, Amsterdam, 1992 .
M. H. Hettler, A. N. Tahvildar-Zadeh, M. Jarrell, T. Pruschke and H. R. Krishnamurthy, Phys. Rev. B [**61**]{}, 12739 (1998). M. H. Hettler, M. Mukherjee, M. Jarrell, H. R. Krishnamurthy, Phys. Rev. B [**61**]{}, 12739 (2000). M. Jarrell and H. R. Krishnamurthy, Phys. Rev. B [**63**]{}, 125102 (2001).
G. Kotliar, S. Y. Savrasov, G. Palsson and G. Biroli, Phys. Rev. Lett. [**87**]{}, 186401 (2001). Moradian Rostam, Balazs. L. Györffy, James. F. Annett, Phys. Rev. Lett. [**89**]{}, 287002 (2002). Moradian Rostam, Phys. Rev. B [**70**]{}, 205425 (2004). Neil. W Ashcroft and N David N. Mermin, [*Solid State Physics*]{}, (HRW International Edition, Hong Kong 1987). D. J. Thouless, Phys. Rep. [**13C**]{}, 94 (1974) .
A. J. McKane and M. Stone, Ann. Phys. N.Y., [**36**]{}, 131 (1981).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present general techniques for proving inapproximability results for several paradigmatic truthful multidimensional mechanism design problems. In particular, we demonstrate the strength of our techniques by exhibiting a lower bound of $2-\frac{1}{m}$ for the scheduling problem with $m$ unrelated machines (formulated as a mechanism design problem in the seminal paper of Nisan and Ronen on Algorithmic Mechanism Design). Our lower bound applies to truthful randomized mechanisms, regardless of any computational assumptions on the running time of these mechanisms. Moreover, it holds even for the wider class of truthfulness-in-expectation mechanisms. This lower bound nearly matches the known 1.58606 randomized truthful upper bound for the case of two machines (a non-truthful FPTAS exists).
Recently, Daskalakis and Weinberg [@Costis-2015-sched] show that there is a polynomial-time 2-approximately optimal *Bayesian* mechanism for makespan minimization for unrelated machines. We complement this result by showing an appropriate lower bound of 1.25 for deterministic incentive compatible Bayesian mechanisms.
We then show an application of our techniques to the workload-minimization problem in networks. We prove our lower bounds for this problem in the inter-domain routing setting presented by Feigenbaum, Papadimitriou, Sami, and Shenker. Finally, we discuss several notions of non-utilitarian fairness (Max-Min fairness, Min-Max fairness, and envy minimization) and show how our techniques can be used to prove lower bounds for these notions. No lower bounds for truthful mechanisms in multidimensional probabilistic settings were previously known.[^1]
author:
- 'Ahuva Mu’alem[^2]'
- 'Michael Schapira[^3]'
title: Setting Lower Bounds on Truthfulness
---
\#1[to \#1[to \#1]{}]{}
Introduction
============
Inapproximability Issues in Algorithmic Mechanism Design {#intro-1}
--------------------------------------------------------
The field of *Mechanism Design* deals with designing protocols for optimizing global goals that require interaction with selfish players. *Algorithmic Mechanism Design* [@NR] combines an economic perspective that takes into account the strategic behavior of the players, with a theoretical computer-science perspective that focuses on aspects such as computational-efficiency and approximability. Let us now describe, more formally, the nature of the problems that Algorithmic Mechanism Design attempts to solve: There is a finite set of *alternatives* $A=\{a,b,c,...\}$, and a set of *strategic players* $N=\{1,...,n\}$. Each player $i$ has a *valuation function* $v_i: A\rightarrow \mathbb{R}$ that is his private information. The players are self-interested and only wish to maximize their own gain. The global goal is expressed by a *social choice function* $f$ that assigns every possible $n$-tuple of players’ valuations $(v_1,...,v_n)$ an alternative $a \in A$. Mechanisms are said to *truthfully implement* a social choice function if their outcome for every $n$-tuple of players’ valuations matches that of the social choice function, and if they enforce payments of the different players in a way that motivates truthfully report of their valuations (no matter what the other players do).[^4]
A canonical social choice function is the *utilitarian* function; A utilitarian function aims to maximize the *social welfare*, i.e. to find the alternative $a$ for which the expression $\Sigma_i \ v_i(a)$ is maximized. Another paradigmatic social choice function is the *Max-Min* function (based on the philosophical work of John Rawls [@Rawls]); For every $n$-tuple of $v_i$ valuations the Max-Min function assigns the alternative $a \in A$ that maximizes the expression $\min_i \: v_i(a)$. Intuitively, the Max-Min function chooses the alternative $a\in A$ in which the least satisfied player has the highest value compared to the least satisfied player in all other alternatives $b\in A$.
While in many computational and economic settings the social choice function we wish to implement in a truthful manner is utilitarian (e.g. combinatorial auctions), often this is not the case. Problems in which the social choice function is non-utilitarian include revenue maximization in auctions (e.g. [@FGHK02]), minimizing the makespan in scheduling (e.g. [@NR; @AT; @Azar; @sched-rfpas; @CK]), fair allocation of resources (e.g. [@LMNB; @BV05; @LMSS]), etc. A classic result of mechanism design – a subfield of economic theory and game-theory (see [@Mas; @Rubi]) – states that for every utilitarian problem there exists a mechanism that truthfully implements it – namely, a member of the celebrated family of *VCG mechanisms* [@Vic61; @Clarke; @Gro73]. No general technique is known for truthfully implementing [*non-utilitarian*]{} social-choice functions. In fact, some non-utilitarian social-choice functions cannot be truthfully implemented [@LMNB; @NR]. Hence, from a computational point of view it is natural to ask how well non-utilitarian social choice functions can be *approximated* in a truthful manner.
Our Results
-----------
In this paper we present and discuss several general techniques for setting lower bounds on the approximability of truthful mechanisms. We obtain the first lower bounds for canonical randomized non-utiltarian settings. Our techniques are powerful in the following sense: Firstly, due to their generality and simplicity they can easily be applied to a variety of problems (as we shall demonstrate). Secondly, they apply to the general case of *multidimensional settings*.[^5] Thirdly, they do not impose any computational assumptions on the mechanism (such as polynomial running-time). Finally, our techniques apply to both deterministic and randomized mechanisms. In particular, we show how to derive lower bounds for both notions of truthfulness for randomized mechanisms - *universal truthfulness* and *truthfulness-in-expectation*.
In Section \[sec-lower\] we present our techniques and demonstrate their use on a non-utilitarian scheduling problem. The single-dimensional version of this scheduling problem has received much attention in recent years [@AT; @Azar; @sched-rfpas; @CK] (and references therein). We deal with the multidimensional version of the problem, formulated as a mechanism design problem by Nisan and Ronen in their seminal paper on Algorithmic Mechanism Design [@NR]: There are $n$ tasks that are to be scheduled on $m$ machines. Every machine $i \in[m]$ is a strategic player with a valuation function $v_i:2^{[n]}\rightarrow \mathbb{R_{+}}$ such that for every task $j\in[n]$, $v_i(\{j\})$ (we shall sometimes simply denote $v_i(j)$) specifies the *cost* of task $j$ on machine $i$. One can think of the cost of task $j$ on machine $i$ as the time it takes $i$ to complete $j$. For every subset of tasks $S\subseteq [n]$, $v_i(S)=\Sigma_{j\in S} \ v_i(j)$. That is, the total cost of a set of tasks on machine $i$ is the additive sum of the costs of the individual tasks on that machine. The global goal is minimizing the makespan of the chosen schedule. I.e., assigning the tasks to the machines in a way that minimizes the latest finishing time. Obviously, the makespan-minimization social choice function is non-utilitarian and hence *might* not be truthfully implemented by any mechanism. Nisan and Ronen prove that not only is it impossible to minimize the makespan in a truthful manner, but that *any approximation strictly better than $2$ cannot be achieved by a truthful deterministic mechanism*. Since a (non-truthful) FPTAS exists [@HS76], this raises a natural question: $$\mbox{Can near-optimality be achieved by using randomization?}$$
Section \[sec-lower\] illustrates our techniques by proving several lower bounds for this problem. In particular, we prove that no randomized truthful mechanism can achieve an approximation ratio better than $2-\frac{1}{m}$. This nearly matches the known truthful upper bound of $1.58606$ for the case in which there are only $2$ machines [@NR; @LY; @Chen]. Hence, $(1+\epsilon)$-approximation is ruled out for unrelated machines (regardless of computational efficiency). Moreover, for non-utilitarian settings randomness *cannot necessarily help* in obtaining approximation ratios that are considerably better than the known lower bound for truthful deterministic mechanisms. Somewhat surprisingly, this lower bound applies even for the substantially weaker notion of truthfulness for randomized mechanisms - truthfulness-in-expectation. We also show a lower bound of 1.25 for incentive compatible Bayesian mechanisms (assuming deterministic mechanisms[^6]). These are the first lower bounds for probabilistic multi-dimensional settings. In fact, to the best of our knowledge these are the first lower bounds for truthful randomized mechanisms, truthful-in-expectation and Bayesian incentive compatible mechanisms in *multidimensional* settings in general.
In addition, we show how to prove lower bounds for the important class of *strongly-monotone* deterministic mechanisms. This is another step towards proving the long-standing conjecture of Ronen and Nisan that *No truthful deterministic mechanism can obtain an approximation ratio better than $m$.*
In Section \[sec-applications\] we show an application of our techniques to another multidimensional non-utilitarian problem – minimizing the workload in communication networks. This problem arises naturally in the design of routing mechanisms. We study the approximability of this problem in the inter-domain routing setting presented by Feigenbaum, Papadimitriou, Sami, and Shenker [@FPSS].
Finally, in Section \[sec-fairness\] we discuss three notions of non-utilitarian fairness – Max-Min fairness, Min-Max fairness, and envy-minimization. We highlight the connections between these notions and the problems studied in this paper and prove several general results using our techniques.
Related Work
------------
In a seminal paper Nisan and Ronen [@NR] introduced the field of Algorithmic Mechanism Design. The main problem presented in [@NR] to illustrate the novelty of this new area of research was *scheduling with unrelated machines*. Nisan and Ronen explored the approximability of this non-utilitarian multidimensional problem and exhibited a lower bound of $2-\epsilon$ for truthful deterministic mechanisms. For this NP-hard scheduling problem there exist an FPTAS [@HS76] (assuming constant number of machines) and a polynomial-time 2-approximation algorithm [@LST87], that are both non-truthful. Additionally, this problem cannot be approximated in polynomial-time within a factor of less than $\frac{3}{2}$ [@LST87].
In recent years Algorithmic Mechanism Design has been the subject of extensive study [@AGT-Book]. A substantial amount of this research has focused on single-dimensional settings (see e.g. [@LOS; @AT02; @MN02; @ATTP; @FGHK02; @KKT]). Nearly-optimal truthful mechanisms were designed for *the single-dimensional problem of minimum makespan for scheduling tasks on related machines* [@AT; @Azar; @sched-rfpas; @CK]. The exploration of truthful mechanisms for multidimensional settings has arguably mainly revolved around the problem of welfare maximization in Multiple-Object auctions [@CA-sur; @Nisan-MU-Survey], that has gained the status of the paradigmatic problem of this field. As this is a utilitarian problem, it can be optimally and truthfully implemented by a VCG mechanism. However, it has been shown that the social welfare in combinatorial auctions cannot be maximized (or even closely approximated) in polynomial time [@LOS; @NS02]. As algorithmic mechanism design seeks time-efficient implementations, the main challenge faced by researchers was devising truthful polynomial-time mechanisms that approximately maximize the social welfare in combinatorial auctions ([@LS05; @DNS05; @DNS06; @BGN; @HKMT; @DD13]).
There are few inapproximability results for truthful mechanisms. This is particularly true in multidimensional settings. Other than Nisan and Ronen’s $2-\epsilon$ lower bound discussed previously, the following inapproximabily results are known: Lavi, Mu’alem and Nisan [@LMN] proved several lower bounds for *polynomial-time* truthful *deterministic* mechanisms. Their work is making several rather restrictive assumptions on the mechanisms beside assuming that they are truthful. Dobzinski and Nisan [@DN06] proved inapproximability results for polynomial-time VCG mechanisms. Recently, Dobzinski [@Dob11] bounds the power of polynomial-time truthful-in-expectation mechanisms in combinatorial auctions with submodular valuations. Several papers use VC dimensionality to prove inapproxiability results for deterministic truthful mechanisms [@MIKE15] (and references therein). We contribute to this ongoing research by presenting methods for deriving the first lower bounds for multidimensional non-utilitarian settings that apply to *general* truthful *randomized* mechanisms. Our lower bounds do not require any assumptions on the running-time of the mechanisms.
Our techniques greatly rely on the work of Bikhchandani et al. [@LMNB]. They characterize truthfulness in multidimensional settings by showing that any truthful deterministic mechanism must maintain a certain *weak monotonicity* property. Using this characterization, [@LMNB] manages to show that while welfare maximization can be truthfully implemented in combinatorial auctions, one cannot truthfully implement the Max-Min social choice function, even in a very restricted type of combinatorial auctions. The weak monotonicity property (and several of its extensions) will play a major role in our inapproximability proofs.
#### Makespan in multi-parameter settings.
Nisan and Ronen showed that no truthful deterministic mechanism can achieve an approximation ratio strictly better than 2 (and also strengthened this lower bound to $m$ for two specific classes of deterministic mechanisms) [@NR]. These lower bounds have been improved and extended in a series of recent results. Christodoulou, Koutsoupias, and Vidali showed a lower bound of $1+\varphi \approx 2.618$ for truthful deterministic mechanisms with $m\rightarrow \infty$ machines [@KoutsoupiasSODA2007; @KoutsoupiasPHI]. An alternative proof for a slight weaker lower bound can be found in [@Gamzu]. An optimal lower bound of $m$ for anonymous truthful mechanisms is shown in [@Ashlagi]. Lavi and Swamy [@Lavi-Swamy-sched] design truthful mechanisms in a multidimensional scheduling setting where the processing time of a task on each machine is either ’low’ or ’high’.
Recently, for Bayesian settings, Daskalakis and Weinberg [@Costis-2015-sched] show that there is a polynomial-time 2-approximately optimal mechanism for makespan minimization for unrelated machines.
In several interesting settings, truthful mechanisms are essentially equivalent to mechanisms that select envy-free allocations with the smallest supporting price vectors [@Gale-Demange-1985]. A natural question to ask is whether envy-free pricing techniques can improve the current striking approximability and inapproximability bounds for truthful mechanisms. Mu’alem [@Mu'alem-FAIR-by-Design] observed that the optimal envy-bounds are far apart from the optimal truthful bounds and therefore concludes that envy-free bounding techniques cannot be applied straightforwardly to tighten the bounds for minimizing the makespan on two unrelated machines.
In a follow-up work, Gamzu [@Gamzu] improved our truthful lower bound for minimizing the workload in inter-domain routing (from $\frac{1 + \sqrt{5}}{2} \approx 1.618$ to 2) and our randomized truthful lower bound (from $\frac{3+\sqrt{5}}{4} \approx 1.309$ to 2).
Open Questions
--------------
- We prove lower bounds for the scheduling problem with unrelated machines (see Section \[sec-lower\]) and for the workload-minimization problem in inter-domain routing (see Section \[sec-applications\]). In both problems, there are very large gaps between the known upper and lower bounds for truthful mechanisms (deterministic and randomized). Narrowing these gaps is an interesting long-standing open question. We believe that our lower bounds are not far from optimal, and suspect that matching (or nearly matching) truthful upper bounds exist.[^7]
- This paper did not make any computational assumptions on mechanisms. Proving (possibly stronger) lower bounds for *polynomial time* truthful mechanisms is a big open question.
The Organization of the Paper
-----------------------------
In Section \[sec-lower\] we present our techniques for setting lower bounds on truthfulness and demonstrate their application to the *scheduling problem with unrelated machines*. In Section \[sec-applications\] we present an applications of our techniques to the problem of *workload-minimization in networks*. In Section \[sec-fairness\] we discuss several notions of non-utilitarian fairness.
A Presentation of Our Techniques Via the Scheduling Problem {#sec-lower}
===========================================================
In this section we present our techniques. To illustrate the use of these techniques we show how they can be used to derive lower bounds for the scheduling problem with $m$ unrelated machines. Nisan and Ronen [@NR] exhibited a truthful $m$-approximation deterministic mechanism for this problem. This mechanism is basically a VCG mechanism, and can easily be shown to be strongly-monotone (see Subsection \[sub-deterministic\] for a formal definition of strong monotonicity). They also proved a lower bound of $2-\epsilon$ for truthful deterministic mechanisms that applies even when there are only two machines and is tight for this case. However, Nisan and Ronen [@NR] conjecture that their lower bound is not tight in general, and that *any truthful deterministic mechanism cannot obtain an approximation ratio better than $m$*.
For the case of two machines, Nisan and Ronen show that randomness helps get an approximation ratio better than $2$; They present a truthful randomized mechanism that has an approximation ratio of $\frac{7}{4}$. We generalize their result by designing a truthful randomized mechanism that obtains an approximation ratio of $\frac{7m}{8}$ (see Appendix \[upper-bound\]). *Thus, we prove that randomness achieves better performances than the known truthful deterministic $m$ upper bound for any number of machines*.[^8]
In Subsection \[sub-deterministic\] we show ways of proving lower bounds for truthful deterministic mechanisms. Using these methods we provide a simple and shorter proof for Nisan and Ronen’s $2-\epsilon$ lower bound. Our proof (unlike the original) relies on exploiting the *weak monotonicity* property defined in [@LMNB]. The techniques of Subsection \[sub-deterministic\] also aid us in deriving a stronger lower bound for the important classes of *strongly-monotone* deterministic mechanisms. We note, that the mechanism in [@NR], which is the best currently known deterministic mechanism for the scheduling problem, is contained in this class. We prove that no approximation ratio better than $m$ is possible for this class of mechanisms ([*thus making another step towards proving the long-standing conjecture of [@NR]*]{}).
After discussing lower bounds for truthful deterministic mechanisms we turn our attention to truthful randomized mechanisms. There are two possible definitions for the truthfulness of a randomized mechanism [@DNS06; @NR]. The first and stronger one is that of *universal truthfulness* that defines a truthful randomized mechanism as a probability distribution over truthful deterministic mechanisms. Thus, this definition requires that for *any* toss of the random coins made by the mechanism, the players still maximize their utility by reporting their true valuations. A considerably weaker definition of truthfulness is that of *truthfulness-in-expectation*. This definition only requires that players maximize their *expected* utility, where the expectation is over the random choices of the mechanism (but still for every behavior of the other players). Unlike universally truthful mechanisms, truthful-in-expectation mechanisms only motivate [*risk-neutral*]{} bidders to act truthfully. Risk-averse bidders may benefit from strategic behavior. In addition, truthful-in-expectation mechanisms induce truthful behavior only as long as players have no information about the outcomes of the random coin flips before they need to act.
In Subsection \[sub-universal\] we prove the first lower bound on the approximability of truthful randomized mechanisms in multidimensional settings. Namely, we show that any universally truthful mechanism for the scheduling problem cannot achieve an approximation ratio better than $2-\frac{1}{m}$. This lower bound nearly matches the universally truthful randomized $1.58606$ upper bound for the case of two machines. To prove this lower bound, we make use of a general technique that is based on Yao’s powerful principle [@Yao]. Our proof for the $2-\epsilon$ lower bound for deterministic mechanisms (in Subsection \[sub-deterministic\]) serves as a building block in the proof of this lower bound.
In Subsection \[sub-expectation\] we strengthen this result by proving that the same lower bound holds even when one is willing to settle for truthfulness-in-expectation. Our proof relies on some of the ideas that appear in the proof of the previous lower bound but takes a different approach. In particular, we generalize the weak monotonicity requirement to fit the class of truthful randomized mechanisms, and explore the implications of this extended monotonicity on the probability distributions over allocations generated by such mechanisms.
In Subsection \[sub-sec-LB-Bayes\] we turn to the notion of Bayesian Mechanisms, where players’ valuations are drawn from a prior distribution that is common knowledge, and show a lower bound of 1.25 (restricting attention to deterministic mechanisms).
Lower Bounds for Truthful Deterministic Mechanisms {#sub-deterministic}
--------------------------------------------------
Bikhchandani et al. [@LMNB] formally define the weak monotonicity property for deterministic mechanisms: Consider an Algorithmic Mechanism Design setting with $n$ strategic players that wish to maximize their personal gain. Before we present the formal definition for monotonicity we will require the following notation: For every $n$-tuple of players’ valuations $v=(v_1,...,v_n)$ we shall denote by $v_{-i}$ the $(n-1)$-tuple of players’ valuations $(v_1,...,v_{i-1},v_{i+1},...,v_n)$. Let $v'_i$ be a valuation function. We denote by $(v'_i,v_{-i})$ the $n$-tuple of valuations $(v_1,...,v_{i-1},v'_i,v_{i+1},...,v_n)$. I.e., $(v'_i,v_{-i})$ is the $n$-tuple of valuations we get by altering the $i$’th coordinate in $v$ from $v_i$ to $v'_i$.
Let $M$ be a deterministic mechanism. Let $i\in [n]$ and let $v=(v_1,...,v_n)$ be an $n$-tuple of players’ valuations. Let $v'_i$ be a valuation function. Denote by $a$ the alternative $M$ outputs for $v$ and by $b$ the alternative that $M$ outputs for $(v'_i,v_{-i})$. $M$ is said to be weakly monotone if for all such $i$, $v$, and $v'_i$ it holds that: $v_i(a)+v'_i(b)\ge
v'_i(a)+v_i(b)$.
This definition of weak monotonicity is for cases in which each player wishes to maximize his value. In problems in which players wish to minimize costs (such as the scheduling and workload minimization problems considered in this paper) the inequality is in the other direction.
Bikhchandani et al. [@LMNB] prove that any truthful deterministic mechanism must be weakly monotone. For completeness, we present this simple proof.
\[monotonicity-lemma\] Any truthful deterministic mechanism must be weakly monotone.
Let $M$ be a truthful deterministic mechanism. Let $i\in [n]$ and let $v=(v_1,...,v_n)$ be an $n$-tuple of players’ valuations. Let $v'_i$ be a valuation function. Denote by $a$ the alternative $M$ outputs for $v$ and by $b$ the alternative that $M$ outputs for $(v'_i,v_{-i})$. Consider player $i$. It is well known that the price a player is charged by the mechanism to ensure his truthfulness cannot depend on the player’s report. Specifically, the payment of player $i$ in $a$ and $b$ is a function of $v_{-i}$ and of $a$ and $b$, respectively. We denote by $p_i(v_{-i},a)$ and by $p_i(v_{-i},b)$ $i$’s payment in $a$ and $b$, respectively. It must hold that $v_i(a)-p_i(v_{-i},a)\ge v_i(b)-p_i(v_{-i},b)$ (for otherwise, if $i$’s valuation function is $v_i$, he would have an incentive to declare his valuation to be $v'_i$). Similarly, $v'_i(b)-p_i(v_{-i},b)\ge v'_i(a)-p_i(v_{-i},a)$. By adding these two inequalities we reach the weak monotonicity requirement.
Relying on the weak monotonicity property we provide an alternative proof for the $2-\epsilon$ lower bound of [@NR] for the scheduling problem with unrelated machines. Our proof shows that an any deterministic mechanism that achieves an approximation ratio better than $2$ violates the weak monotonicity property.
\[deterministic-bound\] Any weakly-monotone mechanism cannot achieve an approximation ratio better than $2$.
Let $\epsilon>0$. Consider the scheduling problem with two machines and three tasks. For every machine $i=1,2$ we define two possible valuation functions $v_i$ and $v'_i$:
$$v_i(t)=\left\{\begin{array}{ll}
1 & t=i\ or\ t=3\\
100 & \mbox{otherwise} \\
\end{array}\right.$$
$$v'_i(t)=\left\{\begin{array}{ll}
0 & t=i \\
1+\epsilon & t=3 \\
100 & \mbox{otherwise.} \\
\end{array}\right.$$
Let $M$ be a deterministic, weakly-monotone, mechanism that achieves an approximation ratio better than $2$. Then, when players $1$ and $2$ have the valuations $v_1$ and $v_2$ respectively, $M$ must assign task $1$ to player $1$, task $2$ to player $2$, and can choose to which player to assign task $3$ (because the optimal makespan is $2$ and any other assignment results in a makespan of at least $100$). W.l.o.g. assume that $M$ assigns task $3$ to player $2$. Now, consider the instance with players’ valuations $(v'_1,v_2)$. Notice that the only task-allocation that guarantees an approximation ratio better than $2$ is assigning tasks $1$ and $2$ to players $1$ and $2$ respectively, and task $3$ to player $1$. However, this turns out to be a violation of the weak monotonicity requirement. Weak monotonicity, in this case, dictates that for every player $i=1,2$ it must hold $v_i(a)+v'_i(b)\le v'_i(a)+v_i(b)$. However, if we look at player $1$ we find that $1+(1+\epsilon)=v_1(1)+v'_1(\{1,2\})>v'_1(1)+v_1(\{1,2\})=2$. A contradiction.
Lavi et al. [@LMN] present and study another property – strong monotonicity.
Let $M$ be a deterministic mechanism. Let $i\in [n]$ and let $v=(v_1,...,v_n)$ be an $n$-tuple of players’ valuations. Let $v'_i$ be a valuation function. Denote by $a$ the alternative $M$ outputs for $v$ and by $b$ the alternative that $M$ outputs for $(v'_i,v_{-i})$. $M$ is said to be strongly-monotone if for all such $i$, $v$, and $v'_i$ it holds that: If $a\neq b$, then $v_i(a)+v'_i(b)
>v'_i(a)+v_i(b)$.
As in the case of weak monotonicity, this definition of strong monotonicity is for cases in which each player wishes to maximize his value. In problems in which players wish to minimize costs (such as the scheduling and workload minimization problems considered in this paper) the inequality is reversed.
We prove that no member of the class of strongly-monotone mechanisms can obtain an approximation better than $m$ for the scheduling problem (even for the case of zero/one valuations). The idea at the heart of our proof of Theorem \[SMON-bound\] is an iterative use of the strong monotonicity property to construct an instance of the problem for which the allocation generated by the mechanisms is very far from optimal.
\[SMON-bound\] Any strongly-monotone mechanism cannot obtain an approximation ratio better than $m$.
Consider an instance of the scheduling problem with $m$ machines and $n=m^2$ tasks. Let $M$ be a deterministic mechanism for which the strong monotonicity property holds. Let $I$ be the instance of the scheduling problem in which every machine $i$ has a valuation function $v_i$ such that $v_i(j)=1$ for all $j\in[n]$. Denote by $S=(S_1,...,S_m)$ the allocation of tasks produced by $M$ for instance $I$. It must be that there is some machine $r$ such that $|S_r|\ge m$. Without loss of generality let $r=m$.
We will now create a new instance $I'$ by altering the valuation function of machine $1$ to $v'_1$ while leaving all the other valuation functions unchanged (in case $S_1=\emptyset$ we skip this part). That is, machine $1$ will have the valuation function $u_1=v'_1$:
$$v'_1(t)=\left\{\begin{array}{ll}
0 & t\in S_1 \\
1 & t\notin S_1 \\
\end{array}\right.$$ and every other machine $i\neq 1$ will have a valuation function $u_i=v_i$. Denote by $T=(T_1,...,T_m)$ the allocation $M$ generates for $I'$. The first step of the proof is showing that $S_1=T_1$. This is guaranteed by the strong monotonicity of $M$. Assume, by contradiction that $S_1\neq T_1$. The monotonicity property ensures that $v_1(S_1)+v'_1(T_1)\le v_1(T_1)+v'_1(S_1)$. By assigning values we get: $$|S_1|+|T_1\setminus S_1|< |T_1| + 0.$$ Observe, that $|S_1|+|T_1\setminus S_1|-|T_1|=|S_1\setminus T_1|$, therefore: $$|S_1\setminus T_1|<0.$$ A contradiction.
We shall now prove that not only does $S_1$ equal $T_1$, but in fact $S_i=T_i$ for every $i\in [m]$; Since $S_1=T_1$ it must be that $v_1(S_1)+v'_1(T_1)=v'_1(S_1)+v_1(T_1)$. However, the strong monotonicity property dictates that if this is true then $S=T$.
In an analogous manner we shall now turn the valuation function of machine $2$ into $v'_2$ while keeping all the other valuation functions in $I'$ unchanged (in case $S_2=\emptyset$ we skip this part). That is, the valuation function of machine $2$ is changed into:
$$v'_2(t)=\left\{\begin{array}{ll}
0 & t\in S_2 \\
1 & t\notin S_2. \\
\end{array}\right.$$
Similar arguments show that the allocation produced by the $M$ for this new instance will remain $S$. We can now iteratively continue to change the valuation functions of machines $3,...,m-1$ into $v'_3,...,v'_{m-1}$ respectively, without changing the allocation the mechanism generates for these new instances. After performing this, we are left with an instance in which every machine $i\in[m-1]$ has the valuation function $v'_i$, and machine $m$ has the valuation function $v_m$. We have shown that the allocation generated by $M$ for this instance is $S$. Recall that $|S_m|\ge m$. Let $R\subseteq S_m$ such that $|R|=m$. We will now create a new instance $INS$ from the previous one by only altering the valuation function $v_m$ into the following valuation function $v'_m$:
$$v'_m(t)=\left\{\begin{array}{ll}
0 & t\in S_m\setminus R \\
1 & \hbox{otherwise.} \\
\end{array}\right.$$
By applying similar arguments to the ones used before, one can show that the allocation generated by $M$ when given instance $INS$ remains $S$. Observe that the finishing time of $S$ is $m$ because all the tasks in $R$ are assigned to machine $m$. Also notice that the finishing time of the optimal allocation of tasks for $INS$ is precisely $1$. The theorem follows.
Lower Bounds for Universally Truthful Mechanisms {#sub-universal}
------------------------------------------------
We now present a technique for deriving lower bounds for universally truthful mechanisms, based on Yao’s principle [@Yao]. Consider a zero-sum game with two players. Let the `"`row player`"`’s strategies be the various different instances of a specific problem, and let the `"`column player`"`’s strategies be all the deterministic truthful mechanisms for solving that problem. Let entry $g_{ij}$ in the matrix $G$ depicting the game be the approximation ratio obtained by the algorithm of column $j$ when given the instance of row $i$.
Recall that every randomized mechanism that is truthful in the universal sense is a probability distribution over deterministic truthful mechanisms. The `"`natural`"` approach for proving a lower bound for such randomized mechanisms is to find an instance of the problem on which every such randomized mechanism cannot achieve (in expectation) a certain approximation factor. By applying the well known Minimax Theorem to the game described above we get that an alternate and just as powerful way for setting lower bounds is to show that there is a probability distribution over instances on which any deterministic mechanism cannot obtain (in expectation) a certain approximation ratio.
We demonstrate this technique by proving a $2-\frac{1}{m}$ lower bound for universally truthful mechanisms for the scheduling problem. Our proof is based on finding a probability distribution over instances of the scheduling problem for which no deterministic truthful mechanism can provide an approximation ratio better than $2-\frac{1}{m}$. To show this, we shall exploit the weak monotonicity property of truthful deterministic mechanisms (as discussed in Subsection \[sub-deterministic\]).
\[universal-bound\] Any randomized mechanism that is truthful in the universal sense cannot achieve an approximation ratio better than $2-\frac{1}{m}$.
Let $\epsilon>0$. Consider the scheduling problem with $m$ machines and $n=m+1$ tasks. For every machine $i\in [m]$ we define two possible valuation functions:
$$v_i(t)=\left\{\begin{array}{ll}
1 & t=i\ or\ t=m+1\\
\frac{4}{\epsilon} & \hbox{otherwise} \\
\end{array}\right.$$
and
$$v'_i(t)=\left\{\begin{array}{ll}
0 & t=i \\
1+\epsilon & t=m+1 \\
\frac{4}{\epsilon} & \hbox{otherwise.} \\
\end{array}\right.$$
Let $I$ be the instance in which the valuation function of every machine $i$ is $v_i$. For every $j\in [m]$, let $I^j$ be the instance in which every machine $i\neq j$ has the valuation function $v_i$, and machine $j$ has the valuation function $v'_j$. We are now ready to define the probability distribution $P$ over instances: instance $I$ is assigned the probability $\epsilon$, and for every $j\in[m]$ instance $I^j$ is picked with probability $\frac{1-\epsilon}{m}$.
We now need to show that any deterministic truthful mechanism $M$ cannot achieve an approximation ratio better than $2-\frac{1}{m}$ on $P$. Let $T^j$ be the allocation of the $m+1$ tasks to the $m$ machines in which every machine $i$ gets task $i$, and machine $j$ is also assigned task $m+1$. Observe, that $T^j$ is the optimal allocation of tasks for instance $I^j$. Also observe, that while the finishing time of the allocation $T^j$ for instance $I^j$ is $1+\epsilon$, the finishing time of any other allocation of tasks is at least $2$. We shall denote the allocation $M$ outputs for instance $I$ by $M(I)$. Similarly, we shall denote the allocation $M$ outputs for instance $I^j$ by $M(I^j)$ (for every $j\in[m]$). We will now examine two distinct cases: The case in which $M(I)\neq T^r$ for any $r\in [m]$, and the case that $M(I)=T^r$ for some $r\in [m]$.
Observe, that in the first case the finishing time is at least $\frac{4}{\epsilon}$ while the optimal finishing time is $2$. Thus, $M$ obtains a $\frac{2}{\epsilon}$-approximation. Since instance $I$ appears in $P$ with probability $\epsilon$ we have that $A$’s expected approximation ratio is at least $\frac{2}{\epsilon}\times \epsilon=2$.
We are left with the case in which $M(I)=T^r$ for some $r\in [m]$. Consider an instance $I^j$ such that $j\neq r$. The following lemma states that $M$ will not output the optimal allocation for $I^j$ (that is, $T^j$).
\[monotone1\] If $M(I)=T^r$ for some $r\in [m]$, then for every $j\neq r$ $M(I^j)\neq T^j$.
Let $j\neq r$. Let us assume by contradiction that $M(I^j)=T^j$. The weak monotonicity property dictates that $v_j({j})+v'_j(\{j,m+1\})\le v'_j({j})+v_j(\{j,m+1\})$. By assigning values we get that $1+(1+\epsilon)\le (1+1)$, and reach a contradiction.
From Lemma \[monotone1\] we learn that if $M(I)=T^r$ (for some $r\in [m]$) then we have that $M(I^j)$ (for every $j\neq r$) is an allocation that is not the optimal one (i.e. not $T^j$). In fact (as mentioned before), any allocation $M$ outputs given $I^j$ will have a finishing time of at least $2$, while the optimal allocation ($T^j$) has a finishing time of $1+\epsilon$. Thus, for every $j\neq r$ the approximation ratio of $M$ for instance $I^j$ is at least $\frac{2}{1+\epsilon}$. The expected approximation ratio of $M$ for $P$ is therefore at least $\frac{(m-1)(1-\epsilon)}{m}\times
\frac{2}{1+\epsilon}+\frac{1-\epsilon}{m}\times 1$. Since this is true for any value of $\epsilon$, the approximation ratio cannot be better than $2-\frac{1}{m}$.
Lower Bounds for Mechanisms that are Truthful-in-Expectation {#sub-expectation}
------------------------------------------------------------
After handling the case of universally truthful mechanisms we now turn to the weaker notion of truthfulness-in-expectation. We start by generalizing the weak monotonicity definition to the case of randomized mechanisms. Any randomized mechanism can be regarded as a mechanism that for every instance of a problem produces a probability distribution over possible alternatives.
A randomized mechanism is a function from $n$-tuples of players’ valuations to probability distributions over the set of alternatives $A$.
The valuation function of each of the players in such mechanisms can therefore be viewed as assigning values to probability distributions over possible alternatives rather than only to the alternatives themselves.
Let $v$ be a valuation function. We define the *extended valuation function* $V_{v}$ as follows. For every probability distribution $P$ over the set of alternatives $A$, $V_{v}(P)=\Sigma_{a\in A}\ Pr_ P[a]\times v(a)$.
Arguments similar to those of Lemma \[monotonicity-lemma\] show that randomized mechanisms that are truthful-in-expectations must be weakly monotone (given the new definition of the valuation functions). This *extended weak monotonicity* is equivalent to the *monotonicity-in-expectation* property defined by Lavi and Swamy [@LS05].
Let $M$ be a randomized mechanism. Let $i\in [n]$ and let $v=(v_1,...,v_n)$ be an $n$-tuple of players’ valuations. Let $v'_i$ be a valuation function. Denote by $P$ the distribution over alternatives $M$ outputs for $v$ and by $Q$ the distribution over alternatives $M$ outputs for $(v'_i,v_{-i})$. $M$ is said to be weakly monotone in the extended sense if for all such $i$, $v$, and $v'_i$ it holds that: $V_{v_i}(P)+V_{v'_i}(Q)\ge
V_{v'_i}(P)+V_{v_i}(Q)$.
As before, if the players wish to minimize costs rather than maximize values, the inequality is reversed.
\[rand-monotonicity-lemma\] Any truthful-in-expectation mechanism must be weakly monotone in the extended sense.
We can exploit this extended definition of weak monotonicity to prove inapproximability results. We show how this is done by strengthening our $2-\frac{1}{m}$ lower bound for universally truthful mechanisms by showing that it applies even for the case of truthfulness-in-expectation. To do this, we show that the extended weak monotonicity of truthful randomized mechanisms implies non-trivial connections between the probability distributions over allocations they produce for different instances of the scheduling problem.
A key element in the proof of Theorem \[expectation-bound\] is the observation that instead of regarding a randomized mechanism for the scheduling problem as generating probability distributions over allocations of tasks, it can be regarded as generating, for each task, a probability distribution over the machines it is assigned to by the mechanism. This is true due to the linearity (additivity) of the valuation functions. This different view of a randomized mechanism for this specific problem, enables us to analyze the contribution of each task to the expected makespan.
The main lemma in the proof of Theorem \[expectation-bound\], namely Lemma \[monotone2\], makes use of this fact together with the extended weak monotonicity of truthful randomized mechanisms. Lemma \[monotone2\] essentially proves that for two carefully chosen instances of the problem, the probability that a specific task is assigned to a specific machine by $M$ in one of the instances, cannot be considerably higher than the probability it is assigned to the same machine in the other. Thus, we show that even though allocating this task to that machine in one of the instances leads to a good approximation, any truthful-in-expectation mechanism will fail to do so.
\[expectation-bound\] Any mechanism that is weakly monotone in the extended sense cannot achieve an approximation ratio better than $2-\frac{1}{m}$.
Let $\epsilon>0$. Consider the scheduling problem with $m$ machines and $n=m+1$ tasks. For every machine $i\in [m]$ we define two possible valuation functions:
$$v_i(t)=\left\{\begin{array}{ll}
1 & t=i\ or\ t=m+1\\
\frac{4}{\epsilon^2} & \hbox{otherwise} \\
\end{array}\right.$$
and
$$v'_i(t)=\left\{\begin{array}{ll}
0 & t=i \\
1+\epsilon & t=m+1 \\
\frac{4}{\epsilon^2} & \hbox{otherwise.} \\
\end{array}\right.$$
Let $I$ be the instance in which the valuation function of every machine $i$ is $v_i$. For every $j\in [m]$ let $I^j$ be the instance in which every machine $i\neq j$ has the valuation function $v_i$, and machine $j$ has the valuation function $v'_j$. Let $T^j$ be the allocation of the $m+1$ tasks to the $m$ machines in which every machine $i$ gets task $i$, and machine $j$ is also assigned task $m+1$.
Let $M$ be a mechanism that is weakly monotone in the extended sense. We shall denote by $P$ the distribution over all possible allocations produced by $M$ when given instance $I$, and by $P^j$ the distribution over all possible allocations $M$ produces when given instance $I^j$. Let $R$ be some distribution over the possible allocations. Fix a machine $i$ and a task $t$, we define $p_{i,t}(R)$ to be the probability that machine $i$ gets task $t$ given $R$. Formally, $p_{i,t}(R)=\Sigma_{a|t\in a_i}\ Pr_R[a]$. Observe that $V_{v_i}(R)=\Sigma_{t\in[n]}\ p_{i,t}(R)v_i(t)$ and $V_{v'_i}(R)=\Sigma_{t\in[n]}\ p_{i,t}(R)v'_i(t)$.
We are now ready to prove the theorem. In order to do so, we prove that for every mechanism $M$, as defined above, one can find an instance of the scheduling problem for which $M$ fails to give an approximation ratio better than $2-\frac{1}{m}$. Consider instance $I$.
If $p_{i,i}(P)<1-\epsilon^2$ (for some $i\in[m]$) then machine $i$ does not get task $i$ with probability of at least $\epsilon^2$. However, when machine $i$ does not get task $i$, the finishing time of a schedule for $I$ cannot be less than $\frac{4}{\epsilon^2}$, while the optimal finish time is $2$. Therefore, with probability of at least $\epsilon^2$ the approximation ratio obtained by the algorithm is at least $\frac{2}{\epsilon^2}$. If this is the case then, in expectation, the approximation ratio is at least $2$ (and the theorem follows). Hence, from now on we will only deal with the case in which for every $i\in[m]$, $$p_{i,i}(P)\ge 1-\epsilon^2.$$
Let $r$ be some machine such that $p_{r,m+1}(P)\le \frac{1}{m}$. Intuitively, $r$ is a machine that is hardly assigned task $m+1$ in $P$. We will show that in this case we can choose the instance $I^r$ to prove our lower bound. The main idea of the proof is showing that machine $r$ will not be assigned task $m+1$ in $P^r$ with probability that is significantly higher than the probability it was assigned the task in $P$. Thus, even though assigning task $m+1$ to machine $r$ is a smart step approximation-wise, the extended weak monotonicity of the mechanism will prevent it from doing so.
\[monotone2\] Let $r$ be some machine such that $p_{r,m+1}(P)\le \frac{1}{m}$. It holds that $p_{r,m+1}(P^r)\le \frac{1}{m}+\epsilon$.
As $M$ is weakly monotone in the extended sense we have that $V_{v_r}(P)+V_{v'_r}(P^r)\le V_{v'_r}(P)+V_v{v_r}(P^r)$. That is:
$$\;\; \Sigma_{t\in[n]} \ p_{r,t}(P)v_r(t)+\Sigma_{t\in[n]} \ p_{r,t}(P^r)v'_r(t)\le$$ $$\Sigma_{t\in[n]} \ p_{r,t}(P)v'_r(t)+\Sigma_{t\in[n]} \ p_{r,t}(P^r)v_r(t)$$
After subtracting identical summands from both sides of the equation we get:
$$\;\;p_{r,r}(P)v_r(r)+p_{r,m+1}(P)v_r(m+1)+p_{r,r}(P^r)v'_r(r)+p_{r,m+1}(P^r)v'_r(m+1)\le$$ $$p_{r,r}(P)v'_r(r)+p_{r,m+1}(P)v'_r(m+1)+p_{r,r}(P^r)v_r(r)+p_{r,m+1}(P^r)v_r(m+1)$$
By assigning values we reach the following inequality:
$$p_{r,r}(P)+p_{r,m+1}(P)+p_{r,m+1}(P^r)\times (1+\epsilon)\le
p_{r,m+1}(P)\times (1+\epsilon)+p_{r,r}(P^r)+p_{r,m+1}(P^r)$$
Therefore:
$$p_{r,r}(P)+p_{r,m+1}(P^r)\times\epsilon\le
p_{r,m+1}(P)\times\epsilon+p_{r,r}(P^r)$$
Because $p_{r,r}(P)\ge 1-\epsilon^2$ and $p_{r,r}(P^r)\le 1$ we get:
$$(1-\epsilon^2)+p_{r,m+1}(P^r)\times \epsilon\le
p_{r,m+1}(P)\times \epsilon+1$$
$$p_{r,m+1}(P^r)\times \epsilon-\epsilon^2<p_{r,m+1}(P)\times
\epsilon$$
$$p_{r,m+1}(P^r)\le p_{r,m+1}(P)+\epsilon$$
Since $p_{r,m+1}(P)\le \frac{1}{m}$ we have that:
$$p_{r,m+1}(P^r)\le p_{r,m+1}(P)+\epsilon \le \frac{1}{m}+\epsilon$$
This concludes the proof of the lemma.
From Lemma \[monotone2\] we learn that if $r$ is a machine such that $p_{r,m+1}(P)\le \frac{1}{m}$ then $p_{r,m+1}(P^r)\le
\frac{1}{m}+\epsilon$. Relying on this fact, we can choose $I^r$ as our instance and show that $M$ fails to provide an approximation ratio better than $2-\frac{1}{m}$ for $I^r$. The optimal allocation for $I^r$ is $T^r$, which has a finishing time of $1+\epsilon$. Any other allocation has a finishing time of at least $2$. Hence, when $T^r$ is not reached by $A$, the approximation ratio obtained is a at least $\frac{2}{1+\epsilon}$. However, we know that with high probability $T^r$ is not reached by $M$; Since $p_{r,m+1}\le
\frac{1}{m}+\epsilon$, and since machine $r$ gets task $m+1$ in $T^r$, we know the probability that $M$ outputs $T^r$ is at most $\frac{1}{m}+\epsilon$. The expected approximation ratio of $M$ is therefore at least $(1-(\frac{1}{m}+\epsilon))\times\frac{2}{1+\epsilon}+(\frac{1}{m}+\epsilon)\times
1$. Since this is true for any value of $\epsilon$ the theorem follows.
A Lower Bound for Bayesian Mechanisms {#sub-sec-LB-Bayes}
-------------------------------------
We now turn to the notion of Bayesian Mechanisms, where players’ valuations are drawn from a prior distribution that is common knowledge. Recently, Daskalakis and Weinberg [@Costis-2015-sched] show that there is a polynomial-time 2-approximately optimal Bayesian mechanism for makespan minimization for unrelated machines. We complement this result by showing an appropriate lower bound of 1.25. No lower bound for Bayesian mechanisms in multi-parameter settings was previously known. In what follows we restrict our attention to deterministic mechanisms.
A mechanism $M(f, p)$ is Bayesian Incentive Compatible (given the common knowledge distribution $D$) if for all $i$
$$E_{v_{-i} \sim D_{-i}} \left[ v _i(f(v_i,v_{-i})) - p_i(v_i, v_{-i}) \right] \ge
E_{v_{-i} \sim D_{-i}} \left[ v_i(f(v'_i, v_{-i})) - p_i(v'_i, v_{-i})\right].$$
That is, player $i$’s expected utility from reporting his true valuation $v_i$ is no less than his expected utility from reporting a different type $v'_i$ when others’ true values are drawn from the joint distribution $D_{-i}$.
As before, if the players wish to minimize costs rather than maximize values, the inequality is reversed.
\[bayesian-bound\] Any deterministic Bayesian Incentive Compatible mechanism $M$ cannot achieve an approximation ratio strictly better than $1.25$.
Let $\epsilon>0$. Consider a setting with $m$ machines and $3$ tasks. For every machine $i=1,2$ we define two possible *equally likely* valuation functions:
$$v_i(t)=\left\{\begin{array}{ll}
1 & t=i\ or\ t=3\\
\frac{4}{\epsilon} & \hbox{otherwise} \\
\end{array}\right.$$
and
$$v'_i(t)=\left\{\begin{array}{ll}
0 & t=i \\
1+\epsilon & t=3\\
\frac{4}{\epsilon} & \hbox{otherwise.} \\
\end{array}\right.$$ For $i =3,...,m, \; t=1,2,3$ we define $v_i(t)=\infty$.
Assume by contradiction that there exists a *deterministic* Bayesian mechanism $M$ with an expected approximation ratio $1.25 -\delta$, where $\delta > 0$. Clearly, it is enough to focus from now on machines 1 and 2 and ignore all other machines. We shall denote the allocation $M$ outputs for instance $I$ by $M(I)$. Let $T^j$ be the allocation of the tasks to machines in which task $i$ is assigned to machine $i$, and task $3$ is assigned to machine $j$ (where $i, j \in \{1,\ 2\}$). Observe that $T^1, T^2$ are optimal w.r.t. makespan for the instances $(v'_1, v_2)$, $(v_1, v'_2)$, respectively. Furthermore, we must have that $M(v'_1, v_2)=T^1, M(v_1, v'_2)=T^2$ under the assumption that an expected approximation ratio strictly better than $1.25$ can be achieved. More formally, observe that the finishing time of any other allocation of tasks is at least $2$, and clearly for a small enough $\epsilon$ we have $\frac34\times 1 + \frac14 \times \frac {2}{1+\epsilon} > 1.25 - \delta$. A similar argument shows that $M(v_1, v_2) , M(v'_1, v'_2) \in \{T^1, T^2\}$.
Without loss of generality assume that $M(v_1, v_2) = T^2$. By symmetry it is enough to examine two distinct cases: $M(v'_1, v'_2)=T^1$ and $M(v'_1, v'_2)=T^2$.
In the first case we have that $M(v_1, v_2)=M(v_1, v'_2)=T^2, M(v'_1, v_2)=M(v'_1, v'_2)=T^1$. Notice that the report of machine 1 solely dictates the chosen allocation. Here we can assume without loss of generality that $p_1(v_1, v_2)=p_2(v_1, v'_2)$, (otherwise, we can simply take the average) and similarly we can assume that $p_1(v'_1, v_2)=p_1(v'_1, v'_2)$.
and by Bayesian incentive compatibility we have that $$\frac12 (p_1(v_1, v_2) - v_1(T^2)) + \frac12 (p_1(v_1, v'_2) - v_1(T^2)) \geq
\frac12 (p_1(v'_1, v_2) - v_1(T^1)) + \frac12 (p_1(v'_1, v'_2) - v_1(T^1)),$$
and $$\frac12 (p_1(v'_1, v_2) - v'_1(T^1)) + \frac12 (p_1(v'_1, v'_2) - v'_1(T^1)) \geq
\frac12 (p_1(v_1, v_2) - v'_1(T^2)) + \frac12 (p_1(v_1, v'_2) - v'_1(T^2)).$$
Putting all together we have $$(p_1(v_1, v_2) - v_1(T^2)) + (p_1(v'_1, v_2) - v'_1(T^1)) \geq
(p_1(v'_1, v_2) - v_1(T^1)) + (p_1(v_1, v_2) - v'_1(T^2)).$$
By rearranging: $v_1(T^1) + v'_1(T^2) \ge v_1(T^2) + v'_1(T^1)$ we reach a contradiction (since $2 + 0 < 1 + (1+\epsilon)$).
We are left with the case in which $M(v_1, v_2)=M(v_1, v'_2)=M(v'_1, v'_2)=T^2, M(v'_1, v_2)=T^1$. Here, the report of player 1 dictates the chosen allocation only if the report of player 2 is $v_2$. By Bayesian incentive compatibility we have that $$\frac12 (p_1(v_1, v_2) - v_1(T^2)) + \frac12 (p_1(v_1, v'_2) - v_1(T^2)) \geq
\frac12 (p_1(v'_1, v_2) - v_1(T^1)) + \frac12 (p_1(v'_1, v'_2) - v_1(T^2)),$$ and $$\frac12 (p_1(v'_1, v_2) - v'_1(T^1)) + \frac12 (p_1(v'_1, v'_2) - v'_1(T^2)) \geq
\frac12 (p_1(v_1, v_2) - v'_1(T^2)) + \frac12 (p_1(v_1, v'_2) - v'_1(T^2)).$$
Once again, putting all together we have $$(p_1(v_1, v_2) - v_1(T^2)) + (p_1(v'_1, v_2) - v'_1(T^1)) \geq
(p_1(v'_1, v_2) - v_1(T^1)) + (p_1(v_1, v_2) - v'_1(T^2)).$$ By rearranging: $v_1(T^1) + v'_1(T^2) \ge v_1(T^2) + v'_1(T^1)$ we reach a contradiction (since $2 + 0 < 1 + (1+\epsilon)$).
Application: Workload Minimization in Inter-Domain Routing {#sec-applications}
==========================================================
In this section, we show an application of our techniques to another non-utilitarian multidimensional problem – *workload minimization in inter-domain routing*. Feigenbaum, Papadimitriou, Sami, and Shenker formulated the inter-domain routing problem as a distributed mechanism design problem [@FPSS] (inspired by the extensive literature on the real-life problem of inter-domain routing in the Internet). In recent years several works that study their model and its extensions have been published [@FKMS; @FRS; @FSS]. All these works deal with the realization of utilitarian social-choice functions (cost minimization, welfare maximization), and focus on the efficient and distributed design of VCG mechanisms.
Workload minimization is a problem that arises naturally in the design of routing protocols, as we wish that no single Autonomous System (AS) will be overloaded with work. It can easily be shown that any such VCG mechanism performs very poorly with regards to workload minimization. Thus, while optimally minimizing the total cost, or maximizing the social welfare, the known truthful mechanisms for this problem can result in workloads that are very far from optimal (in which one AS is burdened by the traffic sent by all other ASes). We initiate the study of truthful workload minimization in inter-domain routing by presenting constant lower bounds that apply to any truthful mechanism (deterministic and randomized).
**Formal Statement of the Problem**
We are given a *directed graph* $G=\langle N,L\rangle$ (called the *AS graph*) in which the set of nodes $N$ corresponds to the Autonomous Systems (ASes) of which the Internet is comprised. $N$ consists of a *destination* node $d$, and $n$ *source* nodes. The set of edges $L$ corresponds to communication links between the ASes. Each source node $i$ is a strategic player. The number of packets (intensity of traffic) originating in source node $i$ and destined for $d$ is denoted by $t_i$.
Let $neighbours(i)$ be all the ASes that are directly linked to $i$ in the AS graph. Each source node $i$ has a *cost function* $c_i:neighbours(i)\rightarrow \mathbb{R_{+}}$ that specifies the per-packet cost incurred by this node for carrying traffic, where $L_i\subseteq L$ is the set of links node $i$ participates in. This cost function represents the additional internal load imposed on the internal AS network when sending a packet from $i$ to an adjacent AS.[^9]
In the *single-dimensional version of this problem* an AS $i$ incurs the same per-packet cost $c_i$ for sending traffic to each of its neighbours (i.e., $c_i(l_1)=c_i(l_2)$ for every $l_1,l_2$).
The goal is to assign all source nodes routes to $d$. This *route allocation* should form a confluent tree to the destination $d$. I.e., no node is allowed to transfer traffic to two adjacent nodes. We seek truthful mechanisms that output routing trees in which the workload imposed on the busiest node is minimized. Formally, let $N^T_i$ be the set of all nodes whose paths in the routing tree $T$ go through node $i$. Let $Next^T(i)$ be the node $i$ transfers traffic to in $T$. We wish to minimize the expression $$max_i\ \Sigma_{j\in N^T_i}\ t_j\times c_i(i,Next^T(i))$$ over all possible routing trees $T$. The problem of load minimization arises naturally in inter-domain routing as we require that no single AS will be overloaded with work.
**Approximability of the Single-Dimensional Case**
It is easy to show (via a simple reduction from Partition) that even the single-dimensional version of the workload-minimization problem is NP-hard. However, is it at all possible to *optimally* solve this problem in a truthful manner? The answer to this question is yes. However, the worst-case running time is exponential (we note that this is also the case in the single-dimensional version of machine-scheduling [@AT]).
There exists a truthful, deterministic, exponential-time mechanism that always finds a workload-minimizing route allocation in the single-dimensional case.
The mechanism $M$ simply goes over all possible route allocations and outputs the optimal one with regards to workload-minimization. As in [@AT], our truthful mechanism outputs the lexicographically-minimal optimal route allocation; That is, let $a$ and $b$ be two distinct optimal route allocations (if two such allocations exist). Let $a_1,...,a_n$ be a *decreasing* order of the workloads of the different nodes in $a$. Similarly, let $b_1,...,b_n$ be *decreasing* order of the workloads of the different nodes in $b$. Let $j\in [n]$ be the first index such that $a_j\neq b_j$ or $j=n$ if no such index exists. The mechanism will choose $a$ if $a_j<b_j$, $b$ if $b_j<a_j$, and otherwise according to a predefined deterministic tie breaking rule.
Obviously, the mechanism always outputs an optimal solution. We are left with proving the truthfulness of the mechanism. It is well known that a mechanism is truthful in a single-dimensional setting such as ours *if and only if* it is weakly monotone [@AT]. Let $a$ be the route allocation $M$ outputs when the per-packet cost of $i$ is $c_i$, and the per-packet costs of the other nodes are $c_{-i}=c_1,...,c_{i-1},c_{i+1},...,c_n$. Let $b$ be the route allocation $M$ outputs when the per-packet cost of $i$ is $c'_i$, and the per-packet costs of the other nodes are $c_{-i}$. Weak monotonicity states that if $c_i< c'_i$ then $k_i\ge k'_i$, where $k_i$ and $k'_i$ are the number of packets that go through $i$ in $a$ and $b$, respectively (and this is true for every node $i$, for every vector of costs per-packet $c_{-i}$ of the other nodes, and for every two costs per-packet $c_i\neq c'_i$).
Fix a node $i$. Assume, by contradiction, that there are $c_i <
c'_i$, and $c_{-i}$ such that $k_i < k'_i$. Let $a_1,...,a_n$ and $b_1,...,b_n$ be defined as before. Let $j\in [n]$ be the first index such that $a_j\neq b_j$ (as before). If the two allocations are identical then no such $j$ exists. However, note that in this case if node $i$ declares $c'_i$ then the allocation $b$ will not be chosen (because $c'_i k_i<c'_i k'_i$ and so $a$ comes before $b$ in the lexicographic order). This contradicts the definition of $b$. We now turn to a sketch of a case by case analysis.
[*Case 1:*]{} $a_j < b_j$. It is not hard to verify that if $a_j<b_j$ and $i$ declares $c'_i$ then $a$ comes before $b$ in the lexicographic order (a contradiction to the definition of $b$).
[*Case 2:*]{} $b_j < a_j$. The are four sub-cases to consider:
- $c'_i k'_i \le b_j < a_j$: In this case it can be shown that if $i$ declares $c_i$ then $a$ will not be chosen (simply by showing that $b$ comes before $a$ in the lexicographic order). This contradicts the definition of $a$.
- $b_j < c'_i k'_i < a_j$: The same analysis as in the previous subcase applies to this sub-case as well.
- $b_j < a_j < c'_i k'_i$: If $i$ declares $c'_i$ then $a$ is chosen and not $b$ (it can easily be shown that $a$ comes before $b$ in the lexicographic order). This contradicts the definition of $b$.
- $b_j < a_j = c'_i k'_i$: Clearly, here $c'_i k_i < a_j = c'_i k'_i$. If $i$ declares $c'_i$ and $b$ is chosen, then $a_{j+1}\ge b_j$. However, if this is the case then $a$ will not be chosen if $i$ decreases its cost value and declares $c_i < c'_i$. A contradiction.
**Approximability of the Multidimensional Case**
Feigenbaum et al. [@FPSS] present a truthful polynomial-time VCG mechanism that always outputs the *cost-minimizing tree* (a tree that minimizes the total sum of costs incurred for the packets sent to $d$). We begin our discussion on the multidimensional version of the workload minimization problem by showing that this VCG mechanism obtains an $n$-approximation for the multidimensional version of our problem (and hence also for the single-dimensional version) in polynomial time.
There is a truthful polynomial-time deterministic $n$-approximation mechanism for the workload minimization problem in inter-domain routing.
We prove that any mechanism that minimizes the total cost provides an $n$-approximation to the minimal workload. Hence, the mechanism of [@FPSS] obtains the required approximation ratio.
Denote by $T$ the cost-minimizing routing-tree and by $T^*$ the workload-minimizing routing-tree. Let $C(T)$ and $C(T^*)$ be the total costs of $T$ and $T^*$, respectively. Let $W_i(T^*)$ be the workload on node $i$ in $T^*$, and let $W(T^*)= \max_i W_i(T^*)$. Notice that $W(T^*)$ is the value of optimal solution for the workload-minimization problem (by the definition of $T^*$). For the sake of contradiction, assume that $W(T^*) < \frac{C(T)}{n}$. Observe, that $C(T^*) = \Sigma_i \ W_i(T^*)$ (by a simple summation argument). However, if this is the case then $C(T^*)=\Sigma_i \ W_i(T^*) \le n\times W(T^*) = C(T)$. This contradicts the optimality of $T$ for the cost-minimization problem.
Unfortunately, it can be shown that any mechanism that minimizes the total cost (and in particular the mechanism in [@FPSS]) cannot obtain a good approximation ratio.
Any mechanism that minimizes the total cost of the routing tree cannot achieve an approximation ratio strictly better than $n$ for the workload minimization problem in inter-domain routing.
Consider the routing instance in figure $1$. Each source node has a single packet it wishes to send to the destination. The number beside every directed link $(u,v)$ in these figures represents the cost $u$ incurs for transferring a packet to $v$. Observe, that any total-cost minimizing mechanism would choose the routing tree in which both $y$ and $z$ send packets through $x$, and $x$ forwards packets directly to $d$. This means that the workload on $x$ is $3$. However, if all nodes chose to send their packets directly to $d$ we would reach a workload of $1+\epsilon$. Clearly, the example in figure $1$ can be generalized to $n$ source nodes. Notice also that a similar example can be used to show that the claim holds for singe-dimensional problems. The idea is to replace the link $(y, d)$ with the links $(y, y'), (y', d)$ and the link $(z, d)$ with $(z, z'), (z', d)$ while $c_y(y, y')=c_z(z, z')=0$, and $c_{y'}=c_{z'}=1+\epsilon$ assuming $x, y, z$ have each a single packet to send, and $y', z'$ have no packets to send to the destination.
Therefore, there exists a tradeoff between the goal of minimizing the total-cost and the goal of minimizing the workload. It would be interesting to construct a truthful mechanism that optimizes (or at least closely approximates) the minimal workload. We present two negative results for this problem (see the Appendix \[LBs-Inter-Domain\]). In a follow-up work, Gamzu [@Gamzu] improved our truthful lower bound for minimizing the workload in inter-domain routing from $\frac{1 + \sqrt{5}}{2} \approx 1.618$ to 2, and our randomized truthful lower bound from $\frac{3+\sqrt{5}}{4} \approx 1.309$ to 2.
On Non-Utilitarian Fairness {#sec-fairness}
===========================
In many scenarios, we desire to implement a utilitarian social choice function. A well studied example of such a scenario is social-welfare maximization in combinatorial auctions. In a combinatorial auction we wish to allocate $m$ indivisible items to $n$ players. Each player $i$ is defined by a valuation function $v_i:2^{[m]}\rightarrow \mathbb{R_{+}}$. We assume that for every $i$ $v_i(\emptyset)=0$ (free disposal) and for every two bundles of items $S,T\subseteq [m]$ such that $S\subseteq T$ $v_i(S)\le v_i(T)$ (monotonicity). The goal is to partition the items into disjoint sets $S_1,...,S_n$ such that the expression $\Sigma_i \ v_i(S_i)$ is maximized.
Utilitarian functions represent the overall satisfaction of the players, as they maximize the sum of players’ values. This notion of *fairness* is but one of several that have been considered (explicitly and implicitly) in mathematical, economic and computational literature. A well known example of non-utilitarian fairness is the cake-cutting problem, presented by the Polish school of mathematicians in the 1950’s (Steinhaus, Banach, Knaster [@Stein]). In recent years, fair allocations of indivisible items (other than social-welfare maximization) have also been studied [@LMN; @LMSS] (these can be regarded as discrete versions of the cake-cutting problem).
In this section, we discuss three general notions of non-utilitarian fairness – Max-Min fairness, Min-Max fairness, and envy-minimization. We prove several general results; In particular, we show that Max-Min fairness is inapproximable within *any ratio*, even for extremely restricted special cases. In contrast, we show that Min-Max fairness (which is a generalization of both the scheduling and workload-minimization problems considered in this paper) can always be truthfully approximated via a simple VCG mechanism. Finally, we make use of our techniques to prove a lower bound for the envy-minimization problem.
**Max-Min Fairness**
The Max-Min social choice function is concerned with maximizing the value of the least satisfied player. Formally, for every $n$-tuple of $v_i$ valuations the Max-Min function assigns the alternative $a$ that maximizes the expression $\min_i \:
v_i(a)$.
[@LMN] proved that Max-Min fairness in allocations of indivisible items cannot be *optimally* implemented in a truthful manner. In recent years, non-truthful algorithms for this problem were designed [@BS06], as well as algorithms that settle for restricted notions of truthfulness [@BV05; @Golo]. We prove that no truthful deterministic mechanism can obtain *any* approximation ratio to the Max-Min fairness value. We prove this lower bound even for the case of $2$ players and $2$ items.
No truthful deterministic mechanism can obtain *any* approximation to the Max-Min fairness value in the allocation of indivisible items. This holds even for the case of $2$ players and $2$ items.
Let $c>0$. Consider an instance with two players $1,2$ and two goods $a,b$. Each player $i=1,2$ has an additive valuation function. $v_1(a)=2\ \ v_1(b)=\frac{1}{c}\ \ v_2(a)=4-\epsilon\ \
v_2(b)=1+\epsilon$. Note, that the optimal allocation assigns $a$ to player $1$ and $b$ to player $2$, thus obtaining a Max-Min value of $1+\epsilon$. Also note, that this allocation will also be chosen by any $c$-approximation mechanism.
We alter the valuation of player $2$ into $v'_2$ such that $v'_2(a)=\frac{1}{c}\ \ v'_2(b)=\frac{1}{c^2}-\epsilon$. The optimal Max-Min value is now $\frac{1}{c}$. Observe that any $c$-approximation mechanism must assign item $b$ to player $1$ and item $a$ to player $2$. However, if this happens we have that:
$$(1+\epsilon)+\frac{1}{c}=v_2(b)+v'_2(a)<v_2(a)+v'_2(b)=(4-\epsilon)+\frac{1}{c^2}$$
This violates weak monotonicity, and so no truthful $c$-approximation mechanism exists. Since this is true for any $c>0$ the theorem follows
**Min-Max Fairness**
Min-Max fairness can be thought of as the dual notion of Max-Min fairness. It is relevant in settings in which each player incurs a cost for every chosen alternative. The Min-Max social choice function is concerned with minimizing the cost incurred by the least satisfied player. Formally, for every $n$-tuple of $v_i$ valuations the Max-Min function assigns the alternative $a$ that minimizes the expression $\max_i \: v_i(a)$.
Observe, that both the scheduling problem and the workload-minimization problem discussed in this paper, are in fact special cases of this notion of fairness. Studying Max-Min fairness in this more abstract setting enables us to state this simple observation – any Min-Max social-choice function can be truthfully approximated within a factor of $n$ (recall that $n$ is the number of players) by a simple VCG mechanism. Since the best currently known approximation-mechanisms for both scheduling and workload-minimization are VCG-based, this result can be viewed as a generalization of both.
Let $f$ be a Min-Max social choice function. Then, there exists a truthful deterministic mechanism that for every $n$-tuple of valuations $v_1,...,v_n$ outputs an alternative $a$ such that $\max_i\: v_i(a)$ is an $n$-approximation to the value of the solution $f$ outputs for these valuations.
Let $v_1,...,v_n$ be the valuation function of the players, and let $A$ be the set of alternatives. Let $b$ be the allocation $f$ outputs for $v_1,...,v_n$. Consider the VCG mechanism that minimizes the total cost the players incur. The truthfulness of this mechanism is guaranteed by the VCG technique. Let $a$ be the allocation this mechanism outputs.
Assume, by contradiction, that $\max_i \: v_i(a)>n\times \max_i \: v_i(b)$. If this is the case, then $$\Sigma_i \ v_i(b)\le n\times \max_i \: v_i(b)< \max_i \: v_i(a)\le
\Sigma_i \ v_i(a)$$
However, this contradicts the optimality of $a$ with regards to cost-minimization.
**Envy-Minimization**
Lipton, Markakis, Mossel, and Saberi [@LMSS] presented the problem of finding envy-minimizing allocations of indivisible items. An envy-minimizing allocation of items is a partition of the $m$ items into disjoint sets $S_1,...,S_n$ (player $i$ is assigned $S_i$) that minimizes the expression $\max_{i,j}\ v_i(S_j)-v_i(S_i)$ (over all possible allocations). Intuitively, we wish to minimize the maximal envy a player might feel by comparing his value for a bundle of items given to another player to the value he assigns the items allocated to him. [@LMSS] proves several approximability results for this problem. The parameter considered in [@LMSS] is the *maximal marginal utility*.
The maximal marginal utility $\alpha$ is defined as follows:
$\alpha=\max_{i\in [n], j\in [m], S\subseteq [m]}\ v_i(S\bigcup
\{j\})-v_i(S)$.
That is, $\alpha$ is the maximal value by which the value of a player increases when one good is added to his bundle. [@LMSS] proves that there always exists an allocation of items with an envy value of at most $\alpha$. [@LMSS] also exhibits a universally truthful randomized mechanism that obtains an approximation of $O(\sqrt{\alpha}n^{\frac{1}{2}+\epsilon})$ w.h.p. for large values of $n$.
Lipton et al. are interested in the question of whether there are truthful mechanisms that produce allocations with minimal or bounded envy. [@LMSS] shows that no truthful mechanism can guarantee a perfect solution (minimum envy). We strengthen this lower bound by showing that no truthful deterministic mechanism can guarantee an allocation that has an envy value within $\alpha$ from optimal.
\[envy-thm\] No truthful deterministic mechanism can guarantee an allocation that has an envy within $\alpha$ from optimal.
Let $M$ be a truthful deterministic mechanism for this problem. Consider an instance with $2$ players and $3$ goods. Each player $i=1,2$ has the same additive valuation function $v_i$ that assigns any of the single items a value of $1$. Observe, that in this case $\alpha=1$. Notice, that the minimal envy for this instance is $1$ (simply assign two items to one of the players and one item to the other). Hence, if $M$ assigns all items to one of the players the envy of the other is precisely $3$, which is a $2\alpha$ distance from optimal.
We are left with the case in which one of the players receives two items and the other is given one item. Assume w.l.o.g. that player $1$ is given items $1,2$ and player $2$ is given item $3$. We now change the valuation function of player $1$ into the following additive valuation:
$$v_1(j)=\left\{\begin{array}{ll}
1+\epsilon, & j=1,2 \\
\epsilon, & j=3 \\
\end{array}\right.$$
Observe, that now $\alpha=1+\epsilon$. Also observe that the minimal envy for this new instance is $0$ (e.g. assign item $1$ to player $1$ and items $2,3$ to player $2$). However, the reader is encouraged to verify that the monotonicity of $M$ dictates that the allocation remain the same even after the alteration of the valuation of player $1$. Therefore, we end up with an allocation in which the envy of player $2$ is $1$. As $\alpha=1+\epsilon$ this is arbitrarily close to $\alpha$.
[@LMSS] also considers the social-choice function that aims to minimize the envy-ratio (defined therein) of the chosen allocation (over all possible allocations of goods). Using similar arguments to those in the proof of Theorem \[envy-thm\] one can easily show that no truthful deterministic mechanism for the envy-ratio minimization problem has an approximation ratio better than $2$. This result too can easily be extended to a weaker lower bound for truthful randomized mechanisms.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Ron Lavi, Noam Nisan, Chaitanya Swamy, and Amir Ronen for helpful discussions.
[10]{}
The communication requirements of efficient allocations and supporting prices. , 129(1):192 – 224, 2006.
Nir Andelman, Yossi Azar, and Motti Sorani. Truthful approximation mechanisms for scheduling selfish related machines. , 40(4):423–436, 2007.
Aaron Archer, Christos H. Papadimitriou, Kunal Talwar, and [É]{}va Tardos. An approximate truthful mechanism for combinatorial auctions with single parameter agents. , 1:129–150.
Aaron Archer and [É]{}va Tardos. Truthful mechanisms for one-parameter agents. In [*[IEEE]{} Symposium on Foundations of Computer Science*]{}, pages 482–491, 2001.
Aaron Archer and [É]{}va Tardos. Frugal path mechanisms. , 3(1), 2007.
Itai Ashlagi, Shahar Dobzinski, and Ron Lavi. Optimal lower bounds for anonymous scheduling mechanisms. , 37(2):244–258, 2012.
Nikhil Bansal and Maxim Sviridenko. The santa claus problem. In [*Proceedings of the 38th Annual [ACM]{} Symposium on Theory of Computing, Seattle, WA, USA, May 21-23, 2006*]{}, pages 31–40, 2006.
Yair Bartal, Rica Gonen, and Noam Nisan. Incentive compatible multi unit combinatorial auctions. In [*Proceedings of the 9th Conference on Theoretical Aspects of Rationality and Knowledge (TARK-2003), Bloomington, Indiana, USA, June 20-22, 2003*]{}, pages 72–87, 2003.
Ivona Bez[á]{}kov[á]{} and Varsha Dani. Allocating indivisible goods. , 5(3):11–18, 2005.
S. Bikhchandani, S. Chatterji, R. Lavi, A. Mu’alem, N. Nisan, and A. Sen. Weak monotonicity characterizes deterministic dominant strategy implementation. , 74(4):1109–1132, July 2006.
Shuchi Chawla, David L. Malec, and Balasubramanian Sivan. The power of randomness in bayesian optimal mechanism design. , 91:297–317, 2015.
Xujin Chen, Donglei Du, and Luis Fernando Zuluaga. Copula-based randomized mechanisms for truthful scheduling on two unrelated machines. In [*Algorithmic Game Theory - 6th International Symposium, [SAGT]{} 2013, Aachen, Germany, October 21-23, 2013. Proceedings*]{}, pages 231–242, 2013.
George Christodoulou, Elias Koutsoupias, and Angelina Vidali. A lower bound for scheduling mechanisms. In [*Proceedings of the Eighteenth Annual [ACM-SIAM]{} Symposium on Discrete Algorithms, [SODA]{} 2007, New Orleans, Louisiana, USA, January 7-9, 2007*]{}, pages 1163–1170, 2007.
George Christodoulou and Annam[á]{}ria Kov[á]{}cs. A deterministic truthful [PTAS]{} for scheduling related machines. , 42(4):1572–1595, 2013.
E. H. Clarke. Multipart pricing of public goods. , 11:17–33, 1971.
Amit Daniely, Michael Schapira, and Gal Shahaf. Inapproximability of truthful mechanisms via generalizations of the [VC]{} dimension. In [*Proceedings of the Forty-Seventh Annual [ACM]{} on Symposium on Theory of Computing, [STOC]{} 2015, Portland, OR, USA, June 14-17, 2015*]{}, pages 401–408, 2015.
Constantinos Daskalakis and S. Matthew Weinberg. Bayesian truthful *Mechanisms* for job scheduling from bi-criterion approximation *Algorithms*. In [*Proceedings of the Twenty-Sixth Annual [ACM-SIAM]{} Symposium on Discrete Algorithms, [SODA]{} 2015, San Diego, CA, USA, January 4-6, 2015*]{}, pages 1934–1952, 2015.
Gabrielle Demange and David Gale. The strategy structure of two-sided matching markets. , 53(4):873–88, July 1985.
Peerapong Dhangwatnotai, Shahar Dobzinski, Shaddin Dughmi, and Tim Roughgarden. Truthful approximation schemes for single-parameter agents. , 40(3):915–933, 2011.
Shahar Dobzinski. An impossibility result for truthful combinatorial auctions with submodular valuations. In [*Proceedings of the 43rd [ACM]{} Symposium on Theory of Computing, [STOC]{} 2011, San Jose, CA, USA, 6-8 June 2011*]{}, pages 139–148, 2011.
Shahar Dobzinski and Shaddin Dughmi. On the power of randomization in algorithmic mechanism design. , 42(6):2287–2304, 2013.
Shahar Dobzinski and Noam Nisan. Mechanisms for multi-unit auctions. In [*Proceedings 8th [ACM]{} Conference on Electronic Commerce (EC-2007), San Diego, California, USA, June 11-15, 2007*]{}, pages 346–351, 2007.
Shahar Dobzinski, Noam Nisan, and Michael Schapira. Approximation algorithms for combinatorial auctions with complement-free bidders. , 35(1):1–13, 2010.
Shahar Dobzinski, Noam Nisan, and Michael Schapira. Truthful randomized mechanisms for combinatorial auctions. , 78(1):15–25, 2012.
Joan Feigenbaum, David R. Karger, Vahab S. Mirrokni, and Rahul Sami. Subjective-cost policy routing. , 378(2):175–189, 2007.
Joan Feigenbaum, Christos Papadimitriou, Rahul Sami, and Scott Shenker. A $\mathrm{BGP}$-based mechanism for lowest-cost routing. , 18:61–72, 2005.
Joan Feigenbaum, Vijay Ramachandran, and Michael Schapira. Incentive-compatible interdomain routing. , 23(5-6):301–319, 2011.
Joan Feigenbaum, Rahul Sami, and Scott Shenker. Mechanism design for policy routing. , 18:293–305, 2006.
Amos Fiat, Andrew Goldberg, Jason Hartline, and Anna Karlin. Competitive generalized auctions. In [*Proc. of the 34th ACM Symposium on Theory of Computing (STOC’02)*]{}, 2002.
Iftah Gamzu. Improved lower bounds for non-utilitarian truthfulness. , 412(7):626–632, 2011.
Daniel Golovin. Max-min fair allocation of indivisible goods, 2005. Technical Report CMU-CS-05-144, Carnegie Mellon University.
Pavlov Gregory. Optimal mechanism for selling two goods. , 11(1):1–35, February 2011.
Theodore Groves. Incentives in teams. , 41(4):617–631, 1973.
Ron Holzman, Noa Kfir-Dahav, Dov Monderer, and Moshe Tennenholtz. Bundling equilibrium in combinatorial auctions. , 47:104–123, 2004.
E. Horowitz and S. Sahni. Exact and approximate algorithms for scheduling nonidentical processors. , 23:317–327, 1976.
Anna R. Karlin, David Kempe, and Tami Tamir. Beyond [VCG:]{} frugality of truthful mechanisms. In [*46th Annual [IEEE]{} Symposium on Foundations of Computer Science [(FOCS]{} 2005), 23-25 October 2005, Pittsburgh, PA, USA, Proceedings*]{}, pages 615–626, 2005.
Elias Koutsoupias and Angelina Vidali. A lower bound of 1+*[$\varphi$]{}* for truthful scheduling mechanisms. , 66(1):211–223, 2013.
Ron Lavi, Ahuva Mu’alem, and Noam Nisan. Towards a characterization of truthful combinatorial auctions. In [*44th Symposium on Foundations of Computer Science [(FOCS]{} 2003), 11-14 October 2003, Cambridge, MA, USA, Proceedings*]{}, pages 574–583, 2003.
Ron Lavi and Chaitanya Swamy. . , 67(1):99–124, 2009. Special Section of Games and Economic Behavior Dedicated to the 8th ACM Conference on Electronic Commerce.
Ron Lavi and Chaitanya Swamy. Truthful and near-optimal mechanism design via linear programming. , 58(6):25, 2011.
Daniel Lehmann, Liadan O’Callaghan, and Yoav Shoham. Truth revelation in approximately efficient combinatorial auctions. , 49(5):577–602, 2002.
Jan Karel Lenstra, David B. Shmoys, and [É]{}va Tardos. Approximation algorithms for scheduling unrelated parallel machines. , 46:259–271, 1990.
Richard J. Lipton, Evangelos Markakis, Elchanan Mossel, and Amin Saberi. On approximately fair allocations of indivisible goods. In [*Proceedings 5th [ACM]{} Conference on Electronic Commerce (EC-2004), New York, NY, USA, May 17-20, 2004*]{}, pages 125–131, 2004.
Pinyan Lu and Changyuan Yu. Randomized truthful mechanisms for scheduling unrelated machines. In [*Internet and Network Economics, 4th International Workshop, [WINE]{} 2008, Shanghai, China, December 17-20, 2008. Proceedings*]{}, pages 402–413, 2008.
A. Mas-Collel, W. Whinston, and J. Green. . Oxford university press, 1995.
Ahuva Mu’alem. Fair by design: Multidimensional envy-free mechanisms. , 88:29 – 46, 2014.
Ahuva Mu’alem and Noam Nisan. Truthful approximation mechanisms for restricted combinatorial auctions. , 64:612–631, 2008.
Noam Nisan. Chapter 9 - algorithmic mechanism design: Through the lens of multiunit auctions. volume 4 of [*Handbook of Game Theory with Economic Applications*]{}, pages 477 – 515. Elsevier, 2015.
Noam Nisan and Amir Ronen. Algorithmic mechanism design. , 35:166–196, 2001.
Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V. Vazirani (eds.). . Cambridge University Press, 2007.
M. J. Osborne and A. Rubinstein. . MIT press, 1994.
Y. Shoham P. Cramton and R. Steinberg (eds.). . MIT Press, 2006.
John Rawls. . Cambridge, MA: Belknap Press of Harvard University Press, 1971.
H. Steinhaus. The problem of fair division. , 16:101–104, 1948.
W. Vickrey. Counterspeculation, auctions and competitive sealed tenders. , pages 8–37, 1961.
Andrew Chi[-]{}Chih Yao. Probabilistic computations: Toward a unified measure of complexity (extended abstract). In [*18th Annual Symposium on Foundations of Computer Science, Providence, Rhode Island, USA, 31 October - 1 November 1977*]{}, pages 222–227, 1977.
Appendix
========
A Randomized Truthful Approximation Mechanism for the Scheduling Problem with Unrelated Machines {#upper-bound}
------------------------------------------------------------------------------------------------
Nisan and Ronen [@NR] present a truthful deterministic mechanism that obtains an $m$-approximation. For the case of $2$ machines, they exhibit a universally-truthful randomized mechanism that obtains an approximation of $\frac{7}{4}$ (this bound was later improved to 1.58606 by [@LY; @Chen]). We generalize this result by presenting a universally-truthful randomized mechanism that obtains an approximation-ratio of $0.79303m$. We now turn to the description of our mechanism:
**Input:** $m$ valuations $v_i$.
**Output:** An allocation $T=T_1$,...,$T_m$ of tasks, and payments $p_1,...,p_m$ such that $T$ has a makespan value which is a $\frac{7m}{8}$-approximation to the optimal makespan value, and the payments induce truthfulness.
[**The Mechanism:**]{}
1. For every machine $i$ let $T_i=\emptyset$ and $p_i=0$.
2. Partition the set of machines into two sets $S_1=\{1,...,\frac{m}{2}\}$ and $S_2=\{\frac{m}{2}+1,...,m\}$.
3. For each task $j=1,...,n$ perform the following actions:
- Let $v^1=\min_{i\in S_1}\ v_i(j)$ and let $I=argmin_{i\in S_1}\ v_i(j)$.
- Let $v'^1=\min_{i\in S_1-\{I\}}\ v_i(j)$.
- Let $v^2=\min_{i\in S_2}\ v_i(j)$ and let $II=argmin_{i\in S_2}\ v_i(j)$.
- Let $v'^2=\min_{i\in S_2-\{II\}}\ v_i(j)$.
- Randomly and uniformly choose a value $R\in \{0,1\}$.
- If $R=0$ and $v^1\le \frac{4}{3}v^2$ set $T_I= T_I \bigcup\; \{ j \}$ and set $p_{I}=p_{I}+\min\{v'^1,\frac{4}{3}v^2\}$.
- If $R=0$ and $v^1> \frac{4}{3}v^2$ set $T_{II}=T_{II}\bigcup\; \{ j \}$ and set $p_{II}=p_{II}+\min\{v'^2,\frac{3}{4}v^1\}$.
- If $R=1$ and $v^2\le \frac{4}{3}v^1$ set $T_{II}= T_{II}\bigcup\; \{ j \}$ and set $p_{II}=p_{II}+\min\{v'^2,\frac{4}{3}v^1\}$.
- If $R=1$ and $v^2> \frac{4}{3}v^1$ set $T_I=\{T_I\bigcup\;\{ j \}$ and set $p_{I}=p_{I}+\min\{v'^1,\frac{3}{4}v^2\}$.
4. Allocate each machine $i$ the tasks in $T_i$, and pay it a sum of $p_i$.
If $m$ cannot be divided by $2$ simply add the extra machine to either $S_1$ or $S_2$.
There exists a universally truthful randomized mechanism for the scheduling problem that obtains an approximation ratio of $\frac{7m}{8}$.
We prove the theorem for the case that $m$ can be divided by $2$. The proof for the other case is similar. Our proof relies on the proof of Nisan and Ronen [@NR]. Observe, that the utility of each machine after the algorithm finishes is the sum of its utilities for the different tasks. Hence, it is sufficient to prove that for each individual task a machine has no incentive to lie. As in [@NR], this is guaranteed because the allocation of each task is in fact a weighted VCG mechanism (see [@NR] for further explanations), which is known to be truthful. Hence, this mechanism is universally truthful.
We now need to prove that the approximation ratio guaranteed by the mechanism is indeed $\frac{7m}{8}$. Let $A$ be an instance of the scheduling problem with $n$ tasks, and with $m$ machines that have the valuation functions $v_1,...,v_m$. We define an instance $B$ of scheduling problem with $n$ tasks, and with $2$ machines that have the valuation function $v'_1,v'_2$, in the following way: For all $j\in [n]$ $v'_1(j)=min_{i\in S_1}\ v_i(j)$. Similarly, for all $j\in [n]$ $v'_2(j)=min_{i\in S_2}\ v_i(j)$. We denote by $M(A)$ and by $M(B)$ the makespan values our mechanism generates for $A$ and $B$ respectively. We denote by $O(A)$ and by $O(B)$ the optimal makespan values for $A$ and $B$ respectively.
First, notice that $M(A)\le M(B)$. This is because applying our mechanism to $B$ results in the same makespan value as applying it to $A$ in the worst-case scenario in which tasks are always assigned to the same machines in $S_1$ and in $S_2$. It also holds that $M(B)\le \frac{7}{4}O(B)$ because in the case that there are only two machines our mechanism is precisely that of [@NR], which guarantees a $\frac{7}{4}$ approximation ratio. We now have that $M(A)\le \frac{7}{4}O(B)$. All that is left to show is that $O(B)\le \frac{m}{2}O(A)$. Consider the optimal allocation of tasks for $A$. By giving all tasks assigned to machines in $S_1$ to machine $1$ in $B$, and allocating all tasks assigned to machines in $S_2$ to machine $2$ in $B$, we end up with a makespan value for $B$ that is at most $\frac{m}{2}O(A)$. The theorem follows.
The above mechanism can be straightforwardly improved to obtain an approximation-ratio of $0.79303m$ by using [@LY; @Chen].
Lower Bounds for minimizing the workload in inter-domain routing {#LBs-Inter-Domain}
----------------------------------------------------------------
\[thm-work\] No truthful deterministic mechanism for minimizing the workload in inter-domain routing can obtain an approximation ratio better than $\frac{1 + \sqrt{5}}{2} \approx 1.618$.
This proof is similar to the proof of Theorem \[deterministic-bound\]. To prove the lower bound consider the instances of the workload-minimization problem with $3$ source nodes $I,II,III$ depicted in figures $2$ and $3$. Each source node has a single packet it wishes do send to the destination. The number beside every directed link $(u,v)$ in these figures represents the cost $u$ incurs for transferring a packet to $v$. Denote the instance in figure $2$ by $INS$ and the instance in figure $3$ by $INS'$. Observe that only the cost function of node $I$ is different in $INS$ and $INS'$. We denote the cost function of $I$ in $INS$ by $c_I$ and his cost function in $INS'$ by $c'_I$.
Assume, by contradiction, that $M$ is a truthful deterministic mechanism that obtains an approximation ratio better than $\frac{1 + \sqrt{5}}{2}$. Observe, that for instance $INS$ $M$ must direct the traffic originating in node $I$ through node $II$ (otherwise this contradicts the fact that $M$ obtains an approximation ratio better than $\frac{1+\sqrt{5}}{2}$). Similarly, for instance $INS'$ $M$ must direct the traffic originating in node $I$ through node $III$. However, this violates the monotonicity of $M$ as $1+1.618=c_I((I,II))+c'_I((I,III))>c_I((I,II))+c'_I((I,III))=(1.618)^2 -\epsilon$.
No universally truthful randomized mechanism for minimizing the workload in inter-domain routing can obtain an approximation ratio better than $\frac{3+\sqrt{5}}{4}$.
This proof is similar to the proof of Theorem \[universal-bound\]. We define $INS$ and $INS'$ as in the proof of Theorem \[thm-work\]. Consider the uniform distribution over $INS$ and $INS'$. Let $M$ be a truthful deterministic mechanism. As shown in the proof of Theorem \[thm-work\], $M$ cannot achieve an approximation better than $\frac{1+\sqrt{5}}{2}$ on both $INS$ and $INS'$ due to its monotonicity. Therefore, the expected approximation of $M$ is at least $\frac{1}{2}\times 1 + \frac{1}{2}\times \frac{1 + \sqrt{5}}{2} \approx 1.309$.
[^1]: The current paper supersedes `"`Setting Lower Bounds on Truthfulness`"` that appeared as an extended abstract in the [*Proceedings of the 18th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA-07)*]{}, pages 1143-1152, 2007. The current version includes a new lower bound result for Bayesian Incentive compatible Mechanisms.
[^2]: Computer Science Department, Technion, Haifa, Israel. [email protected].
[^3]: School of Computer Science and Engineering, Hebrew University, Jerusalem, Israel. [email protected].
[^4]: It is well known (e.g., [@Mas]) that, without loss of generality, we can limit ourselves to only considering direct-revelation incentive-compatible mechanisms, also known as truthful mechanisms. In such mechanisms participants are always rationally motivated to correctly report their private information.
[^5]: As opposed to single-dimensional settings in which the private information of each player essentially consists of a single numerical parameter.
[^6]: This is arguably a mild restriction in multidimensional settings (see e.g. [@Pavlov-2011; @Chawla-GEB-15] and references therein).
[^7]: Notice that our lower bound for randomized truthful mechanisms is almost tight for the two-machine case (Theorem \[universal-bound\] vs. [@Chen]).
[^8]: Notice that our upper bounding technique is quite general: by using [@LY; @Chen], the ratio of $\frac{7m}{8}$ can be improved to $0.79303m$.
[^9]: In the formulation of the problem in [@FPSS], a node does not incur a cost for packets that originate in that node. However, since we are interested in workload minimization, this is not the case in our formulation. Additionally, as we are interested in proving lower bounds we can restrict our attention to the model in which the number of packets $t_i$s are common knowledge.
| {
"pile_set_name": "ArXiv"
} |
=1.0cm
[**Phase Transitions: Summary of Discussion Session II of Camerino 2005**]{}
Abstract
[ *A brief report of the topics which received attention during the discussion session II of the International Workshop on Symmetries and Low-Energy Phase Transitions in Nuclear-Structure Physics, held in Camerino on 9-11 October 2005, is given. These include special solutions of the Bohr Hamiltonian for various potentials, the study of triaxial shapes and of degrees of freedom other than the quadrupole one (octupole, scissors), as well as the search for experimental manifestations of the critical point symmetries E(5) and X(5), and of the recently proposed critical point supersymmetry E(5/4).* ]{}
[**1. Solutions of the Bohr Hamiltonian with different potentials**]{}
In the original E(5) [@IacE5] and X(5) [@IacX5] papers, an infinite square well potential in $\beta$ is used in accordance to the expectation that the potential at the critical point has to be flat and scale invariant.
[**1.1 Solutions related to E(5)**]{}
Different potentials used in the E(5) framework include a well of finite depth [@Capriow], the sextic oscillator [@sextic], Coulomb-like and Kratzer-like potentials [@FortE5], a linear potential [@Fortrev], Davidson potentials [@Dav; @Elliott; @Rowe] of the form $\beta^2 + \beta_0^4 /\beta^2$, where $\beta_0$ is the position of the minimum of the potential [@varPLB; @varPRC], the $\beta^4$ potential [@Ariasb4; @Ariasb4b], as well as the $\beta^6$ and $\beta^8$ potentials [@E5] (the $\beta^2$ potential [@Bohr; @Wilets; @Dussel] corresponding to the well known U(5) case [@IBM]). A hybrid model employing a harmonic oscillator for $L\leq 2$ and an infinite square well potential for $L\geq 4$ has also been developed [@Raduta]. A recent review of this topic has been given in [@Fortrev].
L. Fortunato reported on a detailed study of the potential [@vanRoos] $$u(\beta) = {(1-\eta)\over 2} \beta^2 + {\eta\over 4} (1-\beta^2)^2,$$ which is known [@vanRoos] to correspond to the U(5)-O(6) transition region of IBM, U(5) being obtained for $\eta=0$ and O(6) being obtained for $\eta=\infty$, the critical point occuring at $\eta_C=0.5$ and corresponding to a $\beta^4$ potential, in agreement with the findings of [@Ariasb4; @Ariasb4b]. This potential has also been considered in [@IacE5; @E5], but only within the region $0\leq \eta \leq 1$. Fortunato allowed $\eta$ to obtain higher values, observing that best agreement between E(5) and the potential of Eq. (1) is obtained for $\eta\simeq 5$, in which case the potential develops a bump at the center. It was remarked that B(E2)s should be calculated before final conclusions can be drawn. It is however remarkable that a bump results in the X(5) framework when using an effective $\beta$ deformation, determined by variation after angular momentum projection and two-level mixing [@LeviX5], as well as in Nilsson-Strutinsky-BCS calculations [@ZhangSm] for $^{152}$Sm and $^{154}$Gd, which are good examples of X(5), as will be seen in subsec. 3.2 .
[**1.2 Solutions related to X(5)**]{}
Different potentials used in the X(5) framework include a potential with linear sloped walls [@Capriosl], the confined $\beta$-soft (CBS) rotor model [@Pietr1; @Pietr2] (which utilizes an infinite square well potential displaced from zero), Coulomb-like and Kratzer-like potentials [@FortX5], Davidson potentials [@varPLB; @varPRC], as well as the $\beta^2$, $\beta^4$, $\beta^6$, and $\beta^8$ potentials [@X5]. A recent review of this topic has been given in [@Fortrev].
The special case in which $\gamma$ is frozen to $\gamma=0$, while an infinite square well potential is used in $\beta$, leads to an exactly separable three-dimensional model, which has been called X(3) [@X3].
The approximate separation of variables used in X(5) has been tested recently through exact numerical diagonalization of the Bohr Hamiltonian [@Caprio72], using a recently introduced [@Rowe735; @Rowe45; @Rowe753] computationally tractable version of the Bohr–Mottelson collective model.
[**1.3 Triaxial solutions**]{}
In the E(5) case [@IacE5] the potential is supposed to be $\gamma$-independent, while in the X(5) case [@IacX5] the potential is supposed to be of the form $u(\beta)+u(\gamma)$, with $\gamma$ obtaining values close to $\gamma=0$. Another family of solutions of the Bohr Hamiltonian can be obtained by allowing the potential to be of the form $u(\beta)+u(\gamma)$ and confining $\gamma$ around $\gamma=30^{\rm o}$. An infinite square well potential in this case leads to the Z(5) solution [@Z5] if $\gamma$ is allowed to vary around the $\gamma=30^{\rm o}$ value (the $\gamma$-soft case), and to the Z(4) solution [@Z4] if $\gamma$ is fixed to this value, being a parameter and not a variable any more (the $\gamma$-rigid case). In the $\gamma$-soft case Coulomb-like and Kratzer-like potentials have been used [@Forttri], while a solution similar to Z(5), but with a different $\gamma$-potential, has been given in [@Jolos]. A recent review of the various solutions of the Bohr Hamiltonian has been given in [@Fortrev].
The characteristic differences among the various models were discussed in detail. It was pointed out that E(5) and Z(4) possess very similar ground state bands and $\beta_1$ bands, as well as similar intraband and interband B(E2)s obeying the same selection rules. The main difference between the two models occurs in the $\gamma_1$ band, which exhibits opposite odd-even staggering in the two models, since in E(5) its levels are exactly grouped as $2^+$, ($3^+$, $4^+$), ($5^+$, $6^+$), …, which is a feature of the underlying O(5) subalgebra, while in Z(4) the approximate grouping is ($2^+$, $3^+$), ($4^+$, $5^+$), …, which is a feature of rigid triaxial models [@Davydov]. It is known [@Casten] that $\gamma$-soft and $\gamma$-rigid models provide very similar results if $\gamma_{rms}$ of the former equals $\gamma_{rigid}$ of the latter. Since no clear examples of triaxial nuclei have been identified, it is expected that Z(4) \[and Z(5)\] would provide results in reasonable agreement with experiment only when this condition is approximately fulfilled.
The proton-neutron triaxiality occuring in the SU(3)$^*$ limit [@Dieperink] of IBM-2 was also discussed, in connection to relevant shape phase transitions occuring in the study of the phase structure of IBM-2 [@AriasPRL; @CaprioPRL; @CaprioAP]. It was pointed out that the main features of proton-neutron triaxiality are a low-lying K=2 band and B(E2)s resembling very closely the predictions of the Davydov model [@Davydov] with $\gamma=30^{\rm o}$. These findings are in agreement with the assumption of $\gamma\simeq 30^{\rm o}$ made in Z(5) and Z(4), as well as with the prediction of low-lying $\gamma_1$-bands in these models.
For experimental manifestations of the transition towards triaxiality, $^{168,170}$Er were suggested as possible candidates.
It was also pointed out that Z(4), as well as X(3), are solutions in which the separation of variables is exact. Therefore it might be easier to clarify their algebraic structure, while this task seems more difficult in the cases of X(5) and Z(5), where separation of variables is approximate.
[**2. Additional degrees of freedom**]{}
[**2.1 Octupole deformation**]{}
It is generally accepted that octupole deformation ($\beta_3$) has to be taken into account simultaneously with the quadrupole deformation ($\beta_2$), which plays a dominant role. In the approaches developed so far for the description of the transition from octupole vibrations to octupole deformation in the light actinides, either only axially symmetric shapes are taken into account [@AQOA], or slight triaxiality is allowed [@Bizzoct]. In the Analytic Quadrupole Octupole Axially symmetric (AQOA) model [@AQOA] the angles are left out from the very beginning and a single axis is used for both the quadrupole and the octupole deformations, with no relative angle in between. Both methods lead to a satisfactory description of $^{226}$Th and $^{226}$Ra involving an infinite square well potential [@AQOA; @Bizzoct]. In these cases the $\beta_1$ bandhead is higher than the X(5) bandhead by a factor of almost two, in agreement with experiment. The question of a transition from octupole deformation to octupole vibrations in the rare earth region has also been raised, along with suggestions to examine if the N=90 isotones $^{150}$Nd and $^{152}$Sm, which are known to be the best examples of X(5) (see subsec. 3.2), are also critical with respect to the transition from octupole deformation to octupole vibrations, a question which remains open.
It was pointed out that one has to be very careful in separating out the intrinsic variables, as in the old work by Rohoziński [@Rohozinski]. When building the spdf-IBM [@Engel1; @Engel2] it became clear that separation of the intrinsic and angular variables was very difficult when taking into account only the $s$, $d$, and $f$ bosons, while it became easy when the $p$ boson was also included. The $p$ boson was also helpful for taking into account the center of mass motion. The detailed study of the phase space of spdf-IBM, in analogy to the similar study carried out recently for IBM-2 [@AriasPRL; @CaprioPRL; @CaprioAP] would be very helpful in clarifying shape phase transitions involving the octupole degree of freedom, but it is technically quite demanding [@Engel2].
It was also pointed out that one has to be very careful with the terminology, since octupole deformation (the merging of the ground state band and the nearby negative parity band into a single band with levels of alternating parity) occurs in nuclei which are near-vibrational or transitional from the quadrupole point of view, while octupole vibrations (negative parity bands built on bandheads corresponding to one or more quanta of octupole vibration, their levels thus lying systematically higher than the levels of the ground state band with similar angular momenta) are observed in nuclei which are well deformed from the quadrupole point of view.
Furthermore, it was pointed out that there is no [*a priori*]{} reason for criticality in the transition from octupole deformation to octupole vibrations to occur in the same nuclei, in which criticality in the transition from quadrupole vibrations to axial quadrupole deformation occurs. Such a conclusion has to be based on experimental evidence, since the relevant models contain many drastic approximations, made in order to make them analytically soluble.
[**2.2 Scissors mode**]{}
It is known [@Ziegler; @Ranga; @Herz] that there is a strong correlation between the low lying dipole excitations, referred to as the [*scissors mode*]{} [@Richter] and quadrupole deformation. On the basis of systematics of experimental data in the Xe, Ba, Nd, Sm, and Gd isotopic chains it has been argued [@Scheck] recently that the M1 scissors mode strength distribution exhibits a transitional behaviour in the same nuclei in which the transition from spherical to quadrupole deformed shapes is seen. In other words, it is expected that the low-lying dipole mode would exhibit critical behaviour in the same nuclei which are critical with respect to the quadrupole mode, since the former is driven by the latter, while no similar coincidence is expected [*a priori*]{} for the octupole mode, as already remarked. Recent experimental data on $^{124-136}$Xe [@Kneissl] reinforce this argument.
No analytically soluble model describing a phase shape transition of the scissors mode has been constructed so far. The two-rotor model [@LoIudice; @Palumbo] and its reformulation [@DeFran] could be an appropriate starting point for this effort.
[**3. Experimental manifestations of critical point symmetries and supersymmetries**]{}
[**3.1 Experimental manifestations of E(5)**]{}
The first nucleus to be identified as exhibiting E(5) behaviour was $^{134}$Ba [@CZE5], while $^{102}$Pd [@Zamfir] also seems to provide a very good candidate. Further studies on $^{134}$Ba [@AriasE2] reinforced this conclusion. $^{104}$Ru [@Frank], $^{108}$Pd [@Zhang], $^{114}$Cd [@Long], and $^{130}$Xe [@Liu] have also been suggested as possible candidates. A systematic search [@ClarkE5; @Kirson] on available data on energy levels and B(E2) transition rates suggested $^{102}$Pd, $^{106,108}$Cd, $^{124}$Te, $^{128}$Xe, and $^{134}$Ba as possible candidates, singling out $^{128}$Xe as the best one, in addition to $^{134}$Ba.
S. F. Ashley reported on recent measurements, performed at Yale of B(E2)s in $^{106}$Cd, and planned measurements of low-lying B(E2)s in $^{102-108}$Cd at Cologne, since earlier B(E2) measurements, by Coulomb excitation, in $^{106,108}$Cd seem unreliable. Detailed information on B(E2)s will clarify the presence of E(5) examples in this region, suggested in [@ClarkE5].
S. Harissopulos reported on recent measurements carried out at Legnaro on B(E2)s of $^{102}$Pd [@Kalyva], singling out this nucleus as the best example of E(5) so far, in agreement with earlier work [@Zamfir], among other reasons because no backbending occurs in its ground state band, which remains in excellent agreement with the parameter-free E(5) predictions up to high angular momenta.
S. Harissopulos also reported on measurements to be carried out in Jyväskylä on $^{128}$Xe, which has been suggested [@ClarkE5] as a very good candidate of E(5), but also is not far from a Z(4) behaviour [@Z4]. This is in agreement with a recent report [@Kneissl] on measurements of E1 and M1 strengths of $^{124-136}$Xe carried out at Stuttgart, which provides evidence for a shape phase transition around $A\simeq130$.
S. Harissopulos also mentioned, as possible E(5) candidates, $^{140}$Xe, which lies far from stability, and $^{48}$Ti, for which Coulomb excitation can be performed for the B(E2)s.
[**3.2 Experimental manifestations of X(5)**]{}
The first nucleus to be identified as exhibiting X(5) behaviour was $^{152}$Sm [@CZX5], followed by $^{150}$Nd [@Kruecken]. Further work on $^{152}$Sm [@ZamfirSm; @Clark; @CZK; @Bijker] and $^{150}$Nd [@Clark; @CZK; @Zhao] reinforced this conclusion. The neighbouring N=90 isotones $^{154}$Gd [@Tonev; @Dewald] and $^{156}$Dy [@Dewald; @CaprioDy] were also seen to provide good X(5) examples, the latter being of inferior quality. In the heavier region, $^{162}$Yb [@McC162Yb] and $^{166}$Hf [@McC166Hf] have been considered as possible candidates. A systematic study [@ClarkX5] of available experimental data on energy levels and B(E2) transition rates suggested $^{126}$Ba and $^{130}$Ce as possible good candidates, in addition to the N=90 isotones of Nd, Sm, Gd, and Dy. A similar study in lighter nuclei [@Brenner] suggested $^{76}$Sr, $^{78}$Sr and $^{80}$Zr as possible candidates.
D. Balabanski reported on recent measurements carried out at Yale on $^{128}$Ce, the first results of which on the B(E2)s within the ground state band are promising, while the analysis is going on. R. F. Casten remarked that this is a nucleus with a $P$-factor [@Casten] of 4.8, which is close to 5, thus it is expected to be a very good candidate for X(5). F. Iachello remarked that $^{128}$Ce, having 8 valence protons and 12 valence neutron holes, matches $^{152}$Sm, possessing 12 valence protons and 8 valence neutrons, thus it is expected to be a very good example of X(5), as $^{152}$Sm is.
P. G. Bizzeti referred to the current situation in the lighter region. $^{122}$Ba [@Fransen] has been identified as a possible X(5) candidate, based on experimental information for the energy levels. However, preliminary B(E2) measurements performed in Legnaro indicate that the B(E2)s within the ground state band appear to be closer to the rotational values and not to the X(5) ones. A similar situation seems to occur in $^{124}$Ba, according to a recent analysis, communicated by A. Dewald, of an earlier measurement performed at Cologne. The energy levels of $^{104}$Mo also suggested a X(5) interpretation [@Brenner; @Bizzeti], but B(E2) values turned out to favour the rotor interpretation [@Hutter].
S. Harissopulos reported on recent GASP measurements on $^{178}$Os [@DewaldJPG], for which B(E2)s indicate that it is the first good example of X(5) in the $A\simeq 180$ region, and on $^{176}$Os [@DewaldJPG], for which the analysis is still going on.
[**3.3 Experimental manifestations of E(5/4)**]{}
E(5/4) [@IacE54] corresponds to a particle with $j=3/2$ coupled to an E(5) core. Therefore possible candidates should be searched for among odd nuclei with the odd nucleon lying in a $j=3/2$ level. The first suspects were the neighbours of $^{134}$Ba, a good example of E(5). $^{135}$Ba and $^{133}$Ba have been measured at Yale. It is already clear that $^{135}$Ba is not an example of E(5/4) [@Fetea], probably because of the presence of a $j=1/2$ level close to the $j=3/2$ one, while the analysis of $^{133}$Ba is ongoing. As pointed out by F. Iachello, mixing between the multiplets based on the 1/2 and 3/2 levels leads to raising of one of the multiplets and lowering of the other.
S. Harissopulos pointed out that a multiplet in $^{129}$Xe, based on a 3/2 level lying slightly above the 1/2 ground state, exhibits strong similarities to E(5/4). J. Jolie suggested that a U(6/20) [@U620] supersymmetry \[corresponding to single-particle orbits with $j=1/2$, 3/2, 5/2, 7/2 coupled to an O(6) core\] might be more appropriate for an overall description of this nucleus, since it corresponds to a 1/2 ground state and has been found appropriate for describing nuclei in the A$\simeq$130 region [@U620].
F. Iachello suggested that the Ir-Au region near closed shells is maybe appropriate for looking for experimental manifestations of E(5/4), since the U(6/4) supersymmetry was found there [@IVI; @FVI]. $^{63}$Cu, although of small size, could also be a candidate, as discussed in [@IacE54].
J. Jolie pointed out a recent study [@JolieU612] bringing together the concepts of supersymmetry and shape phase transitions through the use of an IBFM [@IVI; @FVI] Hamiltonian including a vibrational term and a quadrupole-quadrupole interaction, using single-particle orbits with $j=1/2$, 3/2, 5/2, as in the framework of the U(6/12) supersymmetry [@IVI; @FVI]. This approach has given good results in the Os-Hg region [@JolieU612].
[**Acknowledgements**]{}
The author is grateful to S. F. Ashley, D. Balabanski, P. G. Bizzeti and L. Fortunato for providing written descriptions of their contributions.
[99]{}
F. Iachello, [*Phys. Rev. Lett.*]{} 85 (2000) 3580.
F. Iachello, [*Phys. Rev. Lett.*]{} 87 (2001) 052502.
M. A. Caprio, [*Phys. Rev. C*]{} 65 (2002) 031304.
G. Lèvai and J. M. Arias, [*Phys. Rev. C*]{} 69 (2004) 014304.
L. Fortunato and A. Vitturi, [*J. Phys. G*]{} 29 (2003) 1341.
L. Fortunato, [*Eur. Phys. J. A*]{} 26, s01 (2005) 1.
P. M. Davidson, [*Proc. R. Soc.*]{} 135 (1932) 459.
J. P. Elliott, J. A. Evans, and P. Park, [*Phys. Lett. B*]{} 169 (1986) 309.
D. J. Rowe and C. Bahri, [*J. Phys. A*]{} 31 (1998) 4947.
D. Bonatsos, D. Lenis, N. Minkov, D. Petrellis, P. P. Raychev, and P. A. Terziev, [*Phys. Lett. B*]{} 584 (2004) 40.
D. Bonatsos, D. Lenis, N. Minkov, D. Petrellis, P. P. Raychev, and P. A. Terziev, [*Phys. Rev. C*]{} 70 (2004) 024305.
J. M. Arias, C. E. Alonso, A. Vitturi, J. E. García-Ramos, J. Dukelsky, and A. Frank, [*Phys. Rev. C*]{} 68 (2003) 041302.
J. E. García-Ramos, J. Dukelsky, and J. M. Arias, [*Phys. Rev. C*]{} 72 (2005) 037301.
D. Bonatsos, D. Lenis, N. Minkov, P. P. Raychev, and P. A. Terziev, [*Phys. Rev. C*]{} 69 (2004) 044316.
A. Bohr, [*Mat. Fys. Medd. K. Dan. Vidensk. Selsk.*]{} 26, no. 14 (1952).
L. Wilets and M. Jean, [*Phys. Rev.*]{} 102 (1956) 788.
G. G. Dussel and D. R. Bès, [*Nucl. Phys. A*]{} 143 (1970) 623.
F. Iachello and A. Arima, [*The Interacting Boson Model*]{} (Cambridge University Press, Cambridge, 1987).
A. A. Raduta, A. C. Gheorghe, and A. Faessler, [*J. Phys. G*]{} 31 (2005) 337.
O. S. van Roosmalen, [*Ph.D. thesis*]{}, U. Groningen, The Netherlands (1982).
A. Leviatan, [*Phys. Rev. C*]{} 72 (2005) 031305.
J.-Y. Zhang, M. A. Caprio, N. V. Zamfir, and R. F. Casten, [*Phys. Rev. C*]{} 60 (1999) 061304.
M. A. Caprio, [*Phys. Rev. C*]{} 69 (2004) 044307.
N. Pietralla and O. M. Gorbachenko, [*Phys. Rev. C*]{} 70 (2004) 011304.
K. Dusling and N. Pietralla, [*Phys. Rev. C*]{} 72 (2005) 011303.
L. Fortunato and A. Vitturi, [*J. Phys. G*]{} 30 (2004) 627.
D. Bonatsos, D. Lenis, N. Minkov, P. P. Raychev, and P. A. Terziev, [*Phys. Rev. C*]{} 69 (2004) 014302.
D. Bonatsos, D. Lenis, D. Petrellis, P. A. Terziev, and I. Yigitoglu, [*Phys. Lett. B*]{} (2005) in press.
M. A. Caprio, [*Phys. Rev. C*]{} 72 (2005) 054323.
D. J. Rowe, [*Nucl. Phys. A*]{} 735 (2004) 372.
D. J. Rowe, P. S. Turner, and J. Repka, [*J. Math. Phys.*]{} 45 (2004) 2761.
D. J. Rowe and P. S. Turner, [*Nucl. Phys. A*]{} 753 (2005) 94.
D. Bonatsos, D. Lenis, D. Petrellis, and P. A. Terziev, [*Phys. Lett. B*]{} 588 (2004) 172.
D. Bonatsos, D. Lenis, D. Petrellis, P. A. Terziev, and I. Yigitoglu, [*Phys. Lett. B*]{} 621 (2005) 102.
L. Fortunato, [*Phys. Rev. C*]{} 70 (2004) 011302.
R. V. Jolos, [*Yad. Fiz.*]{} 67 (2004) 955 \[[*Phys. At. Nucl.*]{} 67 (2004) 931\].
A. S. Davydov and G. F. Filippov, [*Nucl. Phys.*]{} 8 (1958) 237.
R. F. Casten, [*Nuclear Structure from a Simple Perspective*]{} (Oxford University Press, Oxford, 1990).
A. E. L. Dieperink and R. Bijker, [*Phys. Lett. B*]{} 116 (1982) 77.
J. M. Arias, J. E. García-Ramos, and J. Dukelsky, [*Phys. Rev. Lett.*]{} 93 (2004) 212501.
M. A. Caprio and F. Iachello, [*Phys. Rev. Lett.*]{} 93 (2004) 242502.
M. A. Caprio and F. Iachello, [*Ann. Phys. (N.Y.)*]{} 318 (2005) 454.
D. Bonatsos, D. Lenis, N. Minkov, D. Petrellis, and P. Yotov, [*Phys. Rev. C*]{} 71 (2005) 064309.
P. G. Bizzeti and A. M. Bizzeti-Sona, [*Phys. Rev. C*]{} 70 (2004) 064319.
S. G. Rohoziński, [*Rep. Prog. Phys.*]{} 51 (1988) 541.
J. Engel and F. Iachello, [*Phys. Rev. Lett.*]{} 54 (1985) 1126.
J. Engel and F. Iachello, [*Nucl. Phys. A*]{} 472 (1987) 61.
W. Ziegler, C. Rangacharyulu, A. Richter, and C. Spielel, [*Phys. Rev. Lett.*]{} 65 (1990) 2515.
C. Rangacharyulu, A. Richter, H. J. Wörtche, W. Ziegler, and R. F. Casten, [*Phys. Rev. C*]{} 43 (1991) 949.
N. Pietralla, P. von Brentano, R.-D. Herzberg, U. Kneissl, J. Margraf, H. Maser, H. H. Pitz, and A. Zilges, [*Phys. Rev. C*]{} 52 (1995) 2317.
A. Richter, [*Prog. Part. Nucl. Phys.*]{} 34 (1995) 261.
M. Scheck, H. von Garrel, N. Tsoneva, D. Belic, P. von Brentano, C. Fransen, A. Gade, J. Jolie, U. Kneissl, C. Kohstall, A. Linnemann, A. Nord, N. Pietralla, H. H. Pitz, F. Stedile, C. Stoyanov, and V. Werner, [*Phys. Rev. C*]{} 70 (2004) 044319.
U. Kneissl, in [*Key Topics in Nuclear Structure (Paestum 2004)*]{}, ed. A. Covello (World Scientific, Singapore, 2005) p. 399.
N. Lo Iudice and F. Palumbo, [*Phys. Rev. Lett.*]{} 41 (1978) 1532.
N. Lo Iudice and F. Palumbo, [*Nucl. Phys. A*]{} 326 (1979) 193.
G. De Franceschi, F. Palumbo, and N. Lo Iudice, [*Phys. Rev. C*]{} 29 (1984) 1496.
R. F. Casten and N. V. Zamfir, [*Phys. Rev. Lett.*]{} 85 (2000) 3584.
N. V. Zamfir, M. A. Caprio, R. F. Casten, C. J. Barton, C. W. Beausang, Z. Berant, D. S. Brenner, W. T. Chou, J. R. Cooper, A. A. Hecht, R. Krücken, H. Newman, J. R. Novak, N. Pietralla, A. Wolf, and K. E. Zyromski, [*Phys. Rev. C*]{} 65 (2002) 044325.
J. M. Arias, [*Phys. Rev. C*]{} 63 (2001) 034308.
A. Frank, C. E. Alonso, and J. M. Arias, [*Phys. Rev. C*]{} 65 (2001) 014301.
D.-L. Zhang and Y.-X. Liu, [*Phys. Rev. C*]{} 65 (2002) 057301.
J.-F. Zhang, G.-L. Long, Y. Sun, S.-J. Zhu, F.-Y. Liu, and Y. Jia, [*Chin. Phys. Lett.*]{} 20 (2003) 1231.
D.-L. Zhang and Y.-X. Liu, [*Chin. Phys. Lett.*]{} 20 (2003) 1028.
R. M. Clark, M. Cromaz, M. A. Deleplanque, M. Descovich, R. M. Diamond, P. Fallon, I. Y. Lee, A. O. Macchiavelli, H. Mahmud, E. Rodriguez-Vieitez, F. S. Stephens, and D. Ward, [*Phys. Rev. C*]{} 69 (2004) 064322.
M. W. Kirson, [*Phys. Rev. C*]{} 70 (2004) 049801.
G. Kalyva, A. Spyrou, M. Axiotis, S. Harissopulos, A. Dewald, A. Fitzler, B. Saha, A. Linnemann, O. Möller, R. Vlastou, D. R. Napoli, N. Marginean, C. Rusu, G. de Angelis, C. Ur, D. Bazzacco, E. Farnea, D. L. Balabanski, and R. Julin, in [*Frontiers in Nuclear Structure, Astrophysics and Reactions (Kos 2005)*]{}, ed. R. Julin and S. Harissopulos (2005).
R. F. Casten and N. V. Zamfir, [*Phys. Rev. Lett.*]{} 87 (2001) 052503.
R. Krücken, B. Albanna, C. Bialik, R. F. Casten, J. R. Cooper, A. Dewald, N. V. Zamfir, C. J. Barton, C. W. Beausang, M. A. Caprio, A. A. Hecht, T. Klug, J. R. Novak, N. Pietralla, and P. von Brentano, [*Phys. Rev. Lett.*]{} 88 (2002) 232501.
N. V. Zamfir, H. G. Börner, N. Pietralla, R. F. Casten, Z. Berant, C. J. Barton, C. W. Beausang, D. S. Brenner, M. A. Caprio, J. R. Cooper, A. A. Hecht, K. Krtička, R. Krücken, P. Mutti, J. R. Novak, and A. Wolf, [*Phys. Rev. C*]{} 65 (2002) 067305.
R. M. Clark, M. Cromaz, M. A. Deleplanque, R. M. Diamond, P. Fallon, A. Görgen, I. Y. Lee, A. O. Macchiavelli, F. S. Stephens, and D. Ward, [*Phys. Rev. C*]{} 67 (2003) 041302.
R. F. Casten, N. V. Zamfir, and R. Krücken, [*Phys. Rev. C*]{} 68 (2003) 059801.
R. Bijker, R. F. Casten, N. V. Zamfir, and E. A. McCutchan, [*Phys. Rev. C*]{} 68 (2003) 064304. Erratum: [*Phys. Rev. C*]{} 69 (2004) 059901.
D.-L. Zhang and H.-Y. Zhao, [*Chin. Phys. Lett.*]{} 19 (2002) 779.
D. Tonev, A. Dewald, T. Klug, P. Petkov, J. Jolie, A. Fitzler, O. Möller, S. Heinze, P. von Brentano, and R. F. Casten, [*Phys. Rev. C*]{} 69 (2004) 034334.
A. Dewald, O. Möller, D. Tonev, A. Fitzler, B. Saha, K. Jessen, S. Heinze, A. Linnemann, J. Jolie, K. O. Zell, P. von Brentano, P. Petkov, R. F. Casten, M. Caprio, J. R. Cooper, R. Krücken, V. Zamfir, D. Bazzacco, S. Lunardi, C. Rossi Alvarez, F. Brandolini, C. Ur, G. de Angelis, D. R. Napoli, E. Farnea, N. Marginean, T. Martinez, and M. Axiotis, [*Eur. Phys. J. A*]{} 20 (2004) 173.
M. A. Caprio, N. V. Zamfir, R. F. Casten, C. J. Barton, C. W. Beausang, J. R. Cooper, A. A. Hecht, R. Krücken, H. Newman, J. R. Novak, N. Pietralla, A. Wolf, and K. E. Zyromski, [*Phys. Rev. C*]{} 66 (2002) 054310.
E. A. McCutchan, N. V. Zamfir, M. A. Caprio, R. F. Casten, H. Amro, C. W. Beausang, D. S. Brenner, A. A. Hecht, C. Hutter, S. D. Langdown, D. A. Meyer, P. H. Regan, J. J. Ressler, and A. D. Yamamoto, [*Phys. Rev. C*]{} 69 (2004) 024308.
E. A. McCutchan, N. V. Zamfir, R. F. Casten, M. A. Caprio, H. Ai, H. Amro, C. W. Beausang, A. A. Hecht, D. A. Meyer, and J. J. Ressler, [*Phys. Rev. C*]{} 71 (2005) 024309.
R. M. Clark, M. Cromaz, M. A. Deleplanque, M. Descovich, R. M. Diamond, P. Fallon, R. B. Fierstone, I. Y. Lee, A. O. Macchiavelli, H. Mahmud, E. Rodriguez-Vieitez, F. S. Stephens, and D. Ward, [*Phys. Rev. C*]{} 68 (2003) 037301.
D. S. Brenner, in [*Mapping the Triangle*]{}, ed. A. Aprahamian, J. A. Cizewski, S. Pittel, and N. V. Zamfir, [*AIP CP*]{} 638 (2002) 223.
C. Fransen, N. Pietralla, A. Linnemann, V. Werner, and R. Bijker, [*Phys. Rev. C*]{} 69 (2004) 014313.
P. G. Bizzeti and A. M. Bizzeti-Sona, [*Phys. Rev. C*]{} 66 (2002) 031301.
C. Hutter, R. Krücken, A. Aprahamian, C. J. Barton, C. W. Beausang, M. A. Caprio, R. F. Casten, W.-T. Chou, R. M. Clark, D. Cline, J. R. Cooper, M. Cromaz, A. A. Hecht, A. O. Macchiavelli, N. Pietralla, M. Shawcross, M. A. Stoyer, C. Y. Wu, and N. V. Zamfir, [*Phys. Rev. C*]{} 67 (2003) 054315.
A. Dewald, O. Möller, B. Saha, K. Jessen, A. Fitzler, B. Melon, T. Pissulla, S. Heinze, J. Jolie, K. O. Zell, P. von Brentano, P. Petkov, S. Harissopulos, G. De Angelis, T. Martinez, D. R. Napoli, N. Marginean, M. Axiotis, C. Rusu, D. Tonev, A. Gadea, Y. H. Zhang, D. Bazzacco, S. Lunardi, C. A. Ur, R. Menegazzo, and E. Farnea, [*J. Phys. G*]{} 31 (2005) S1427.
F. Iachello, [*Phys. Rev. Lett.*]{} 95 (2005) 052503.
M. S. Fetea, R. B. Cakirli, R. F. Casten, and D. D. Warner, in [*Exotic Nuclei and Nuclear/Particle Astrophysics (Constanta 2005)*]{}, ed. S. Stoica, L. Trache, and F. Carstoiu (IFIN-HH, Bucharest, 2005) p. 20.
J. Jolie, K. Heyde, P. Van Isacker, and A. Frank, [*Nucl. Phys. A*]{} 466 (1987) 1.
F. Iachello and P. Van Isacker, [*The Interacting Boson-Fermion Model*]{} (Cambridge U. Press, Cambridge, 1991).
A. Frank and P. Van Isacker, [*Algebraic Methods in Molecular and Nuclear Structure Physics*]{} (Wiley, New York, 1994).
J. Jolie, S. Heinze, P. Van Isacker, and R. F. Casten, [*Phys. Rev. C*]{} 70 (2004) 011305.
| {
"pile_set_name": "ArXiv"
} |
Thin films of elongated molecules with tilt ordering, including Smectic C liquid crystals and dense fluid phases of amphiphiles deposited on water, very often possess fascinating distributions of the tilt azimuth. The organization of the tilt azimuth is referred to as the texture and can be observed by polarized light microscopy or Brewster angle microscopy. Many classes of these textures have been found experimentally, like stripes[@stripe], stars[@star], boojums[@boojums] and hedgehogs[@deMul]. These examples are observed in Langmuir monolayers composed of molecules which are symmetric under in-plane reflection, or achiral. Recently, more attention has been paid to the hedgehog patterns, obtained in a circular domain with a central point defect. The reversing spiral is one of the spectacular hedgehog textures discovered in a chiral tilted Smectic C liquid crystal film on water [@Kraus]. Let us emphasize that both systems described above are polar, they are not symmetric under 180$^\circ$ rotation about an in-plane axis. The tilt azimuth can be represented by a two-dimensional vector $\hat{c}$, the projection onto the film of the elongated molecules, analogous to the order parameter of an $XY$-model. Textures in two-dimensional $XY$-like systems have attracted a good deal of attention. A brief theoretical account based on a perturbative approach of hedgehogs in achiral monolayers can be found in Ref. [@FiscBru]. The stability of the hedgehog configuration in an isotropic and achiral system has also been investigated [@PettyLuben]. Spiral textures in an achiral system have been studied in the small distortion regime [@Williams]. Although a theoretical approach to solutions for hedgehog textures is given in Ref. [@Kraus], to the best of our knowledge, a systematic discussion of hedgehog textures in a chiral system has not been presented.
In this work, we study a generic model [@LangSeth] for the texture in a circular domain of a two-dimensional $XY$-like system. We have found a family of equilibrium configurations that can be classified by a new topological number analogous to the winding number that classifies a two-dimensional point defect[@Mermin]. When a system is achiral, these metastable configurations can be chiral as a result of spontaneously broken chiral symmetry. When chirality is explicitly introduced, more complex equilibrium textures result. The crossed polarizer images generated from the theoretical textures are in excellent agreement with the pictures taken experimentally. The starting point of our investigation is the following elastic energy of a chiral polar film with tilt ordering, $$\begin{aligned}
H[\hat{c}]&=&\frac{1}{2}\int_\Omega dA \left[K_s|\nabla\cdot\hat{c}|^2
+K_{sb}(\nabla\cdot\hat{c})(\nabla\times\hat{c}\cdot\hat{z})+
\right.\nonumber\\
&& \left.K_b|\nabla\times\hat{c}\cdot\hat{z}|^2\right]+
\oint_\Gamma ds\;\sigma(\vartheta-\Theta).\end{aligned}$$ It is computed for a circular area $\Omega$ containing the ordered medium enclosed by the boundary $\Gamma$. The unit vector $\hat{z}$ points normal to the film. The quantities $K_b$ and $K_s$ are, respectively, the bend and splay elastic moduli, $\hat{c}\equiv\hat{x}\cos\Theta+\hat{y}\sin\Theta$ is the order parameter of the system, $\Theta$ is the angle between $\hat{c}$ and the $x$ axis, $\vartheta$ is the angle between the outward normal to $\Gamma$ and the $x$ axis. The anisotropic line tension $\sigma(\phi)$ can formally be expanded as $\sigma_0+\sum_n(a_n\cos n\phi+c_n \sin n\phi)$. We will consider only the first few coefficients in the expansion. The first order terms are $a_1$ for polar films and $c_1$ for chiral systems. When the system is nonpolar, $a_1$ vanishes and $a_2$ must be considered. The coefficient $c_2$ is relevant when the system is both polar and chiral, but it will be neglected because $a_1$ and $c_1$ are non zero for such a system. The cross term$(\nabla\cdot\hat{c})(\nabla\times\hat{c}\cdot\hat{z})$ has to be included when the film is both chiral and polar. It is required that the coefficient $|K_{sb}|<2\sqrt{K_sK_b}$ so that the elastic energy density remains positive for arbitrary splay and bend distortions.
We will restrict ourselves to the hedgehog textures, each containing a central point defect of winding number $+1$ with core radius $\xi$. The defect core corresponds to the region in which $\hat{c}$ is not defined. We assume that its presence affects the elastic energy only through the inner boundary condition at $r=\xi$ and we neglect its energetic contribution when $r<\xi$ [@deGennes]. The boundary condition is taken to be $\Theta|_{r=\xi}=\varphi+\varphi_d$, where $r$ is the radial distance, $\varphi$ is the polar angle in plane-polar coordinates and $\varphi_d$ is a constant. This is indeed justifiable for $\pm1$ defects with the structure discussed in Ref. [@Tabe]. We further assume that the defect is stable when it is located at the center of $\Omega$ and that the system is cylindrically symmetric, i.e., $\Theta=\varphi+f(\ln r)$, where $f(\ln r)$ is the radial distribution of $\hat{c}$. These assumptions can be shown to be valid for all the textures to be discussed. We shall compute the possible expressions for $f(\ln r)$ that minimize the elastic energy and give the equilibrium textures. In terms of $k\equiv\ln r$, the elastic energy reduces to $$\begin{aligned}
H[f, f^\prime]&=&\pi\kappa\int_{\ln\xi}^{\ln R_0}dk\left\{1+f^{\prime2}-
\right.\nonumber\\
&&\mu\left[(1-f^{\prime2})\cos(2f-2\varsigma_+)-\right.\nonumber\\
&&\left.\left.2f^\prime\sin(2f-2\varsigma_+)\right]
\right\}+\left.2\pi R_0\sigma(-f)\right|_{\ln R_0},\label{reducedH}\end{aligned}$$ where $2\kappa\equiv K_s+K_b$, $2\kappa\beta\equiv K_s-K_b$, $2\kappa\tau\equiv K_{sb}$, $\mu\equiv\sqrt{\beta^2+\tau^2}$ and $2\varsigma_\pm\equiv\tan^{-1}\tau/\beta\pm\pi$. The equilibrium condition for $f(k)$ is $$\begin{aligned}
-f^{\prime\prime}-\mu\left[f^{\prime\prime}\cos(2f-2\varsigma_+)-
\right.&&\nonumber\\
\left.(f^{\prime2}+1)\sin(2f-2\varsigma_+)\right]&=&0\label{bulkr}\end{aligned}$$ and the boundary condition at $k=\ln R_0$ is $$\begin{aligned}
\kappa\left\{f^\prime+\mu\left[f^\prime\cos(2f-2\varsigma_+)+\sin(2f-2
\varsigma_+)\right]\right\}-&&\nonumber\\
R_0\sigma^\prime(-f)&=&0.\end{aligned}$$ Depending on the choice of the parameters, $f$ may possess more than one or, at times, numerous solutions. We use linear stability analysis to determine if these solutions are local minima. The elastic energy is expanded to second order in small variations $\chi\equiv f-f_0$ about an equilibrium configuration $f_0$ as $\delta H=\int dk\chi{\cal L}\chi + {\psi}^T{\cal B}{\psi}$ where $$\begin{aligned}
{\cal L}&=&-\pi\kappa\left\{\left[1+\mu\cos(2f_0-2\varsigma_+)
\right]\frac{d^2}{dk^2}-\right.\nonumber\\
&&\left.\mu\left[f^{\prime\prime}_0\sin(2f_0-2\varsigma_+)+
2\cos(2f_0-2\varsigma_+)\right]
\right\},\\
{\cal B}_{\chi^\prime\chi^\prime}&=&0,\\
{\cal B}_{\chi\chi^\prime}&=&{\cal B}_{\chi^\prime\chi}=\frac{\pi\kappa}{2}
\left[1+\mu\cos(2f_0-2\varsigma_+)\right]_{\ln R_0},\\
{\cal B}_{\chi\chi}&=&\pi\left\{-\kappa\mu\left[f^\prime_0\sin(2f_0-2\varsigma_+)-2
\cos(2f_0-2\varsigma_+)\right]+\right.\nonumber\\
&&\left.R_0\sigma^{\prime\prime}(-f_0)\right\}_{\ln R_0},\end{aligned}$$ and $\psi^T\equiv(\chi^\prime,\chi)|_{\ln R_0}$. The deviation of the elastic energy $\delta H$ from its equilibrium value can be examined in terms of the eigenvalue $\lambda$ and associated eigenfunction $\phi_\lambda$ satisfying ${\cal L} \phi_\lambda=\lambda\phi_\lambda$. The eigenfunctions $\phi_\lambda$ are normalized so that $\int dk \phi_\lambda^2=1$ and $\phi_\lambda|_{\ln\xi}=0$ for infinitely strong anchoring at the inner boundary. It is typical in linear stability analysis that there is another boundary condition on $\phi_\lambda$ that isolates a set of eigenvalues $\lambda$ in which the sign of the lowest one is to be tested. For our particular case, there is no other restriction on $\phi_\lambda$ that can be imposed and all $\lambda$’s are allowed. The deviation of energy associated with the fluctuational mode $\phi_\lambda$ is no longer $\lambda$ but $$\begin{aligned}
\delta H_\lambda=\lambda+\psi^T_\lambda{\cal B}\psi_\lambda,\end{aligned}$$ where $\psi^T_\lambda=(\phi^\prime_\lambda, \phi_\lambda)|_{\ln R_0}$. Instability is signified by $\delta H_\lambda$ becoming negative. The asymptotic behavior of $\delta H_\lambda$ can be shown to be $\delta H_\lambda\sim -|\lambda| + R_0^{(1+2\sqrt{|\lambda}|)}$ when $\lambda\rightarrow-\infty$, and $\delta H_\lambda> (1-{\cal
B}_{\chi\chi^\prime}/2{\cal B}_{\chi\chi})\lambda$ when $\lambda\rightarrow\infty$ for $ R_0\sqrt{a_1^2+c_1^2}\gg\kappa$ and $\sigma^{\prime\prime} (-f_0)|_{k=\ln R_0} \approx
\sqrt{a_1^2+c_1^2}$. We see that $\delta H_\lambda$ is positive in both limits and can only change sign for small $\lambda$.
We have outlined the procedures to obtain the equilibrium textures and to examine their stability against infinitesimal fluctuations. To understand these equilibrium textures, we first look at the simplest case $\beta=\tau=c_1=a_2=\varphi_d=0$ and $a_1<0$. It corresponds to a polar film made of achiral molecules, with isotropic elastic constants and having fixed anchoring at the inner boundary such that $\hat{c}$ points normal into the bulk. The condition $a_1<0$ indicates that $\hat{c}$ favors the outward normal direction at the outer boundary, without being locked. In this case, there are analytic solutions. It is obvious that $f_0(k)=0$ is the lowest energy configuration: all $\hat{c}$ vectors point along the radial direction. The general solution to $f_0(k)$ is $$\begin{aligned}
f_0(k)&=&\frac{\varphi_b}{\ln R_0-\ln\xi}(k-\ln\xi),\end{aligned}$$ and $\varphi_b$ satisfies $$\begin{aligned}
\varphi_b&=&\frac{R_0a_1(\ln R_0-\ln\xi)}{\kappa}\sin\varphi_b.\label{bc1}\end{aligned}$$ The quantity $\varphi_b$ is the anchoring angle of $\hat{c}$ measured with respect to $\varphi$ at $k=\ln R_0$. It is easy to see that the system supports numerous equilibrium solutions when the amplitude $|R_0a_1(\ln
R_0-\ln\xi)/\kappa|\gg1$. Figure \[mltspr\] shows the plots of $\varphi_b$ and $R_0a_1(\ln R_0-\ln\xi)\sin\varphi_b/\kappa$. We denote the solutions of Eq. (\[bc1\]) by $\varphi_b^{(i)}$ with an index $i$. Not all the solutions are stable. The stable ones are indicated by open circles in Fig. \[mltspr\]. When $\varphi_b^{(i)}\neq0$, we have spirals in which $\hat{c}$ points in the radial direction at the inner boundary and rotates counterclockwise through $\varphi_b^{(i)}$ along a radial path ending at the outer boundary. These metastable textures do not have in-plane reflection symmetry although the system in question is achiral. We find, by inspecting the boundary condition Eq. (\[bc1\]), that there is a solution $\varphi_b^{(j)}=-\varphi_b^{(i)}$ for any solution $\varphi_b^{(i)}$ and these configurations have the same energy. Hence, chiral symmetry is spontaneously broken for the higher energy metastable configurations. The family of solutions can be classified in terms of a topological index $m_i$, which is defined to be the nearest integer around $\varphi_b^{(i)}/2\pi$. This classification can also be drawn topologically by examining the possible order-parameter distributions along a line in the radial direction connecting the inner and the outer boundaries[@Mermin]. All integers are allowed topologically, but only some are supported energetically in a bounded system.
The situation is different when $a_1>0$, i.e., when the preferred direction of $\hat{c}$ at the outer boundary is in the inward normal direction. The solution $f_0(k)=0$ is now unstable and we can not have a pure splay texture anymore. The lowest energy configuration corresponds to the first nonzero $\varphi_b^{(i)}$ satisfying Eq. (\[bc1\]). It is a spiral texture as are all the other metastable configurations allowed. Let us emphasize that they all break the chiral symmetry spontaneously. The spiral textures we have just discussed arise from spontaneously broken chiral symmetry in an isotropic achiral system. The actual system may not be isotropic and the parameter $\beta$ is not necessarily zero. Further, it is possible to explicitly break the chiral and polar symmetries ($\tau \neq 0$) by adding chiral molecules to the system and by putting the film on water, respectively. When $\beta$ or $\tau$ are nonzero, analytic solutions for Eq. (\[bulkr\]) are not obvious and we resort to numerical methods. More complex equilibrium textures with $m_i=0$ are supported in addition to those with higher $m_i$’s. The behavior of the $\hat{c}$-distribution in such cases can be understood intuitively by noting the existence of two preferred bulk directions for $\hat{c}$ when $\Theta-\varphi\in[-\pi, \pi)$, namely $\Theta_{\pm}=\varphi+\varsigma_\pm$. Right at the inner boundary of $\Omega$, $\Theta$ is set to $\varphi+ \varphi_d$. Along a trajectory traversing radially into the bulk, $\Theta$ settles smoothly near one of $\Theta_\pm$, depending on which of these gives the lower overall energy of the system ($\Theta_+$ for our case). When it gets to the outer boundary, $\Theta$ gradually sets itself at $\varphi+\varphi_b$. The $f^{(1)}_0(k)$ and texture for the lowest energy configuration taking $\tau \neq 0$ are displayed in Fig. \[lwscfg\]. The texture is depicted as the density plot of $\sin^2\Theta\cos^2\Theta$, which simulates the images obtained by a set of crossed polarizers. Figure \[lwscfg\](c) shows an experimental image of a chiral Smectic C liquid crystal domain in a free standing film. It is very similar to the image shown in Fig. \[lwscfg\](b), computed using the appropriate parameters for a chiral free standing film.
Apparently, these images do not resemble the remarkable and dramatic reversing spiral texture reported in Ref. [@Kraus] and shown in Fig. \[revspr\](c). Nevertheless, Fig. \[lwscfg\] does represent a reversing spiral, since $f^{(1)}_0(k)$ goes through a maximum at about $k=2.0$. However, the initial rotation of the director near the core is so rapid that it is barely visible. We are able to locate among the various solutions an equilibrium texture that resembles the “dramatic reversing spiral” (or DRS texture), illustrated in Fig. \[revspr\]. The distinguishing feature of the DRS texture is the extended region near the core in which the director rotates slowly. This configuration is constructed by choosing the values of $\Theta$ at the inner boundary and at the outer boundary to be around $\Theta_- =\varphi+\varsigma_-+2\pi$. Starting from the inner boundary, $\Theta$ first remains near $\Theta_-$, then swings nearly $\pi$ to a value close to $\Theta_+$ and finally returns rapidly to $\varphi+\varphi_b$ (which has been set the be around $\Theta_-$) near the outer boundary. The function $f^{(2)}_0(k)$ and its texture are depicted in Fig. \[revspr\]. There is strong resemblance between the density plot of $\cos^2\Theta$, simulating the images obtained with slightly uncrossed polarizers, and the experimental picture of the DRS.
Remarkably, the DRS is only a metastable texture! It is easy to see that $f^{(2)}_0(k)$ is not the lowest energy solution. The configuration, in which $\Theta$ sets smoothly near $\Theta_-$ without going through the rapid changes to $\Theta_+$ and then back to $\Theta_-$, has the lowest elastic energy. Since $f^{(2)}_0(k)$ is not the global minimum of the elastic energy Eq. (\[reducedH\]), its stability against fluctuations is in question. We have examined $\delta H_\lambda$, the deviation in energy associated with the reversing spiral solution, in a wide range of $\lambda$, and conclude that the DRS is a metastable configuration for the present choice of the parameters. In general, the DRS texture is not even metastable for an arbitrary set of parameters. As noted in Ref. [@Kraus], the DRS texture is rare, which is consistent with these findings.
In conclusion, we have analyzed a generic model for textures associated with a $+1$ defect in domains with $XY$-like ordering. Restricting our analysis to the class of textures with the defect fixed at the center, we find the equilibrium configurations and examine the stability of these configurations against infinitesimal fluctuations. Many metastable configurations are supported at some suitable choice of parameters. These can be classified by a new radial topological number analogous to the winding number classifying the point defect. It is also found that metastable configurations are chiral, including the stable lowest energy texture when $a_1>0$, even if the system possesses in-plane reflection symmetry. When elastic anisotropy, chirality and polarity are introduced explicitly, more equilibrium solutions exist. Among the equilibrium textures, we have found two kinds of reversing spirals, simple ones that are absolutely stable textures, and metastable ones of a more dramatic appearance, that resemble closely the reversing spiral reported in Ref. [@Kraus].
K.-K. L. would like to express his gratitude to Professor Joseph Rudnick for ideas, suggestions and support, and thank Professor Charles Knobler and Professor Robijn Bruinsma for numerous stimulating discussions. This research was supported in part by the NSF through grant DMR-9974388, and by Brandeis University.
J. Ruiz-Garcia, X. Qiu, M.-W. Tsao, G. Marshall, C.M.Knobler, G. A. Overbeck and D. Möbius, J. Phys. Chem. [**97**]{}, 6955 (1993). X. Qiu, J. Ruiz-Garcia, K. J. Stine, C.M. Knobler, J.V. Selinger, Phys. Rev. Lett. [**67**]{}, 703 (1991). S. Rivière and J. Meunier, Phys. Rev. Lett. [**74**]{}, 2495 (1995); J. Fang, E. Teer, C.M. Knobler, K.-K. Loh and J. Rudnick, Phys. Rev. E [**56**]{}, 1859 (1997). M.N.G. de Mul and J.A. Mann Jr., Langmuir [**11**]{}, 3292 (1995). I. Kraus and R.B. Meyer, Phys. Rev. Lett. [**82**]{}, 3815 (1999). T.M. Fischer, R.F. Bruinsma, and C.M. Knobler, Phys. Rev. E [**50**]{}, 413 (1994). D. Pettey and T.C. Lubensky, Phys. Rev. E [**59**]{}, 1834 (1999). D.R.M. Williams, Phys. Rev. E [**50**]{}, 1686 (1994). S.A. Langer and J.P. Sethna, Phys. Rev. A [**34**]{}, 5035 (1986). N.D. Mermin, Rev. Mod. Phys. [**51**]{}, 591 (1979). P.G. de Gennes and J. Prost, The Physics of Liquid Crystals (Clarendon Press, Oxford, 1993), p.171. Y. Tabe, N. Shen, E. Mazur, and H. Yokoyama, Phys. Rev. Lett. [**82**]{}, 759 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Taking seriously the interpretation of black hole entropy as the logarithm of the number of microstates, we argue that thermal gravitons may undergo a phase transition to a kind of black hole condensate. The phase transition proceeds via nucleation of black holes at a rate governed by a saddlepoint configuration whose free energy is of order the inverse temperature in Planck units. Whether the universe remains in a low entropy state as opposed to the high entropy black hole condensate depends sensitively on its thermal history. Our results may clarify an old observation of Penrose regarding the very low entropy state of the universe.'
author:
- 'Stephen D.H. Hsu'
- 'Brian M. Murray'
title: 'Thermal gravity, black holes and cosmological entropy'
---
Introduction {#intro}
============
Years ago Penrose noticed that the universe must have begun in a very low entropy state [@penrose]. By considering the entropy of black holes, he argued that the current state of the universe has significantly lower entropy than the maximum possible entropy state. For example, while holding the number of baryons fixed one could increase the total entropy tremendously by letting matter collapse into black holes [@grentropy]. Indeed, it seems that while the matter degrees of freedom were born hot, i.e., in a maximum entropy thermal state, the gravitational degrees of freedom were born in a very special low entropy state. Interpreting entropy as the logarithm of phase space volume, a low entropy state is an exponentially unlikely state and hence can only result from fine-tuned initial conditions [@ergodic]. Reasoning along these lines suggests that spacetimes with numerous horizons, perhaps resembling a dense agglomeration of black holes, occupy an exponentially larger fraction of gravitational phase space than smooth spacetimes like the usual Friedmann-Robertson-Walker (FRW) cosmologies. For related discussions, see, e.g., [@carroll; @wald] and references contained therein.
One may ask whether special initial conditions at the Planck scale are sufficient to produce the low entropy universe we see today. It might be the case that interactions with thermal matter in the early universe inevitably cause the gravitational degrees of freedom to thermalize as well. Such a thermal state, assuming ergodicity of gravity, would likely evolve to a configuration of much higher entropy, and hence a cosmology very different from the one we observe.
Since black holes are our only hint at the highly entropic configurations of gravity [@horizon], they should play a prominent role in the transition from low entropy to high entropy spacetimes. In this paper we suggest a specific mechanism involving the nucleation of black holes from a thermal graviton state. We note that the corresponding nucleation rate from a thermal matter state is much smaller, and probably irrelevant cosmologically. The mechanism we describe provides a plausible means by which Penrose’s ergodic evolution could proceed. We examine whether the transition to a new, highly entropic, phase of condensed black holes can occur in standard big bang cosmology. The result depends sensitively on the thermal history of the universe at early times. Moreover, the relevant energy scales are all higher than the energy scale at which an inflationary epoch is usually assumed to take place. Therefore, we are considering a phase transition which only may take place before and not after inflation. Presumably, the probability for a given patch to inflate would be affected by whether or not that patch has undergone a phase transition to the high entropy phase.
We should note that there is not a consensus on the issue of whether gravity is ergodic nor on the interpretation of the gap between the maximum allowed and the actual entropy in an FRW spacetime. Tipler [@tipler] showed that under a reasonable set of assumptions, closed universes are technically not ergodic, i.e., there is no Poincaré recurrence. Moreover, Barrow [@barrow] has pointed out that in a spacetime restricted to be FRW there is necessarily an entropy gap, i.e., the entropy in thermal radiation is much less than the entropy associated with a black hole of horizon size. However, the phenomena we described in the previous paragraph, which are investigated in this paper, are independent of these larger questions about general relativity. That is, the mechanism by which black holes are nucleated occurs on sub-horizon time and length scales. The statistical approach we take below is justified by the presence of a thermal bath of gravitons or other particles, whose existence is not in dispute. In this sense we do not require any assumption of ergodicity, [*except in some small sub-horizon patch.*]{}
In Sec. \[stat\] we consider black hole nucleation in a system of thermal gravitons and compare to a thermal system of matter. In Sec. \[therm\] we show that gravitons may thermalize in the early universe even if they started out cold. We determine the conditions necessary for a phase transition to a black hole condensate via percolation in Sec. \[perc\]. Finally, in Sec. \[disc\] we relate these results to Penrose’s observation. We use Planck units throughout, i.e., $\hbar = c = G = k_{\rm B} = 1$.
Statistical mechanics of gravitons {#stat}
==================================
Consider a box of hot gravitons. The probability for a fluctuation to lead to a black hole of radius $R$ is $$\begin{aligned}
P(R) \sim N e^{-E/T}, \label{prob}\end{aligned}$$ where $E=R/2$ is the energy of the black hole, and $N$ is the multiplicity of microstates which, when coarse grained, appear as a black hole of radius $R$. The probability can be written as $$\begin{aligned}
P(R) \sim e^{-F/T}, \label{free_prob}\end{aligned}$$ where $F = E - TS$ is the free energy, and $S = \ln N$. We assume that black hole entropy is accounted for by gravitational microstates, as suggested by results from string theory [@strominger]. Using the Bekenstein-Hawking formula [@bh_entropy] for black hole entropy $S_{\rm BH} = A/4$, where $A=4 \pi R^2$ is the area, we see that $$F(R) = R/2 - \pi T R^2. \label{free_energy}$$ Strictly speaking, we want the free energy [*relative*]{} to that of hot, flat space. This means we should subtract from (\[free\_energy\]) a correction $F_0 (R) \sim - R^3 T^4$. It is easy to see that, near the saddlepoint found below, $F_0$ is a negligible correction as long as the saddlepoint radius $R_*$ is much smaller than the horizon size.
The radius that maximizes the free energy, the saddlepoint radius, is given by $R_* = (4\pi T)^{-1}$. We obtain $1 \ll R_* \ll H^{-1}$, where $H^{-1} \sim 1/T^2$ is the horizon size for a radiation dominated FRW universe, so our analysis encounters no difficulties from quantum gravity or causality. The corresponding maximum free energy is (see Fig. \[fig\_energy\]) $$\begin{aligned}
F_* = (16 \pi T)^{-1}. \label{f_star}\end{aligned}$$
![Free energy versus black hole radius for a gas of hot gravitons. Black holes of size $R < R_*$ shrink, leading to a weak-field gravity phase. For $R > R_*$, however, black holes grow, possibly giving rise to a new nonperturbative phase of gravity.[]{data-label="fig_energy"}](fig_energy){width="8cm"}
At the saddlepoint radius the black hole temperature is just equal to that of the heat bath. Black holes with $0 < R < R_*$ shrink to zero size, leaving a weak-field phase with a thermal population of graviton states (gravity waves) on a smooth background metric. However, for $R
> R_*$, black holes grow without bound (they are colder than the environment), and for $R > R_0 = (2\pi T)^{-1}$ the free energy is actually negative, less than that of $R=0$. This instability may indicate a new nonperturbative phase of gravity, which is not asymptotically flat, and in which spacetime is all or partially filled with black holes. Such a phase is highly entropic and occupies an exponentially larger phase space volume than the smooth weak-field phase.
The nucleation rate for supercritical black holes (which might be thought of as bubbles of the new nonperturbative gravitational phase) is controlled by the free energy at the saddlepoint, as in the usual case of a first order phase transition (for early papers on nucleation theory, see [@langer]). This yields $$\begin{aligned}
\lambda(T) \sim T^4 e^{-F_*/T}. \label{lambda}~\end{aligned}$$ (Strictly speaking, the dimensional prefactor could be modified by subleading terms in the exponent.) The physics is similar to that of nonperturbative baryon number violation in the standard model. There, the rate is controlled by the free energy of the electroweak sphaleron, which is the saddlepoint configuration separating vacua of different winding number [@sphaleron].
Now consider a box of hot photons. If a fluctuation of size $R$ and energy $E$ satisfies $E > R$ (we ignore factors of order one), it will inevitably evolve into a black hole [@hoop]. Therefore, Eqns. (\[prob\]) and (\[free\_prob\]) for the probability for a fluctuation to lead to a black hole of radius $R$ still hold. In this case, however, the entropy is not proportional to $A$. In order to evaluate the multiplicity $N$, we make use of a bound on the entropy of a region of size $R$ filled with thermal radiation, originally derived by ’t Hooft [@'t; @Hooft]. By noting that matter in thermal equilibrium, energy scales as $E \sim R^3 T^4$, and entropy as $S \sim
R^3 T^3$, and further requiring that the system not have already undergone gravitational collapse, i.e., $E < R$, ’t Hooft obtained the bound $S < A^{3/4}$ (again, we ignore numerical factors). Matter configurations that lead to black holes saturate this bound. Therefore, we see that the free energy of relativistic matter configurations which evolve into black holes is of the form $F(R) \sim
R - T R^{3/2}$.
Once a fluctuation of sufficient size to lead to a black hole has occurred, the evolution is then governed by the same physics as in the original graviton case. We therefore compare the multiplicities of configurations of photons and gravitons, $N_{\gamma}$ and $N_g$, respectively, that will lead to black holes of critical size $R_*$. For temperatures well below the Planck scale, $A \gg A^{3/4}$, and so fluctuations of critical size in the photon gas are suppressed relative to the case of a graviton gas by a factor of roughly $$\begin{aligned}
\frac{N_{\gamma}}{N_g} \sim
\exp\left(-\frac{1}{16 \pi T^2}\right).\end{aligned}$$
The key difference between the two cases considered is the entropy limit on thermal degrees of freedom. Ordinary matter cannot achieve the entropic density of gravitational degrees of freedom, under the assumption that black holes are coarse grained objects with $e^{A/4}$ gravitational microstates. Nucleation of black holes is much more likely if these gravitational microstates are thermally occupied (i.e., in a graviton heat bath), than if one starts with hot matter and cold gravitons.
Graviton thermalization {#therm}
=======================
Natural initial conditions for the universe might have both gravitons and matter in thermal equilibrium. In some cases, however, thermal matter leads to thermal graviton populations even when the gravitons are initially cold. We can investigate the thermalization of gravitons using a long-wavelength effective field theory (EFT) for quantum gravity [@eft; @donoghue]. If one is only interested in processes occurring at energies sufficiently below the Planck scale, as is the case here, there is no obstacle to using the standard EFT approach of including all terms in the effective lagrangian that are consistent with the symmetries of the system, in this case general coordinate transformation invariance.
Without knowledge of the fundamental theory of quantum gravity, we are not able to write down the Boltzmann equation for the evolution of the phase space distribution for the gravitational degrees of freedom. Boltzmann heuristics, however, motivate using $\Gamma(T) = H(T)$ as a reasonable criterion for freeze-out of a given particle species, in this case the graviton. $\Gamma(T)$ is the interaction rate of gravitons with the heat bath at temperature $T$, and $H(T)$ is the Hubble expansion rate at temperature $T$. Below the decoupling temperature $T_{\rm dec}$ the Hubble expansion rate is greater than the interaction rate of gravitons with the heat bath, and gravitons are decoupled.
Let us estimate the graviton decoupling temperature based on two different scattering processes. First, consider the interaction $X_i
X_i \rightarrow gg$, where $X_i$ is any one of $N$ scalar, fermion or vector particles, and $g$ is the graviton. In a relativistic gas, Boltzmann heuristics suggest that the interaction rate is roughly $\Gamma_{X_i X_i \rightarrow gg} \sim n_i \sigma_i$, where $n_i$ and $\sigma_i$ are the number density and cross section, respectively, of species $i$. At energies not too far below the Planck scale all particle species are relativistic so $n_i \sim T^3$. The matrix element for this process goes as $T^2$ and the Hubble rate as $H \sim
N^{1/2}T^2$. We see that the decoupling temperature obtained from considering the process $\sum_i X_i X_i \rightarrow gg$ is $T_{\rm
dec} \sim N^{-3/2}$. For a model with a large number of particle species this scale may be well below the Planck scale.
This estimate, however, does not take into account interactions between particle species. A similar argument based on considering the process $X_i X_i \rightarrow \gamma g$, where $\gamma$ is another particle species, leads to a decoupling temperature of roughly $T_{\rm
dec}^{\gamma} \sim N^{-1/2}\alpha^{-1}$, where $\alpha^{1/2}$ is the coupling between $X_i$ and $\gamma$. Again $T_{\rm dec}^{\gamma}$ may be significantly below the Planck scale.
Let us obtain a more careful estimate of $T_{\rm dec}^{\gamma}$ in a specific model: long-wavelength quantum gravity (EFT) coupled to scalar QED (see, e.g., [@scalar_qed]). For a tree level calculation, the relevant terms in the effective lagrangian are: $$\begin{aligned}
{{\cal L}}&=&
|g|^{\frac{1}{2}}\left[g^{\mu\nu}(D_{\mu}\phi)^*(D_{\nu}\phi)
- \tfrac{1}{4}g^{\mu\nu}g^{\rho\sigma}
F_{\mu\rho}F_{\nu\sigma} \right] {\nonumber}\\
&+&\frac{1}{16 \pi}|g|^{\frac{1}{2}}{{\cal R}},
\label{L_qed}\end{aligned}$$ where $D_{\mu} = {\partial}_{\mu} + ieA_{\mu}$, $F_{\mu\nu} = {\partial}_{\mu}A_{\nu}
- {\partial}_{\nu}A_{\mu}$, and ${{\cal R}}$ is the scalar curvature. The metric $g_{\mu\nu}$ is expanded about the Minkowski metric $\eta_{\mu\nu} =
{\rm diag}(1,-1,-1,-1)$, i.e., $g_{\mu\nu}=\eta_{\mu\nu} + \sqrt{32
\pi} h_{\mu\nu}$, and Feynman rules are obtained by the standard procedure. The cross section for the interaction $\phi\phi \rightarrow
\gamma g$ is found to be $$\begin{aligned}
\sigma_{\phi\phi \rightarrow \gamma g} = \frac{32 \pi}{6}\alpha,
\label{sigma}\end{aligned}$$ where $\alpha = e^2 / 4\pi$. Assuming local thermal equilibrium, the scalar number density is $n_{\phi} = (2 \zeta(3)/
\pi^2)T^3$. Moreover, if there are $N_s$ species of scalar particles, the graviton decoupling temperature is $$\begin{aligned}
T_{\rm dec}^{\gamma} \simeq \frac{3}{5\alpha N_s^{1/2}}.
\label{T_dec}\end{aligned}$$
For consistency we must also estimate the temperature below which we can trust the above calculation, i.e., estimate the energy scale below which the EFT description is valid. Let us look at the next order contribution to the matrix element for this process: $$\begin{aligned}
{{\cal M}}_{\phi\phi \rightarrow \gamma g} =
{{\cal M}}_{\phi\phi \rightarrow \gamma g}^{(1)}
+ {{\cal M}}_{\phi\phi \rightarrow \gamma g}^{(3)} + \ldots.
\label{matrix_el}\end{aligned}$$ Notice ${{\cal M}}_{\phi\phi \rightarrow \gamma g}^{(1)} = {{\cal O}}(e\sqrt{32\pi}q)$, where $q$ is some typical energy scale of the interaction. In order to obtain a conservative estimate of ${{\cal M}}_{\phi\phi \rightarrow \gamma g}^{(3)}$, we will assume that all diagrams interfere constructively. Then ${{\cal M}}_{\phi\phi \rightarrow
\gamma g}^{(3)} = {{\cal O}}(FLe(\sqrt{32\pi}q)^3)$, where $F$ is the number of graphs at this order and $L$ is a loop factor associated with each graph. The EFT is valid for $$\begin{aligned}
q \ll q_{\rm pert} = (32 \pi FL)^{-1/2}.
\label{q_pert}\end{aligned}$$ Taking $F=27$ and $L = 1/4\pi^2$ we get $q_{\rm pert} \simeq 10^{-1}$. By comparing Eqs. (\[T\_dec\]) and (\[q\_pert\]), we see that $T_{\rm dec}^{\gamma} \ll q_{\rm pert}$ in the large $N_s$ limit, due to the fact that $T_{\rm dec}^{\gamma} \sim N_s^{-1/2}$, while $q_{\rm
pert}$ is independent of $N_s$. Thus, there is a class of models (those with large numbers of matter fields) in which gravitons are copiously produced and interact frequently, at energy scales where the effective field theory applies.
One could perform a similar calculation with a more realistic model for the matter content of the early universe. At the energy scales of interest, say $10^{16}$ GeV, the matter interactions may, for example, be described by a grand unified theory (GUT) [@gut] with $N_s \sim
10^2$ and couplings not much smaller than unity. A calculation in such a model would likely yield the same conclusion, i.e., that there was an epoch in which even initially cold gravitons interacted strongly enough with the heat bath that they were themselves thermalized.
Percolation of black holes {#perc}
==========================
In light of the previous sections, it is interesting to consider the possibility of a first order phase transition in an FRW universe with a thermal population of gravitons. Assuming that gravitons are in thermal equilibrium in the early universe between an initial time $t_0$ and a later time $t_1$, the volume fraction in the weak-field gravity phase is given by: $$\begin{aligned}
p(t_0,t_1) = \exp\left[
-\int_{t_0}^{t_1} dt' V(t',t_1)\lambda(t') \right].
\label{vol_frac}\end{aligned}$$ In flat FRW spacetime $$\begin{aligned}
V(t',t_1) = \frac{4\pi}{3}\left[a(t')b
\int_{t'}^{t_1} \frac{dt''}{a(t'')}\right]^3.\end{aligned}$$ Here $a$ is the scale factor, and $b$ is the (constant) speed at which the black holes expand. Eq. (\[vol\_frac\]) is a standard formula from old inflationary cosmology [@guth]. $b$ is governed by relativistic physics and, therefore, should not be significantly smaller than unity. During a radiation dominated epoch the scale factor goes as $a(t) \sim t^{1/2}$ and temperature and time are related by $t \simeq 0.3 g_*^{-1/2}T^{-2}$. The volume fraction (now as a function of temperature) is then of the form: $$\begin{aligned}
p(T_0,T_1) = \exp\left[
-\frac{b^3}{g_*^2}f(T_0,T_1) \right].\end{aligned}$$ We define the critical initial temperature $T_0^c$, so that if the initial graviton temperature is larger than $T_0^c$, the volume fraction in the weak-field phase is nearly zero. See Fig. \[fig\_perc\] for a plot of $p(T_0,T_1)$ for $b = 10^{-1}$, $10^{-2}$, and $10^{-3}$, with $T_1 = 10^{-6}$ and $g_* =
10^2$. Decreasing the expansion speed by two orders of magnitude has the effect of increasing $T_0^c$ by less than a factor of two. Increasing the number of effective degrees of freedom or the cut-off temperature $T_1$ both have similar effects on $T_0^c$, i.e., they cause it to increase by a relatively small amount.
The critical initial temperature $T_0^c$ is of order $10^{-2}$ to $10^{-1}$, which from Eq. (\[q\_pert\]) is the same order of magnitude as the temperature at which quantum gravity becomes perturbative. Thus, taking the initial temperature of the big bang to be roughly the scale at which gravity becomes perturbative, the universe could avoid a phase transition to the nonperturbative black hole phase, assuming it begins in the weak-field phase.
![Volume fraction in the weak-field phase of gravity as a function of the temperature $T_0$ at which gravitons enter thermal equilibrium. Shown are plots for three different values of the expansion speed, $b = 10^{-1}$, $10^{-2}$, $10^{-3}$ (left to right). Here the cut-off temperature $T_1 = 10^{-6}$, and the effective number of degrees of freedom $g_* = 10^2$.[]{data-label="fig_perc"}](fig_perc){width="8cm"}
Discussion {#disc}
==========
We examined a possible first order phase transition of spacetime to a black hole phase with high entropy. Percolation of the high entropy phase occurs if gravitons are ever in a thermal state with temperature above $T_0^c$, either because they were born hot at the Planck epoch or because they were thermalized due to interactions with thermal matter. It seems possible, as suggested by entropic arguments, that almost all of gravitational phase space is accounted for by the nonperturbative phase. However, we find that the low entropy phase is metastable over timescales which are exponentially sensitive to the temperature, and potentially quite long.
One may wonder how inflation changes this conclusion. We note that $T_0^c$ is higher than the energy scale at which inflation is usually assumed to take place. If gravitons are never thermalized above a temperature of $T_0^c$ then presumably inflation would simply take place as originally envisioned. However, if gravitons are ever thermal with a temperature above $T_0^c$ then we would speculate that it may be less probable for a given patch to inflate, although depending on the details of the model some non-zero probability may remain, even if a phase transition to the high entropy black hole phase does occur.
Hot gravitons with temperature slightly below $T_0^c$ will not lead to a phase transition; they will simply be red-shifted away. Both gravitons and matter may be born hot, as long as the temperature of the universe (either initially or after a period of inflation) is never greater than $T_0^c$. This does not require fine tuning because $T_0^c$ is of the same order as $q_{\rm pert}$, the energy scale below which quantum gravity effects are small. It may still be the case that the initial conditions represent a subset of measure zero in the total phase space, which is dominated by the nonperturbative black hole phase [@anthropic]. However, our analysis does show that once the initial choice of the low entropy phase is made, no transition to the high entropy phase need occur.
These conclusions remain unchanged in a spacetime of arbitrary dimension $d$. For hot gravitons in $d$ dimensions, the exponent governing the nucleation rate of Eq. (\[lambda\]) goes as $F_*/T
\sim T^{-(d-2)}$, while for matter $F_*/T \sim T^{-(d-1)(d-2)/2}$. For $d > 3$, black hole nucleation is suppressed more strongly in the matter system than in the gravitational one. Moreover, as $d
\rightarrow \infty$, $T_0^c$ increases, meaning a transition to the black hole condensate phase is less likely in a universe with a large number of spacetime dimensions.
[**Note added**]{}: After this paper was completed we became aware of earlier work using Euclidean path integral methods in which Eqs. (\[free\_energy\]), (\[f\_star\]), and (\[lambda\]) were independently derived [@gross]. In these calculations imaginary time boundary conditions are applied to the gravitational field. Therefore, those authors were also studying thermal gravity and not only thermal matter. For discussion of black hole phase transitions, see [@hu].
**Acknowledgments**
The authors thank R. Buniy and Y. S. Myung for useful comments. This work was supported by the Department of Energy under DE-FG06-85ER40224.
[99]{}
R. Penrose, [*The Emperor’s New Mind*]{}, Oxford University Press, Oxford, 1989; R. Penrose, [*The Road to Reality: A Complete Guide to the Laws of the Universe*]{}, Knopf, New York, 2005.
Gravitational entropy is maximized by clumping because gravity is a long range, unscreened, attractive force. This is quite unlike non-gravitational systems (with screened or limited range interactions) whose entropy is maximized when matter and energy are uniformly distributed.
The use of terms such as entropy, ergodicity or phase space volume to describe configurations in general relativity (spacetimes) might be unfamiliar. However, classical general relativity can be considered a dynamical system (i.e., a set of rules for evolving a set of initial data given on some spacelike slice) like any other. Then, if the usual assumptions of statistical mechanics hold, time averaging is equivalent to ensemble averaging (ergodicity), and the system is overwhelmingly likely to be found in states of maximum entropy.
S. M. Carroll and J. Chen, arXiv:hep-th/0410270; S. M. Carroll and J. Chen, arXiv:gr-qc/0505037; R. Holman and L. Mersini-Houghton, arXiv:hep-th/0511102; R. Holman and L. Mersini-Houghton, arXiv:hep-th/0511112.
R. M. Wald, arXiv:gr-qc/0507094.
F. Tipler, Nature [**280**]{}, 203 (1979).
J. D. Barrow, New Astron. [**4**]{}, 333 (1999) \[arXiv:astro-ph/9903225\].
One can also associate an entropy with the presence of a cosmological horizon. Interestingly, when a black hole is added to de Sitter space the total entropy associated with both the black hole and cosmological horizons decreases; see, e.g., D. Klemm and L. Vanzo, JCAP [**0411**]{}, 006 (2004) \[arXiv:hep-th/0407255\]. However, the interpretation of the entropy associated with a cosmological horizon seems to us even more challenging than that of black holes, and so we choose to focus on the latter.
K. S. Thorne, [*Nonspherical gravitational collapse: A short review*]{}, in J. R. Klauder, [*Magic Without Magic*]{}, San Francisco 1972, 231-258; D. M. Eardley and S. B. Giddings, Phys. Rev. D [**66**]{}, 044011 (2002) \[arXiv:gr-qc/0201034\]; S. D. H. Hsu, Phys. Lett. B [**555**]{}, 92 (2003) \[arXiv:hep-ph/0203154\].
G. ’t Hooft, arXiv:gr-qc/9310026.
A. Strominger and C. Vafa, Phys. Lett. B [**379**]{}, 99 (1996), \[arXiv:hep-th/9601029\].
J. D. Bekenstein, Phys. Rev. D [**7**]{}, 2333 (1973); S. W. Hawking, Commun. Math. Phys. [**43**]{}, 199 (1975) \[Erratum-ibid. [**46**]{}, 206 (1976)\].
J. S. Langer, Annals Phys. [**41**]{}, 108 (1967) \[Annals Phys. [**281**]{}, 941 (2000)\]; J. S. Langer, Annals Phys. [**54**]{}, 258 (1969).
P. Arnold and L. D. McLerran, Phys. Rev. D [**36**]{}, 581 (1987); P. Arnold and L. D. McLerran, Phys. Rev. D [**37**]{}, 1020 (1988).
S. Weinberg, PhysicaA [**96**]{}, 327 (1979).
J. F. Donoghue, Phys. Rev. D [**50**]{}, 3874 (1994), \[arXiv:gr-qc/9405057\].
N. E. J. Bjerrum-Bohr, Phys. Rev. D [**66**]{}, 084023 (2002), \[arXiv:hep-th/0206236\].
G. G. Ross, [*Grand Unified Theories*]{}, Benjamin/Cummings, 1984.
A. H. Guth and E. J. Weinberg, Nucl. Phys. B [**212**]{}, 321 (1983).
Of course, one could always invoke anthropic arguments to exclude the black hole phase, as it is hard to imagine how life might evolve there.
D. J. Gross, M. J. Perry and L. G. Yaffe, Phys. Rev. D [**25**]{}, 330 (1982); B. F. Whiting and J. W. York, Phys. Rev. Lett. [**61**]{}, 1336 (1988).
G. J. Stephens and B. L. Hu, Int. J. Theor. Phys. [**40**]{}, 2183 (2001) \[arXiv:gr-qc/0102052\].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Jeffrey Burdges[^1]'
bibliography:
- 'burdges.bib'
- 'fMr.bib'
title: |
On Frattini arguments in $L$-groups\
of finite Morley rank
---
\[p:invCarter\] Let $\hat{G}$ be a group of finite Morley rank, let $G$ itself be a definable connected normal subgroup group of $\hat{G}$, and let $S$ be a Sylow 2-subgroup of $\hat{G}$. Then $G$ has an $S$-invariant Carter subgroup.
We use the following facts, as well as the -Jaligot construction and the conjugacy of descent tori.
\[nildecomp\] Let $H$ be a nilpotent group of finite Morley rank. Then $H = D * B$ is a central product of definable characteristic subgroups $D,B \leq H$ where $D$ is divisible and $B$ has bounded exponent (which is connected iff $H$ is connected). Let $T$ be the torsion part of $D$. Then we have decompositions of $D$ and $B$ as follows. $$\begin{aligned}
D &=& d(T) * U_{0,1}(H) * U_{0,2}(H) * \cdots \\
B &=& U_2(H) \oplus U_3(H) \oplus U_5(H) \oplus \cdots\end{aligned}$$
\[Uz\_on\_Up\] Let $H$ be a connected solvable group of finite Morley rank. Suppose that $S$ is a nilpotent ${U_{0,r}}$-subgroup of $H$, and that $H = U_p(H) S$ for some $p$ prime. Then $H$ is nilpotent, and $[U_p(H),S]=1$.
\[nilpotencepre2\] Let $H = K T$ be a group of finite Morley rank with $K \normal H$ a nilpotent $U_{0,r}$-group and $T$ a nilpotent $U_{0,s}$-group for some $s \geq r$. Then $H$ is nilpotent.
We take $G$ to be a minimal counterexample, i.e. proper connected $S$-invariant subgroups of $G$ have $S$-invariant Carter subgroups. Then $G$ is nonsolvable by a Frattini argument. We observe that Carter subgroups of $G$ correspond to Carter subgroups of $G/Z^\o(G)$. So $G$ is centerless too. We may also assume that $C_S(G) = 1$.
If $G$ contains divisible torsion, we take $T$ to be a maximal descent torus of $G$ (exists by [@Ch05]). By [@Ch05], all maximal descent tori are conjugate, so $G = N_G(T) G$ by a Frattini argument. Hence a conjugate or $T$ is $S$-invariant. There is a Carter subgroup $Q$ containing $T$, by a -Jaligot construction. By Fact [@BN 6.16], $T$ is central in $M := N^\o_G(T)$. Now we have $M < G$, since $G$ is centerless. Since $M$ is $S$-invariant, there is an $S$-invariant Carter subgroup $Q$ of $M$. Since $T$ is a characteristic subgroup of $Q$, $N^\o_G(Q) \leq N^\o_G(T) \leq M$. So $Q$ is a Carter subgroup of $G$. Hence we may assume that $G$ has no divisible torsion.
We find a maximal $S$-invariant group $R$ with “minimal degree of unipotence” $r$. Let $r$ be the minimal reduced rank such that there is an $S$-invariant nilpotent ${U_{0,r}}$-subgroup of $G$, or take $r = \infty$ if there are no $S$-invariant nilpotent ${U_{0,r}}$-subgroups. Let $i$ be a central involution of $S$. Since $G$ is nonabelian, $C^\o_G(i)$ is infinite by [@BN Ex. 13 & 15 on p. 79]. and clearly $S$-invariant. Since $C^\o_G(i) < G$, $C^\o_G(i)$ has a definable $S$-invariant nilpotent subgroup, by induction. So, if $r = \infty$, there is an $S$-invariant nilpotent group of bounded exponent, by Fact \[nildecomp\]. Let $R$ be a maximal $S$-invariant nilpotent ${U_{0,r}}$-subgroup of $G$, or else a maximal $S$-invariant $p$-unipotent subgroup of $G$ if $r = \infty$. Now consider the $S$-invariant group $M := N^\o_G(R)$.
First suppose that $M < G$. Then there is an $S$-invariant Carter subgroup $Q$ of $M$, by induction. Now $R Q$ is solvable. We show that $R$ is characteristic in $Q$. If $r = \infty$, $R Q$ is nilpotent by Fact \[Uz\_on\_Up\], and $R Q$ has bounded exponent. So $R = Q$ by maximality. If $r < \infty$, $r$ is the minimal reduced rank found in $Q$, since $U_{0,t}(Q)$ is $S$-invariant for all $t$. So $R Q$ is nilpotent by Fact \[nilpotencepre2\]. $R \leq Q$, and hence ${U_{0,r}}(Q) = R$. In either case, $N^\o_G(Q) \leq M$. So $Q$ is a Carter subgroup of $G$. Next suppose that $M = G$, i.e. $R \normal G$. Then there is an $S$-invariant Carter subgroup $\bar{Q}$ of $\bar{G} = G/R$. Now the pullback $Q$ of $\bar{Q}$ to $G$ is solvable, and $R \leq Q$. Since $G$ is nonsolvable, $\bar{Q} < \bar{G}$, and $Q < G$. Since $Q$ is $S$-invariant, there is an $S$-invariant Carter subgroup $L$ of $Q$, by induction. We show that $Q = R L$ is nilpotent. If $r = \infty$, $R L$ is nilpotent by Fact \[Uz\_on\_Up\]. If $r < \infty$, $r$ is the minimal reduced rank found in $L$ (or $Q$), since $U_{0,t}(L)$ is $S$-invariant for all $t$. So $R L$ is nilpotent by Fact \[nilpotencepre2\], and $R \leq L$. By maximality of $R$, ${U_{0,r}}(L) = R$. By Fact \[nildecomp\]. $U_{0,t}(L)$ centralizes $R$ for $t > r$. In either case, $Q = R L$ is nilpotent. We observe that $N_G(Q) / R \leq N_{\bar{G}}(\bar{Q})$ has the same rank as $\bar{Q}$. So $\rk(N_G(Q)) = \rk(Q)$, and $N^\o_G(Q) = Q$. Thus $Q$ is an $S$-invariant Carter subgroup of $G$.
[^1]: Supported by NSF grant DMS-0100794, and Deutsche Forschungsgemeinschaft grant Te 242/3-1.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the relationship between the quenching of star formation and the structural transformation of massive galaxies, using a large sample of photometrically-selected post-starburst galaxies in the UKIDSS UDS field. We find that post-starburst galaxies at high-redshift ($z>1$) show high Sérsic indices, significantly higher than those of active star-forming galaxies, but with a distribution that is indistinguishable from the old quiescent population. We conclude that the morphological transformation occurs before (or during) the quenching of star formation. Recently quenched galaxies are also the most compact; we find evidence that massive post-starburst galaxies (M$_{\ast}> 10^{10.5} ~$M$_{\sun}$) at high redshift ($z>1$) are on average smaller than comparable quiescent galaxies at the same epoch. Our findings are consistent with a scenario in which massive passive galaxies are formed from three distinct phases: (1) gas-rich dissipative collapse to very high densities, forming the proto-spheroid; (2) rapid quenching of star formation, to create the “red nugget” with post-starburst features; (3) a gradual growth in size as the population ages, perhaps as a result of minor mergers.'
author:
- |
Omar Almaini$^{1}$[^1], Vivienne Wild$^{2}$, David T. Maltby$^{1}$, William G. Hartley$^{3}$, Chris Simpson$^{4}$, Nina A. Hatch$^{1}$, Ross J. McLure$^{5}$, James S. Dunlop$^{5}$, Kate Rowlands$^{2}$\
$^{1}$School of Physics and Astronomy, University of Nottingham, University Park, Nottingham NG7 2RD, U.K.\
$^2$School of Physics and Astronomy, University of St Andrews, North Haugh, St Andrews, KY16 9SS, U.K.\
$^3$ETH Zürich, Institut für Astronomie, HIT J 11.3, Wolfgang-Pauli-Str. 27, 8093 Zürich, Switzerland\
$^4$Gemini Observatory, Northern Operations Center, 670 N. A’ohuku Place, Hilo, HI96720, USA\
$^5$Institute for Astronomy, University of Edinburgh, Royal Observatory, Blackford Hill, Edinburgh, EH9 3HJ, U.K.\
date: 'Accepted 2017 July 28. Received 2017 July 24; in original form 2016 February 3.'
title: 'Massive post-starburst galaxies at $z>1$ are compact proto-spheroids'
---
\[firstpage\]
galaxies: evolution – galaxies: formation – galaxies: fundamental parameters – galaxies: structure – galaxies: high-redshift
Introduction {#sec:intro}
============
Galaxies in the local Universe display a striking bimodality in their morphological and spectral characteristics; massive galaxies (M$_{\ast}> 10^{10.5} ~$M$_{\sun}$) are typically spheroidal with old stellar populations, while lower mass galaxies are typically disc-dominated with blue, younger stellar populations (e.g. Strateva et al. 2001; Hogg et al. 2002). Deep surveys have revealed that the most massive galaxies were formed at high redshift ($z>1$; e.g. Fontana et al. 2004; Kodama et al. 2004; Cirasuolo et al. 2010), but we still do not understand why their star formation was abruptly terminated. Feedback from AGN (e.g. Silk & Rees 1998; Hopkins et al. 2005) or starburst-driven superwinds (e.g. Diamond-Stanic et al. 2012) are leading contenders for rapidly quenching distant galaxies, while jet-mode AGN feedback may be required to maintain the “red and dead” phase (e.g. Best et al. 2006).
In addition to the quenching of star formation, massive galaxies must also undergo a dramatic structural transformation, to produce the spheroid-dominated population we see today. The transition appears to occur at $z>1$ for most galaxies with M$_{\ast}>
10^{10.5} ~$M$_{\sun}$ (Mortlock et al. 2013; Bruce et al. 2014), but it is unclear if quenching and structural transformation occurred during the same event. Over the last 10 years it has also emerged that quiescent galaxies were significantly more compact in the early Universe compared to the present day (e.g. Daddi et al. 2005; Trujillo et al. 2006; van Dokkum et al. 2008; Belli et al. 2014). Plausible explanations for the dramatic size growth include minor mergers (e.g. Bezanson et al. 2009; Naab, Johansson & Ostriker 2009) or expansion due to mass loss (Fan et al. 2008). For the population of quiescent galaxies as a whole, however, there may also be an element of progenitor bias; galaxies quenched at lower redshift tend to be larger than their counterparts at early times, which may drive much of the observed size evolution (e.g. Poggianti et al. 2013; Carollo et al. 2013).
From a theoretical perspective, the formation of ultra-compact massive spheroids requires the concentration of vast reservoirs of cool gas via dissipation, which can radiate and collapse to very high densities. A variety of models have arisen to explain spheroid formation in detail, with some invoking gas-rich mergers (e.g. Hopkins et al. 2009; Wellons et al. 2015) while others use the inflow of gas through cold streams, feeding an extended disc that eventually becomes unstable and contracts (e.g. Dekel et al. 2009; Zolotov et al. 2015). Outflows driven by AGN or star formation may then terminate the star formation by expelling the remainder of the gas (e.g. Hopkins et al. 2005).
In this work we explore the relationship between the quenching of distant galaxies and their structural transformation. We focus in particular on the rare class of “post-starburst” (PSB) galaxies, which are observed a few hundred Myr after a major episode of star formation was rapidly quenched. In the local Universe PSBs are identified from characteristic strong Balmer absorption lines (Dressler & Gunn 1983; Wild et al. 2009), due to the strong contribution from A stars, but until recently very few were spectroscopically identified at $z>1$ (e.g. Vergani et al. 2010). Two photometric methods have therefore been developed to identify this population. Whitaker et al. (2012) used medium-band photometry from the NEWFIRM Medium-Band Survey to identify “young red-sequence” galaxies using rest-frame UVJ colour-colour diagrams. In Wild et al. (2014) an alternative approach was used, based on a Principal Component Analysis (PCA) applied to the deep multi-wavelength photometry in the UDS field. Three spectral shape parameters (“supercolours”) were found to provide a compact representation for a wide range of photometric SEDs. In addition to cleanly separating quiescent and star-forming galaxies, the PCA method identifies “post-starburst” galaxy candidates in a distinct region of supercolour space, corresponding to galaxies in which a significant amount of mass was formed within the last Gyr but then rapidly quenched. The method was recently verified with deep 8m spectroscopy from VLT, which established that between 60% and 80% of photometric candidates are spectroscopically confirmed post-starburst galaxies, depending on the specific criteria adopted (Maltby et al. 2016)[^2]. In terms of completeness, the photometric method was found to identify approximately 60% of galaxies that would be spectroscopically classified as PSBs. Overall, these figures confirm that photometric PCA techniques can be used to identify large and relatively clean samples of recently quenched galaxies. So far the spectroscopic confirmation is restricted to $z<1.4$, given the requirement to detect H$\delta$ with optical spectroscopy. Future near-infrared spectroscopy (e.g. with the [*James Webb Space Telescope*]{}) will allow a detailed investigation of the population at higher redshift.
The identification of large PSB samples has allowed the first study of the PSB galaxy mass function and its evolution to $z=2$ (Wild et al. 2016). Strong evolution was observed, with the implication that a large fraction of massive galaxies are rapidly quenched and pass through a PSB phase. In this paper we use the unique PCA sample described in Wild et al. (2016) to explore the structural properties of post-starburst galaxies. As newly quenched systems, our primary aim is to investigate if this population is structurally similar to star-forming galaxies at the same epoch, or if they already show evidence for the compact spheroid-dominated morphology of well-established quiescent galaxies.
We assume a cosmology with $\Omega_M=0.3$, $\Omega_\Lambda=0.7$ and $h=0.7$. All magnitudes are given in the AB system.
Data and sample selection
=========================
The UDS K-band galaxy sample
----------------------------
Our study is based on deep $K$-band imaging from the UKIRT Infrared Deep Sky Survey (UKIDSS; Lawrence et al. 2007) Ultra-Deep Survey (UDS; Almaini et al., in preparation). The UDS is the deepest of the UKIDSS surveys, covering $0.77$ square degrees in the $J$, $H$ and $K$ bands. We use the 8th UDS data release (Hartley et al. 2013), reaching depths of $J=24.9$, $H=24.2$ and $K=24.6$ (AB, 5$\sigma$, 2-arcsec apertures). The final UDS data release (June 2016) achieved estimated depths $J=25.4, H=24.8, K=25.3$, and will be used to extend our PSB studies in future work.
To complement the near-infrared imaging from UKIDSS, the UDS has deep optical coverage from [*Subaru*]{} Suprime-CAM, to depths of $B= 27.6,
~ V= 27.2, ~ R = 27.0, ~ i^{\prime} = 27.0$ and $ z^{\prime} = 26.0$ (AB, 5$\sigma$, $2$ arcsec), as described in Furusawa et al. (2008). Additional $u^{\prime}$-band imaging is provided by the Canada-France-Hawaii Telescope (CFHT) MegaCam instrument, reaching $ u^{\prime} = 26.75$ (AB, 5$\sigma$, $2$ arcsec). Deep imaging at longer near-infrared wavelengths is provided by the [*Spitzer*]{} UDS Legacy Program (SpUDS; PI: Dunlop), achieving depths of 24.2 and 24.0 (AB) using the IRAC camera at 3.6$\mu$m and 4.5$\mu$m respectively. The resulting area with full multiwavelength coverage, following the masking of bright stars and artefacts, is 0.62 square degrees. Further details on the construction of the multiwavelength DR8 catalogue can be found in Hartley et al. (2013).
We determined photometric redshifts using the techniques outlined in Simpson et al. (2013). The 11-band photometric data were fit using a grid of galaxy templates, assembled using simple stellar populations from Bruzual & Charlot (2003; hereafter BC03), with a logarithmic spacing of ages between 30 Myr and 10 Gyr, and the addition of younger templates with dust-reddened spectral energy distributions (SEDs). The additional templates consist of a mildly reddened ($A_V=0.25$ mag) version of the two youngest templates, plus a version of the 30 Myr template with heavier reddening ($A_V=1.0$ mag). The resulting photometric redshifts show a normalized median absolute deviation $\sigma_{{\rm NMAD}} = 0.027$. The stellar masses used in this work differ slightly from those presented in Simpson et al. (2013), and instead are based on supercolour templates (see Section \[sec:masses\]). As noted in Section \[sec:uncertainties\], we investigated a range of alternative stellar masses, including those from Simpson et al. (2013), and found no significant impact on our conclusions.
Classification using PCA supercolours
-------------------------------------
We classify galaxies using a Principal Component Analysis (PCA) method applied to the broad-band photometric data, using the techniques outlined in Wild et al. (2014; hereafter W14). The aim of the PCA analysis is to describe the variation in galaxy SEDs using the linear combination of only a small number of components. The components are derived using a library of spectral synthesis models from BC03, using a wide range of stochastic star formation histories, metallicities, and dust-reddening. We found that only three principal components (effectively low-resolution “eigenspectra”) are required, in linear combination, to account for $>$99.9% of the variance in photometric SEDs. The amplitude of each component defines a “supercolour”, analogous to a traditional colour but defined using all available photometric bands. Supercolours allow the comparison of SEDs without the need for model fitting, and galaxies with extreme properties are free to have colours that differ substantially from any of the input model components.
Using the supercolour technique, we separated the UDS galaxy population into three categories; passive galaxies (with low specific star-formation rates), star-forming galaxies, and post-starburst galaxies. Post-starburst galaxies are identified as a well-defined stream of galaxies in supercolour space, consistent with quiescent stellar populations in which a large fraction ($>$10%) of the stellar mass was formed within the last $\sim$1 Gyr, with the star formation then rapidly quenched (W14; see also Wild et al. 2016). The precise boundary between the passive and post-starburst population is set by the ability to observe strong Balmer absorption lines in optical spectroscopy (W14; Maltby et al. 2016). As outlined in Wild et al. (2016), evolutionary tracks suggest that not all photometric PSBs may have necessarily undergone a short-lived ‘burst’ of star formation. The population may also include galaxies that have undergone more extended ($\la$ 3 Gyr) periods of star formation, but the key characteristic is the rapid quenching of star formation within the last $\sim$1 Gyr.
In W14, four categories of star-forming galaxies were identified (SF1, SF2, SF3 and dusty), which we combine for the purposes of this work. As outlined in W14 (see also Section \[sec:uvj\]) the classification by supercolours is in very good agreement with the separation of galaxies using the more traditional rest-frame UVJ technique for separating star-forming and passive galaxies (Labbé et al. 2005; Wuyts et al. 2007). There is also good agreement with the UVJ method of Whitaker et al. (2012), who identified recently-quenched candidates at the blue end of the passive UVJ sequence.
![The distribution of stellar mass as a function of photometric redshift for the three primary galaxy populations. The curves show the corresponding 95% completeness limits, determined using the method of Pozzetti et al. (2010). Details of the galaxy classification and stellar mass determination can be found in Wild et al. (2016). In this paper we focus on the structural properties of galaxies in the redshift range $1<z<2$.[]{data-label="fig:zmass"}](zmass.pdf){width="0.9\linewidth"}
![A comparison of size measurements for galaxies measured with ground-based $K$-band imaging and HST $H$-band imaging, using the subsample of $\sim$9% of galaxies within the UDS CANDELS mosaic. The upper panel compares half-light radii for individual galaxies. We find that ground-based imaging can provide robust measurements of galaxy sizes, although the relationship is not precisely 1:1. A linear fit to all points suggests that ground-based sizes are $\sim$13% smaller on average (dashed line). A characteristic uncertainty on individual measurements is shown, representing the median error on the PSB sizes. The lower (binned) distribution compares mean sizes for the three key populations, with unfilled symbols representing mean values determined from fewer than 5 galaxies. We find no evidence for a significant systematic bias in ground-based size determinations with spectral type. Galaxies shown have photometric redshifts in the range $1<z<2$ and stellar masses with log (M$_{\ast}$/M$_{\sun}$)$>10$, to match the samples used in this work.[]{data-label="fig:size-test"}](rtest.pdf "fig:"){width="0.9\linewidth"} ![A comparison of size measurements for galaxies measured with ground-based $K$-band imaging and HST $H$-band imaging, using the subsample of $\sim$9% of galaxies within the UDS CANDELS mosaic. The upper panel compares half-light radii for individual galaxies. We find that ground-based imaging can provide robust measurements of galaxy sizes, although the relationship is not precisely 1:1. A linear fit to all points suggests that ground-based sizes are $\sim$13% smaller on average (dashed line). A characteristic uncertainty on individual measurements is shown, representing the median error on the PSB sizes. The lower (binned) distribution compares mean sizes for the three key populations, with unfilled symbols representing mean values determined from fewer than 5 galaxies. We find no evidence for a significant systematic bias in ground-based size determinations with spectral type. Galaxies shown have photometric redshifts in the range $1<z<2$ and stellar masses with log (M$_{\ast}$/M$_{\sun}$)$>10$, to match the samples used in this work.[]{data-label="fig:size-test"}](rtest_bin_mean.pdf "fig:"){width="0.9\linewidth"}
Stellar masses {#sec:masses}
--------------
We determined stellar masses for each galaxy using a Bayesian analysis, to account for the degeneracy between physical parameters. Further details may be found in Wild et al. (2016). A library of 10’s of thousands of population synthesis models was created using BC03 and fit to the supercolours to obtain a probability density function for each physical property. A wide range of star formation histories, dust properties, and metallicities were explored, including exponentially declining star formation rates with superimposed stochastic starbursts. Stellar masses were calculated assuming a Chabrier initial mass function (IMF), defined as the stellar mass at the time of observation, i.e. allowing for the fraction of mass in stars returned to the interstellar medium due to mass loss and supernovae. The resulting stellar mass uncertainties from the Bayesian fits are typically $\pm ~0.1$ dex for all populations, assuming BC03 stellar population synthesis models and allowing for uncertainties in the photometric redshifts. The potential impact of stellar mass errors (random and systematic) is discussed further in Section \[sec:uncertainties\].
Figure \[fig:zmass\] shows the resulting distribution of stellar mass as a function of redshift, separated into the three primary populations. Mass completeness limits (95%) were determined as a function of redshift using the method of Pozzetti et al. (2010). We note that the 95% completeness limit for star-forming galaxies appears to be surprisingly high in this diagram (formally slightly higher than the PSB population). This is caused by the wide range in mass-to-light ratios within the star-forming population.
As outlined in Wild et al. (2016), the PSB mass function evolves strongly with redshift, so that the comoving space density of massive PSBs (M$_{\ast}> 10^{10} ~$M$_{\sun}$) is several times higher at $z\sim 2$ than at $z\sim 0.5$. This trend is apparent in Figure \[fig:zmass\], which shows a sharp decline in the number of massive PSBs at $z<1$. The majority of PSBs at $z<1$ are close to the 95% completeness limit (see Figure \[fig:zmass\]) and typically very faint; the median $K$-band magnitude for PSBs at $z<1$ is $K=23.0$, compared to $K=21.8$ at $z>1$. In this work we therefore concentrate on the structural properties of PSBs at $z>1$, and defer an examination of the low-redshift ground-based sample to future work using deeper $K$-band imaging. In the redshift range $1<z<2$, our initial sample consists of 24,880 star-forming galaxies, 2043 passive galaxies and 502 PSBs, of which 9183, 2001 and 385, respectively, have stellar masses M$_{\ast}> 10^{10} ~$M$_{\sun}$.
![ A comparison of Sérsic index measurements for galaxies measured with ground-based $K$-band imaging and HST $H$-band imaging, using the subsample of $\sim$9% of galaxies within the UDS CANDELS mosaic. The upper panel compares individual galaxies, with a characteristic error bar denoting the median uncertainty in $\log_{10}(n)$ for the PSB population. We find that ground-based determinations of Sérsic indices are sufficient to broadly distinguish populations with high average values from those with low average values. The lower (binned) distribution compares the mean Sérsic indices, binned as a function of $n_{\rm
CANDELS}$. Open symbols denote mean values determined from fewer than 5 galaxies. There are deviations from a 1:1 relation, but no evidence for a systematic bias in ground-based Sérsic determinations with spectral type. Galaxies shown have photometric redshifts in the range $1<z<2$ and stellar masses with log (M$_{\ast}$/M$_{\sun}$)$>10$, to match the samples used in this work. []{data-label="fig:sersic-test"}](ntest.pdf "fig:"){width="0.9\linewidth"} ![ A comparison of Sérsic index measurements for galaxies measured with ground-based $K$-band imaging and HST $H$-band imaging, using the subsample of $\sim$9% of galaxies within the UDS CANDELS mosaic. The upper panel compares individual galaxies, with a characteristic error bar denoting the median uncertainty in $\log_{10}(n)$ for the PSB population. We find that ground-based determinations of Sérsic indices are sufficient to broadly distinguish populations with high average values from those with low average values. The lower (binned) distribution compares the mean Sérsic indices, binned as a function of $n_{\rm
CANDELS}$. Open symbols denote mean values determined from fewer than 5 galaxies. There are deviations from a 1:1 relation, but no evidence for a systematic bias in ground-based Sérsic determinations with spectral type. Galaxies shown have photometric redshifts in the range $1<z<2$ and stellar masses with log (M$_{\ast}$/M$_{\sun}$)$>10$, to match the samples used in this work. []{data-label="fig:sersic-test"}](ntest_bin_mean.pdf "fig:"){width="0.9\linewidth"}
CANDELS-UDS
-----------
Throughout this work we compare our ground-based determinations of galaxy size and Sérsic index with measurements from the Hubble Space Telescope (HST) CANDELS survey (Grogin et al. 2011; Koekemoer et al. 2011), using measurements provided in van der Wel et al. (2012). The UDS is one of the three targets for the CANDELS Wide survey, with imaging in the $J$ and $H$ bands taken with the Wide Field Camera 3 (WFC3). The CANDELS imaging covers only $\sim$7% of the UDS field ($\sim$9% of the area used in our analysis), but this is sufficient to provide an independent test and calibration for our ground-based structural parameters. Full details of the <span style="font-variant:small-caps;">GALFIT</span> measurement of structural parameters within CANDELS are given in van der Wel et al. (2012).
{width="0.45\linewidth"} {width="0.45\linewidth"} {width="0.45\linewidth"} {width="0.45\linewidth"}
![We compare the distribution of $\Sigma_{1.5}$ values for the three galaxy populations in the redshift range $1<z<2$, with stellar masses M$_{\ast}> 10^{10.5} ~$M$_{\sun}$. This parameter ($\Sigma_{1.5} \equiv$M$_{\ast}/R_e^{1.5}$), defined by Barro et al. (2013), effectively removes the slope of the galaxy mass/size relation; high values of $\Sigma_{1.5}$ correspond to galaxies that are compact for their stellar mass. A KS test rejects the null hypothesis that passive galaxies and PSBs are drawn from the same underlying distribution in $\Sigma_{1.5}$, with a significance of $>99.99\%$.[]{data-label="fig:sigma15"}](sigma15.pdf){width="0.9\linewidth"}
{width="0.45\linewidth"} {width="0.45\linewidth"}
Ground-based measurements of size and Sérsic index {#sec:sizesersic}
==================================================
We determined structural parameters for the $K$-band galaxy sample using the <span style="font-variant:small-caps;">GALAPAGOS</span> software (Barden et al. 2012). The package allows the automated use of <span style="font-variant:small-caps;">GALFIT</span> (Peng et al. 2002) to fit Sérsic light profiles (Sérsic 1968) to all galaxies in the UDS, parameterized with a Sérsic index, $n$, and effective radius, $R_e$, measured along the semi-major axis. We acknowledge that many high-redshift galaxies are described by more complex morphologies (see Bruce et al. 2014), but single Sérsic fits provide a simple parameterization to allow us to compare the bulk properties of the galaxy populations.
An accurate determination of the point-spread function (PSF) is critical for this process, as galaxies at $z>1$ typically have half-light radii below 0.5 arcsec. Following the work of Lani et al. (2013), we investigated PSF variations across the UDS field and found that most variation occurred between WFCAM detector boundaries within the UDS mosaic (Casali et al. 2007). Testing revealed that we could obtain consistent results by splitting the UDS field into 16 overlapping sub-regions, corresponding to the $4\times 4$ WFCAM tiling pattern. Within each region the light profiles of approximately 100 stars were stacked to provide the local PSF measurement, with variations across the field in the range $0.75-0.81$ arcsec (FWHM). Considerable care was taken to mask sources in the vicinity of the stars used for PSF measurement.
From the resulting measurements of size and Sérsic index we rejected $\sim$9% of galaxies where GALFIT failed to converge on a Sérsic solution. The rejection rate was similar for the star-forming, passive and post-starburst populations. A further $\sim$1% of galaxies were rejected (a-priori) if the fits were formally very poor ($\chi^2_\nu>100$), which typically corresponded to highly blended objects on the $K$-band image. Matching the output from <span style="font-variant:small-caps;">GALAPAGOS</span> with our Supercolour catalogue, we obtain a final sample of 8098 star-forming galaxies, 1829 passive galaxies, and 348 PSBs in the redshift range $1<z<2$ with M$_{\ast}> 10^{10}
~$M$_{\sun}$.
In Figures \[fig:size-test\] and \[fig:sersic-test\] we display the resulting size and Sérsic measurements for the subset of UDS galaxies within the HST CANDELS survey. Ground-based measurements are compared with those obtained using H-band CANDELS measurements, as published in van der Wel et al. (2012).
For the size measurements, we find a tight relationship between the ground-based $K$-band and CANDELS $H$-band sizes, as previously found by Lani et al. (2013). On average, the sizes obtained from CANDELS are systematically $\sim$13% larger, which is consistent with previous comparisons of size measurements as a function of waveband (e.g. Kelvin et al. 2012). The systematic offset is consistent among the three galaxy populations studied here so we apply no corrections for this effect. The characteristic scatter in $\delta
R/R$, given by the normalized median absolute deviation ($\sigma_{{\rm
NMAD}}$) is 17%, 16% and 24% for the star-forming, passive and PSB populations respectively.
Ground-based measurements of Sérsic index are far more uncertain for a given galaxy (Figure \[fig:sersic-test\]), and we find a significant degree of scatter when comparing ground-based and CANDELS measurements. Formally, the characteristic scatter in $\delta n/n$, given by the normalized median absolute deviation ($\sigma_{{\rm
NMAD}}$) is 45%, 39%, and 35% for the star-forming, passive and PSB populations respectively. Nevertheless, there is a clear correlation, and we find that ground-based determinations are sufficient to distinguish populations with “high” Sérsic indices (e.g. $n>2$) from those with “low” Sérsic indices ($n<2$). Comparing the three primary galaxy types, we find consistent results; the passive and post-starburst populations show consistently high Sérsic indices (from ground or spaced-based measurements), while star-forming galaxies are concentrated at lower values. The binned distribution demonstrates that the correlation between ground and space-based Sérsic indices is not perfectly 1:1, but we see no systematic differences in this relation between the three populations. We conclude that ground-based measurements can be used to broadly compare the Sérsic indices for our galaxy populations. In addition, we will use the subset of galaxies with HST measurements ($\sim$9%) to verify any conclusions drawn from the larger ground-based sample. The Sérsic distributions will be compared further (and as a function of stellar mass) in Section \[sec:sersic\].
The sizes of post-starburst galaxies
====================================
The size–mass relation {#sec:sizemass}
----------------------
In Figure \[fig:size-mass\] we compare the size versus stellar-mass relation for galaxies in the redshift range $1<z<2$. Individual galaxies are shown, along with mean values as a function of stellar mass (in bins of 0.25 dex). In the upper panels we show the results from the ground-based $K$-band imaging, while the lower panels are based on independent sizes from CANDELS $H$-band imaging (covering $\sim$9% of the sample). The 95% mass completeness limits are shown (see Section \[sec:masses\]), determined at $z=2$ to provide conservative limits. In determining mean values for the ground-based sample we applied a 5$\sigma$ clip (with one iteration), to remove extreme outliers, but removing this constraint has no significant influence.
Representative error bounds for individual galaxies are shown on the left panels, based on the median errors on the PSB sample. For the CANDELS sizes, data are taken from van der Wel et al. (2012), which include estimates for random and systematic errors from GALFIT. For the ground-based individual errors, we add in quadrature the scatter in $\delta R/R$ determined by the comparison with CANDELS (Section \[sec:sizesersic\]). The representative uncertainty on stellar masses is based on our Bayesian mass-fitting analysis described in Wild et al. (2016), as briefly outlined in Section \[sec:masses\].
The size–mass relations show the expected trends for star-forming and passive galaxies, consistent with previous studies (e.g. Daddi et al. 2005; Trujillo et al. 2006; van Dokkum et al. 2008; van der Wel et al. 2014; McLure et al. 2013). On average, passive galaxies appear significantly more compact than star-forming galaxies of equivalent stellar mass, but show a steeper size–mass relation, leading to convergence at the highest masses (M$_{\ast} \sim 10^{11.5}
~$M$_{\sun}$). Intriguingly, we find that post-starburst galaxies at this epoch are also extremely compact; they are comparable in size to the established passive galaxies, with evidence that they are smaller on average at high mass (M$_{\ast}> 10^{10.5} ~$M$_{\sun}$). These trends are apparent with the large ground-based sample and with the smaller CANDELS sample.
We performed a bootstrap analysis as a simple test of significance, randomly sampling the (ground-based) populations within each mass bin, with replacement. For the 4 bins above $10^{10.5} ~$M$_{\sun}$, the fraction of the resampled populations in which the passive galaxies show mean sizes equal to (or smaller than) the mean of the PSB population are $0.07$, $5\times10^{-5}$, $1.2\times 10^{-4}$, $4\times10^{-5}$ (low to high mass, respectively). We note that the final bin contains 86 passive galaxies, but only 3 PSBs, so the bootstrap comparison for this bin may be unreliable. Overall, assuming no systematic errors, we find evidence that massive post-starburst galaxies at $z>1$ are significantly more compact, on average, than passive galaxies of comparable mass. Repeating the analysis using the median sizes produced very similar trends. The median analysis and further tests of robustness are presented in Appendix A. Our results are consistent with the findings of Yano et al. (2016), who also found evidence that high-redshift post-starburst galaxies are very compact.
As an additional comparison, in Figure \[fig:sigma15\] we display the distribution of $\Sigma_{1.5}\equiv M_{\ast}/R_e^{1.5}$ for the three galaxy populations, measured for the redshift range $1<z<2$ and stellar masses M$_{\ast}> 10^{10.5} ~$M$_{\sun}$. Following Barro et al. (2013), we use this parameter to effectively remove the slope in the galaxy size/mass relation. Fitting a function of the form M$_{\ast} = \Sigma ~ R_e^{\alpha}$ to the passive population, we find a best fit with $\alpha=1.55$, in very good agreement with the value $\alpha=1.5$ assumed by Barro et al. (2013). A simple Kolmogorov-Smirnov (KS) test rejects the null hypothesis that passive galaxies and PSBs are drawn from the same underlying distribution in $\Sigma_{1.5}$ with a significance $>99.99$%, with the same significance obtained with either value of $\alpha$.
Given the strong evolution of the PSB mass function (Wild et al. 2016), a concern is that massive PSBs are more common at higher redshifts, which may bias the size–mass comparison when measured over a wide redshift range. In Figure \[fig:size-mass-z12\] we therefore display the size–mass relation in two narrower redshift bins, $1.0<z<1.5$ and $1.5<z<2.0$. With two independent samples, the results confirm that post-starburst galaxies, on average, show smaller half-light radii than the passive population at high mass (M$_{\ast}> 10^{10.5} ~$M$_{\sun}$). An additional test was performed using a weighted mean, with a redshift-dependent weight determined for each PSB using the ratio of the $n(z)$ distributions for passive galaxies and PSBs. The resulting size–mass relations were barely changed, with only a slight reduction in the significance of the differences reported above.
A natural interpretation of our findings is that quiescent galaxies are most compact when they are newly-quenched, but then grow with cosmic time. Given the short-lived nature of the PSB phase, we expect the majority of passive galaxies to have gone through a similar stage in their past (Wild et al. 2016). Our results therefore provide evidence for the genuine growth of individual galaxies, suggesting that the growth of the population as a whole is not purely caused by a progenitor bias. We discuss the implications further in Section \[sec:discussion\].
A simple calculation allows us to compare the size differences with the observed cosmological growth. Based on population synthesis models, we estimate that the established passive population quenched approximately 0.5–1 Gyr before the PSB population at these redshifts (Wild et al. 2016), and the characteristic difference in size (e.g. based on the shift in $\Sigma_{1.5}$) is approximately 25% on average at M$_{\ast}> 10^{10.5} ~$M$_{\sun}$. The implied growth rate is similar to the observed [*cosmological*]{} growth, parameterized by van Dokkum et al. (2010) in the form $r_e
\propto (1+z)^{-1.3}$ (i.e. $\sim$25% per Gyr at $z=1.5-2$). An improvement on these tentative conclusions will require more accurate age-dating of the stellar populations, which will soon be possible with growing spectroscopic samples.
{width="0.45\linewidth"} {width="0.45\linewidth"}
Stellar mass uncertainties {#sec:uncertainties}
--------------------------
In this section we explore the potential impact of both random and systematic errors on our stellar masses.
The typical uncertainty from our Bayesian stellar mass fitting is $\sigma \simeq$ 0.1 dex for all galaxy types, allowing for the degeneracy between fitted parameters and the uncertainties on photometric redshifts. To investigate the impact on our conclusions, we performed Monte Carlo realisations, allowing the stellar masses to shift randomly within a Gaussian probability distribution in log M$_{\ast}$. We found no impact on any of the results presented in Section \[sec:sizemass\]. Comparing the resulting distributions in $\Sigma_{1.5}$, the significance of the difference between the PSB and passive galaxy populations was unchanged, suggesting that the sample overall is large enough to minimise the influence of random errors.
To investigate the influence of fitting methods, we re-evaluated the size–mass relations using two independent sets of stellar masses derived by Simpson et al. (2013) and Hartley et al. (2013), the latter also using an independent set of photometric redshifts. No significant differences were found.
As a note of caution, however, we acknowledge that the observed differences between PSBs and passive galaxies could arise if our stellar masses are [*systematically*]{} overestimated for younger stellar populations. Our stellar masses are based on population synthesis models from BC03, which may underestimate the influence of thermally-pulsing asymptotic giant branch (TP-AGB) stars (Maraston et al. 2006). Such stars may have a major contribution to the rest-frame near-infrared light for stellar populations in the age range 0.2 to 2 Gyr, potentially leading to overestimated stellar masses for passive galaxy populations. The influence of TP-AGB stars is discussed further in Wild et al. (2016), where it was found that the influence of TP-AGB stars is strongest for galaxies with BC03-determined ages $>1$ Gyr, for which the ages and mass-to-light ratios are reduced using the models of Maraston et al. (2006). The net effect would be to move our passive galaxies to lower masses relative to the younger PSB population, which would [*enhance*]{} the differences in the size–mass relation.
In summary, our findings appear robust to known sources of random and systematic error, but we acknowledge the possibility that unknown systematic uncertainties in the stellar mass determination may contribute to the difference in size–mass relations presented in our work. Future deep infrared spectroscopy may allow a more detailed investigation of the inherent uncertainties in determining stellar masses from photometric data.
A comparison with UVJ selection {#sec:uvj}
-------------------------------
To allow a comparison of our supercolour technique with previous work, in Figure \[fig:uvjfig\] (left) we present a rest-frame UVJ colour-colour diagram for UDS galaxies in the redshift range $1<z<2$. As previously shown in W14, the classification of galaxies using supercolours agrees very well with the more traditional UVJ colour selection (Labbé et al. 2005; Wuyts et al. 2007). Post-starburst galaxies are generally found at the blue end of the passive UVJ region, in good agreement with the findings of Whitaker et al. (2012). The distribution in $U-V$ alone confirms that PSBs predominantly lie in the classic “green valley”, intermediate between passive and star-forming galaxies, but the addition of the $V-J$ colour isolates this population in the distinct region corresponding to the youngest red-sequence galaxies.
In the right-hand panel of Figure \[fig:uvjfig\] we illustrate the average effective radii in colour-colour bins. We select the subset of galaxies over the mass range $10.5< \log
~($M$_{*}/$M$_{\sun})<11.5$ to minimise the effects of the size–mass relation. The trends confirm that “young quiescent” galaxies selected by the UVJ technique show smaller average sizes, in good agreement with our analysis based on supercolours. Our results are consistent with a similar recent analysis by Yano et al. (2016).
The Sérsic indices of post-starburst galaxies {#sec:sersic}
=============================================
{width="0.32\linewidth"} {width="0.32\linewidth"} {width="0.32\linewidth"}
In Figure \[fig:sersicfig\] we compare the Sérsic indices for star-forming, passive and post-starburst galaxies at $z>1$. The distributions obtained from the ground-based $K$-band data are consistent with those obtained for the smaller CANDELS sample. In both cases, we find that star-forming galaxies show a distribution peaking sharply at $n\simeq 1$, while passive and post-starburst galaxies show very different distributions, peaking at significantly higher values. A Kolmogorov-Smirnov test confirms these findings, rejecting the null hypothesis that either passive or post-starburst galaxies are drawn from the same distribution as star-forming galaxies to a high level of significance ($>>99.99$%, using the ground-based sample). In contrast, the Sérsic distributions for passive and post-starburst galaxies do not appear significantly different.
The right-hand panel in Figure \[fig:sersicfig\] presents the median Sérsic indices as a function of stellar mass. We find evidence for a slight increase in the median Sérsic index with stellar mass for all populations, but in all mass bins the post-starburst galaxies show significantly higher Sérsic indices than star-forming galaxies, and values consistent with the passive population. Using the mean produced very similar trends, but Sérsic indices were slightly higher in all cases (by $\delta n\simeq 0.5$).
We caution that single Sérsic fits provide only a crude parameterisation of the data, as it is now well-established that most galaxies at this epoch have more complex morphologies (e.g. Bruce et al. 2014). A high Sérsic index does not necessarily imply a purely spheroidal system, and low Sérsic indices do not necessarily imply the presence of an established disc (Mortlock et al. 2013). Nevertheless, it is clear from our data that the post-starburst galaxies are structurally very different to the actively star-forming population, and more comparable to ultra-compact equivalents of the passive population. We explore multiple-component fitting in future work (Maltby et al., in preparation).
Discussion {#sec:discussion}
==========
We have presented evidence that massive (M$_{\ast}> 10^{10.5}
~$M$_{\sun}$) recently-quenched (post-starburst) galaxies at high redshift ($z>1$) are exceptionally compact. Furthermore, they show high Sérsic indices, indistinguishable from the established passive population at the same epoch. We conclude that the structural transformation must have occurred before (or during) the event that quenched their star formation. Given that the majority of massive passive galaxies at $z>1$ are thought to have passed through a post-starburst phase (Wild et al. 2016), our findings suggest a strong link between quenching and the formation of a compact spheroid.
Our results confirm the findings of Whitaker et al. (2012), who found evidence that younger passive galaxies are more compact at $z>1$, and the more recent HST CANDELS study by Yano et al. (2016). At intermediate redshifts ($z\sim 1$) previous studies have found conflicting results on the relationship between stellar age and the compactness of passive galaxies, with indications that progenitor bias may be playing a role at this epoch (Keating et al. 2015; Williams et al. 2017).
Our findings may be explained if high-redshift post-starburst galaxies are formed from the dramatic collapse of gas at high redshift, formed from either a gas-rich merger (e.g. Hopkins et al. 2009; Wellons et al. 2015), or from gas inflow feeding a massive disc, which becomes unstable and collapses by “compaction” (e.g. Dekel et al. 2009; Zolotov et al. 2015; Tacchella et al. 2016). Star formation must then be rapidly quenched, either by a central AGN or feedback from highly nucleated star formation (e.g. Hopkins et al. 2005; Diamond-Stanic et al. 2012), leaving an ultra-compact post-starburst remnant. It may be possible to test these evolutionary scenarios by comparing the properties of post-starburst galaxies with their likely active progenitors, i.e. star-forming galaxies caught during the merging or “compaction” phase. Current candidates include submillimetre galaxies, many of which appear to be highly compact at $\sim$250 $\mu$m in the rest-frame (e.g. Simpson et al. 2015), and the high-redshift “blue nuggets” (e.g. Barro et al. 2013; Mei et al. 2015; Barro et al. 2017). Whatever the true progenitors for our PSBs, the most likely explanation is that structural transformation occurred immediately prior to quenching. The fact that the PSBs and passive galaxies have indistinguishable Sérsic indices (Figure \[fig:sersicfig\]) would suggest that most of the structural change is already established when the star formation is quenched, unless the structural transformation occurs on a much shorter timescale than the $\sim$500 Myr lifetime for the PSB phase.
Following the formation of the proto-spheroid, there are currently two leading explanations for the observed growth in passive galaxies with cosmic time. Minor gas-free mergers provide a plausible physical mechanism (e.g. Bezanson et al. 2009; Naab, Johansson & Ostriker 2009), and indeed there is evidence that high-redshift passive galaxies are larger in dense environments, where such interactions are more likely (e.g. Lani et al. 2013). Alternatively, progenitor bias may mimic the observed growth, since passive galaxies formed at later times are typically larger (e.g. Poggianti et al. 2013; Carollo et al. 2013).
Our finding that post-starburst galaxies are more compact than older passive galaxies, on average, would suggest that we are observing an earlier phase in the lifetime of steadily-growing spheroids. Thus progenitor bias is unlikely to be the primary cause for the observed growth at early times; our observed trends show precisely the opposite behaviour (with younger galaxies being more compact). On the other hand, if massive post-starburst galaxies represent newly-formed “red nuggets”, it is notable that their abundance is a strong function of redshift; massive (M$_{\ast}>
10^{10.5} ~$M$_{\sun}$) post-starburst galaxies are several times more abundant at $z\sim 2$ compared to $z\sim 0.5$ (Wild et al. 2016). Thus, while the majority of high-mass passive galaxies at $z\sim 2$ are likely to have been through the ultra-compact post-starburst phase (Wild et al. 2016), this may become an increasingly less dominant channel towards lower redshifts, when progenitor bias may play a more significant role in explaining size evolution (e.g. see Fagioli et al. 2016, Williams et al. 2017). Even at low redshift, however, there is evidence that the most compact quiescent galaxies have evolved from post-starburst progenitors (Zahid et al. 2016).
Finally, we note that the mass function for PSBs shows a very distinctive evolution in shape (Wild et al. 2016). At high redshift ($z\sim
2$) the mass function resembles that of quiescent galaxies, dominated by high-mass systems with a sharp decline in space density above a mass of M$_{\ast}\sim 10^{10.5} ~$M$_{\sun}$. At low redshift ($z<1$) the population is dominated by lower-mass systems, with a shape resembling the mass function for star-forming galaxies. These features are interpreted as evidence for two distinct formation channels for post-starburst galaxies; high-mass systems formed from gas-rich dissipative collapse, and low-mass systems formed from environmental quenching or the merging of normal disc galaxies (Wild et al. 2016). Our structural findings are in good agreement with this scenario, with the PSBs above the same characteristic mass displaying distinctive, ultra-compact morphologies, consistent with a highly-dissipative, gas-rich origin. We will present a detailed study of the structural parameters for low-mass PSBs in future work (Maltby et al., in preparation).
Conclusions
===========
We present a study of the structural parameters for a large sample of photometrically-selected post-starburst galaxies in the redshift range $1<z<2$, recently identified in the UKIDSS UDS field. These rare transition objects provide the ideal sample for understanding the links between the quenching of star formation and the structural transformation of massive galaxies.
We demonstrate that deep ground-based near-infrared imaging can be used to obtain robust sizes and Sérsic indices for large samples of high-redshift galaxies. From the resulting size–mass relation, we find that massive (M$_{\ast}> 10^{10.5} ~$M$_{\sun}$) post-starburst galaxies are exceptionally compact at $z>1$, with evidence that they are more compact on average than established passive galaxies at the same epoch. Since most high-mass passive galaxies at $z>1$ are likely to have been through a post-starburst phase (Wild et al. 2016), the implication is that quiescent galaxies are most compact when they are newly quenched, thereafter growing with cosmic time. An important caveat, however, is to acknowledge the considerable uncertainty in stellar mass estimation, as discussed in Section \[sec:uncertainties\]. As an avenue for future research, it will be important to determine whether stellar masses are systematically overestimated for recently quenched stellar populations.
We also find that post-starburst galaxies show high Sérsic indices, significantly higher than star-forming galaxies on average, but statistically indistinguishable from the Sérsic indices of established passive galaxies at the same epoch. We conclude that massive post-starburst galaxies represent newly-formed compact proto-spheroids. Furthermore, the structural transformation of these galaxies must have occurred before (or during) the event that quenched their star formation.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Louis Abramson, Steven Bamford, Dale Kocevski, Mike Merrifield, and Ian Smail for useful discussions. We extend our gratitude to the staff at UKIRT for their tireless efforts in ensuring the success of the UDS project. We also wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has within the indigenous Hawaiian community. We were most fortunate to have the opportunity to conduct observations from this mountain. V.W. and K.R. and acknowledge support from the European Research Council Starting Grant (PI Wild). RJM acknowledges the support of the European Research Council via the award of a consolidator grant (PI McLure).
Barden M., H[ä]{}u[ß]{}ler B., Peng C. Y., McIntosh D. H., Guo Y., 2012, MNRAS, 422, 449
Barro G., et al., 2013, ApJ, 765, 104
Barro G., et al., 2017, ApJ, 840, 47
Belli S., Newman A. B., Ellis R. S., Konidaris N. P., 2014, ApJ, 788, L29
Best P. N., Kaiser C. R., Heckman T. M., Kauffmann G., 2006, MNRAS, 368, L67
Bezanson R., van Dokkum P. G., Tal T., Marchesini D., Kriek M., Franx M., Coppi P., 2009, ApJ, 697, 1290
Bruce V. A., et al., 2014, MNRAS, 444, 1001
Bruzual G., Charlot S., 2003, MNRAS, 344, 1000
Carollo C. M., et al., 2013, ApJ, 773, 112
Casali M., et al., 2007, A&A, 467, 777
Cirasuolo M., McLure R. J., Dunlop J. S., Almaini O., Foucaud S., Simpson C., 2010, MNRAS, 401, 1166
Daddi E., et al., 2005, ApJ, 626, 680
Dekel A., et al., 2009, Nature, 457, 451
Diamond-Stanic A. M., Moustakas J., Tremonti C. A., Coil A. L., Hickox R. C., Robaina A. R., Rudnick G. H., Sell P. H., 2012, ApJ, 755, L26
Dressler A., Gunn J. E., 1983, ApJ, 270, 7
Fagioli M., Carollo C. M., Renzini A., Lilly S. J., Onodera M., Tacchella S., 2016, ApJ, 831, 173
Fan L., Lapi A., De Zotti G., Danese L., 2008, ApJ, 689, L101
Fontana A., et al., 2004, A&A, 424, 23
Furusawa H., et al., 2008, ApJS, 176, 1
Grogin N. A., et al., 2011, ApJS, 197, 35
Hartley W. G., et al., 2013, MNRAS, 431, 3045
Hogg D. W., et al., 2002, AJ, 124, 646
Hopkins P. F., Hernquist L., Cox T. J., Di Matteo T., Martini P., Robertson B., Springel V., 2005, ApJ, 630, 705
Hopkins P. F., Hernquist L., Cox T. J., Keres D., Wuyts S., 2009, ApJ, 691, 1424
Keating S. K., Abraham R. G., Schiavon R., Graves G., Damjanov I., Yan R., Newman J., Simard L., 2015, ApJ, 798, 26
Kelvin L. S., et al., 2012, MNRAS, 421, 1007
Kodama T., et al., 2004, MNRAS, 350, 1005
Koekemoer A. M., et al., 2011, ApJS, 197, 36
Labb[é]{} I., et al., 2005, ApJ, 624, L81
Lani C., et al., 2013, MNRAS, 435, 207
Lawrence A., et al., 2007, MNRAS, 379, 1599
McLure R. J., et al., 2013, MNRAS, 428, 1088
Maltby D. T., et al., 2016, MNRAS, 459, L114
Maraston C., Daddi E., Renzini A., Cimatti A., Dickinson M., Papovich C., Pasquali A., Pirzkal N., 2006, ApJ, 652, 85
Mei S., et al., 2015, ApJ, 804, 117
Mortlock A., et al., 2013, MNRAS, 433, 1185
Naab T., Johansson P. H., Ostriker J. P., 2009, ApJ, 699, L178
Peng C. Y., Ho L. C., Impey C. D., Rix H.-W., 2002, AJ, 124, 266
Poggianti B. M., et al., 2013, ApJ, 762, 77
Pozzetti L., et al., 2010, A&A, 523, A13
Sersic J. L., 1968, Atlas de Galaxias Australes. Observatorio Astronomico, Cordoba, Argentina
Silk J., Rees M. J., 1998, A&A, 331, L1
Simpson C., Westoby P., Arumugam V., Ivison R., Hartley W., Almaini O., 2013, MNRAS, 433, 2647
Simpson J. M., et al., 2015, ApJ, 799, 81
Strateva I., et al., 2001, AJ, 122, 1861
Tacchella S., Dekel A., Carollo C. M., Ceverino D., DeGraf C., Lapiner S., Mandelker N., Primack J. R., 2016, MNRAS, 458, 242
Trujillo I., et al., 2006, MNRAS, 373, L36
van der Wel A., et al., 2012, ApJS, 203, 24
van der Wel A., et al., 2014, ApJ, 788, 28
van Dokkum P. G., et al., 2008, ApJ, 677, L5
van Dokkum P. G., et al., 2010, ApJ, 709, 1018
Vergani D., et al., 2010, A&A, 509, A42
Wellons S., et al., 2015, MNRAS, 449, 361
Whitaker K. E., Kriek M., van Dokkum P. G., Bezanson R., Brammer G., Franx M., Labb[é]{} I., 2012, ApJ, 745, 179
Wild V., Walcher C. J., Johansson P. H., Tresse L., Charlot S., Pollo A., Le F[è]{}vre O., de Ravel L., 2009, MNRAS, 395, 144
Wild V., et al., 2014, MNRAS, 440, 1880
Wild V., Almaini O., Dunlop J., Simpson C., Rowlands K., Bowler R., Maltby D., McLure R., 2016, MNRAS, 463, 832
Williams C. C., et al., 2017, ApJ, 838, 94
Wuyts S., et al., 2007, ApJ, 655, 51
Yano M., Kriek M., van der Wel A., Whitaker K. E., 2016, ApJ, 817, L21
Zahid H. J., Baeza Hochmuth N., Geller M. J., Damjanov I., Chilingarian I. V., Sohn J., Salmi F., Hwang H. S., 2016, ApJ, 831, 146
Zolotov A., et al., 2015, MNRAS, 450, 2327
Robustness tests {#sec:appendix}
================
{width="0.45\linewidth"} {width="0.45\linewidth"}
In this section we discuss a number of additional tests that were performed to investigate the robustness of the size–mass relations presented in this paper.
Throughout this work we have measured average sizes when comparing galaxy populations, yielding evidence that post-starburst galaxies are typically smaller than passive galaxies of comparable stellar mass. The use of a mean may produce misleading results, however, if either population is skewed by a significant number of outliers (e.g. misclassified interlopers from other galaxy categories). In Figure \[fig:appenfig1\] (left panel) we therefore reproduce the size–mass relation from Figure \[fig:size-mass\], but this time using median size values. In most bins the median sizes are slightly smaller than the mean, as the size distributions show a slight tail to high values. The significant differences between the passive and post-starburst galaxies remain, however. At high-mass (M$_{\ast}>
10^{10.5} ~$M$_{\sun}$) the overall significance of the trends is essentially unchanged compared to the results outlined in Section \[sec:sizemass\]. We conclude that the use of a mean has not biased our primary conclusions.
As a further source of error, we considered the possibility that the passive and post-starburst samples are contaminated by star-forming galaxies, e.g. due to uncertainties in classification. If the contaminating fraction is higher for the passive population (for reasons unknown), this could skew the size measurements upwards. We tested for this effect by applying a cut in Sérsic index. Noting the very different distributions in Sérsic index between star-forming and passive samples (see Figure \[fig:sersicfig\]), we re-evaluated the size–mass relations using only passive and post-starburst galaxies with $n>2$ (see Figure \[fig:appenfig1\], right panel), again using the median size to further reduce the impact of interlopers. The striking difference in sizes remains, with a negligible reduction in significance. Using a mean estimator yields the same result, with a size–mass relation that is almost identical to the upper-right panel of Figure \[fig:size-mass\]. We conclude that “contamination” from galaxies with low Sérsic indices (whether passive or star-forming) is not affecting our conclusions.
As an additional test, we investigated the influence of using a cleaner (though less complete) sample of post-starburst galaxies. As outlined in Section \[sec:intro\] and Maltby et al. (2016), the primary source of contamination is between post-starburst galaxies and “normal” passive galaxies. Depending on the precise selection criteria, between 20–40% galaxies in the PSB category would be classified as passive (rather than PSB) using spectroscopy, while 6–10% of the passive category would be spectroscopically classified as PSBs (Maltby; private communication). Based on Figure 3 in Maltby et al. (2016), we therefore identify a “cleaner” PSB sample by selecting galaxies further from the passive/PSB boundary, with supercolours $SC2$>6. In this regime, formally 100% of PSB candidates (15/15) were confirmed with $W_{\rm H\delta}>5$Å. Using this new sub-sample, the resulting size–mass relation (evaluated using the standard mean estimator) is shown in Figure \[fig:appenfig2\] (left panel). We find that the difference in size compared to passive galaxies remains, and in fact is slightly enhanced; the mean PSB size is formally smaller in five out of six bins. We conclude that contamination of the PSB category by older passive galaxies can only act to dilute the differences we observe.
As another test for contamination, we combined the PCA classification with the classic UVJ criteria (e.g. see Figure \[fig:uvjfig\]), to exclude PSBs and passive galaxies that are classified as “star-forming” using rest-frame UVJ colours. The aim is to remove red star-forming galaxies that may have been misclassified by the PCA technique. Using these joint criteria removes 26% of passive galaxies and 34% of PSBs from our primary sample ($z>1$, M$_{\ast}>10^{10} ~$M$_{\sun}$). The resulting size–mass relations are shown in Figure \[fig:appenfig2\] (right panel). We find that the difference between the PSBs and passive galaxies at high mass remains, and in fact is slightly enhanced.
Finally, we performed a variety of tests using more stringent cuts on the structural parameters derived for our ground-based galaxy sample, using the CANDELS dataset as a calibrating sample. No significant differences were found. We noted, however, that ground-based size measurements become increasingly unreliable when GALFIT assigns a very low axis ratio, $q<0.1$. Approximately 6% of the galaxies in our sample are affected, mostly among the star-forming galaxies, but also affecting 3–4% of the passive and post-starburst sample (mostly at low mass; M$_{\ast}< 10^{10.5} ~$M$_{\sun}$). Removing these galaxies had no major influence on the size-mass relations; in fact, the difference in size between passive and post-starburst galaxies became marginally more significant.
{width="0.45\linewidth"} {width="0.45\linewidth"}
\[lastpage\]
[^1]: E-mail: [email protected]
[^2]: From a sample of 24 PSB candidates, 19 galaxies ($\sim$80%) showed strong Balmer absorption lines ($W_{\rm H\delta}>5$Å), dropping to 14 confirmations ($\sim$60%) if stricter criteria are used to exclude galaxies with significant \[O<span style="font-variant:small-caps;">ii</span>\] emission. The fraction of spectroscopic PSBs among the passive and star-forming PCA classes was estimated to be $<$10% and $<$1% respectively (Maltby; private communication).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The conflict between canonical commutation relation and gauge invariance, which both the momentum and angular momentum of quark and gluon should satisfy, is clarified. The quantum version of gauge invariance is studied. The gauge independence of the matrix elements of quark momentum and angular momentum operators between physical states are proved. We suggest to use the canonical quark momentum and angular momentum distributions to describe the nucleon internal structure in order to establish an internal consistent description of hadron spectroscopy and hadron structure. The same problem for the atomic spectroscopy and structure is discussed.'
address:
- 'Department of Physics, Nanjing University, Nanjing, 210093, China'
- 'Department of Physics, Nanjing University, Nanjing, 210093, China'
- 'Department of Physics, Sichuan University, Chengdu, 610064, China'
author:
- Fan Wang
- 'Wei-Min Sun'
- 'Xiao-Fu Lü'
title: QUANTUM VERSION OF GAUGE INVARIANCE AND NUCLEON INTERNAL STRUCTURE
---
Conflict between Canonical Commutation Relation and Gauge Invariance
====================================================================
The nucleon (atom) is a QCD (QED) gauge field system. The momentum and angular momentum of the nucleon (atom) is the sum of contributions from quark (electron) and gluon (photon) respectively: $$\begin{aligned}
\vec{P}&=&{\int}d^3x{\psi}^{\dag}\frac{\vec{\nabla}}{i}{\psi}+{\int}d^3xE_i{\vec{\nabla}}A_i.
\label{pequ}\end{aligned}$$ $$\begin{aligned}
\vec{J}&=&\frac{1}{2}{\int}d^3x{\psi}^{\dag}\vec{\Sigma}\psi+
{\int}d^3x{\psi}^{\dag}{\vec{r}{\times}\frac{\vec{\nabla}}{i}{\psi}}+
{\int}d^3x{\vec{E}{\times}\vec{A}}+
{\int}d^3x{E_i\vec{r}{\times}{\vec{\nabla}}A_i} \label{jequ}\end{aligned}$$ In the above equations, $\psi$ is the quark (electron) field, $\vec{E}$ is the color electric (ordinary electric) fields, $\vec{A}$ is the vector potential. In QCD case a summation over color indices is understood. The good side of the above decomposition is that each term in Eq.(1,2) satisfies the canonical commutation relation of the momentum and angular momentum operator, so they are quark and gluon momentum, quark spin and quark orbital angular momentum, gluon spin and gluon orbital angular momentum, respectively. However they are not gauge invariant individually except the quark (electron) spin term.
Alternatively one can derive a gauge invariant decomposition, $$\begin{aligned}
\vec{P}={\int}d^3x(\psi^{\dag}\frac{\vec{D}}{i}\psi+\vec{E}\times\vec{B}),\end{aligned}$$ $$\begin{aligned}
\vec{J}={\int}d^3x(\frac{1}{2}\psi^{\dag}\vec{\Sigma}\psi+
\psi^{\dag}\vec{r}\times\frac{\vec{D}}{i}\psi+\vec{r}\times(\vec{E}\times\vec{B})).\end{aligned}$$ The good side of this decomposition is that each term is gauge invariant. However they do not satisfy the canonical commutation relation individually except the quark (electron) spin term[@cw].
In classical gauge field theory, only gauge invariant quantities are physically meaningful. In the study of nucleon internal parton momentum and angular momentum structure, also only the gauge invariant operators related to quark and gluon momenta, spin and orbital angular momenta are appreciated[@ji]. In hadron spectroscopy, partial wave analysis and multi-pole radiation are widely used where the gauge non-invariant canonical momentum and orbital angular momentum must be used accordingly.
Canonical momentum and orbital angular momentum have been used in describing atomic structure for almost a century already. Are the atomic electron momentum and orbital angular momentum not measurable ones? Can these operators be used to describe the nucleon internal structure? In this report we show that the canonical quark (electron) momentum and orbital angular momentum have gauge independent matrix elements between physical states and so is observable, which should be used to establish an internal consistent description of hadron spectroscopy and hadron internal structure.
Quantum Version of Gauge Invariance
===================================
F.Strocchi and A.S.Wightman studied the quantum version of gauge invariance[@sw].
A gauge (or a quantization scheme) in a quantum gauge field theory is specified by
\(a) field operators: *A$_{\mu}$*,the gauge potential; *j$_{\mu}$*, the gauge interaction current; $\psi$, the fermion field and other fields of the gauge in a Hilbert space *H*;
\(b) a representation *U* of the Poincar$\acute{e}$ group in *H*;
\(c) a sesquilinear form (Gupta scalar product) $\langle\Phi,\Psi\rangle$ on *H* with respect to which *U* is unitary, $\Phi$ and $\Psi$ are vectors in *H*;
\(d) a distinguished subspace $\emph{H}' \subset \emph{H}$ such that
\(i) The restriction of the sesquilinear form to $\emph{H}'$ is bounded and nonnegative
$\langle{\Psi} ,{\Psi}\rangle\geq{0}$ for${\Psi}\in\emph{H}'$.
\(ii) The analogue of the Maxwell equation holds in the sense that $$\langle \Phi,(\partial_{\mu}F^{\mu\nu}-j^{\nu})\Psi \rangle=0$$ for all $\Phi,\Psi \in \emph{H}'$.
\(iii) $\emph{H}'$ has a subspace $\emph{H}''$ consisting of vectors $\Phi$ in $\emph{H}'$ of zero length $\langle \Phi,\Phi \rangle=0$. The physical Hilbert space is $\emph{H}_{phys}=\emph{H}'/\emph{H}''$.
There exists a unique vector $\Psi_0$, called the vacuum, which is invariant under the translation subgroup of the Poincar$\acute{e}$ group. The vector $\Psi_0$ lies in $\emph{H}'$.
This is a generalization of the Gupta-Bleuler quantization scheme of QED.
A generalized gauge transformation is an ordered pair consisting of two gauges
$<{\emph{A}_{1\mu},\emph{H}_1,<\cdot,\cdot>_1,\emph{H}'_1}>$ and $<{\emph{A}_{2\mu},\emph{H}_2,<\cdot,\cdot>_2,\emph{H}'_2}>$
together with a bijection *g* of $\emph(H)_{1phys}$ onto $\emph(H)_{2phys}$
$[\Psi_2]=g[\Psi_1],[\Phi_2]=g[\Phi_1]$
$[\Psi_{20}]=g[\Psi_{10}]$
Note that in the quantum version there is no need of the form of classical gauge invariance, such as $F_{1\mu\nu}=F_{2\mu\nu}$. Only under some special gauge transformation, one has such a gauge invariant form. Instead the gauge invariance of an operator is classified into four categories, i.e., gauge independence, weak gauge invariance, gauge invariance, and strict gauge invariance.
An operator *O*, mapping *H* into *H*, is called gauge independent if $$\langle\Phi,\emph{O}\Psi\rangle=\langle\Phi+\chi_1,\emph{O}(\Psi+\chi_2)\rangle$$ for all $\Phi, \Psi \in \emph{H}'$ and all $\chi_1, \chi_2 \in
\emph{H}''$. In other words, the matrix elements $\langle\Phi,
\emph{O}\Psi\rangle$ for $\Phi, \Psi \in\emph{H}'$ depend only on the equivalence classes $[\Phi], [\Psi] \in \emph{H$_{phys}$}$. Such a gauge independent operator is an observable. The stronger restricted operators, weakly gauge invariant, gauge invariant and strictly gauge invariant ones are all gauge independent ones, obviously they are observable. The classically gauge invariant ones belong to the strictly gauge invariant category. Quantum gauge field theory includes more observable operators than the classical one.
Quark (Electron) Momentum and Orbital Angular Momentum Are Observable
======================================================================
To prove the quark (electron) momentum and orbital angular momentum are observable, one has to prove they are gauge independent, i.e., one has to prove $$\begin{aligned}
\langle(\Phi+\chi_1),\vec{P}_q(\Psi+\chi_2)\rangle=\langle\Phi,\vec{P}_q\Psi\rangle,\nonumber\\
\langle(\Phi+\chi_1),\vec{J}_q(\Psi+\chi_2)\rangle=\langle\Phi,\vec{J}_q\Psi\rangle,\end{aligned}$$ or equivalently to prove $$\begin{aligned}
\langle\chi_1,\vec{P}_q\chi_2\rangle=0,~~~\langle\Psi,\vec{P}_q\chi_2\rangle=0,~~~\langle\chi_1,\vec{P}_q\Psi\rangle=0,\end{aligned}$$ and the same for $\vec{J}_q$, where $\Psi\in\emph{H}'$ and $\chi_1, \chi_2\in\emph{H}''$. All states $\chi$ of $\emph{H}''$ can be expressed as $\sum_n
a_n(\partial_{\mu}\emph{A}^{\mu})^n|\Psi_{phys}\rangle$, so one has to consider $$_{out}\langle\Phi|{\int}d^3y\psi^{\dag}(y)\frac{\vec{\nabla}}{i}\psi(y)\partial_{\mu}A^{\mu}(x)|\Psi\rangle_{in},$$ which in the interaction representation can be written as $$\langle\Phi|\emph{T}(\int
d^3y\psi^{\dag}(y)\frac{\vec{\nabla}}{i}\psi(y)\partial_{\mu}A^{\mu}(x)\emph{S})|\Psi\rangle,$$ where $\emph{T}$ is time-ordering operator, $\emph{S}$ is the scattering operator. Expanding the scattering operator as usual, one has $$\emph{S}=\sum_n \frac{(-i)^n}{n!} \int dx_1{\cdots}dx_n\emph{T}
\mathcal{H}_I (x_1) \cdots \mathcal{H}_I (x_n),$$ where $\mathcal{H}_I(x)=-j_{\mu}A^{\mu}(x)$. The physical states contain only transverse gluons (photons), so the operator $\partial_{\mu}A^{\mu}(x)$ must be contracted with one $A^{\nu}(y)$ in the interaction term $\mathcal{H}_I(x)$. This will give rise to a term $$\partial_{\mu}^x\overbrace{{A^{\mu}(x)A^{\nu}(y)}}=\partial_{\mu}^x\{g^{\mu\nu}D(x-y)\}=\partial^{\nu}_xD(x-y)
=-\partial^{\nu}_yD(x-y),$$ here the symbol $\overbrace{A^{\mu}(x)A^{\nu}(y)}=g^{\mu\nu}D(x-y)$ means contraction. Then using integration by parts one can move the differential operator $\partial_{\nu}$ to act on $j_{\nu}$ in the interaction term $\mathcal{H}_I(y)$ and use current conservation $\partial_{\nu}j^{\nu}=0$ to prove $Eq.(10)=0$. The other terms of Eq.(8) can be proved to be zero in the same way. Thus one has proved the quark (electron) momentum and orbital angular momentum operators are gauge independent and so are observable.
Conclusion
==========
The conflict between canonical commutation relation and gauge invariance of the quark (electron) momentum and angular momentum operators in a nucleon (atom) can be remedied by using gauge non-invariant canonical quark (electron) momentum and orbital angular momentum operators because they are gauge independent ones. This problem has been discussed by us since 1998[@chen]. Here we verified the results there by an alternative argument.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by NSFC through grant 10375030, 10435080.
[0]{} X.S. Chen and F. Wang, [*Commun. Theor. Phys.*]{} [**27**]{}, 121 (1997).
X. Ji, [*Phys. Rev. Lett.*]{} [**78**]{}, 610 (1997).
F. Strocchi and A.S. Wightman, [*J. Math. Phys.*]{} [**15**]{}, 2198 (1974).
X. Chen and F. Wang, hep-ph/9802346.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The time-evolution and space-distribution of internal electromagnetic fields in heavy-ion reactions at beam energies between 200 and 2000 MeV/nucleon are studied within an Isospin-dependent Boltzmann-Uhling-Uhlenbeck transport model IBUU11. While the magnetic field can reach about $7\times 10^{16}$ G which is significantly higher than the estimated surface magnetic field ($\sim 10^{15}$ G) of magnetars, it has almost no effect on nucleon observables as the Lorentz force is normally much weaker than the nuclear force. Very interestingly, however, the magnetic field generated by the projectile-like (target-like) spectator has a strong focusing/diverging effect on positive/negative pions at forward (backward) rapidities. Consequently, the differential $\pi^-/\pi^+$ ratio as a function of rapidity is significantly altered by the magnetic field while the total multiplicities of both positive and negative pions remain about the same. At beam energies above about 1 GeV/nucleon, while the integrated ratio of total $\pi^-$ to $\pi^+$ multiplicities is not, the differential $\pi^-/\pi^+$ ratio is sensitive to the density dependence of nuclear symmetry energy $E_{\rm{sym}}(\rho)$. Our findings suggest that magnetic effects should be carefully considered in future studies of using the differential $\pi^-/\pi^+$ ratio as a probe of the $E_{\rm{sym}}(\rho)$ at supra-saturation densities.'
author:
- Li Ou
- 'Bao-An Li[^1]'
title: 'Magnetic effects in heavy-ion collisions at intermediate energies'
---
Introduction
============
Magnetic fields exist everywhere in the Universe. To set the scale and appreciate the strong magnetic fields created during heavy-ion collisions, we first recall the magnitudes of several typical magnetic fields from various sources. Many spiral galaxies have magnetic fields with a typical strength of $\sim 3\times 10^{-6}$ G [@Sofue86] and it is estimated that the intergalactic magnetic fields presently have an intensity of about $\leq 10^{-9}$ G [@Kawabata69]. Some people believe that the present magnetic field of the Universe is amplified from a seed about $10^{-20}$ G by the dynamo mechanism [@Dimopoulosa97; @Grasso01] while magnetic fields up to $10^{24}$ G might appear in the early Universe [@Grasso01]. The strongest magnetic field of about $10^{15}$ G near the surfaces of magnetars [@Kouveliotou98; @Ibrahim02] or even higher ($10^{16}$-10$^{17}$ G) associated with the cosmological gamma-ray bursts [@Ruderman00] have been found from astrophysical observations. Due to the limit of tensile strength of terrestrial materials, the strongest man-made steady magnetic field is only about $4.5 \times 10^{5}$ G. To our best knowledge, it was first pointed out by Rafelski and M$\rm{\ddot{u}}$ller that, in addition to strong electrical fields, unusually strong magnetic fields are also created in heavy-ions collisions (HICs). In sub-Coulomb barrier U+U collisions, the magnetic field was estimated to be on the order of $10^{14}$ G [@Rafelski76]. More recently, it has been shown by Kharzeev et al. that HICs at RHIC and LHC can create the strongest magnetic field ever achieved in a terrestrial laboratory [@Kharzeev08]. For example, in noncentral Au+Au collisions at 100 GeV/nucleon, the maximal magnetic field can reach about $10^{17}$ G [@Kharzeev08; @Skokov09]. It thus provides a unique environment to investigate the Quantum Chromodynamics (QCD) at the limit of high magnetic field. Indeed, the study of quark-gluon-plasma under strong magnetic field has attracted much attention by the high energy heavy-ion community, see, e.g., ref. [@Kha11] and references therein. In particular, it has been shown theoretically that [@Kharzeev06; @Kharzeev07; @Kharzeev08; @Voloshin10] QCD topological effects in the presence of very intense electromagnetic fields, i.e., the “Chiral Magnetic Effect”, may be an evidence of local parity violation in strong interactions. Experimentally, interesting indications have been reported, see, e.g., refs. [@Abelev09; @Abelev10].
Stimulated by the interesting findings at RHIC and realizing that all transport model studies of magnetic effects have so far focused on high energy HICs [@Skokov09; @Voronyuk11], we investigate in this work first the strength, duration and distribution of internal magnetic fields created in HICs at beam energies between 200 and 2000 MeV/nucleon. This is the beam energy range covered by several accelerators in the world. We then focus on identifying possible magnetic effects on experimental observables using an isospin-dependent Boltzmann-Uhling-Uhlenbeck (BUU) transport model IBUU11 [@BALi04; @Li11]. We find that while the magnetic field can reach about $7\times 10^{16}$ G in these reactions, it has almost no effect on nucleon observables as the Lorentz force is negligibly small compared to the nuclear force. Very interestingly, however, the magnetic field generated by the projectile-like (target-like) spectator moving forward (backward) in the center of mass frame has a strong focusing/diverging effect on positive/negative pions moving forward (backward). As a result, the differential $\pi^-/\pi^+$ ratio as a function of rapidity is significantly altered by the magnetic field while the total $\pi^-$ and $\pi^+$ multiplicities remain about the same.
The paper is organized as follows. In the next section, we outline how the internal electromagnetic fields in HICs are calculated in the IBUU11 transport model. The characteristics of the electromagnetic fields and their effects on several experimental observables in intermediate energy HICs are then discussed in Section III. Finally, a summary is given at the end.
The model
=========
In the presence of electrical and magnetic fields $\bm{E}$ and $\bm{B}$, the BUU equation can be written as $$\begin{aligned}
\label{BUU}
\left[\frac{\partial}{\partial
t}+\frac{\bm{P}}{E}\nabla_r-(\nabla_r\bm{U}-q\bm{v}\times
\bm{B}-q\bm{E})\nabla_p\right]f(\bm{r},\bm{p},t)
=I(\bm{r},\bm{p},t)\end{aligned}$$ where $I(\bm{r},\bm{p},t)$ is the collision integral simulated by using the Monte Carlo method. The electrical field $\bm{E}$ (Coulomb field) has already been considered in most transport models. To include consistently both the electrical and magnetic fields satisfying Maxwell’s equations, the Li$\rm{\acute{e}}$nard-Wiechert potentials at a position $\bm{r}$ and time $t$ are evaluated according to $$\begin{aligned}
\label{Ef}
e \bm{E}(\bm{r},t) &=&\frac{e^2}{4\pi \epsilon_0}\sum_n
Z_n\frac{c^2-v_n^2} {(c R_n-\bm{R}_n \cdot \bm{v}_n )^3 }
(c\bm{R}_n - R_n\bm{v}_n)\end{aligned}$$ and $$\begin{aligned}
\label{Bf}
e \bm{B}(\bm{r},t) &=&\frac{e^2}{4\pi \epsilon_0 c}\sum_n Z_n\frac{c^2-v_n^2}
{(c R_n-\bm{R}_n \cdot \bm{v}_n )^3 } \bm{v}_n \times \bm{R}_n\end{aligned}$$ where $Z_n$ is the charge number of the $n$th particle. $\bm{R}_n=\bm{r}-\bm{r}'_n$ is the relative position of the field point $\bm{r}$ with respect to the position $\bm{r}'_n$ of particle $n$ moving with velocity $\bm{v}_n$ at the retarded time $t_{{\rm
r}n}=t-|\bm{r}-\bm{r}'_n(t_{{\rm r} n})|/c$. The summation runs over all charged particles in the reaction system. In non-relativistic cases, i.e., all particles satisfy the condition $v\ll c$, the Eq. and reduce to the classical expressions $$\begin{aligned}
\label{Efc}
e \bm{E}(\bm{r},t) &=&\frac{e^2}{4\pi \epsilon_0}\sum_n Z_n\frac{1}
{R_n^3 } \bm{R}_n\end{aligned}$$ and $$\begin{aligned}
\label{Bfc}
e \bm{B}(\bm{r},t) &=&\frac{e^2}{4\pi \epsilon_0 c^2}\sum_n Z_n\frac{1} {R_n^3 } \bm{v}_n \times
\bm{R}_n.\end{aligned}$$ The first equation is essentially the Coulomb’s law, and the latter is the Bio-Savart law for a system of moving charges.
To take into account accurately the retardation effects, the phase space information of all nucleons before the moment $t$ are required to calculate the electromagnetic fields at that moment. Some special care is thus necessary in initializing the reaction. In principle, the two colliding nuclei should be initialized to come from infinitely far away towards each other on their Coulomb trajectories. In practice, considering the need of keeping the initial nuclei stable and the computing time low, the initial distance between the surfaces of the two colliding nuclei is taken as 3 fm in our calculations. We make a pre-collision phase space history for all nucleons assuming that they are frozen in the projectile/target moving with a center of mass velocity $\bm{v}_{p/t}$, i.e., $\bm{r}_i=\bm{r}_i^0+\bm{v}_{p/t}\cdot t$, where $\bm{r}_i^0$ is the initial coordinate of the nucleon. As we shall show, comparisons of our transport model calculations with analytical estimates for two moving charges (target and projectile) in both relativistic and non-relativistic cases indicate that our method of handing the pre-collision phase-space histories of all nucleons is reasonable.
![Density dependence of nuclear symmetry energy used in the IBUU11 calculations.[]{data-label="SymE"}](Esym.eps){width="50.00000%"}
We refer the BUU code used in this study IBUU11. Compared to the IBUU04 [@BALi04] where the MDI (Momentum-Dependent-Interaction) is used [@Das03], besides the electromagnetic fields with retardation effects, an isospin-dependent three-body force [@Xu10b] (instead of the standard one used in the MDI, Gogny and Skyrme effective interactions) is used. Moreover, the high-momentum tail of the MDI isoscalar potential is readjusted to better fit the nucleon optical potential from nucleon-nucleus scattering experiments. Details of these modifications and their effects on experimental observables will be presented in a forthcoming publication [@Li11]. In this work, we focus on the magnetic aspect of HICs at intermediate energies. Since one of our main motivations here is to see whether experimental observables known to be sensitive to the $E_{\rm{sym}}(\rho)$ is affected by the magnetic effects, we notice here that in the IBUU11 the $E_{\rm{sym}}(\rho)$ is controlled by a parameter $x$ introduced in the three-body part of the MDI interaction [@Das03; @Xu10b]. By adjusting the parameter $x$ one can mimic diverse behaviors of the $E_{\rm{sym}}(\rho)$ predicted by various microscopic many-body theories [@LCK08]. As an example, shown in Fig. \[SymE\] are the $E_{\rm{sym}}(\rho)$ with $x=1, 0$ and $-1$, respectively.
Results and discussions
=======================
In this section, we first illustrate and discuss the beam energy and impact parameter dependence of the time-evolution and space-distribution of magnetic field. To help understand the magnetic effect in HICs, we shall also compare the Lorentz force with the Coulomb and nuclear forces. We then present and discuss magnetic effects on experimental observables.
Characteristics of internal electromagnetic fields in heavy-ion reactions
-------------------------------------------------------------------------
Features of the internal electromagnetic fields are independent of the symmetry energy parameter $x$. In this subsection, unless otherwise specified a value of $x=1/3$ is used. We take the $z$ ($x$) axis as the beam (impact parameter) direction. Based on the formula of magnetic field strength in Eq. , the dominant component of the internal magnetic field is in the $y$ axis perpendicular to the reaction plane ($z-x$). The component in the reaction plane is negligible because of the slow motions of nucleons in the $x$ or $y$ directions especially in the early phase of the reaction. To test our approach used in calculating the electromagnetic fields, we first compare the magnetic field $B_y(0)$ at the center of mass of the reaction system calculated using the full IBUU11 dynamically with those obtained under some limiting conditions for idealized situations. Shown in Fig. \[bb0520\] are the values of $B_y(0)$ for Au+Au reactions at a beam energy of 500 AMeV and an impact parameter of $b$=5 and 20 fm, respectively. As a reference, the approximate magnetic field of $10^{15}$ G on the surfaces of magnetars is also indicated. The legend “classical” and “relativistic” indicate results obtained using Eq. and Eq. , respectively. For a comparison, we have also performed calculations using both Eq. and Eq. assuming that the projectile and target are two point charges located at their individual centers of masses and are moving with their initial velocities only. Results of this calculation are denoted by the “kine.”.
{width="50.00000%"}
[\[bb0520\]]{}
Several interesting observations can be made. Firstly, it is seen that the $B_y(0)$ calculated with the classical and relativistic formulas are very close to each other, for both the kinematic and dynamical calculations, as one expects for reactions at relatively low beam energies. Secondly, the dynamical IBUU11 results and the kinematic estimates are very close at the beginning and the end of the reaction, but they are very different during the reaction phase spanned by the small balls of the same color. The magnetic field has contributions from the projectile-like and target-like spectators as well as charged particles in the participant region. Contributions from the latter, however, are very weak because of the approximately isotropic nucleon momentum distribution there. Once the projectile and target begins overlapping, nucleon-nucleon collisions will start transferring the participants’ longitudinal momenta into transverse directions. Thus, the $B_y(0)$ from the IBUU11 is weaker than the kinematic estimate during the reaction phase. We notice that the magnetic field in the $x$ and $z$ directions are rather weak because they only come from charged participants which are moving essentially randomly in all possible directions. For the very peripheral reactions with $b$=20 fm, the two nuclei do not overlap. As one expects, thus there is almost no difference between the kinematic and dynamical results. The above comparisons enhance our confidence in using the IBUU11 model to study the internal electromagnetic fields and their effects in HICs. In the following, we only present results calculated with the relativistic formula and the dynamical IBUU11 model.
{width="100.00000%"}
[\[fig3D\]]{}
The contours of the nucleon density $\rho/\rho_{0}$ , the magnetic field strength $eB_y$, and the electric field strength $eE_x$ in the $x-z$ plane at $t$=10, 20, 30 and 40 fm/$c$ for the 500 AMeV Au+Au collisions at an impact parameter of $b$=10 fm are shown in Fig. \[fig3D\]. We notice that both the $eB_y$ and $eB_x$ are plotted here in unit of MeV$^2$ which is equal to $1.44\times
10^{13}$ G. For discussing the spatial distribution of the electromagnetic fields, we can divide the space into three zones in terms of the $x$ coordinate: the outside-zone where $|x|>$ 15 fm; the spectator-zone where 5 fm $\leq |x|\leq$ 15 fm; and the overlap-zone where $|x|<$ 5 fm. As mentioned above, the electromagnetic fields come from both the spectators and participants. In the outside-zone, the spectator near the field point generates a stronger magnetic field in the negative $y$-direction while the other spectator farther away generates a weaker magnetic field in the positive $y$-direction. The superposition leads to a magnetic field points to the negative $y$-direction. On the other hand, the electric field $eE_x$ in the outside-zone includes contributions from all charges. Its sign is the same as the sign of the $x$-coordinate of the field point. In the overlap-zone, the magnetic fields generated by the two spectators will superimpose constructively since they are all in the positive $y$-direction, while the magnetic fields generated there by the moving charges in the participant region will largely cancel each other. The strength of the magnetic field peaks when the two nuclei have reached the maximum compression. It then drops when the spectators depart from each other. The signs of the electric filed in the $x$-direction generated by the two spectators are always opposite, leading to the very weak electrical field in the participant region where the magnetic field is the strongest.
Next, we explore the impact parameter and beam energy dependence of the magnetic field at the center of mass of the reaction system. Shown in the right panel of Fig. \[figeb\] is the impact parameter dependence of $eB_y(0)$.
{width="1.\textwidth"}
[\[figeb\]]{}
The strength of magnetic field grows with increasing impact parameter $b$ up to about $b=12$ fm. It then starts decreasing with larger $b$. This is easily understandable. There are basically two factors determining the magnetic field strength for a given beam energy. One is the position vector $\bm R$ from the moving charges to the field point, and the other one is the charge number of the spectator $N_s$. Their competition determines the strength of the magnetic field. For head-on collisions, equivalently there are two counter currents leading to an almost zero magnetic field at the center of the reaction. For off-central collisions, as the impact parameter increases, while the spectators are farther away from the center they carry more charges. The net result is that the magnetic field becomes stronger with increasing impact parameter. However, as the impact parameter becomes larger than the sum of the radius of the projectile and target, e.g., when $b>12$ fm for the Au+Au reaction, almost all charges are with the spectators, the magnetic field is thus only determined by the $\bm R$. Therefore, the reactions with lager impact parameters create weaker magnetic fields at the center of the reaction. Based on the IBUU11 results, off-central collisions with $b=8\sim 10$ fm seem to be the most suitable impact parameter range to produce the strongest magnetic effect. These reactions create strong magnetic fields and also enough light charged particles moving in the magnetic fields to be detected in experiments. Another factor determining the strength of magnetic field is the velocity of spectators, i.e., the beam energy of the reaction. Shown in the left panel of Fig. \[figeb\] is the beam energy dependence of $eB_y(0)$. As one expects, while the maximum strength of the magnetic field increases with beam energy the duration of the strong magnetic field decreases since the spectators leave the collision region quickly at higher beam energies. Compared to reactions at RHIC, the strength of the magnetic field is about 10 times lower but the reaction lasts about 10 times longer. Since observable effects of any force depend on not only its strength but also its duration, magnetic effects in HICs at intermediate energies are thus worth an investigation.
Magnetic effects on observables in heavy-ion collisions
-------------------------------------------------------
While no chiral magnetic effect is expected in HICs at intermediate energies, it is still interesting to examine magnetic effects on hadronic observables. First of all, we would like to mention that the effects of strong magnetic fields on the Equation of State (EOS) of cold hadronic and quark matter including the Landau quantization and the nucleon anomalous magnetic moment in neutron stars have been studied extensively, see, e.g., refs. [@Bro; @Car; @Ban; @Rab08; @Rab11]. It has been shown consistently that the magnetic effects become significant only for magnetic fields stronger than about $10^{18}$ G. Moreover, at finite temperature some of the magnetic effects get mostly washed out [@Rab11]. Since the temperature is high and the maximum strength of the magnetic field created is still below $10^{18}$ G even at RHIC energies, it is not necessary to consider effects of the magnetic field on the nuclear EOS. Instead, we focus directly on magnetic effects due to the Lorentz force acting on moving charges. In the following, we examine separately magnetic effects on nucleons and pions.
### Lorentz force compared with the Coulomb and nuclear forces
{width="45.00000%"} {width="45.00000%"}\
{width="45.00000%"} {width="45.00000%"}\
[\[fben\]]{}
For nucleons, the magnetic effects are expected to be negligible as the Lorentz force is known to be very small compared to the nuclear force. On the other hand, while the electrical and magnetic fields are strongly correlated, the Coulomb force has been routinely taken into account but the Lorentz force is normally neglected in modeling HICs. To check the validity of this practice and obtain a more quantitative understanding about the relative importance of the Lorentz, Coulomb and nuclear forces, we examine in Fig. \[fben\] the ratios of the Lorentz force over the Coulomb and nuclear forces for a test-charge. To be specific, we calculate the ratio $R_x^{ME}$ of the $x$-component of the Lorentz force over that of the Coulomb force for a test-charge in the outside-zone. As a reference, we first make an analytical analysis for a simplified case. For a test-charge located at the surface of the projectile moving on the trajectory of $\bm{r}$($-\frac{b}{2}-R$, 0, $z_0+v_{0}
t$), where $R$, $z_0$ and $v_0$ is the radii, initial $z$-coordinate and the beam velocity, assuming the electromagnetic fields are due to two moving point charges (projectile and target) given by Eqs. and , the $R_x^{ME}$ is simply $$\begin{aligned}
\label{rx}
R_x^{ME}=\frac{F^M_x}{F^E_x}=\frac{e v_zB_y }{e
E_x}=\left(\frac{v_0}{c}\right)^2.\end{aligned}$$ Thus, it is clear that only for fast moving particles likely existing in reactions at high beam energies, the Lorentz force is expected to be significant compared to the Coulomb force. We now examine numerically the $R_x^{ME}$ for the test-charge using the electromagnetic fields calculated with the IBUU11. In window (a), the time evolution of $R_x^{ME}$ is shown for several impact parameters for the 500 AMeV Au+Au reactions. The evolution can be approximately divided into four periods. Before the two nuclei get in touch, $R_x^{ME}=$0.21 which is exactly the same as the prediction of Eq. \[rx\]. In the compression phase, since the magnetic fields in the outside region generated by the projectile-like and target-like spectators are in the opposite directions, the net magnetic field decreases whereas the electric field there becomes stronger. Consequently, the $R_x^{ME}$ drops until about 15 fm/c. In the expansion phase, the situation is reversed. After the collisions are over, the $R_x^{ME}$ keeps approximately a constant value smaller than $(v_0/c)^2$ depending on the impact parameter. The beam energy dependence shown in window (b) for the Au+Au reactions with $b$=10 fm can be similarly understood. We notice that $R_x^{ME}=(v_0/c)^2$ at the beginning of the collision is satisfied at all beam energies. As the incident energy increases, the Lorentz force becomes closer to the Coulomb force.
We now turn to the ratio between the $x$-components of the nuclear and Lorentz forces, i.e., $R_x^{NM}=F_x^N/F_x^M$, for a test-proton at the center of mass with a constant velocity of $v_z=v_{0}$. Shown in windows (c) and (d) are the impact parameter and beam energy dependences of the $R_x^{NM}$. Because the nuclear force is proportional to the gradient of the single-nucleon potential, i.e., $F^{M}=-\nabla_r U$, large fluctuations are seen in the $R_x^{NM}$. It is seen that the nuclear force is several 10 to 10$^2$ times larger than the Lorentz force. The magnetic field is thus not expected to affect the reaction dynamics and nucleon observables. Therefore, it is not surprising that nuclear reaction models can describe most experimental data without considering any magnetic effect at all.
### Magnetic effects on collective observables of nucleons and pions
![(Top pane) The average in-plane transverse momentum of free protons and pions as a function of rapidity. (Bottom pane) Elliptic flow for free protons and pions as a function of transverse momentum for the 2 AGeV Au+Au reaction at an impact parameter of 5 fm with $x$=0. []{data-label="fv1v2"}](v1v2x0.EPS){width="100.00000%"}
While the magnetic effects on nucleon observables are expected to be very small, to be quantitative it is still necessary to examine how small the effects are. From the expression of the Lorentz force $F^{M}= q \bm{v} \times \bm{B}$, it is easy to see that the main component of the Lorentz force is in the reaction plane (especially in the $x$-direction). The average transverse momentum in the reaction plane, i.e., $<p_x>$, is thus a good candidate. Shown in the top panels of Fig. \[fv1v2\] are the average in-plane transverse momentum as a function of rapidity, the so-called in-plane transverse flow [@Danielewicz85], for free protons and pions, respectively. Indeed, there is essentially no magnetic effect on nucleons. It is seen that both negative and positive pions flow in the same direction as nucleons but with much lower transverse momentum in the reaction plane [@Li94]. Interestingly, there is a very weak indication of some magnetic effects on the $<p_x(y)>$ of pions at forward/backward rapidities. This is qualitatively understandable because the Lorentz force influences pions motion easily as they are light compared to nucleons. Moreover, it also indicates that the magnetic field decreases (increases) very slightly the magnitude of $<p_x>$ for positive (negative) pions at both forward and backward rapidities due to the magnetic focusing/diverging effects as we shall discuss in detail in the next subsection. Next, we investigate in the lower panels of Fig. \[fv1v2\] the so-called differential elliptic flow as a function of transverse momentum [@Ollitrault98; @Poskanzer98], $$\begin{aligned}
\left<v_2(p_t)\right>=\frac{1}{N}\sum_{i=1}^{N}\frac{p_{ix}^2-p_{iy}^2}{p_{ix}^2+p_{iy}^2}\end{aligned}$$ where $N$ is the total number of free particles. The $p_{iy}$ is $i$th particle’s transverse momentum perpendicular to the reaction plane. Again, there is essentially no magnetic effect on the differential elliptical flow of both nucleons and pions.
### Magnetic effects on the $\pi^-/\pi^+$ and neuteron/proton ratio
It is well known that the Coulomb force affects significantly the $\pi^-/\pi^+$ ratio in HICs. The so-called Coulomb peak often appears near the projectile and/or target rapidities. This phenomenon has been studied extensively both experimentally [@wolf79; @chiba79; @benenson; @radi82; @schn82] and theoretically [@koonin79; @bertsch80; @gyulassy81; @bonasera; @nore88; @li94] since the 1970’s, see. e.g., ref. [@stock86] for a review. However, magnetic effects were not considered in any of these studies. While the Lorentz force on pions is normally smaller than the Coulomb force, they have the same order of magnitude. Moreover, compared to nucleons pions are light with relatively higher speeds and are thus more easily affected by the Lorentz force. Furthermore, there is no nuclear force acting on pions once they are produced at least in most model simulations where pions change their momenta only through pion-hadron collisions and the Coulomb field. To our best knowledge, theoretical studies on the mean-field (in-medium dispersion relation) for pions are still rather inconclusive [@Ko10]. Considering all of the above, the magnetic force on pions can be significant. In fact, we expect the Lorentz and Coulomb forces to have the opposite effects on the $\pi^-/\pi^+$ ratio. Namely, near the projectile/targer rapidity the Coulomb force increase the $\pi^-/\pi^+$ ratio while the Lorentz force reduces it. Effects of the Lorentz forces on positive and negative pions are illustrated in Fig. \[Bfocus\] using the projectile-like spectator as an example.
![A sketch of magnetic effect on the $\pi^-/\pi^+$ ratio near the projectile rapidity.[]{data-label="Bfocus"}](Bfocus.EPS){width="50.00000%"}
The moving track of the spectator can be regarded as a current. Above/below the current, the magnetic field is perpendicular to the reaction plane and points outward/inward. The Lorentz force focuses the $\pi^{+}$ into smaller forward (backward) polar angles while disperses the $\pi^{-}$ to larger forward (backward) polar angles. So the $\pi^{-}/\pi^{+}$ ratios at large rapidities are reduced by the Lorentz force. Moreover, due to the magnetic focusing/dispersing effect on positive/negative charges, the changes in transverse momentum for particles above and below the current are opposite. So the total magnetic effect on the average transverse momentum in the reaction plane is very tiny even for pions. This explains why the magnetic effects on the transverse flow $<p_x(y)>$ and the differential elliptical flow $v_2(p_t)$ are negligible for both nucleons and pions.
Why is it so important to understand clearly and precisely the electromagnetic effects on the $\pi^-/\pi^+$ ratio? One special reason is that the $\pi^{-}/\pi^{+}$ ratio has been predicted as one of the most promising probes of the nuclear symmetry energy at supra-saturation densities [@LiBA02]. While comparisons of transport model predictions [@xia; @Feng10; @Pra] with exiting data [@Rei07] are still inconclusive, all models have consistently shown that the $\pi^-/\pi^+$ ratio is rather sensitive to the high density behavior of the nuclear symmetry energy. The latter is rather poorly known as indicated in Fig. \[SymE\]. In fact, even the trend of the symmetry energy at supra-saturation densities, namely, whether it increases or decreases with increasing density, is still controversial partially because of our poor knowledge about the isospin-dependence of strong interaction. To extract reliably accurate information from the $\pi^-/\pi^+$ ratio about the high-density symmetry energy, it is thus necessary to understand precisely effects from the well-known electromagnetic interactions. So, how strong is the magnetic effect on the $\pi^{-}/\pi^{+}$ ratio in comparison to the symmetry energy effect? To answer this question and give a quantitative example, we show in Fig. \[ry\] the $\pi^{-}/\pi^{+}$ ratio as a function of rapidity with and without the magnetic field calculated with three different values of the symmetry energy parameter $x$ for the 2 AGeV Au+Au reactions at an impact parameter of b=0 and 5 fm, respectively. In each case considered here, 200,000 IBUU11 events are used.
![The $\pi^{-}/\pi^{+}$ ratio (top windows), the neutron/proton $n/p$ ratio of free (middle windows) and all (bottom windows) nucleons as a function of rapidity with and without the magnetic field calculated with the three different values of symmetry energy parameter $x$ for the reactions of 2 AGeV Au+Au at an impact parameter b=0 and 5 fm, respectively.[]{data-label="ry"}](ryX.EPS){width="70.00000%"}
Comparing the results obtained with and without the magnetic field using any of the $x$ parameter considered, it is seen that significant magnetic effects on the $\pi^{-}/\pi^{+}$ ratio are obvious especially at forward and backward rapidities particularly for mid-central collisions. Quantitatively, the $\pi^-/\pi^+$ ratio obtained with the magnetic field is significantly lower at forward and backward rapidities (polar angles) due to the magnetic focusing/diverging effects on the positive/negative pions as we illustrated in Fig. \[Bfocus\]. Pions at higher rapidities have larger longitudinal momenta and thus feel stronger Lorentz forces compared to those at mid-rapidity. For the head-on collisions, the magnetic effect is small but still appreciable especially in the early phase of the reactions when most of the pions are produced. From peripheral to head-on collisions, the $\pi^-/\pi^+$ ratio changes gradually from forward-backward peaked to center-peaked distributions. In peripheral collisions, there are significant Coulomb effects due to the spectators. One thus expects the $\pi^-/\pi^+$ ratio to peak at forward-backward rapidities. It is seen that the magnetic effect at forward-backward rapidities is compatible with the symmetry energy effect from changing the $x$ parameter by one unit. Overall, the $\pi^-/\pi^+$ ratio decreases as the symmetry energy at supra-saturation densities becomes stiffer when the parameter $x$ changes from 1 to -1.
It is worth noticing that so far only the integrated $\pi^{-}/\pi^{+}$ ratio, i.e., the ratio of total $\pi^-$ to $\pi^+$ multiplicities, has been used in attempts to constrain the symmetry energy at high densities without considering the magnetic effects. While the integrated $\pi^{-}/\pi^{+}$ ratio is rather sensitive to the symmetry energy parameter $x$ in reactions near the pion production threshold, as the beam energy becomes higher than about 1 Gev/nucleon, the sensitivity gradually disappears [@xia]. It is thus interesting to see that the rapidity distribution of the $\pi^{-}/\pi^{+}$ ratio shows a strong sensitivity to the parameter $x$ even in the reactions at a beam energy of 2 GeV/nucleon where the baryon density can reach about $3.5\rho_0$. Since the strongest sensitivity to the symmetry energy is at forward and backward rapidities where the $\pi^{-}/\pi^{+}$ ratio is also strongly affected by the magnetic field, special cares have to be taken in both model calculations and the data analysis. Most of the available detectors including the one used by the FOPI Collaboration [@Rei07] do not provide full coverage at very forward/backward angles. The integrated $\pi^{-}/\pi^{+}$ ratio is normally obtained by extrapolating the angular distributions of pions measured in a limited angular range to all polar angles. By doing so, however, the magnetic effects on the angular distribution were neglected. Previous conclusions on the high-density symmetry energy based on comparing various transport model calculations with the experimental data without considering the magnetic effects thus need to be taken with caution. For comparisons, the neutron/proton ratio $n/p$ of free (selected as those with local density less than $\rho_0/8$ at freeze-out) and all nucleons are shown as functions of rapidity in the middle and bottom windows of Fig. \[ry\], respectively. It is seen that there is essentially no noticeable magnetic effects within error bars on the $n/p$ ratios. This is consistent with our expectation and the results on the transverse and elliptical flows discussed earlier. The non-uniform $n/p$ and $\pi^{-}/\pi^{+}$ ratios as functions of rapidity indicates the lack of complete isospin equilibrium for both the nucleon and pion components. This is the so-called isospin translucency expected in heavy-ion reactions at the beam energies studied here [@LiBA95].
[@\*[5]{}[c]{} ]{} Ratio &$b$ (fm) &$x$=1 &$x$=0&$x$=-1\
& 0 & 2.02/1.97& 1.81/1.78& 1.68/1.67\
&5 & 1.87/1.86& 1.79/1.79 & 1.73/1.73\
&0 &1.23/1.23 &1.24/1.24 &1.25/1.25\
&5 &1.28/1.28 &1.29/1.29 &1.29/1.29\
&0 &1.23/1.23 &1.24/1.24 &1.25/1.25\
&5 &1.31/1.31 &1.31/1.31 &1.32/1.32\
------------------------------------------------------------------------
Shown in Table \[table1\] are the integrated $\pi^{-}/\pi^{+}$ and neutron/proton ratios calculated without/with the magnetic field. It is seen that the integrated ratios are not affected much by the magnetic field. This is what we expected as the Lorentz force affects differently only the angular distributions of positively and negatively charged particles, but not their total multiplicities. Also, consistent with previous findings [@xia] the integrated $\pi^{-}/\pi^{+}$ ratio at beam energies higher than about 1 GeV/nucleon is not so sensitive to the variation of the symmetry energy while there is a clear indication that a higher $\pi^{-}/\pi^{+}$ ratio is obtained with a softer $E_{\rm{sym}}(\rho)$ at supra-saturation densities. Thus, the differential $\pi^{-}/\pi^{+}$ ratio as a function of rapidity, as we discussed earlier, is a better probe of the symmetry energy at supra-saturation densities after taking care of the magnetic effects.
Summary
=======
In summary, within the transport model IBUU11, the time-evolution and space-distribution of internal electromagnetic fields in HICs at beam energies between 200 and 2000 MeV/nucleon are studied. While the magnetic field can reach about $7\times 10^{16}$ G, it has almost no effect on nucleon observables as the Lorentz force is normally much weaker than the nuclear force. On the other hand, the magnetic field has a strong focusing/diverging effect on positive/negative pions at forward/backward rapidities. Consequently, the differential $\pi^-/\pi^+$ ratio as a function of rapidity, but not the integrated one, is significantly altered by the magnetic field. At beam energies above about 1 GeV/nucleon, the differential $\pi^-/\pi^+$ ratio is more sensitive to the $E_{\rm{sym}}(\rho)$ than the integrated $\pi^-/\pi^+$ ratio. Our findings suggest that magnetic effects should be carefully considered in future studies of using the differential $\pi^-/\pi^+$ ratio as a probe of the $E_{\rm{sym}}(\rho)$ at supra-saturation densities.
Acknowledgements
================
We would like to thank Drs. N. Chamel, W. G. Newton and C. Providencia for helpful discussions and information on magnetic effects in neutron stars, Dr. Lie-Wen Chen and Dr. Chang Xu for collaborations in developing the IBUU11 code used in this study. We would also like to thank Dr. Derek Harter who made our very intensive calculations possible within a rather short time by providing us access to the high-performance Computational Science Research Cluster at Texas A&M University-Commerce. This work was supported in part by the NSF under grants PHY-0757839 and PHY-1068022 and NASA under grant NNX11AC41G issued through the Science Mission Directorate, and the National Natural Science Foundation of China under Grant Nos 11005022, 10847004 and 11075215.
[99]{}
Y. Sofue, M. Fujimoto and R. Wielebinski, Annu. Rev. Astron. Astrophys. [**24**]{}, 459 (1986).
K. Kawabata, M. Fujimoto, Y. Sofue and M. Fukui, Publ. Astron. Soc. Jpn. [**21**]{}, 239 (1969).
K. Dimopoulosa and A.-C. Davis, Phys. Lett. B [**390**]{}, 87 (1997)
D. Grasso and H. R. Rubinstein, Phys. Rep. [**348**]{}, 163 (2001).
C. Kouveliotou, S. Dieters and T. Strohmayer, [*et al.*]{}, Nature [**393**]{}, 235 (1998).
A. I. Ibrahim, S. Safi-Harb and J. H. Swank [*et al.*]{}, Astrophys. J., [**574**]{}, L51¨CL55 (2002).
M. A. Ruderman, L. Tao and W. Kluzniak, Astrophys. J. [**542**]{}, 243 (2000).
J. Rafelski and B. M$\rm{\ddot{u}}$ller, Phys. Rev. Lett. [**36**]{}, 517 (1976).
Dmitri E. Kharzeev and Ho-Ung Yee, Phys. Rev. D[**83**]{}, 085007 (2011) and references therein.
D. E. Kharzeev, L. D. McLerran and H. J. Warringa, Nucl. Phys. A [**803**]{}, 227 (2008).
D. Kharzeev, Phys. Lett. B [**633**]{}, 260 (2006).
D. Kharzeev and Zhitnitky, Nucl. Phys. A [**797**]{}, 67 (2007).
S. A. Voloshin, Phys. Rev. Lett. [**105**]{}, 172301 (2010).
V. Skokov, A. Illarionov, V. Toneev, Int. J. Mod. Phys. A [**24**]{}, 5925 (2009).
B. I. Abelev [*et al.*]{} (STAR Collaboration), Phys. Rev. Lett. [**103**]{}, 251601 (2009).
B. I. Abelev [*et al.*]{} (STAR Collaboration), Phys. Rev. C [**81**]{}, 054908 (2010).
V. Voronyuk, V.D Toneev, W. Cassing, E.L. Bratkovskaya, V.P. Konchakovski and S.A. Voloshin, Phys. Rev. C [**83**]{}, 054911 (2011).
B. A. Li, C. B. Das, S. Das Gupta and C. Gale, Phys. Rev. C [**69**]{}, 011603(R) (2004); Nucl. Phys. A [**735**]{}, 563 (2004).
B. A. Li, L.W. Chen, L. Ou and C. Xu in preparation.
C. B. Das, S. Das Gupta, C. Gale, B. A. Li, Phys. Rev. C **67**, 034611 (2003).
C. Xu and Bao-An Li, Phys. Rev. C[**81**]{}, 044603 (2010).
A. Broderick, M. Prakash and J. M. Lattimer, APJ [**537**]{}, 351 (2000).
C. Y. Cardall, M. Prakash and J. M. Lattimer, APJ [**554**]{}, 322 (2001).
D. Bandyopadhyas, S. Chakrabarty and S. Pal, Phys. Rev. Lett. [**79**]{}, 2176 (1997).
A. Rabhi and C. Providencia, J. Phys. G: Nucl. Part. Phys. [**35**]{}, 125201 (2008).
A. Rabhi, P. K. Panda and C. Providencia, arXiv:1105.0254.
P. Danielewicz and G. Odyniec, Phys. Lett. B [**157**]{}, 146 (1985).
Bao-An Li, Nucl. Phys. [**A570**]{}, 797 (1994).
J.-Y. Ollitrault, Nucl. Phys. A [**638**]{}, 195c (1998).
A. Poskanzer and S. A. Voloshin, Phys. Rev. C [**58**]{}, 1671 (1998).
K. L. Wolf et. al. Phys. Rev. Lett. [**42**]{}, 1448 (1979).
J. Chiba et al. , Phys. Rev. C [**20**]{}, 2210 (1979).
W. Benenson et al. , Phys. Rev. Lett. [**10**]{}, 683 (1979).
H. M. A. Radi et al. , Phys. Rev. C[**25**]{}, 1518 (1982).
S. Schnetzer et al. , Phys. Rev. Lett. [**49**]{}, 989 (1982).
J. P. Sullivan et al. Phys. Rev. C [**25**]{}, 1499 (1982).
K. G. Libbrecht and S. E. Koonin, Phys. Rev. Lett. [**43**]{}, 1581 (1979).
G. F. Bertsch, Nature [**283**]{}, 280 (1980).
M. Gyulassy and S. K. Kauffmann, Nucl. Phys. [**A362**]{}, 503 (1981).
A. Bonasera and G. F. Bertsch, Phys. Lett. [**B195**]{}, 521 (1987).
B. Noren et al. , Nucl. Phys. [**A489**]{}, 763 (1988).
B. A. Li, Phys. Rev. C[**50**]{}, 2144 (1994); Phys. Lett. [**B346**]{}, 5 (1995).
R. Stock, Phys. Report, [**135**]{}, 259 (1986).
Jun Xu, Che Ming Ko and Yongseok Oh, Phys. Rev. C[**81**]{}, 024901 (2010).
B. A. Li, Phys. Rev. Lett. [**88**]{}, 192701 (2002); Nucl. Phys. [**A708**]{}, 365 (2002).
Z. G. Xiao, B. A. Li, L. W. Chen, G. C. Yong and M. Zhang, Phys. Rev. Lett. [**102**]{}, 062502 (2009).
Z. Q. Feng and G. M. Jin, Phys. Lett. [**B683**]{}, 140 (2010).
V. Prassa, G. Ferini, T. Gaitanos, H.H. Wolter, G.A. Lalazissis and M. Di Toro, Nucl. Phys. [**A789**]{}, 311 (2007).
W. Reisdorf et al., Nucl. Phys. A [**781**]{}, 459 (2007).
Bao-An Li, Lie-Wen Chen and Che Ming Ko, Phys. Rep. [**464**]{}, 113 (2008).
B. A. Li and S. J. Yennello, Phys. Rev. C 52, 1746(R) (1995).
[^1]: Corresponding author. Bao-An\[email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We formulate the Hubbard model for the simple cubic lattice in the representation of interacting dimers applying the exact solution of the dimer problem. By eliminating from the considerations unoccupied dimer energy levels in the large $U$ limit (it is the only assumption) we analytically derive the Hubbard Hamiltonian for the dimer (analogous to the well-known $t-J$ model), as well as, the Hubbard Hamiltonian for the crystal as a whole by means of the projection technique. Using this approach we can better visualize the complexity of the model, so deeply hidden in its original form. The resulting Hamiltonian is a mixture of many multiple ferromagnetic, antiferromagnetic and more exotic interactions competing one with another. The interplay between different competitive interactions has a decisive influence on the resulting thermodynamic properties of the model, depending on temperature, model parameters and assumed average number of electrons per lattice site. A simplified form of the derived Hamiltonian can be obtained using additionally Taylor expansion with respect to $x=\frac{t}{U}$ ($t$-hopping integral between nearest neighbours, $U$-Coulomb repulsion). As an example, we present the expansion including all terms proportional to $t$ and to $\frac{t^2}U$ and we reproduce the exact form of the Hubbard Hamiltonian in the limit $U\rightarrow \infty $.
The nonperturbative approach, presented in this paper, can, in principle, be applied to clusters of any size, as well as, to another types of model Hamiltonians.
author:
- |
M.Matlak[^1], J.Aksamit, B.Grabiec\
Institute of Physics, Silesian University,\
4 Uniwersytecka, PL-40-007 Katowice,Poland\
and\
W.Nolting\
Institute of Physics, Humboldt University,\
110 Invalidenstr., D-10115 Berlin, Germany
title: 'Hubbard Hamiltonian in the dimer representation. Large $U$ limit'
---
Introduction
============
The single-band Hubbard model, Ref. \[1\], plays in the solid state physics a similar principal role as the hydrogen atom in the atomic physics. This explains a continuous interest in its properties. The Hubbard Hamiltonian reads $$H=\sum\limits_{i,j,\sigma }t_{i,j}c_{i,\sigma }^{+}c_{j,\sigma
}+U\sum\limits_in_{i,\uparrow }n_{i,\downarrow }.$$ Here $c_{i,\sigma }$ $(c_{i,\sigma }^{+})$ are annihilation (creation) operators of an electron with spin $\sigma =\uparrow $, $\downarrow $ in the Wannier representation at the lattice site $\mathbf{R}_i$ and $n_{i,\sigma
}=c_{i,\sigma }^{+}c_{i,\sigma }$. Moreover, $t_{i,j}$ is the hopping integral between different lattice sites $i$ and $j$ ($t_{i,i}=0)$ and $U$ is the intrasite Coulomb repulsion. The Bloch conduction band energy $%
\varepsilon _{\mathbf{k}}$ is given by
$$\varepsilon _{\mathbf{k}}=\sum\limits_{i-j}t_{i,j}e^{-i\mathbf{k\cdot (R}_i-%
\mathbf{R}_j)}.$$
In the following we restrict ourselves to the simple cubic (sc) lattice and assume that
$$t_{i,j}=\left\{
\begin{array}{ll}
-t & i,j\mbox{-nearest neighbours} \\
0 & \mbox{otherwise}.
\end{array}
\right.$$ Then
$$\varepsilon _{\mathbf{k}}=-2t(\cos k_xa+\cos k_ya+\cos k_za).$$
The hopping parameter $t$ is simply related to the bandwidth $W$ of the Bloch band (2), e. g. $W=12t$ for the sc lattice$.$ The interplay between the two model parameters, $W$ and $U,$ is decisive for the properties of the model resulting in strong electron correlations, leading to band magnetism (see e.g. Refs \[2\], \[3\] for a review), insulator-to-metal transition (see e.g. Refs \[3\], \[4\] and papers cited therein) and high-$T_c$ superconductivity (negative $U$-model, see e.g. Ref. \[5\]). The Hubbard model very often plays also a role of a submodel for many other more complicated models (as e.g. Anderson model (see Ref. \[6\]), s-f model (see Refs \[7,8\]) and so on). Especially interesting, but difficult to handle are the properties of the Hubbard model in such a regime of the model parameters where the bandwidth $W$ is comparable to the Coulomb repulsion $U$. In a large number of papers \[9-30\] the authors tried to solve this model using many sophisticated methods. The exact solution, however, does not exist till now. Many authors tried to change the situation in this field by introducing the expansion parameter $x=\frac{t}{U}$ $(x\ll 1)$. This idea (cf Refs \[31,32\]) consists in replacing ”difficult physics” connected with the model by ”difficult mathematics” obtained by a laborious expansion with respect to $x.$ Different methods connected with this problem have been applied as e.g. the perturbation expansion (see Refs \[18\], \[21\]), canonical transformation (see Refs \[9\], \[12\], \[17\], \[22\], \[23\]) or ab initio derivations (see Refs \[25\], \[28-30\]). Most of the methods, leading to the $%
t-J$ model (or generalized $t-J$ model) are also summarized in Refs \[27\], \[4\].
The goal of the present paper is just to show that we can take another, nonperturbative way. In the first step we divide the crystal lattice into a set of interacting dimers. In other words, we can rewrite the Hubbard Hamiltonian (1) for the sc lattice (see Fig. 1) in the equivalent form [^2]:
$$\begin{array}{ll}
H= & \sum\limits_IH_I^d-t\sum\limits_{I,\sigma }(c_{I,2,\sigma
}^{+}c_{I+1,1,\sigma }+c_{I+1,1,\sigma }^{+}c_{I,2,\sigma }) \\
& -t\sum\limits_{I\neq J,\sigma }(c_{I,1,\sigma }^{+}c_{J,1,\sigma
}+c_{I,2,\sigma }^{+}c_{J,2,\sigma })
\end{array}$$
where
$$H_I^d=-t\sum\limits_{\sigma} (c_{I,1,\sigma }^{+}c_{I,2,\sigma
}+c_{I,2,\sigma }^{+}c_{I,1,\sigma })+U(n_{I,1,\uparrow }n_{I,1,\downarrow
}+n_{I,2,\uparrow }n_{I,2,\downarrow }).$$
The indices $I$ and $J$ enumerate the dimers and $H_I^d$ is the dimer Hamiltonian. The second term in (4) describes the hopping between nearest dimers in the $z$-direction, the third one represents the hopping between nearest dimers (($y,z$)-plane) and between different dimer planes (see Fig. 1).
The equivalent form of the Hubbard Hamiltonian (4) is especially suitable because we can apply in the following the exact solution of the dimer problem (5). In the next step we express the construction operators $%
a_{\sigma}$ ($a_{\sigma}^{+}$) as linear combinations of the transition operators between different dimer states. This, in turn, allows to find the exact dimer representation of the whole Hamiltonian (4). The space of the dimer eigenvectors consists, however, of two subspaces. One of them corresponds to the lowest lying energy levels, the second one contains the levels with energies which in the large $U$ limit take on large, positive values. These levels in a reasonable temperature range cannot be occupied by electrons and therefore we can exclude them from further considerations. It is interesting to note that this approch, without any additional assumptions, applied to the dimer Hamiltonian (5) produces the analogy of the well known $t-J$ model (cf Refs \[31\], \[32\]). A similar approach can be applied to the Hamiltonian (4), describing the crystal as a whole. By eliminating from the considerations the unoccupied dimer energy levels in the large $U$ limit (it is the only assumption) with the use of the projection technique we can find in a straightforward way the final form of the Hamiltonian in this limit without using perturbation expansion or canonical transformation. The resulting Hamiltonian, obtained in this way, is very complicated. It, however, explicitly shows all possible magnetic, nonmagnetic or more complex competitive interaction processes very deeply hidden in the original Hamiltonian written in the site representation (1). It is the aim of this paper just to reveal these important but normally invisible elementary interactions. One important advantage might be that a given approach to the unsolvable Hubbard problem can be tested with respect to the types of neglected interaction processes. Besides, the new and straightforward method, presented in this paper, can easily be adopted to clusters of any size and also to another types of model Hamiltonians.
The paper is organized as follows. In Sec. 2 we find the exact solution of the Hubbard dimer (5) and give the exact expressions for the annihilation operators $c_{I,1(2),\sigma }$ in the dimer representation. In Sec. 3 we derive the Hubbard dimer Hamilonian (5) in the large $U$ limit. With the use of the projection technique onto the lowest lying dimer states we derive in Sec. 4 the Hubbard Hamiltonian for the crystal in this limit (central formula of this paper). A simplified version of the derived Hamiltonian with the use of the Taylor expansion with respect to $x=\frac{t}{U}$ $(x\ll 1)$ and the case $U\rightarrow \infty $ is discussed in Sec. 5.
Exact solution of the Hubbard dimer
===================================
The eigenvalues and eigenvectors of the dimer Hamiltonian (5) can be found by using a standard procedure (cf Refs \[33-35\]). We start with the vectors $%
|n_{1,\uparrow },n_{1,\downarrow };n_{2,\uparrow },n_{2,\downarrow }\rangle $ $(n_{i,\sigma }=0,1;$ $i=1,2$; $\sigma =\uparrow ,\downarrow )$ which constitute the Fock basis of the single-dimer space of states:
$$\begin{array}{lll}
\begin{array}{l}
\begin{array}{l}
|0\rangle =|0,0;0,0\rangle ,
\end{array}
\\
\\
\begin{array}{l}
|11\rangle =|1,0;0,0\rangle , \\
|12\rangle =|0,1;0,0\rangle , \\
|13\rangle =|0,0;1,0\rangle , \\
|14\rangle =|0,0;0,1\rangle ,
\end{array}
\end{array}
&
\begin{array}{l}
|21\rangle =|1,1;0,0\rangle , \\
|22\rangle =|1,0;1,0\rangle , \\
|23\rangle =|1,0;0,1\rangle , \\
|24\rangle =|0,1;1,0\rangle , \\
|25\rangle =|0,1;0,1\rangle , \\
|26\rangle =|0,0;1,1\rangle ,
\end{array}
&
\begin{array}{l}
\begin{array}{l}
|31\rangle =|0,1;1,1\rangle , \\
|32\rangle =|1,0;1,1\rangle , \\
|33\rangle =|1,1;0,1\rangle , \\
|34\rangle =|1,1;1,0\rangle ,
\end{array}
\\
\\
\begin{array}{l}
|4\rangle =|1,1;1,1\rangle .
\end{array}
\end{array}
\end{array}$$
Starting with the vectors (6) we easily get the eigenvalues $E_\alpha $ and the eigenvectors $|E_{\alpha} \rangle $ of the Hubbard dimer (5). Here, we only mention that the space of 16 eigenvectors $|E_{\alpha} \rangle $ can be devided into some subspaces numbered by $n=\sum\limits_{i,\sigma
}n_{i,\sigma }$. The subspace, belonging to $n=0$ and $n=4$ is 1-dimensional ($E_0=0,$ $|E_0\rangle =|0\rangle ;$ $E_4=2U,$ $|E_4\rangle =|4\rangle )$. There are, however, two 2-dimensional subspaces corresponding to $n=1$ (as e.g. $E_{11}=-t,$ $|E_{11}\rangle =\frac 1{\sqrt{2}}(|11\rangle +|13\rangle
),$ etc.) and two 2-dimensional subspaces corresponding to $n=3$ (as e.g. $%
E_{31}=t+U,$ $|E_{31}\rangle =\frac 1{\sqrt{2}}(|31\rangle +|33\rangle ),$ etc.). The subspace, belonging to $n=2$ consists of one 4-dimensional subspace (as e.g. $E_{21}=0,$ $|E_{21}\rangle =\frac 1{\sqrt{2}}(|23\rangle
+|24\rangle ),$ etc.) and two 1-dimensional subspaces ($E_{25}=E_{26}=0,$ $%
|E_{25}\rangle =|22\rangle ,$ $|E_{26}\rangle =|25\rangle $). The complete set of the eigenvalues $E_{\alpha} $ and eigenvectors $|E_{\alpha} \rangle $ of the Hubbard dimer is given in the Appendix A. It allows to express the dimer Hamiltonian (5) in the equivalent form
$$H^d=\sum\limits_{\alpha} E_{\alpha} |E_{\alpha} \rangle \langle E_{\alpha} |.$$
The next important step in our calculations is the possibility to express the annihilation operators $c_{1(2),\sigma }$ as linear combinations of transition operators between dimer states $P_{\alpha ,\beta }=|E_{\alpha}
\rangle \langle E_{\beta} |.$ This procedure can easily be performed when acting with the annihilation operators on the basis vectors (6) and using the reciprocal relations to (A.1). In this way we can obtain the dimer representation of the annihilation (creation) operators, given in Appedix B. After this operation we can insert so prepared $c_{I,1(2),\sigma }$ $%
(c_{I,1(2),\sigma }^{+})$ into the Hubbard Hamiltonian (4) to obtain the Hubbard model in the dimer representation for the sc lattice. This representation will be used later to derive the Hubbard Hamiltonian in the large $U$ limit (see Sec. 4).
Hubbard dimer for large $U$
===========================
Looking at the eigenvalues (A.1) of the Hubbard dimer it is easy to see that for large $U$ $(U\gg t)$ the energies $E_{\alpha \text{ }%
}=E_{22},E_{23},E_{31},E_{32},E_{33},E_{34}$ and $E_4$ take on large, positive values, producing in the partition function the terms which can practically be neglected. It means that the mentioned energies are not occupied in the reasonable temperature range ($1$ $eV\sim 11604.5$ $K$) and can be excluded from our considerations. Therefore the dimer Hamiltonian, given by (7), reduces to
$$\overline{H}^d=-tP_{11,11}+tP_{12,12}-tP_{13,13}+tP_{14,14}+\left( -C+\frac
U2\right) P_{24,24}.$$
To bring this expression into a compact (second quanization) form we introduce the Hubbard operators
$$\begin{aligned}
a_{i,\sigma } &=&c_{i,\sigma }(1-n_{i,-\sigma }), \\
b_{i,\sigma } &=&c_{i,\sigma }n_{i,-\sigma }\end{aligned}$$
and spin operators
$$\begin{aligned}
S_i^z &=&\frac 12(n_{i,\uparrow }-n_{i,\downarrow })=\frac 12(n_{i,\uparrow
}^a-n_{i,\downarrow }^a), \\
S_i^{+} &=&c_{i,\uparrow }^{+}c_{i,\downarrow }=a_{i,\uparrow
}^{+}a_{i,\downarrow }, \\
S_i^{-} &=&c_{i,\downarrow }^{+}c_{i,\uparrow }=a_{i,\downarrow
}^{+}a_{i,\uparrow }\end{aligned}$$
where $n_{i,\sigma }^a=a_{i,\sigma }^{+}a_{i,\sigma }$ $(i=1,2;$ $\sigma
=\uparrow $, $\downarrow ).$
With the use of (9)-(13) the Hamiltonian of the Hubbard dimer (8) for large $%
U$ can be presented in the form (we introduce the omitted earlier dimer index $I$)
$$\begin{aligned}
\overline{H}_I^d &=&-t\sum\limits_{\sigma} [a_{I,1,\sigma }^{+}a_{I,2,\sigma
}+a_{I,2,\sigma }^{+}a_{I,1,\sigma }] \nonumber \\
&& \nonumber \\
&&+\frac{4t^2}{U\sqrt{1+(\frac{4t}U)^2}}[\overrightarrow{S}_{I,1}\cdot
\overrightarrow{S}_{I,2}-\frac{n_{I,1}^an_{I,2}^a}4] \nonumber \\
&& \nonumber \\
&&+\frac{t^2(1-\sqrt{1+(\frac{4t}U)^2})}{U(1+\sqrt{1+(\frac{4t}U)^2})\sqrt{%
1+(\frac{4t}U)^2}}[2(b_{I,1,\uparrow }^{+}a_{I,1,\downarrow
}^{+}a_{I,2,\downarrow }b_{I,2,\uparrow } \nonumber \\
&& \nonumber \\
&&+b_{I,2,\uparrow }^{+}a_{I,2,\downarrow }^{+}a_{I,1,\downarrow
}b_{I,1,\uparrow }) \\
&& \nonumber \\
&&+(1-n_{I,1}^a)n_{I,2}^b+(1-n_{I,2}^a)n_{I,1}^b-n_{I,1}^bn_{I,2}^b]
\nonumber \\
&& \nonumber \\
&&+\frac{t(1-\sqrt{1+(\frac{4t}U})^2)}{2\sqrt{1+(\frac{4t}U)^2}}%
\sum\limits_{\sigma} \sum\limits_{\alpha =1,2}[a_{I,\alpha ,\sigma }^{+}b_{I,%
\overline{\alpha },\sigma }+b_{I,\alpha ,\sigma }^{+}a_{I,\overline{\alpha }%
,\sigma }] \nonumber\end{aligned}$$
where $n_{I,\alpha }^{a,b}=\sum\limits_{\sigma} n_{I,\alpha ,\sigma }^{a,b},$ $n_{I,\alpha ,\sigma }^b=b_{I,\alpha ,\sigma }^{+}b_{I,\alpha ,\sigma
}=n_{I,\alpha ,\sigma }n_{I,\alpha ,-\sigma }$ $(\alpha =1,2)$, $\overline{%
\alpha }=1$ when $\alpha =2$ and $\overline{\alpha }=2$ when $\alpha =1$. It is very important to stress that the formula (14) has been obtained in a nonperturbative way, starting from the exact form of the dimer Hamiltonian (see (5) or (7)), and excluding from the considerations unoccupied dimer energy levels in the large $U$ limit. Let us note that using the Taylor expansion in (14) with respect to $x=\frac{t}{U}$ and retaining the terms proportional to $\frac{t^2}U$ we obtain
$$\overline{H}_I^d=-t\sum\limits_\sigma [a_{I,1,\sigma }^{+}a_{I,2,\sigma
}+a_{I,2,\sigma }^{+}a_{I,1,\sigma }]+\frac{4t^2}U[\overrightarrow{S}%
_{I,1}\cdot \overrightarrow{S}_{I,2}-\frac{n_{I,1}^an_{I,2}^a}4].$$
The formula (15) is the well-known $t-J$ model for the Hubbard dimer where the first part in (15), similarly to (14), represents the exact form of the dimer Hamiltonian (5) in the limit $U\rightarrow \infty $.
The same result (14) can also be obtained applying a more general approach which will be used later to derive the Hubbard Hamiltonian for large $U$ in the case of a crystal. Let us note that after elimination of the unoccupied levels the subspace of the eigenvectors of the Hubbard dimer (A.1) for large $U$ consists of the following eigenvectors: $|E_0\rangle $, $|E_{11}\rangle $, $|E_{12}\rangle $, $|E_{13}\rangle $, $|E_{14}\rangle $, $|E_{21}\rangle $, $|E_{24}\rangle $, $|E_{25}\rangle $, and $|E_{26}\rangle $. It means that we can define a projection operator onto this subspace
$$\begin{array}{ll}
P_I= &
P_{0,0}^{(I)}+P_{11,11}^{(I)}+P_{12,12}^{(I)}+P_{13,13}^{(I)}+P_{14,14}^{(I)}
\\
& +P_{21,21}^{(I)}+P_{24,24}^{(I)}+P_{25,25}^{(I)}+P_{26,26}^{(I)}
\end{array}$$
which, in the second quantization form, reads
$$\begin{array}{ll}
P_I= & 1-\frac 12(n_{I,1}^b+n_{I,2}^b)+\frac 14n_{I,1}^bn_{I,2}^b \\
& \\
&
\begin{array}{ll}
& +\frac 18(1-\frac 1{\sqrt{1+(\frac{4t}U)^2}})[4(\overrightarrow{S}%
_{I,1}\cdot \overrightarrow{S}_{I,2}-\frac 14n_{I,1}^an_{I,2}^a) \\
& \\
& +2(b_{I,1,\uparrow }^{+}a_{I,1,\downarrow }^{+}a_{I,2,\downarrow
}b_{I,2,\uparrow }+b_{I,2,\uparrow }^{+}a_{I,2,\downarrow
}^{+}a_{I,1,\downarrow }b_{I,1,\uparrow }) \\
& \\
& +(1-n_{I,1}^a)n_{I,2}^b+(1-n_{I,2}^a)n_{I,1}^b-n_{I,1}^bn_{I,2}^b]
\end{array}
\\
& \\
& +\frac t{U\sqrt{1+(\frac{4t}U)^2}}\sum\limits_{\sigma} \sum\limits_{\alpha
=1,2}(a_{I,\alpha ,\sigma }^{+}b_{I,\overline{\alpha },\sigma }+b_{I,\alpha
,\sigma }^{+}a_{I,\overline{\alpha },\sigma }).
\end{array}$$
Now, the Hamiltonian (14) can also be obtained from (5) (or (7)) with the use of the projection operator (16) or (17) using the relation
$$\overline{H}_I^d=P_IH_I^dP_I$$
and applying a straightforward but laborious algebraic calculation.
Hubbard model for large $U$
===========================
The Hamiltonian of the whole crystal (4) can be expressed (similar to (7)) in the form $$H=\sum\limits_{\gamma} \overline{E}_{\gamma} |\overline{E}_{\gamma} \rangle
\langle \overline{E}_{\gamma} |$$ with unknown energies $\overline{E}_{\gamma} $ and eigenvectors $|\overline{E%
}_{\gamma} \rangle $. We can, however, expand the eigenvectors $|\overline{E}%
_{\gamma} \rangle $ in the series of the dimer eigenvectors (see (A.1))
$$|\overline{E}_{\gamma} \rangle =\sum\limits_{\gamma _1,..,\gamma
_M}c_{\gamma _{1,}.._{,}\gamma _M}^{\gamma} |E_{\gamma _1}\rangle
..|E_{\gamma _M}\rangle$$
assuming that the crystal consists of $M$ dimers. Using (19) and (20) we obtain
$$H=\sum\limits_{\gamma} \overline{E}_{\gamma} \sum\limits_{\gamma
_1,..,\gamma _M}\sum\limits_{\gamma _1^{^{,}},..,\gamma _M^{,}}c_{\gamma
_{1,}.._{,}\gamma _M}^{\gamma} c_{\gamma _{1,}^{,}.._{,}\gamma
_M^{,}}^{\gamma *}|E_{\gamma _1}\rangle ..|E_{\gamma _M}\rangle \langle
E_{\gamma _1^{,}}|..\langle E_{\gamma _M^{,}}|.$$
It is clear that to obtain the Hubbard Hamiltonian for large $U$ we have to project (21) onto the subspace of the lowest lying dimer states with the use of the projection operator
$$P=P_1P_2...P_M$$
where $P_{I}$ is given by (16) or (17). In analogy to (14) we denote the Hubbard Hamiltonian in the large $U$ limit by $\overline{H}$. Similar to (18) we write
$$\overline{H}=PHP$$
and instead of the form (21) for the Hamiltonian $H$ we can use (4). Taking into account that $P^2=P$ $(P_I^2=P_I,$ $\left[ P_I,P_J\right] =0)$ we obtain
$$\begin{array}{ll}
\overline{H}= & P[\sum\limits_I\overline{H}_I^d-t\sum\limits_{I,\sigma
}\left( \overline{c}_{I,2,\sigma }^{+}\overline{c}_{I+1,1,\sigma }+\overline{%
c}_{I+1,1,\sigma }^{+}\overline{c}_{I,2,\sigma }\right) \\
& -t\sum\limits_{I\neq J,\sigma }\left( \overline{c}_{I,1,\sigma }^{+}%
\overline{c}_{J,1\sigma }+\overline{c}_{I,2,\sigma }^{+}\overline{c}%
_{J,2,\sigma }\right) ]\equiv P\overline{\overline{H}}
\end{array}$$
where $\overline{H}_I^d$ is given by (14) and $(\alpha =1,2)$
$$\overline{c}_{I,\alpha ,\sigma }=c_{I,\alpha ,\sigma }P_I,$$
$$\overline{c}_{I,\alpha ,\sigma }^{+}=P_Ic_{I,\alpha ,\sigma }^{+}.$$
Applying the projection operator (16) or (17) to (B.1) - (B.4) and introducing Hubbard- and spin operators (9) - (13) we obtain
$$\overline{c}_{I,1,\uparrow }=\underline{\underline{a}}_{I,1,\uparrow }+\beta
\left[ S_{I,2}^za_{I,1,\uparrow }+S_{I,2}^{-}a_{I,1,\downarrow }\right]
-\delta \left[ S_{I,1}^za_{I,2,\uparrow }+S_{I,1}^{-}a_{I,2,\downarrow
}\right] ,$$
$$\overline{c}_{I,1,\downarrow }=\underline{\underline{a}}_{I,1,\downarrow
}-\beta \left[ S_{I,2}^za_{I,1,\downarrow }-S_{I,2}^{+}a_{I,1,\uparrow
}\right] +\delta \left[ S_{I,1}^za_{I,2,\downarrow
}-S_{I,1}^{+}a_{I,2,\uparrow }\right]$$
where the corresponding expressions for $\overline{c}_{I,2,\sigma }$ ($%
\sigma =\uparrow ,\downarrow $) can easily be obtained by changing the internal dimer index $1\Leftrightarrow 2$ in (27) and (28).
The new operators $\underline{\underline{a}}_{I,\alpha ,\sigma }\left(
\alpha =1,2;\sigma =\uparrow ,\downarrow \right) $ are introduced to obtain a relatively compact form of (27) and (28). They are defined as follows
$$\begin{array}{ll}
\underline{\underline{a}}_{I,\alpha ,\sigma }= & \underline{a}_{I,\alpha
,\sigma }+\beta \left( \underline{b}_{I,\alpha ,\sigma }+a_{I,\alpha
,-\sigma }^{+}a_{I,\overline{\alpha },-\sigma }b_{I,\overline{\alpha }%
,\sigma }\right) \\
& +\delta \left[ \underline{b}_{I,\overline{\alpha },\sigma }+a_{I,\overline{%
\alpha },-\sigma }^{+}a_{I,\alpha ,-\sigma }b_{I,\alpha ,\sigma }+a_{I,%
\overline{\alpha },\sigma }\frac{n_{I,\alpha }^a}2\right]
\end{array}$$
where
$$\underline{a}_{I,\alpha ,\sigma }=a_{I,\alpha ,\sigma }\left( 1-\frac \beta
2n_{I,\overline{\alpha }}^a-\frac 12n_{I,\overline{\alpha }}^b\right) ,$$
$$\underline{b}_{I,\alpha ,\sigma }=b_{I,\alpha ,\sigma }\left( 1-n_{I,%
\overline{\alpha }}^a-\frac 12n_{I,\overline{\alpha }}^b\right)$$
and
$$\beta =\frac 14\left( 1-\frac 1{\sqrt{1+\left( \frac{4t}U\right) ^2}}\right)
,$$
$$\delta =\frac{t}{U}\frac 1{\sqrt{1+\left( \frac{4t}U\right) ^2}}.$$
To write down the explicit form of the Hubbard Hamiltonian $\overline{%
\overline{H}}$ (see (24)) we have to insert (27) and (28) into $\overline{%
\overline{H}}$. This operation leads, however, to a very complicated form of $\overline{\overline{H}},$ given in the Appendix C (a simplified form of this Hamiltonian is discussed in the next Section). Here again the only assumption made to derive $\overline{\overline{H}}$ was the reduction of the whole dimer space (A.1) to the subspace of the dimer eigenvectors $\left|
E_0\right\rangle $, $\left| E_{11}\right\rangle $, $\left|
E_{12}\right\rangle $, $\left| E_{13}\right\rangle $, $\left|
E_{14}\right\rangle $, $\left| E_{21}\right\rangle $, $\left|
E_{24}\right\rangle $, $\left| E_{25}\right\rangle $ and $\left|
E_{26}\right\rangle $, corresponding to the lowest lying dimer energy levels because in the large $U$ limit only these levels can be occupied. The Hamiltonian $\overline{\overline{H}}$, obtained in this way, contains many competing, magnetic, nonmagnetic and more complex interactions. Among them we can find a direct antiferromagnetic interaction generated by the term $%
\stackrel{\rightarrow }{S}_{I,1}\cdot \stackrel{\rightarrow }{S}_{I,2}$ (Heisenberg exchange interaction) multiplied by the positive coupling constant. Such a term appears in $\overline{H}_I^d$ (see (C.6) and (14)). Inside of $\overline{\overline{H}}$ (see (C.6) and (C.2) - (C.5)) a kind of ferromagnetic interactions between spins from different dimers, represented by (C.3), appears within the terms proportional to $\beta ^2$ and $\delta ^2$ (negative coupling constants). The antiferromagnetic interactions, however, appear again in terms proportional to $\beta \delta $. There are also many other magnetic, more exotic interactions, represented by (C.2), (C.4) and (C.5), entering into (C.6). The situation is, however, much more complicated when we consider the total Hamiltonian $\overline{H}$ (see (24)) in the large $U$ limit. $\overline{H}$ differs from $\overline{\overline{H}}$ by the multiplicative factor $P$ (a product of the projection operators $P_I$ (see (22) and (17)), standing on the left. Inside of each $P_I$ the mentioned antiferromagnetic interaction also appears (see (17)). In other words, the total Hamiltonian $\overline{H}$, we are interested in, is actually a sum of the products of many competitive ferromagnetic, antiferromagnetic and more exotic interactions. The thermodynamic properties of the system, described by the Hamiltoniam $\overline{H}$ (24), are then a result of the competition between all of them. Which interaction wins in such a competition it certainly depends on temperature, model parameters ($%
t,U$) and on the average number of electrons per lattice site, determining the chemical potential of the system.
The formalism presented in this paper is also applicable to a more complicated decomposition of the Hubbard Hamiltonian (1) into a set od interacting clusters consisting e.g. of one central atom and $z$ its nearest neighbours. We, however, know (see e.g. Refs \[36\]-\[39\] and papers cited therein) that, unfortunately, the mathematical problems in this case exponentially grows up with the size of the cluster.
Taylor expansion
================
The complicated form of the Hubbard Hamiltonian $\overline{H}$ in the large $%
U$ limit (see (24), (C.6)) where $P$ is given by (22) (see also (17)) can essentially be reduced when applying the Taylor expansion with respect to the parameter $x=\frac{t}{U}\ll 1$. To do it we have to expand all the coefficients in (14), (17) and (C.6) including also $\beta $ and $\delta $ (see (32), (33)). Such an expansion can be performed to any power of $x$, however, the most simple form we obtain when we restrict ourselves to the linear approximation, resulting in the terms proportional to $t$ and $\frac{%
t^2}U.$ The accuracy of this expansion can easily be verified assumming e.g. a typical value of the ratio $\frac{W}{U}=\frac 15$ (or less). Because the bandwidth of the conduction band for the sc lattice is $W=12t,$ it results in a small value of the expansion parameter $x=\frac tU=\frac 1{60}$ in this case. It, however, means that the linear approximation is quite reasonable because all higher terms in the expansion, proportional to $x^n$ ($%
n=2,3,...),$ produce $60$ times smaller contribution.
To present the results of the Taylor expansion including all the terms proportional to $t$ and $\frac{t^2}U$ let us first define several auxiliary quantities
$$P^{(1)}=\prod\limits_{I=1}^MP_I^{(1)},$$
$$P^{(2)}=\frac tU\sum\limits_{I=1}^MP_1^{(1)}\cdot ...\cdot
P_{I-1}^{(1)}P_I^{(2)}P_{I+1}^{(1)}\cdot ...\cdot P_M^{(1)}$$
where
$$P_I^{(1)}=1-\frac 12\left( n_{I,1}^b+n_{I,2}^b\right) +\frac
14n_{I,1}^bn_{I,2}^b,$$
$$P_I^{(2)}=\sum\limits_{\sigma} \sum\limits_{\alpha =1}^2\left( a_{I,\alpha
,\sigma }^{+}b_{I,\overline{\alpha },\sigma }+b_{I,\alpha ,\sigma }^{+}a_{I,%
\overline{\alpha },\sigma }\right)$$
and ($\alpha =1,$ $2$)
$$\widetilde{a}_{I,\alpha ,\sigma }=a_{I,\alpha ,\sigma }\left( 1-\frac 12n_{I,%
\overline{\alpha }}^b\right) ,$$
$$\widetilde{\widetilde{a}}_{I,\alpha ,\sigma }=\underline{b}_{I,\overline{%
\alpha },\sigma }+a_{I,\overline{\alpha },-\sigma }^{+}a_{I,\alpha ,-\sigma
}b_{I,\alpha ,\sigma }+a_{I,\overline{\alpha },\sigma }\frac{n_{I,\alpha }^a}%
2$$
where $\underline{b}_{I,\alpha ,\sigma }$ is given by (31).
The Hamiltonian $\overline{\overline{H}}$ (C.6), including all the terms proportional to $t$ and $\frac{t^2}U$, takes on the form
$$\overline{\overline{H}}=\overline{\overline{H}}^{(1)}+\overline{\overline{H}}%
^{(2)}$$
where $$\begin{aligned}
\overline{\overline{H}}^{(1)} &=&-t\sum\limits_{I,\sigma }\left[
a_{I,1,\sigma }^{+}a_{I,2,\sigma }+a_{I,2,\sigma }^{+}a_{I,1,\sigma }\right]
\nonumber \\[0.2cm]
&&-t\sum\limits_{I,\sigma }\left[ \widetilde{a}_{I,2,\sigma }^{+}\widetilde{%
a}_{I+1,1,\sigma }+\widetilde{a}_{I+1,1,\sigma }^{+}\widetilde{a}%
_{I,2,\sigma }\right] \\[0.2cm]
&&-t\sum\limits_{I\neq J,\sigma }\sum\limits_{\alpha =1}^2\widetilde{a}%
_{I,\alpha ,\sigma }^{+}\widetilde{a}_{J,\alpha ,\sigma } \nonumber\end{aligned}$$
and $$\begin{aligned}
\overline{\overline{H}}^{(2)} &=&\frac{4t^2}U\sum\limits_I[\overrightarrow{S%
}_{I,1}\cdot \overrightarrow{S}_{I,2}-\frac 14n_{I,1}^an_{I,2}^a] \nonumber
\\
&&-\frac{t^2}U\sum\limits_{I,\sigma }[\widetilde{a}_{I,2,\sigma }^{+}%
\widetilde{\widetilde{a}}_{I+1,1,\sigma }+\widetilde{\widetilde{a}}%
_{I,2,\sigma }^{+}\widetilde{a}_{I+1,1,\sigma } \nonumber \\
&&+\widetilde{a}_{I+1,1,\sigma }^{+}\widetilde{\widetilde{a}}_{I,2,\sigma }+%
\widetilde{\widetilde{a}}_{I+1,1,\sigma }^{+}\widetilde{a}_{I,2,\sigma }]
\nonumber \\
&&-\frac{t^2}U\sum\limits_{I\neq J,\sigma }\sum\limits_{\alpha =1}^2[%
\widetilde{a}_{I,\alpha ,\sigma }^{+}\widetilde{\widetilde{a}}_{J,\alpha
,\sigma }+\widetilde{\widetilde{a}}_{I,\alpha ,\sigma }^{+}\widetilde{a}%
_{J,\alpha ,\sigma }] \\
&&+\frac{2t^2}U\sum\limits_I[\overrightarrow{S}_{I,2}\cdot (\overrightarrow{%
\underline{\underline{s}}}_{I,1;I+1,1}+\overrightarrow{\underline{s}}%
_{I+1,1;I,1}) \nonumber \\
&&+\overrightarrow{S}_{I+1,1}\cdot (\overrightarrow{\underline{\underline{s}}%
}_{I+1,2;I,2}+\overrightarrow{\underline{s}}_{I,2;I+1,2})] \nonumber \\
&&+\frac{2t^2}U\sum\limits_{I\neq J}[\overrightarrow{S}_{I,1}\cdot (%
\overrightarrow{\underline{\underline{s}}}_{I,2;J,1}+\overrightarrow{%
\underline{s}}_{J,1;I,2}) \nonumber \\
&&+\overrightarrow{S}_{I,2}\cdot (\overrightarrow{\underline{\underline{s}}}%
_{I,1;J,2}+\overrightarrow{\underline{s}}_{J,2;I,1})]. \nonumber\end{aligned}$$
The operators $\underline{\underline{s}}_{I,\mu ;J,\nu }^{z,\pm }$ and $%
\underline{s}_{I,\mu ;J,\nu }^{z,\pm }$ in (42) retain their forms introduced in (C.1) but $\underline{\underline{a}}_{I,\mu ,\sigma }(%
\underline{\underline{a}}_{I,\mu ,\sigma }^{+})$ in (C.1) should actually be replaced by $\widetilde{a}_{I,\mu ,\sigma }(\widetilde{a}_{I,\mu ,\sigma
}^{+})$, defined by (38). The total Hamiltonian $\overline{H}$ in the large $%
U$ limit (24) including all terms proportional to $t$ and $\frac{t^2}U$ can thus be written in the form
$$\overline{H}=P^{(1)}(\overline{\overline{H}}^{(1)}+\overline{\overline{H}}%
^{(2)})+P^{(2)}\overline{\overline{H}}^{(1)}=P^{(1)}\overline{\overline{H}}%
^{(1)}+(P^{(1)}\overline{\overline{H}}^{(2)}+P^{(2)}\overline{\overline{H}}%
^{(1)}).$$
The first part, $P^{(1)}\overline{\overline{H}}^{(1)}$, contains the terms proportional to $t$ whereas the term in the parentheses is proportional to $%
\frac{t^2}U$ (see (35) and (42)). In the expression for $\overline{\overline{%
H}}^{(2)}$(see (42)) the first term describes the antiferromagnetic, Heisenberg intradimer interaction (see also the second term in (15)). The apearence of this term may suggest that such an interaction should also arise between different dimers. This is of course the case. However, because of applied procedure such terms do not appear explicitly. In the last analysis our method treats all interactions within and between dimers, on the same quality level, i.e. all interactions are taken into account. The magnetic interdimer interactions, represented by the fourth and fifth term in (42), have formally the same structure as the Heisenberg interactions but instead of the scalar products of spin operators there are the products of spin operators and ”hopping spin” operators (defined in C.1). All the terms presented in (42) are correct because they originate from the exact decomposition of the Hubbard Hamiltonian (1) into a set of interacting dimers (4) where each dimer problem has been exactly solved (exact dimer representation of the construction operators after applying the projection procedure, given by (27) and (28)). The total Hamiltonian $\overline{H}$ (43), we are interested in, is much more complicated than $\overline{%
\overline{H}}^{(1)}$ and $\overline{\overline{H}}^{(2)}$alone (see (41)-(43)) because of the presence of the projection operators $P^{(1)}$ and $P^{(2)}$ (see (34)-(37)) in the expression (43).
It is also interesting to see what happens in the special case when taking the limit $U\rightarrow \infty $. The second term in the parentheses of Eq.(43) vanishes in this limit (cf (35) and (42)). Besides, each lattice site cannot be at the same time occupied by two electrons what is equivalent to the assumption ($\alpha =1,2)$
$$n_{I,\alpha ,\sigma }^b=b_{I,\alpha ,\sigma }^{+}b_{I,\alpha ,\sigma
}=n_{I,\alpha ,\sigma }n_{I,\alpha ,-\sigma }=0,$$
$$n_{I,\alpha }^b=\sum\limits_{\sigma} n_{I,\alpha ,\sigma }^b=0$$
and (cf (38))
$$\widetilde{a}_{I,\alpha ,\sigma }=a_{I,\alpha ,\sigma }.$$
The total Hamiltonian (43) is thus given by a simple formula
$$\begin{array}{ll}
\overline{H}=P^{(1)}\overline{\overline{H}}= & -t\sum\limits_{I,\sigma
}\left[ a_{I,1,\sigma }^{+}a_{I,2,\sigma }+a_{I,2,\sigma }^{+}a_{I,1,\sigma
}\right] \\
& \\
& -t\sum\limits_{I,\sigma }\left[ a_{I,2,\sigma }^{+}a_{I+1,1,\sigma
}+a_{I+1,1,\sigma }^{+}a_{I,2,\sigma }\right] \\
& \\
& -t\sum\limits_{I\neq J,\sigma }\sum\limits_{\alpha =1}^2a_{I,\alpha
,\sigma }^{+}a_{J,\alpha ,\sigma }.
\end{array}$$
Going back to the original lattice (cf (4),(5) and (1)) it is easy to see that the exact Hubbard Hamiltonian in the limit $U\rightarrow \infty $ takes on the form
$$\overline{H}=-t\sum\limits_{i,j,\sigma }a_{i,\sigma }^{+}a_{j,\sigma }$$
where $i$ and $j$ (as before) number the lattice points and $a_{i,\sigma }$ $%
(a_{i,\sigma }^{+})$ are the Hubbard operators, defined by (9).
Conclusions
===========
Using a new, nonperturbative approach, basing on the equivalent form of the Hubbard Hamiltonian, represented by the collection of interacting dimers (4) where each dimer problem has been exactly solved, we have expressed the annihilation (creation) operators (B.1)-(B.4) as linear combinations of transition operators between different dimer states (dimer representation). This method made it possible to exclude from the considerations the unoccupied dimer energy levels in the large $U$ limit by means of the projection technique resulting in the final Hamiltonian for the dimer itself (14), as well as, for the crystal as a whole (see (23), (24) and (C.6)). It is important to stress that the elimination of the unoccupied dimer energy levels was the only assumption to derive the Hubbard Hamiltonian (24) in the large $U$ limit. Therefore we can be sure that expanding (24) with respect to $x=\frac tU$ (Taylor expansion) we obtain absolutely all terms proportional to $x^n$ $(n=1,2,...).$ In other words all the coefficients proportional to $x^n$ are easy to control what is not always the case when using another methods. The final form of the obtained Hamiltonian in the large $U$ limit (see (23), (24) and (C.6)) visualizes high degree of complexity of the model, deeply hidden in its original version, forming a mixture of many multiple ferromagnetic, antiferromagnetic and more complex interactions competing one with another. This fact seems to be decisive for our final conclusion. Because we are still dealing with the approximate solutions of the model (the exact solution does not exist till now) it may happen that we underestimate in this way some important interactions and overestimate the others. It is the reason why the resulting thermodynamic properties of the Hubbard model, obtained in an approximate way, so strongly depend on the quality of applied approximations.
We note in passing that the exact dimer solution may serve as a novel alloy analogy for the Hubbard model, which could be treated by coherent potential approximation (CPA). It is well-known that the standard alloy analogy, based on the atomic limit, does not allow for ferromagnetism in the Hubbard model. This may change by application of the dimer solution which already accounts for a restricted hopping of the band electrons. A corresponding study is in preparation.
Another example of a dimer approach to Hubbard-like models is the bond operator theory as an extension of the slave bosonic and fermionic operators (see Refs \[40\], \[41\] and papers cited therein).
[Appendix A]{}
The exact solution of the dimer eigenvalue problem (5) reads:
$$\begin{array}{ll}
E_0=0; & |E_0\rangle =|0\rangle ,
\end{array}$$
$$\begin{array}{ll}
E_{11}=-t; & |E_{11}\rangle =\frac{1}{\sqrt{2}}(|11\rangle +|13\rangle ), \\
E_{12}=t; & |E_{12}\rangle =\frac{1}{\sqrt{2}}(|11\rangle -|13\rangle ), \\
E_{13}=-t; & |E_{13}\rangle =\frac{1}{\sqrt{2}}(|12\rangle +|14\rangle ), \\
E_{14}=t; & |E_{14}\rangle =\frac{1}{\sqrt{2}}(|12\rangle -|14\rangle ),
\end{array}$$
$$\begin{array}{ll}
E_{21}=0; & |E_{21}\rangle =\frac 1{\sqrt{2}}(|23\rangle +|24\rangle ), \\
E_{22}=U; & |E_{22}\rangle =\frac 1{\sqrt{2}}(|21\rangle -|26\rangle ), \\
E_{23}=C+\frac U2; & |E_{23}\rangle =a_1(|21\rangle +|26\rangle
)-a_2(|23\rangle -|24\rangle ), \\
E_{24}=-C+\frac U2; & |E_{24}\rangle =a_2(|21\rangle +|26\rangle
)+a_1(|23\rangle -|24\rangle ), \\
E_{25}=0; & |E_{25}\rangle =|22\rangle , \\
E_{26}=0; & |E_{26}\rangle =|25\rangle ,
\end{array}
\tag{A.1}$$
$$\begin{array}{ll}
E_{31}=t+U; & |E_{31}\rangle =\frac 1{\sqrt{2}}(|31\rangle +|33\rangle ), \\
E_{32}=-t+U; & |E_{32}\rangle =\frac 1{\sqrt{2}}(|31\rangle -|33\rangle ),
\\
E_{33}=t+U; & |E_{33}\rangle =\frac 1{\sqrt{2}}(|32\rangle +|34\rangle ), \\
E_{34}=-t+U; & |E_{34}\rangle =\frac 1{\sqrt{2}}(|32\rangle -|34\rangle ),
\end{array}$$
$$\begin{array}{ll}
E_4=2U; & |E_4\rangle =|4\rangle
\end{array}$$ where
$$C=\sqrt{{\left( \frac U2\right) }^2+4t^2}, \tag{A.2}$$
$$a_1=\frac 12\sqrt{1+\frac U{2C}}, \tag{A.3}$$
$$a_2=\frac 12\sqrt{1-\frac U{2C}}. \tag{A.4}$$
[Appendix B]{}
The exact dimer representation of the construction operators is given by the following expressions:
$$\begin{array}[t]{ll}
c_{1,\uparrow }= & \frac 1{\sqrt{2}}\left( P_{0,11}+P_{0,12}\right) +\frac 1{%
\sqrt{2}}P_{11,25}-\frac 1{\sqrt{2}}P_{12,25} \\
& +\frac 12\left( P_{13,21}+P_{13,22}\right) +\frac 1{\sqrt{2}}\left(
bP_{13,23}+aP_{13,24}\right) \\
& -\frac 12\left( P_{14,21}-P_{14,22}\right) +\frac 1{\sqrt{2}}\left(
aP_{14,23}-bP_{14,24}\right) \\
& +\frac 12\left( P_{21,33}-P_{21,34}\right) -\frac 12\left(
P_{22,33}+P_{22,34}\right) \\
& +\frac 1{\sqrt{2}}\left( aP_{23,33}+bP_{23,34}\right) -\frac 1{\sqrt{2}%
}\left( bP_{24,33}-aP_{24,34}\right) \\
& +\frac 1{\sqrt{2}}\left( P_{26,31}-P_{26,32}\right) +\frac 1{\sqrt{2}%
}P_{31,4}+\frac 1{\sqrt{2}}P_{32,4},
\end{array}
\tag{B.1}$$
$$\begin{array}[t]{ll}
c_{1,\downarrow }= & \frac 1{\sqrt{2}}\left( P_{0,13}+P_{0,14}\right) +\frac
1{\sqrt{2}}P_{13,26}-\frac 1{\sqrt{2}}P_{14,26} \\
& +\frac 12\left( P_{11,21}-P_{11,22}\right) -\frac 1{\sqrt{2}}\left(
bP_{11,23}+aP_{11,24}\right) \\
& -\frac 12\left( P_{12,21}+P_{12,22}\right) -\frac 1{\sqrt{2}}\left(
aP_{12,23}-bP_{12,24}\right) \\
& -\frac 12\left( P_{21,31}-P_{21,32}\right) -\frac 12\left(
P_{22,31}+P_{22,32}\right) \\
& +\frac 1{\sqrt{2}}\left( aP_{23,31}+bP_{23,32}\right) -\frac 1{\sqrt{2}%
}\left( bP_{24,31}-aP_{24,32}\right) \\
& -\frac 1{\sqrt{2}}\left( P_{25,33}-P_{25,34}\right) -\frac 1{\sqrt{2}%
}P_{33,4}-\frac 1{\sqrt{2}}P_{34,4},
\end{array}
\tag{B.2}$$
$$\begin{array}[t]{ll}
c_{2,\uparrow }= & \frac 1{\sqrt{2}}\left( P_{0,11}-P_{0,12}\right) -\frac 1{%
\sqrt{2}}P_{11,25}-\frac 1{\sqrt{2}}P_{12,25} \\
& -\frac 12\left( P_{13,21}+P_{13,22}\right) +\frac 1{\sqrt{2}}\left(
bP_{13,23}+aP_{13,24}\right) \\
& -\frac 12\left( P_{14,21}-P_{14,22}\right) -\frac 1{\sqrt{2}}\left(
aP_{14,23}-bP_{14,24}\right) \\
& -\frac 12\left( P_{21,33}+P_{21,34}\right) +\frac 12\left(
P_{22,33}-P_{22,34}\right) \\
& +\frac 1{\sqrt{2}}\left( aP_{23,33}-bP_{23,34}\right) -\frac 1{\sqrt{2}%
}\left( bP_{24,33}+aP_{24,34}\right) \\
& -\frac 1{\sqrt{2}}\left( P_{26,31}+P_{26,32}\right) +\frac 1{\sqrt{2}%
}P_{31,4}-\frac 1{\sqrt{2}}P_{32,4},
\end{array}
\tag{B.3}$$
$$\begin{array}[t]{ll}
c_{2,\downarrow }= & \frac 1{\sqrt{2}}\left( P_{0,13}-P_{0,14}\right) -\frac
1{\sqrt{2}}P_{13,26}-\frac 1{\sqrt{2}}P_{14,26} \\
& -\frac 12\left( P_{11,21}-P_{11,22}\right) -\frac 1{\sqrt{2}}\left(
bP_{11,23}+aP_{11,24}\right) \\
& -\frac 12\left( P_{12,21}+P_{12,22}\right) +\frac 1{\sqrt{2}}\left(
aP_{12,23}-bP_{12,24}\right) \\
& +\frac 12\left( P_{21,31}+P_{21,32}\right) +\frac 12\left(
P_{22,31}-P_{22,32}\right) \\
& +\frac 1{\sqrt{2}}\left( aP_{23,31}-bP_{23,32}\right) -\frac 1{\sqrt{2}%
}\left( bP_{24,31}+aP_{24,32}\right) \\
& +\frac 1{\sqrt{2}}\left( P_{25,33}+P_{25,34}\right) -\frac 1{\sqrt{2}%
}P_{33,4}+\frac 1{\sqrt{2}}P_{34,4}
\end{array}
\tag{B.4}$$
where
$$P_{\alpha ,\beta }=|E_{\alpha} \rangle \langle E_{\beta} | \tag{B.5}$$
and
$$\begin{array}{l}
a=a_1+a_2 \\
b=a_1-a_2.
\end{array}
\tag{B.6}$$
To obtain the annihilation operators of the $I-$th dimer, the index $I$ should be added $(c_{1(2),\sigma }\longrightarrow c_{I,1(2),\sigma }$, $%
P_{\alpha ,\beta }\rightarrow P_{\alpha ,\beta }^{(I)})$ in (B.1)-(B.4).
[Appendix C]{}
To present the Hamiltonian $\overline{\overline{H}}$ in a compact form we first introduce the following operators $\left( \mu ,\nu =1,2\right) $ :
$$\begin{array}{ll}
s_{I,\mu ;J,\nu }^{+}= & a_{I,\mu ,\uparrow }^{+}a_{J,\nu ,\downarrow }, \\
& \\
s_{I,\mu ;J,\nu }^{-}= & a_{J,\nu ,\downarrow }^{+}a_{I,\mu ,\uparrow }, \\
& \\
n_{I,\mu ;J,\nu ;\sigma }= & a_{I,\mu ,\sigma }^{+}a_{J,\nu ,\sigma }, \\
& \\
n_{I,\mu ;J,\nu }= & \sum\limits_{\sigma} n_{I,\mu ;J,\nu ;\sigma }, \\
& \\
s_{I,\mu ;J,\nu }^z= & \frac 12\left( n_{I,\mu ;J,\nu ;\uparrow }-n_{I,\mu
;J,\nu ;\downarrow }\right) ; \\
&
\end{array}$$
$$\begin{array}{ll}
\underline{s}_{I,\mu ;J,\nu }^{+}= & \underline{\underline{a}}_{I,\mu
,\uparrow }^{+}a_{J,\nu ,\downarrow }, \\
& \\
\underline{s}_{I,\mu ;J,\nu }^{-}= & a_{J,\nu ,\downarrow }^{+}\underline{%
\underline{a}}_{I,\mu ,\uparrow }, \\
& \\
\underline{n}_{I,\mu ;J,\nu ;\sigma }= & \underline{\underline{a}}_{I,\mu
,\sigma }^{+}a_{J,\nu ,\sigma }, \\
& \\
\underline{n}_{I,\mu ;J,\nu }= & \sum\limits_{\sigma} \underline{n}_{I,\mu
;J,\nu ;\sigma }, \\
& \\
\underline{s}_{I,\mu ;J,\nu }^z= & \frac 12\left( \underline{n}_{I,\mu
;J,\nu ;\uparrow }-\underline{n}_{I,\mu ;J,\nu ;\downarrow }\right) ; \\
&
\end{array}
\tag{C.1}$$
$$\begin{array}{ll}
\underline{\underline{s}}_{I,\mu ;J,\nu }^{+}= & a_{I,\mu ,\uparrow }^{+}%
\underline{\underline{a}}_{J,\nu ,\downarrow }, \\
& \\
\underline{\underline{s}}_{I,\mu ;J,\nu }^{-}= & \underline{\underline{a}}%
_{J,\nu ,\downarrow }^{+}a_{I,\mu ,\uparrow }, \\
& \\
\underline{\underline{n}}_{I,\mu ;J,\nu ;\sigma }= & a_{I,\mu ,\sigma }^{+}%
\underline{\underline{a}}_{J,\nu ,\sigma }, \\
& \\
\underline{\underline{n}}_{I,\mu ;J,\nu }= & \sum\limits_{\sigma}
\underline{\underline{n}}_{I,\mu ;J,\nu ;\sigma }, \\
& \\
\underline{\underline{s}}_{I,\mu ;J,\nu }^z= & \frac 12\left( \underline{%
\underline{n}}_{I,\mu ;J,\nu ;\uparrow }-\underline{\underline{n}}_{I,\mu
;J,\nu ;\downarrow }\right) .
\end{array}$$
The operators $s_{I,\mu ;J,\nu }^{z,\pm }$, $\underline{s}_{I,\mu ;J,\nu
}^{z,\pm }$, $\underline{\underline{s}}_{I,\mu ;J,\nu }^{z,\pm }$, are not strictly the spin operators, they, however, show some similarities to the true spin operators (as e.g. (11)-(13)) and therefore they can be called ”hopping spin operators”. Moreover, the following abreviations have to be used:
$$Q\left( I,\mu ;J,\nu \right) =\overrightarrow{S}_{I,\mu }\cdot \left(
\overrightarrow{\underline{\underline{s}}}_{I,\overline{\mu };J,\nu }+%
\stackrel{\rightarrow }{\underline{s}}_{J,\nu ;I,\overline{\mu }}\right) ,
\tag{C.2}$$
$$R\left( I,\mu ;J,\nu \right) =\stackrel{\rightarrow }{S}_{I,\mu }\cdot
\stackrel{\rightarrow }{S}_{J,\nu }n_{I,\overline{\mu };J,\overline{\nu }},
\tag{C.3}$$
$$R^{z,\pm }\left( I,\mu ;J,\nu \right) =\left( S_{I,\mu }^zS_{J,\nu }^{\pm
}-S_{I,\mu }^{\pm }S_{J,\nu }^z\right) s_{I,\overline{\mu };J,\overline{\nu }%
}^{\mp }, \tag{C.4}$$
$$R^{-,+}\left( I,\mu ;J,\nu \right) =\left( S_{I,\mu }^{-}S_{J,\nu
}^{+}-S_{I,\mu }^{+}S_{J,\nu }^{-}\right) s_{I,\overline{\mu };J,\overline{%
\nu }}^z. \tag{C.5}$$
The Hamiltonian $\overline{\overline{H}}$ (cf (24)) with the use of (C.1) - (C.5) takes on the form
$$\begin{array}{ll}
\overline{\overline{H}}= & \sum\limits_I\overline{H}_I^d-t\sum\limits_{I,%
\sigma }\left[ \underline{\underline{a}}_{I,2,\sigma }^{+}\underline{%
\underline{a}}_{I+1,1,\sigma }+\underline{\underline{a}}_{I+1,1,\sigma }^{+}%
\underline{\underline{a}}_{I+1,2,\sigma }\right] \\
& \\
& -2t\beta \sum\limits_I\left[ Q\left( I,1;I+1,1\right) +Q\left(
I+1,2;I,2\right) \right] \\
& \\
& +2t\delta \sum\limits_I\left[ Q\left( I,2;I+1,1\right) +Q\left(
I+1,1;I,2\right) \right]
\end{array}$$
$$\begin{array}{ll}
-t\beta ^2\sum\limits_I & [R\left( I,1;I+1,2\right) +R^{z,-}\left(
I,1;I+1,2\right) \\
& \\
& +R^{z,+}\left( I,1;I+1,2\right) +R^{-,+}\left( I,1;I+1,2\right) \\
& \\
& +R\left( I+1,2;I,1\right) +R^{z,-}\left( I+1,2;I,1\right) \\
& \\
& +R^{z,+}\left( I+1,2;I,1\right) +R^{-,+}\left( I+1,2;I,1\right) ]
\end{array}$$
$$\begin{array}{ll}
-t\delta ^2\sum\limits_I & [R\left( I,2;I+1,1\right) +R^{z,-}\left(
I,2;I+1,1\right) \\
& \\
& +R^{z,+}\left( I,2;I+1,1\right) +R^{-,+}\left( I,2;I+1,1\right) \\
& \\
& +R\left( I+1,1;I,2\right) +R^{z,-}\left( I+1,1;I,2\right) \\
& \\
& +R^{z,+}\left( I+1,1;I,2\right) +R^{-,+}\left( I+1,1;I,2\right) ]
\end{array}$$
$$\begin{array}{ll}
+t\beta \delta \sum\limits_I\sum\limits_{\mu =1}^2 & [R\left( I,\mu
;I+1,\mu \right) +R^{z,-}\left( I,\mu ;I+1,\mu \right) \\
& \\
& +R^{z,+}\left( I,\mu ;I+1,\mu \right) +R^{-,+}\left( I,\mu ;I+1,\mu \right)
\\
& \\
& +R\left( I+1,\mu ;I,\mu \right) +R^{z,-}\left( I+1,\mu ;I,\mu \right) \\
& \\
& +R^{z,+}\left( I+1,\mu ;I,\mu \right) +R^{-,+}\left( I+1,\mu ;I,\mu
\right) ]
\end{array}
\tag{C.6}$$
$$\begin{array}{ll}
& -t\sum\limits_{I\neq J,\sigma }\sum\limits_{\mu =1}^2\underline{%
\underline{a}}_{I,\mu ,\sigma }^{+}\underline{\underline{a}}_{J,\mu ,\sigma }
\\
& \\
& -2t\beta \sum\limits_{I\neq J}\sum\limits_{\mu =1}^2Q\left( I,\mu ;J,%
\overline{\mu }\right) +2t\delta \sum\limits_{I\neq J}\sum\limits_{\mu
=1}^2Q\left( I,\mu ;J,\mu \right)
\end{array}$$
$$\begin{array}{ll}
& -t\left( \beta ^2+\delta ^2\right) \sum\limits_{I\neq J}\sum\limits_{\mu
=1}^2[R\left( I,\mu ;J,\mu \right) +R^{z,-}\left( I,\mu ;J,\mu \right) \\
& \\
& +R^{z,+}\left( I,\mu ;J,\mu \right) +R^{-,+}\left( I,\mu ;J,\mu \right) ]
\\
& \\
& +2t\beta \delta \sum\limits_{I\neq J}\sum\limits_{\mu =1}^2[R\left(
I,\mu ;J,\overline{\mu }\right) +R^{z,-}\left( I,\mu ;J,\overline{\mu }%
\right) \\
& \\
& +R^{z,+}\left( I,\mu ;J,\overline{\mu }\right) +R^{-,+}\left( I,\mu ;J,%
\overline{\mu }\right) ]
\end{array}$$ where $\overline{H}_I^d$ is given by (14). It should also be noted that the operators $\underline{\underline{a}}_{I,\alpha ,\sigma }$ $(\underline{%
\underline{a}}_{I,\alpha ,\sigma }^{+}),$ entering in (C.1) and (C.2), are in fact, dependent on $\beta $ and $\delta $ (see (29)). The decomposition of the Hamiltonian $\overline{\overline{H}}$ (C.6) according to the terms proportional to $\beta ,$ $\delta $, $\beta ^2$,$\delta ^2$ and $\beta
\delta $ has thus only a formal character in order to keep the presentation of the Hamiltonian $\overline{\overline{H}}$ in a compact form.
The formula (C.6) is the complete expression.
[99]{} Hubbard J. 1963 *Proc. Roy. Soc. London* A **276** 238 ; 1964 A **281** 401.
Herrmann T. and Nolting W. 1997 *J. Magn. Magn. Mater.* **170** 253.
Fulde P. 1995 *Electron Correlations in Molecules and Solids* (Berlin: Springer).
Gebhard F. 1997 *The Mott Metal-Insulator Transitions* (Springer Tracts in Modern Physics 137 Berlin: Springer).
Micnas R., Ranninger J. and Robaszkiewicz S. 1990 *Rev. Mod. Phys.* **62** 113.
Anderson P. W. 1961 *Phys. Rev.* **124** 41.
Nolting W. 1976 *Phys. Stat. Sol.* (b) **96** 11.
Matlak M. and Nolting W. 1984 *Z. Physik* B **55** 103.
Kohn W. 1964 *Phys. Rev.* **133** 171.
Nagaoka Y. 1966 *Phys. Rev.* **147** 392.
Roth L. 1966 *Phys. Rev.* **149** 306.
Harris A. B. and Lange R. V. 1967 *Phys. Rev.* **157** 295.
Penn D. 1968 *Phys. Lett.* A **26** 509.
Carron L. G. and Pratt G. W. 1968 *Rev. Mod. Phys.* **40** 802.
Langer W., Plischke M. and Mattis D. 1969 *Phys. Rev. Lett.* **23** 1448.
Kaplan T. A. and Bari R. A. 1970 *J. Appl. Phys.* **41** 875.
Sokoloff J. B. 1970 *Phys. Rev.* B **1** 1144.
Klein D. J. and Seitz W. A. 1973 *Phys. Rev.* B **8** 2236.
Visscher P. B. 1974 *Phys. Rev.* B **10** 943.
Ogawa T., Kanda K. and Matsubara T. 1975 *Prog. Theor. Phys.* **53** 614.
Florencio J., Jr. , and Chao K. A. 1975 *Phys. Rev. Lett.* **35** 741.
Florencio J., Jr. , and Chao K. A. 1976 *Phys. Rev.* B **14** 3121.
Chao K. A., Spałek J. and Oleś A. M. 1977 *J. Phys.* C **10** L271.
Takahashi M. 1982 *J. Phys. Soc. Jpn.* **51** 3475.
Irkhin V. Yu. and Katsnelson M. I. 1985 *J. Phys.* C **18** 4173.
Shastry B. S., Krishnamurthy H. R. and Anderson P. W. 1990 *Phys. Rev.* B **41** 2375.
Dagotto E. 1994 *Rev. Mod. Phys.* **66** 763.
Irkhin V. Yu. and Irkhin Yu. P. 1994 *Phys. Stat. Sol.* (b) **183** 9.
Eskes H. and Eder R. 1996 *Phys. Rev.* B **54** R14226.
Irkhin V. Yu. 1998 *Phys. Rev.* B **57** 13375.
Anderson P. W. 1959 *Phys. Rev.* **115** 2.
Anderson P. W. 1963 *Solid State Physics*, edited by F.Seitz and D.Turnbull (Academic Press: New York) Vol. 14, p. 99.
Cheng V. C. and Chen S. H. 1977 *Physica* B **85** 299.
Matlak M. 1980 *Phys. Stat. Sol.* (b) **99** K 97.
Lorentz B. 1983 *Phys. Stat. Sol.* (b) **119** 555.
Falicov L. M. and Victor R. H. 1984 *Phys. Rev.* B **30** 1695.
Callaway J., Chen D. P. and Tang R. 1987 *Phys Rev.* B **35** 3705.
Callaway J., Chen D. P. and Zhang Y. 1987 *Phys. Rev.* B **36** 2084.
Pastor G. M., Hirsch R. and Műhlschlegel B. 1996 *Phys. Rev.* B **53**
10382.
Sachdev S., Bhatt R. N. 1990 *Phys. Rev.* B **41,** 9323.
Park K., Sachdev S. 2001 *Phys. Rev.* B **64,** 1845510.
[^1]: e-mail:[email protected]
[^2]: The dimer Fourier transformation
$c_{I,1,\sigma }=\frac 1{\sqrt{N}}\sum\limits_{\mathbf{k}}c_{\mathbf{k,}%
\sigma }e^{i\mathbf{k\cdot R}_{I,1}},c_{I,2,\sigma }=\frac 1{\sqrt{N}%
}\sum\limits_{\mathbf{k}}c_{\mathbf{k,}\sigma }e^{i\mathbf{k\cdot R}_{I,2}}$
where $N$ is the number of lattice points, applied to (4), gives the well known result
$H=\sum\limits_{\mathbf{k,\sigma }}\varepsilon _{\mathbf{k}}n_{\mathbf{k,}%
\sigma }+\frac UN\sum\limits_{\mathbf{k,k}^{\prime },\mathbf{q}}c_{\mathbf{%
k+q,}\uparrow }^{+}c_{\mathbf{k,}\uparrow }c_{\mathbf{k}^{\prime }-\mathbf{q,%
}\downarrow }^{+}c_{\mathbf{k}^{\prime }\mathbf{,}\downarrow }$ with $%
\varepsilon _{\mathbf{k}}$ given by (3).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Reinforcement learning methods have been developed to achieve great success in training control policies in various automation tasks. However, a main challenge of the wider application of reinforcement learning in practical automation is that the training process is hard and the pretrained policy networks are hardly reusable in other similar cases. To address this problem, we propose the cascade attribute network (CAN), which utilizes its hierarchical structure to decompose a complicated control policy in terms of the requirement constraints, which we call attributes, encoded in the control tasks. We validated the effectiveness of our proposed method on two robot control scenarios with various add-on attributes. For some control tasks with more than one add-on attribute attributes, by directly assembling the attribute modules in cascade, the CAN can provide ideal control policies in a zero-shot manner.'
address:
- 'Dept. of Robotics, University of Michigan, Ann Arbor. (e-mail: [email protected]).'
- 'Dept. of Mechanical Engineering, University of California, Berkeley. (e-mail: {zhuoxu,tomizuka}@berkeley.edu)'
author:
- Haonan Chang
- Zhuo Xu
- Masayoshi Tomizuka
bibliography:
- 'ifacconf.bib'
title: 'Cascade Attribute Network: Decomposing Reinforcement Learning Control Policies using Hierarchical Neural Networks'
---
Reinforcement learning control,Deep Learning, Artificial Intelligence
INTRODUCTION
============
Reinforcement learning (RL) is an artificial intelligence approach that solves for automatic control policies, $\pi(a|s)$, that map the state space inputs $s$ to the control command outputs $a$. The training of the RL policies is through the interaction of the agent and the environment, which is often modeled using a Markov Decision Process (MDP). The RL has been successful in solving many robotics and automation problems in simulations as well as real-world scenarios ([@visuomotor])([@rainbow])([@navigation]). However, the wide application of the RL in automation is slowed down by several challenges, and one of the main challenges is the difficulty of RL policy training. The RL policy generally takes a great amount of computation power to train for complicated control tasks, and what makes it worse is that the RL policies are optimized based on certain fixed MDPs, and the knowledge encoded inside an optimized control policy are hard to transfer to other similar tasks. That is, new RL policies have to be trained from scratch even for those control tasks that are very similar to the pretrained task, which leads to a great amount of computation power waste.
Unlike the existing transfer learning approaches ([@progressive1]) ([@transferable_policy]) that are lack of interpretable transferable features, we propose to decompose the complicated control problems in terms of its consisting requirement constraints, which we call attributes. The conception attributes refer especially to global characteristics or requirements that take effect throughout the task. For example, to solve an autonomous driving problem, we can first decompose the requirements of the task into a base target reaching attribute, an add-on obstacle avoidance and a speed limit attribute. Our methodology includes training an attribute module for each of the attributes and then assembling the attribute modules together to produce the overall policy. To deal with tasks under different attributes, we propose an RL framework called the cascade attribute network (CAN). In CAN, the attribute modules are connected in the cascade series. Our method has two main intriguing advantages:
1. The decomposed attribute modules are much easier and faster to train compared with the overall control policy.
2. The attribute related knowledge is encoded in interpretable modules, which can be built up to create versatile policies that can adjust to changes in the control tasks.
Related Work and Background
===========================
Related Work
------------
There have been lots of attempts to create versatile intelligence that can adjust to changes in the tasks as well. Transfer learning ([@reinforcement_transfer]) is a key tool that makes the use of previously learned knowledge for the better or faster learning of new knowledge. Rusu et al ([@progressive1])([@progressive2]) establish a progressive network, in which newly added columns are laterally connected to previously learned columns for knowledge transfer. Drafty et al ([@transferable_policy]) and Braylan et al ([@reuse_module]) also design network architectures for knowledge transfer in MAV control and video game playing. Barreto et al ([@successor1])([@successor2]) propose a scheme based on successor features, a value function describing the dynamics of the environment and a generalized policy improvement over multiple policies. In this way, the method can provide an efficient transferring among different tasks where the reward function is changed but the dynamics of the system remains the same. As for the combinations of transfer learning and imitation learning, Ammar et al ([@unsupervised_transfer]) use unsupervised learning to map states for transfer, assuming the existence of distance function between different state spaces. Gupta et al ([@invariant_feature]) learn an invariant feature between different dimensional states and use demonstrations to increase the density of the rewards. Our recent works ([@xu2018zero])([@tang2019disturbance]) on transfer learning utilizes interpretable trajectory as transferrable feature, but the human engineered features lack the versatility to be applied in various robotic control areas. There are other methods seeking to learn a globally general policy: Curriculum learning (CL)([@curriculum]) trains a model on a sequence of cognate tasks that get more and more challenging gradually, so as to solve hard tasks that could not be learned from scratch. Florensa et al ([@reverse_curriculum]) apply reverse curriculum generation (RCL) in RL. In the early stage of the training process, the RCL initializes the agent state to be very close to the target state, making the policy very easy to train. They then gradually increase the random level of the initial state as the RL model performs better and better. Sanmit et al ([@CLMDP]) formulate the curriculum learning sequence as an MDP process, which can also be learned from experience. These results show that with a good consideration of the nature of the problem, CL can help reinforcement learning have a better convergence speed. Our policy training strategy is inspired by the idea of CL and achieved satisfying robustness for the policies. There are also researches in training modular neural networks, ([@modular_robot_task]) investigates the combinations of multiple robots and tasks, while ([@modular_subtask]) investigates the combinations of multiple sequential subtasks. Our work, along with our parallel work ([@xu2019toward]) different from those works, looks into modularization in a different dimension. We investigate the modularization of attributes, the characteristics or requirements that take effect throughout the whole task.
The Cascade Attribute Networks
==============================
Problem Formulation
-------------------
We consider an agent with a state-space $A$ performing a class of tasks built up with one base attribute and a series of add-on attributes. We model the control tasks and the attribute in a unified form using MDPs. We denote the attributes using their index $\{ 0,1,2,\ldots \}$, where the $0^{th}$ attribute is the base attribute, which usually corresponds to the most fundamental goal of the task, such as the target reaching attribute in the autonomous driving task. We define the state space of each attribute to be the minimum state space that fully characterizes the attribute, denoted $S=\{S_0, S_1, S_2, S_3\ldots \}$. For example, let the base attribute be the target reaching attribute, and the $1^{st}$ attribute be the obstacle avoidance attribute. Then $S_0$ consists of the states of the agent and the target, while $S_1$ consists of the states of the agent and the obstacle, but does not include the states of the target. Using the attribute as elements, one can build various control tasks, such as pure target reaching, target reaching while avoiding an obstacle, target reaching under external force influence, and so on.
Each attribute has an unique reward function as well, denoted $R=\{R_0,R_1,R_2,R_3\ldots\}$. Each $R_i$ is a function mapping a state action pair to a real number reward, i.e. $R_i: S_i \times A \rightarrow \mathbb{R}$.
Network Architecture
--------------------
The architecture of the CAN is shown in Fig. \[sa\]. The training for a task with more than one attribute is divided into two parts: the training of the base attribute module and the add-on attribute module. In the training phase, first, an RL policy $\pi_0(a_0|s_0)$ is trained to accomplish the goal of the base attribute. The base attribute network takes in $s_0\in S_0$ and outputs $a_0 \in A$, the reward function of the MDP is given by $R_0$. This process is a default RL training process.
Consider the control task with more than one attribute, without loss of generality, the $1^{st}$ add-on attribute module is trained next, in a series of the base attribute module. The $1^{st}$ attribute module consists of a compensation network and a weighted sum operator. The compensation network is fed with state $s_1 \in S_1$, and action $a_0$ chosen by $\pi_0$. The output of the compensation network is the compensation action $a^{c}_1$, which is used to compensate $a_0$ to produce the overall action $a_1$. The reward for the MDP is given by $R_0+R_1$ so that the requirements for both constraints of attributes are satisfied. Since the parameters of the base attribute network are pretrained and fixed, the cascading attribute network would extract the features of the attribute by exploring the new MDP under the guidance of the base policy. It is noted that in the add-on attribute module, the weight of the compensation action $a^{c}_1$ is initiated to be small and increased over the training time. That is, at the early stage of the training process, mainly $a_0$ takes effect, while $a^{c}_1$ gradually gets to influence the overall $a_1$ as the training goes on. For the other attributes, the training of their attribute networks is the same.
Empirical results also show that for cases, where the number of the add-on attributes is small, and different add-on attributes do not entangle with each other, with carefully designed training schemes, the CAN shown in Fig. \[ma\] can be used to provide an ideal control policy. In the CAN shown in Fig. \[ma\], the $i^{th}$ attribute module takes in $s_i$ and $a_{i-1}$, and outputs $a_i$ that satisfies all the attribute before the $i^{th}$ module. The final output $a_j$ is the overall output that satisfies all the constraints in the constraint array.
![The training procedure of an attribute module in the CAN: first train the base attribute module, then train the add-on attribute module based on fixed pretrained base module[]{data-label="sa"}](figure2.jpg "fig:") \[figurelabel\]
![One kind of usage of the CAN in multi-add-on-attributes tasks, by assembling add-on attribute modules in cascade to the base attribute module, the output action of the last attribute module is the outcome of the overall hierarchical policy network.[]{data-label="ma"}](figure3.jpg "fig:") \[figurelabel\]
Experiments
===========
Environmental Setup
-------------------
The experiments are powered by the MuJoCo physics simulator ([@mujoco]). The attribute modules in our experiments are all three-layer fully connected networks which output Gaussian distributed stochastic actions, built using TensorFlow. The baseline RL algorithm we use is the PPO ([@ppo]) method with GAE ([@gae]) as the advantage estimator.
We evaluate the capability of the CAN on two types of robot scenarios in our experiments. One is a point robot with a 2-dimensional action space, and the other is an articulated robot with a 5-dimensional action space. For the point robot, the state space includes the position and velocity of the robot, and the action vector is the driving force applied to it. For the articulated robot the state spaces are the angle and angular velocity at all the joints, and the x-position and speed of the base of the robot, while the action vector includes the torques at all the joints, and the force applied to the base of the robot. For each robot agent, we implement one base attribute and four different add-on attributes. For each different attribute, we designed a reward function $R_i$. The attribute settings and reward functions are described as follow:
 \[figurelabel\]
### Reaching (base attribute)
Target reaching task, a common robot task, is defined to be the base attribute for both robot scenarios in our experiment. Given initial configuration and location of the target, the robots aim to to reach the target zone using their end effector. The reward function is shown below: $$r_b(t)=\left\{
\begin{aligned}
1, &\ reaches\ target \\
0, &\ other\ case
\end{aligned}
\right.\eqno{(4)}$$
### Obstacle (position phase)
In our obstacle avoidance attribute, a moving circular obstacle is placed in the work area. In our implementation, one obstacle avoidance attribute correspond to one circular obstacle. Nevertheless, many obstacle attributes are allowed to be assembled together to create a more complicated environment. The reward function is represented as following: $$r_o^i(t)=\left\{
\begin{aligned}
-0.3, &\ touches\ i^{th} obstacle \\
0, &\ other\ case
\end{aligned}
\right.\eqno{(5)}$$
### Automated door (time phase)
The automated door attribute is a time controlled obstacle. This door is closed in the first and will open at some specific time slots. This attribute is challenging to train using RL since it punishes the agent even if it goes to the right direction at a wrong time. The reward function is designed as following:
$$r_d(t)=\left\{
\begin{aligned}
-0.01, &\ touches\ door \\
0, &\ other\ case
\end{aligned}
\right. \eqno{(6)}$$
### Speed limit (velocity phase)
Practical robots always have dynamic constraints on their joints which limit their speeds. The speed limit at time t, L(t), is designed to be a time-variant function as shown in Fig \[fas\]. Denote the maximum joint velocity as $v_{max}$. Reward function can be written as:
$$r_s(t)= -0.3(max(v_{max} - L(t),0))\eqno{(7)}$$
### Force disturbance (acceleration phase)
In working scenarios, a robot can be influenced by force disturbance or repulsion force on their joints. The force disturbance attribute in our experiment corresponds to a time-invariant force disturbance added to a certain joint of both point robot and articulated robot.
The force disturbance only affects the dynamics function of the robot system, with no additional reward function added. $$r_f(t)=0 \eqno{(8)}$$
Training Schemes
----------------
To guarantee the capacity of the CAN, the attribute modules need to meet two requirements:
1. The base attribute policies should be robust over the state space, rather than being effective only at the states that are close to the optimal trajectory. This enables the base attribute policies to be instructive when a compensation action is added on the top of it.
2. The compensation action for a certain attribute should be close to zero if the agent is in a state where this attribute is not active. This property increases the capability of multi-attribute structures.
For the sake of the robustness of the attribute policies, we apply CL to learn a general policy that can accomplish the task starting from any initial state. The CL algorithm first trains a policy with a fixed initial state. As the training goes on, the random level of the initial state is smoothly increased, until the initial state is randomly sampled from the whole state space. The random level is increased only if the policy is capable enough for the current random level, which is reflected by the increase of the episodic reward. The pseudocode for this process is shown in Algorithm 1.
$RandomLevel = $ Initial Random Level $\lambda = 1 + $ Random Level Increase Rate $N = $ Batch Number $LongTermR = Queue()$ Update the policy using PPO $Rewards \gets RunEpisode(N)$ $LongTermR.append(Rewards)$ $RandomLevel = RandomLevel \times \lambda$ $Clear(LongTermR)$
To guarantee the second requirement, an extra loss term that punishes the magnitude of the compensative action, $l^{c}_i \propto -\|a^{c}_i\|^2$, is added to the reward function so as to reduce $\|a^{c}_i\|$ when attribute $i$ is not active.
CAN Performance
---------------
The first set of experiments test the capability of the CAN to learn attributes and assemble learned attributes. We first train the base attribute module using the baseline RL algorithm with CL and then use the cascading modules to decompose the different add-on attributes based on the pretrained base module. In the actor-critic RL training using PPO, the maximum episodes is 10000. Both the actor network and critic network are trained using the Adam optimizer, with a batch number of 256, an initial learning rate of 0.0001, and are updated for 20 times in each training iteration. After training the two robot scenarios and all the four add-on attributes for each scenario, tasks. Therefore, the results show that the add-on attributes can be successfully added to the base attribute using the CAN. Fig. 4. shows the random level training log for the four add-on attribute modules in the point mass scenario, we ran 10 test episodes for the 8 combination tasks and received 10/10 success rate for all the tasks and Fig. 5 and Fig. 6 show the example episodes of the performing different attributes combinations.
We also show cases where the CAN can assemble more than one add-on attribute modules. Concretely, we first train the obstacle attribute module based on the point mass robot. Then we connect two identically parameterized obstacle attribute modules in series following the base attribute module, each handling one obstacle ball. The CAN structure is the same as the one shown in Fig. 2.. Fig. 7. shows two examples of the CAN zero shooting the task where the moving point robot reaches the target while avoiding two obstacles simultaneously. Our perspective is not learning a universally effective policy for any control task, but utilizing the decomposition of the policy to reuse pre-learned knowledge, so as to achieve easier and faster training compared to training a new policy from scratch for new control task (as shown in Section IV-C).
![Training log on random level for the point mass robot with the four add-on attributes. CL is applied, and the improvement of the random level is based on that the total reward performance being consistently improving.[]{data-label="random_log"}](log.png){width="0.9\linewidth"}
 \[figurelabel\]
 \[figurelabel\]
Comparison with Baseline RL Methods
-----------------------------------
We compare the capability and efficiency of the CAN and the baseline RL on an obstacle task for the point mass robot. The CAN is trained with CL, with the base attribute pretrained and fixed. For the baseline RL, the task is trained from scratch. For baseline RL trained with RCL, we let the initial state be very close to the target in the early stage of the training phase. Using RCL, the baseline RL could gain positive reward in the early stage. The challenge would be whether the RL algorithm can maintain a high reward level as the random level increases.
The resulting comparison of the reward and random level are shown in Fig. 8, with the maximum training iterations exceeding 120000. It is shown that the baseline RL using CL barely learns anything. This is because the reward is too sparse and the agent is consistently receiving punishment from the obstacle and falls into a local minimum of purely avoiding the obstacle and ignoring the target. For the baseline RL using RCL, in the early stage, the average discounted reward in an episode is high as expected. But as the random level rises, the performance of the baseline RL with RCL drops. Therefore, the random level increases slowly as the training goes on. The CAN, on the other hand, is able to overcome the misleading punishments from the obstacle. As a result, the random level of the CAN rises rapidly, and the CAN achieves terminal random level more than 10 times faster than the baseline.
 \[comparision\]
Conclusions and Future Work
===========================
In this paper, we propose the attribute learning and present the advantages of using this novel method to decompose the control policies of complicated control tasks. The RL framework we propose, the CAN, uses cascading attribute modules structure to model the characteristics of the attributes. The attribute modules are trained with the guidance of the pretrained base attribute module. We validate the effectiveness of the CAN of decomposing and assembling attributes and show the advantages of the CAN in solving complicated tasks compared to the baseline RL. Our ongoing future work includes applying the CAN method on more advanced simulation scenarios such as autonomous driving and using a real 2D articulated robot similar to the one in the simulation experiment to verify the capability of the CAN in more general and complicated settings.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper Butson-type complex Hadamard matrices $\mathrm{BH}(n,q)$ of order $n$ and complexity $q$ are classified for small parameters by computer-aided methods. Our main results include the enumeration of $\mathrm{BH}(21,3)$, $\mathrm{BH}(16,4)$, and $\mathrm{BH}(14,6)$ matrices. There are exactly $72$, 1786763, and $167776$ such matrices, up to monomial equivalence. Additionally, we show an example of a $\mathrm{BH}(14,10)$ matrix for the first time, and show the nonexistence of $\mathrm{BH}(8,15)$, $\mathrm{BH}(11,q)$ for $q\in\{10,12,14,15\}$, and $\mathrm{BH}(13,10)$ matrices.'
address: 'P.H.J. L., P.R.J. Ö., and F. Sz.: Department of Communications and Networking, Aalto University School of Electrical Engineering, P.O. Box 15400, 00076 Aalto, Finland'
author:
- 'Pekka H.J. Lampio, Patric R.J. Östergård, and Ferenc Szöllősi'
date: '. Preprint. This research was supported in part by the Academy of Finland, Grant \#289002'
title: Orderly generation of Butson Hadamard matrices
---
Introduction
============
Let $n$ and $q$ be positive integers. A Butson-type complex Hadamard matrix of order $n$ and complexity $q$ is an $n\times n$ matrix $H$ such that $HH^\ast=nI_n$, and each entry of $H$ is some complex $q$th root of unity, where $I_n$ denotes the identity matrix of order $n$, and $H^\ast$ denotes the conjugate transpose of $H$. The rows (and columns) of $H$ are therefore pairwise orthogonal in $\mathbb{C}^n$. For a fixed $n$ and $q$ we denote the set of all Butson-type complex Hadamard matrices by $\mathrm{BH}(n,q)$, and we simply refer to them as a “Butson matrix” for brevity [@cHOR]. The canonical examples are the Fourier matrices $F_n:=[\mathrm{exp}(2\pi\mathbf{i} jk/n)]_{j,k=1}^n\in\mathrm{BH}(n,n)$, frequently appearing in various branches of mathematics [@cKarol].
A major unsolved problem in design theory is “The Hadamard Conjecture” which predicts the existence of $\mathrm{BH}(n,2)$ matrices (real Hadamard matrices) for all orders divisible by $4$. The concept of Butson matrices was introduced to shed some light onto this question from a more general perspective [@cBUT]. Complex Hadamard matrices play an important role in the theory of operator algebras [@cHaa], [@cNic], and they have also applications in harmonic analysis [@cKMat]. Currently there is a renewed interest in complex Hadamard matrices due to their connection to various concepts of quantum information theory, e.g., to quantum teleportation schemes and to mutually unbiased bases [@cBAN], [@cDIT], [@cKA], [@cKarol], [@cWer].
This paper is concerned with the computer-aided generation and classification of Butson matrices. Let $X$ be an $n\times n$ monomial matrix, that is $X$ has exactly one nonzero entry in each of its rows and columns which is a complex $q$th root of unity. The group $G$ of pairs of monomial matrices act on the Butson matrix $H$ by $H^{(X,Y)}\to XHY^\ast$. Two Butson matrices $H_1$ and $H_2$ are called (monomial) equivalent, if they are in the same $G$-orbit. The automorphism group of $H$, denoted by $\mathrm{Aut}(H)$ is the stabilizer subgroup of $G$ with respect to $H$. Note that if $H\in\mathrm{BH}(n,q)$ then naturally $H\in\mathrm{BH}(n,r)$ for any $r$ being a multiple of $q$. Therefore the group $\mathrm{Aut}(H)$ depends on the choice of $q$.
Earlier work predominantly considered the classification of the real case in a series of papers [@cKha], [@cKIM], [@cSPE], see also [@OLDBOOK Section 7.5] for a historical overview. The quaternary case also received some attention in [@cLOS] and [@cS1]. Other papers in the literature dealt with settling the simpler existence problem through combinatorial constructions [@cBAN], [@cSEB], [@cS2], [@cKYO] or focused on the generation of matrices with some special structure [@cAKI], [@cCCdL], [@cCHK], [@cDJ], [@cPAD], [@cHIR], [@cMW].
The outline of this paper is as follows. In Section \[sect2\] we give a short overview of computer representation of Butson matrices, and recall the concept of vanishing sums of root of unity. In Section \[sect3\] we briefly describe the method of orderly generation which serves as the framework used for equivalence-free exhaustive generation. In Section \[sect4\] we present three case studies: the classification of $\mathrm{BH}(16,4)$ matrices; the classification of $\mathrm{BH}(21,3)$ matrices; and the nonexistence of $\mathrm{BH}(n,q)$ matrices for several values $n$ and $q$. An additional notable contribution of this section is Theorem \[newkron\] establishing a connection between unreal $\mathrm{BH}(n,6)$ matrices and $\mathrm{BH}(2n,4)$ matrices. We conclude the paper in Section \[sect99\] with several open problems.
The results of this paper considerably extend the work [@cBAN Theorem 7.10], where the (non)existence of Butson matrices was settled for $n\leq 10$ and $q\leq 14$. The reader might wish to jump ahead to Table \[tableBE\] to get a quick overview of the known number of $\mathrm{BH}(n,q)$ matrices for $n\leq 21$ and $q\leq 17$, including the new results established in this paper for the first time. The generated matrices are available as an electronic supplement on the web.[^1] The interested reader is also referred to [@cKarolweb] where various parametric families of complex Hadamard matrices [@cDIT] can be found, based on the catalog [@cKarol].
Computer representation of Butson Hadamard matrices {#sect2}
===================================================
A Butson matrix $H\in\mathrm{BH}(n,q)$ is conveniently represented in logarithmic form, that is, the matrix $H=[\mathrm{exp}(2\pi\mathbf{i}\varphi_{j,k}/q)]_{j,k=1}^n$ is represented by the matrix $L(H):=[\varphi_{j,k}\ \mathrm{mod}\ q]_{j,k=1}^n$ with the convention that $L_{j,k}\in\mathbb{Z}_q$ for all $j,k\in\{1,\dots,n\}$. Throughout this paper we denote by $\mathbb{Z}_q$ the additive group of integers modulo $q$, where the underlying set is $\{0,\dots,q-1\}$. With this convention $(\mathbb{Z}_q^n,\prec)$ is a linearly ordered set, where for $a,b\in\mathbb{Z}_q^n$ we write $a\prec b$ if and only if $a=b$ or $a$ lexicographically precedes $b$.
\[ex1\] The following is a $\mathrm{BH}(14,10)$ matrix $H$, displayed in logarithmic form.
Observe that the matrix shown in Example \[ex1\] is in dephased form [@cKarol], that is, its first row and column are all $0$ (representing the logarithmic form of $1$). Every matrix can be dephased by using equivalence-preserving operations. Throughout this paper all matrices are assumed to be dephased.
Let $H\in\mathrm{BH}(n,q)$, and let $r_1, r_2\in\mathbb{Z}_q^n$ be row vectors of $L(H)$. Then, by complex orthogonality, the difference row $d:=r_1-r_2\in\mathbb{Z}_q^n$ satisfies $\mathcal{E}_{n,q}(d)=0$, where $$\mathcal{E}_{n,q} \colon \mathbb{Z}_q^n \to \mathbb{C},\qquad \mathcal{E}_{n,q}(x):=\sum_{i=1}^{n}\mathrm{exp}(2\pi\mathbf{i}x_i/q)$$ is the evaluation function. In other words, $d$ represents an $n$-term vanishing sum of $q$th roots of unity [@cLL]. We note that the number $\mathcal{E}_{n,q}(x)$ is algebraic, and its value is invariant up to permutation of the coordinates of $x\in\mathbb{Z}_q^n$. In particular, $\mathcal{E}_{n,q}(x)=\mathcal{E}_{n,q}(\mathrm{Sort}(x))$, where $\mathrm{Sort}(x)=\min\{\sigma(x)\colon \text{$\sigma$ is a permutation on $n$ elements}\}$ (with respect to the ordering $\prec$). We introduce the orthogonality set which contains the representations of the normalized, sorted, $n$-term vanishing sums of $q$th roots of unity: $$\mathcal{O}(n,q):=\{x\in\mathbb{Z}_q^n \colon x_1=0;\ x=\mathrm{Sort}(x);\ \mathcal{E}_{n,q}(x)=0\}.$$ Once precomputed, the set $\mathcal{O}(n,q)$ allows us to determine if two rows of length $n$ of a dephased matrix with elements in $\mathbb{Z}_q$ are complex orthogonal in a combinatorial way, i.e., without relying on the analytic function $\mathcal{E}_{n,q}$. Indeed, for any vector $x\in \mathbb{Z}_q^n$ having at least one $0$ coordinate, $\mathcal{E}_{n,q}(x)=0$ if and only if $\mathrm{Sort}(x)\in\mathcal{O}(n,q)$.
One can observe that for certain values of $n$ and $q$ the set $\mathcal{O}(n,q)$ is empty, that is, it is impossible to find a pair of orthogonal rows in $\mathbb{Z}_q^n$ and consequently $\mathrm{BH}(n,q)$ matrices do not exist. For example, it is easy to see that $|\mathcal{O}(n,2)|=0$ for odd $n>1$. The following recent result characterizes the case when the set $\mathcal{O}(n,q)$ is nonempty, and should be viewed as one of the fundamental necessary conditions on the existence of Butson matrices.
\[LLMAIN\] Let $n$, $r$, and $a_i$, $i\in\{1,\dots, r\}$ be positive integers, and let $q=\prod_{i=1}^rp_i^{a_i}$ with distinct primes $p_i$, $i\in\{1,\dots, r\}$. Then, we have $|\mathcal{O}(n,q)|\geq 1$ if and only if there exist nonnegative integers $w_i$, $i\in\{1,\dots,r\}$ such that $n=\sum_{i=1}^r w_i p_i$.
In order to classify all $\mathrm{BH}(n,q)$ matrices for a given parameters, three tasks have to be completed: the set $\mathcal{O}(n,q)$ has to be determined; vectors $x\in \mathbb{Z}_q^n$ orthogonal to a prescribed set of vectors should be generated; and equivalent matrices should be rejected. In the next section we discuss these three tasks in detail.
Generating Butson Hadamard matrices {#sect3}
===================================
Generating the vanishing sums of roots of unity {#sssectonq}
-----------------------------------------------
For a given $n$ and $q$, our first task is to determine the set $\mathcal{O}(n,q)$ which in essence encodes complex orthogonality of a pair of rows. It turns out that when $q$ is a product of at most two prime powers, then a compact description of the elements of $\mathcal{O}(n,q)$ is possible. The following two results are immediate consequences of [@cLL Corollary 3.4].
\[l1\] Let $a$, $n$ be positive integers, and let $q=p^a$ be a prime power. Let $u=[0,q/p,2q/p,\dots,(p-1)q/p]\in\mathbb{Z}_q^p$, and let $x\in\mathbb{Z}_q^n$. Then $x\in\mathcal{O}(n,q)$ if and only if there exist a positive integer $s$ such that $ps=n$, and $r_i\in\{0,\dots,q/p-1\}$, $i\in\{1,\dots, s-1\}$, such that $x=\mathrm{Sort}([u,r_1+u,\dots,r_{s-1}+u])$.
\[l2\] Let $a$, $b$ and $n$ be positive integers, and let $q=p_1^ap_2^b$ be the product of two distinct prime powers. Let $u=[0,q/p_1,2q/p_1,\dots,(p_1-1)q/p_1]\in\mathbb{Z}_q^{p_1}$, $v=[0,q/p_2,2q/p_2,\dots,(p_2-1)q/p_2]\in\mathbb{Z}_q^{p_2}$, and let $x\in\mathbb{Z}_q^n$. Then $x\in\mathcal{O}(n,q)$ if and only if there exist nonnegative integers $s$, $t$ such that $p_1s+p_2t=n$, and $r_i\in\{0,\dots,q/p_1-1\}$, $i\in\{1,\dots, s\}$, $R_j\in\{0,\dots,q/p_2-1\}$, $j\in\{1,\dots, t\}$ such that $x=\mathrm{Sort}([r_1+u,r_2+u,\dots,r_s+u,R_1+v,R_2+v,\dots,R_t+v])$, and $0\in\{r_1,R_1\}$.
The main point of the rather technical Lemma \[l1\] and Lemma \[l2\] is the following: as long as $q$ is the product of at most two prime powers, the constituents of any $n$-term vanishing sum of $q$th roots of unity are precisely $p$-term vanishing sums, where $p$ is some prime divisor of $q$. These $p$-term vanishing sums are in turn the (scalar multiplied, or, “rotated”) sums of every $p$th root of unity.
The significance of these structural results is that based on them one can design an efficient algorithm to generate the set $\mathcal{O}(n,q)$ as long as $q<30=2\cdot3\cdot5$ in a combinatorial way (i.e., without the need of the analytic function $\mathcal{E}_{n,q}$). In particular, this task can be done relying on exact integer arithmetic. We spare the reader the details.
In certain simple cases it is possible to enumerate (as well as to generate) the set $\mathcal{O}(n,q)$ by hand. We offer the following counting formulae for means of checking consistency.
\[minor1\] Let $a$ and $n$ be positive integers, and let $q=p^a$ be a prime power. Assume that $p$ divides $n$. Then $|\mathcal{O}(n,q)|=\binom{(n+q)/p-2}{n/p-1}$.
By Lemma \[l1\] members of the set $\mathcal{O}(n,q)$ can be partitioned into $n/p$ parts of the form $r_i+[0,q/p,2q/p,\dots,(p-1)q/p]$, each part being identified by the rotation $r_i\in\{0,\dots,q/p-1\}$, $i\in\{0,\dots, n/p-1\}$ with $r_0=0$. The number of ways to assign $q/p$ values to a set of $n/p-1$ variables (up to relabelling) is exactly $\binom{(n+q)/p-2}{n/p-1}$; each of these choices lead to different members of $\mathcal{O}(n,q)$.
A slightly more complicated variant is the following result.
\[minor2\] Let $n\geq 2$ be an integer, let $p$ be an odd prime, and let $q=2p$. Then $$|\mathcal{O}(n,q)|=\frac{1+(-1)^n}{2}\binom{p+\left\lfloor n/2\right\rfloor-2}{\left\lfloor n/2\right\rfloor-1}+\sum_{\substack{2s+pt=n\\ s\geq1,t\geq1}}\binom{p+s-1}{s}+\sum_{\substack{2s+pt=n\\ s\geq1, t\geq1}}\binom{p+s-2}{s-1}+\delta,$$ where $\delta=1$ if $p$ divides $n$, and $\delta=0$ otherwise.
This can be inferred by using Lemma \[l2\]. We count the elements $x\in\mathcal{O}(n,q)$ based on how many pairs of coordinates $[x_i,x_i+p]\in\mathbb{Z}_q^2$ they have. Let us call this number $s$.
If $s=0$, then clearly $p$ divides $n$ and $x$ can be partitioned into $t=n/p$ parts, each being either of the form $[0,2,4,\dots,2p-2]$ or $[1,3,5,\dots,2p-1]$. However, since $s=0$, only one of these two forms could appear, and since $x$ must have a coordinate $0$, this left us with only $\delta=1$ case.
If $s=n/2\geq 1$ then $n$ is necessarily even, and $x$ can be partitioned into $n/2$ parts, each being of the form $[x_i,x_i+p]$ for some $x_i\in\{0,\dots,p-1\}$, $i\in\{1,\dots,n/2\}$. Since $x$ must contain $0$, one of these parts must be $[0,p]$, while the other $n/2-1$ parts can take $p$ different forms. There are a total of $\binom{p+n/2-2}{n/2-1}$ cases.
Finally, if $0<s<n/2$, then there are either $t=(n-2s)/p\geq 1$ parts of the form $[0,2,4,\dots,2p-2]$, or $t$ parts of the form $[1,3,5,\dots,2p-1]$. In the first case there are $\binom{p+s-1}{s}$ ways to assign values to the remaining $s$ parts; in the second case, since $x$ must have a $0$ coordinate, there are $\binom{p+s-2}{s-1}$ ways to assign values to the remaining $s$ parts.
The statements of Lemma \[minor1\] and Lemma \[minor2\] are strong enough to cover all cases $q\leq 17$ except for $q\in\{12,15\}$. We have applied these results to verify that the computer-generated sets $\mathcal{O}(n,q)$ are of the correct cardinality. In the next section we will see a further application of the set $\mathcal{O}(n,q)$.
\[magicsumq\] There is no analogous result to Lemma \[l1\] and Lemma \[l2\] when $q$ has more than two prime factors. Indeed, the reader might amuse themselves by verifying that while $[0,1,7,13,19,20]\in\mathcal{O}(6,30)$, it does not have any $m$-term vanishing subsums with $m\in\{2,3,5\}$. See [@cLL Example 6.7] for examples of similar flavor.
An alternative, algebraic way to generate the set $\mathcal{O}(n,q)$ is to compute for all $x\in\mathbb{Z}_q^n$ with $x_1=0$ and $\mathrm{Sort}(x)=x$ the minimal polynomial $p(t)$ of the algebraic number $\mathcal{E}_{n,q}(x)$. With this terminology, $x\in\mathcal{O}(n,q)$ if and only if $p(t)=t$. The efficiency of this approach can be greatly improved by testing first by fast numerical means whether the Euclidean norm of $\mathcal{E}_{n,q}(x)$ is small, say if $\left\|\mathcal{E}_{n,q}(x)\right\|^2=\mathcal{E}_{n,q}(x)\mathcal{E}_{n,q}(-x)<0.01$ holds.
Orderly generation of rectangular matrices {#subsorder}
------------------------------------------
In this section we briefly recall the method of orderly generation, which is a technique for generating matrices exhaustively in a way that no equivalence tests between different matrices are required [@cKAS Section 4.2.2], [@cREA]. Such a search can be efficiently executed in parallel. The main idea is to select from each equivalence class of Butson matrices a canonical representative, and organize the search in a way to directly aim for this particular matrix. Variations of this basic approach were employed for the classification of $\mathrm{BH}(n,2)$ matrices for $n\leq 32$, see [@cKha], [@cSPE].
Let $n,r\geq1$. We associate to each $r\times n$ matrix $R$ whose elements are complex $q$th roots of unity its vectorization $v(R):=[L(R)_{1,1}, \dots, L(R)_{1,n}, L(R)_{2,1},\dots,L(R)_{r,n}]\in\mathbb{Z}_q^{rn}$ formed by concatenating the rows of its logarithmic form $L(R)$. We say that $R$ is in canonical form, if $v(R)=\min\{v(XRY^\ast)\colon \text{$X$ and $Y$ are $q$th root monomial matrices}\}$, where comparison is done with respect to the ordering $\prec$. Canonical matrices defined in this way have a number of remarkable properties. For example, if $R$ is canonical, and $r_1$ and $r_2$ are consecutive rows of $L(R)$, then $r_1\prec r_2$, and analogously for the columns. Moreover, canonical matrices are necessarily dephased. Let $\sigma$ be a permutation on $r$ elements, and let $i\in\{1,\dots,n\}$. Let us denote by $R^{(\sigma,i)}$ the matrix which can be obtained from $R$ by permuting its rows according to $\sigma$, then swapping its first and $i$th columns, then dephasing it, and finally arranging its columns according to $\prec$.
\[l35\] Let $n,r\geq 1$, and let $R$ be an $r\times n$ matrix. The matrix $R$ is canonical, if and only if $v(R)=\min\{v(R^{(\sigma,i)})\colon\text{$\sigma$ is a permutation on $r$ elements, $i\in\{1,\dots,n\}$}\}$.
This is an immediate consequence of the fact that canonical matrices are dephased and their columns are sorted with respect to $\prec$.
It is possible to further improve the test described in Lemma \[l35\] by the following considerations. Let $k\in\{1,\dots,r\}$ and let $R_k$ denote the leading $k\times n$ submatrix of $R$. If there exists a pair $(\sigma,i)$ such that $v(R_k)\neq v(R^{(\sigma,i)}_k)$ and $v(R_k)\prec v(R^{(\sigma,i)}_k)$ then the same holds for all other permutations whose first $k$ coordinates agree with that of $\sigma$. In particular, all those permutations can be skipped. An efficient algorithm for permutation generation with restricted prefixes is discussed in [@cKNU Algorithm X].
The computational complexity of this method is exponential in the number of rows $r$, polynomial in the number of columns $n$, and independent of the complexity $q$. Testing whether a matrix is in canonical form is the most time-consuming part of the generation.
Finally, we note one more property of canonical matrices.
\[lsecpr\] Let $H\in\mathrm{BH}(n,q)$ in canonical form. Let us denote by $r_2$ the second row of $L(H)$, and by $c_2$ the second column of $L(H)$. Then $r_2\in\mathcal{O}(n,q)$ and $c_2^T\in\mathcal{O}(n,q)$.
This follows from the fact that $H$ is necessarily dephased, and its rows and columns are ordered with respect to the ordering $\prec$.
The significance of Lemma \[lsecpr\] is that if the (transpose of the) logarithmic form of the second column of a rectangular orthogonal matrix is not a prefix of any of the elements of the set $\mathcal{O}(n,q)$, then that matrix can be discarded during the search. We refer to this look-ahead strategy as “pruning the search tree by the second column condition”.
The matrices $H\in\mathrm{BH}(n,q)$ (more precisely, their logarithmic form) are generated in a row-by-row fashion. Every time a new row is appended we first test whether it is orthogonal to all previous rows by checking if the difference vectors belong to the set $\mathcal{O}(n,q)$ as described in Section \[sssectonq\]. If the rows of the matrix are pairwise orthogonal, then we further check whether (the transpose of) its second column is a prefix of an element of the set $\mathcal{O}(n,q)$. Finally, we test whether it is in canonical form. Only canonical matrices will be processed further, the others will be discarded and backtracking takes place.
In a prequel to this work [@cLOS] we employed the method of canonical augmentation [@cPET Section 4.2.3] to solve the more general problem of classification of all rectangular orthogonal matrices. Here we solve the relaxed problem of classification of those matrices which can be a constituent of an orderly-generated Butson matrix. The reader might wish to look at the impact of the second column pruning strategy on the number of $r\times 14$ submatrices in Table \[tabletreecomp\], where we compare the size of the search trees encountered with these two methods during the classification of $\mathrm{BH}(14,4)$ matrices.
We have observed earlier that the computational cost of equivalence testing is independent of the complexity $q$ when orderly generation is used. This is in contrast with the method of canonical augmentation employed earlier in [@cLOS] which relies on graph representation of the $r\times n$ rectangular orthogonal matrices with $q$th root entries on $3q(r+n)+r$ vertices. See [@cLAM], [@pekkadiffm] for more on graph representation of Butson matrices.
Augmenting rectangular orthogonal matrices
------------------------------------------
Let $n,r\geq 1$, and let $R$ be an $r\times n$ canonical matrix with pairwise orthogonal rows. Let $r_i$, $i\in\{1,\dots, r\}$ denote the rows of $L(R)$. The goal of this section is to describe methods for generating the vectors $x\in\mathbb{Z}_q^n$ such that $\mathcal{E}_{n,q}(r_i-x)=0$ hold simultaneously for every $i\in\{1,\dots, r\}$. Note that since we are only interested in canonical Butson matrices, we assume that $x_1=0$.
The most straightforward way of generating the vectors $x$ is to consider the permutations of the elements of the set $\mathcal{O}(n,q)$. Indeed, the following two conditions (i) $x$ has a coordinate $0$; and (ii) $\mathcal{E}_{n,q}(r_1-x)=0$ are together equivalent to $\mathrm{Sort}(x)\in\mathcal{O}(n,q)$. For all such vectors $x$ the remaining conditions $\mathcal{E}_{n,q}(r_i-x)=0$, $i\in\{2,\dots,r\}$ should be verified. This strategy of generating the rows works very well for small matrices, say, up to $n\leq 11$. One advantage of this naïve method is that permutations can be generated one after another, without the need of excessive amount of memory [@cKNU].
Next we describe a more efficient divide-and-conquer strategy [@cKAS p. 157] for generating the vectors $x$. Let $m\in\{1,\dots,n-1\}$ be a parameter, and for every $i\in\{1,\dots, r\}$ write $r_i=[a_i,b_i]$, where $a_i\in\mathbb{Z}_q^{n-m}$, $b_i\in\mathbb{Z}_q^{m}$, and write $x=[c,d]$, where $c\in\mathbb{Z}_q^{n-m}$, $d\in\mathbb{Z}_q^{m}$.
As a first step, we create a lookup table $\mathcal{T}$ indexed by $\iota\in\mathbb{C}^r$, where the value at $\mathcal{T}(\iota)$ is a certain subset of $\mathbb{Z}_q^m$. Formally, consider $\mathcal{T}\colon\mathbb{C}^r\to\mathcal{P}(\mathbb{Z}_q^m)$, where for every $d\in\mathbb{Z}_q^m$ it holds that $d\in\mathcal{T}([\mathcal{E}_{n,q}(b_1-d),\dots,\mathcal{E}_{n,q}(b_r-d)])$. Naturally, we assume that the values form a partition of $\mathcal{P}(\mathbb{Z}_q^m)$. As a second step, for every $c\in\mathbb{Z}_q^{n-m}$ we look the vectors $d\in\mathbb{Z}_q^m$ up (if any) contained in the set $\mathcal{T}([-\mathcal{E}_{n,q}(a_1-c),\dots,-\mathcal{E}_{n,q}(a_r-c)])$. By construction, the vectors $x=[c,d]$ fulfill the desired conditions; if no such $d$ were found, then $c$ cannot be a prefix of $x$.
In practice, however, it is inconvenient to work with complex-valued indices, and therefore one needs to use a hash function $\mathcal{H}\colon \mathbb{C}^r\to \mathbb{Z}^+_0$ to map them to nonnegative integers. This leads to a convenient implementation at the expense of allowing hash collisions to occur. Since it is not at all clear how to come up with a nontrivial hash function (apart from $\mathcal{H}\equiv 0$) we describe here an elegant choice exploiting the number theoretic properties of the Gaussian- and the Eisenstein integers. We assume for the following argument that $q\in\{2,3,4,6\}$. Recall that $\mathcal{T}$ was indexed by complex $r$-tuples of the form $[\mathcal{E}_{n,q}(b_1-d),\dots,\mathcal{E}_{n,q}(b_r-d)]$. Let $p_{\mathrm{big}}$ be a (large) prime, and let $p_i\ll p_{\mathrm{big}}$, $i\in\{1,\dots, r\}$ be $r$ other distinct primes. We define $\mathcal{H}$ through the Euclidean norm of the partial inner products as follows: $\mathcal{H}([\mathcal{E}_{n,q}(b_1-d),\dots,\mathcal{E}_{n,q}(b_r-d)]):=\sum_{i=1}^r\left\|\mathcal{E}_{n,q}(b_i-d)\right\|^2p_i\ (\mathrm{mod}\ p_{\mathrm{big}})$. This gives rise to a table $\mathcal{S}\colon\mathbb{Z}_0^+\to\mathcal{P}(\mathbb{Z}_q^m)$ which is defined through $\mathcal{T}$ and $\mathcal{H}$ as follows: for every $\iota\in\mathbb{C}^r$, let $\mathcal{S}(\mathcal{H}(\iota)):=\mathcal{T}(\iota)$. As for the second step, for every $c\in\mathbb{Z}_q^{n-m}$ we look the vectors $d\in\mathbb{Z}_q^m$ up (if any) contained in the set $\mathcal{S}(k)$, $k\in\{0,\dots,p_{\mathrm{big}}-1\}$, for which the modular equation $k\equiv \sum_{i=1}^r\left\|\mathcal{E}_{n,q}(a_i-c)\right\|^{2}p_i\ (\mathrm{mod}\ p_{\mathrm{big}})$ holds. Finally, for all (if any) vectors $x=[c,d]$ one should test whether they are orthogonal to the rows of $R$.
The table $\mathcal{T}$ is generated once for every matrix $R$, and it is reused again during a depth-first-search. The advantage of this technique is that as long as $n\leq 21$ and $m\approx n/2$ the $q$-ary $m$-tuples can be generated efficiently. For higher sizes, however, precomputing and storing such a table becomes quickly infeasible due to memory constraints, and therefore one needs to carefully choose the value of $m$ in terms of $n$, $q$, and the number of processors accessing the shared memory.
Let $x\in\mathbb{Z}_q^n$, and for every $i\in\mathbb{Z}_q$ let us denote by $f_i$ the frequency distribution of the number $i$ occurring as a coordinate of $x$. We have $\|\mathcal{E}_{n,2}(x)\|^2=(f_0-f_1)^2$; $\|\mathcal{E}_{n,3}(x)\|^2=f_0^2+f_1^2+f_2^2-f_0f_1-f_0f_2-f_1f_2$; $\|\mathcal{E}_{n,4}(x)\|^2=(f_0-f_2)^2+(f_1-f_3)^2$; and finally, $\|\mathcal{E}_{n,6}(x)\|^2=(f_0-f_3)^2+(f_4-f_1)^2+(f_5-f_2)^2-(f_0-f_3)(f_4-f_1)-(f_0-f_3)(f_5-f_2)-(f_4-f_1)(f_5-f_2)$. In particular, these numbers are nonnegative integers.
For $q\not\in\{2,3,4,6\}$ the hash function $\mathcal{H}$ should be replaced by a suitable alternative, as the quantity $\|\mathcal{E}_{n,q}(x)\|^2$ is no longer guaranteed to be an integer. For example, when $q=10$, one may verify that for every $x\in\mathbb{Z}_{10}^n$ we have $2\|\mathcal{E}_{n,10}(x)\|^2=A+\sqrt{5}B$, where $A$ and $B$ are integers. Therefore one can map $\|\mathcal{E}_{n,10}(x)\|^2$ to $A^2+pB^2$ (where $p$ is some large prime). Similar techniques work for certain other values of $q$.
Results and case studies {#sect4}
========================
Main results and discussion
---------------------------
Based on the framework developed in Sections \[sect2\]–\[sect3\] we were able to enumerate the set $\mathrm{BH}(n,q)$ for $n\leq 11$ and $q\leq 17$ up to monomial equivalence (cf. [@cBAN Theorem 7.10]). Several additional cases were also settled.
The known values of the exact number of $\mathrm{BH}(n,q)$ matrices, up to monomial equivalence, is displayed in Table \[tableBE\].
The legend for Table \[tableBE\] is as follows. An entry in the table at position $(n,q)$ indicates the known status of the existence of $\mathrm{BH}(n,q)$ matrices. Empty cells indicate cases where $\mathrm{BH}(n,q)$ matrices do not exist by Theorem \[LLMAIN\]; cells marked by an “E” indicate cases where $\mathrm{BH}(n,q)$ matrices are known to exist, but no full classification is available; cells marked by an “U” indicate that existence is unknown; finally cells displaying a number indicate the exact number of $\mathrm{BH}(n,q)$ matrices up to monomial equivalence.
Next we briefly review the contents of Table \[tableBE\], and comment on the cases based on their complexity $q\in\{2,3,\dots, 17\}$. We note that most of the numbers shown are new.
: This is the real Hadamard case. Complete classification is available up to $n\leq 32$, see [@OLDBOOK Section 7.5], [@cKha]. The number of $\mathrm{BH}(36,2)$ matrices is at least $1.8\times 10^7$ [@cORR], while according to [@cLLT] the number of $\mathrm{BH}(40,2)$ matrices is at least $3.66\times 10^{11}$.
: Complete classification is available up to $n\leq 21$, see Section \[bh213x\]. The case $\mathrm{BH}(18,3)$ was reported in [@cHAR2] and independently in [@cLAM]. Several cases of $\mathrm{BH}(21,3)$ were found by Brock and Murray as reported in [@cAKI] along with additional examples. There are no $\mathrm{BH}(15,3)$ matrices [@cHAR2], [@OLDBOOK Theorem 6.65], [@cLAM Theorem 3.2.2].
: Classification is known up to $n\leq 16$, see [@cLOS], [@cS1] and Section \[seccase1\]. The difference matrices over $\mathbb{Z}_4$ with $\lambda=4$ (essentially: the $\mathrm{BH}(16,4)$ matrices of type-$4$) were reported independently in [@cGM], [@cHLT], [@pekkadiffm]. A $\mathrm{BH}(18,4)$ can be constructed from a symmetric conference matrix [@cSEB Theorem 3], [@cTUR].
: An explicit example of $\mathrm{BH}(20,5)$ can be found in [@cSEB2], while a $\mathrm{BH}(15,5)$ does not exist [@OLDBOOK Theorem 6.65], [@cLAM Theorem 3.2.2].
: Examples of $\mathrm{BH}(7,6)$ matrices were presented in [@cBRO] and independently but slightly later in [@cPET]. A $\mathrm{BH}(10,6)$ was reported in [@cAGA p. 105]. Several unreal $\mathrm{BH}(13,6)$ were reported in [@cCCdL]; additional examples were reported by Nicoară et al. on the web site [@cKarolweb]. A $\mathrm{BH}(19,6)$ was found in [@cS2], based on the approach of [@cPET]. A necessary condition on the existence of a $\mathrm{BH}(n,6)$ matrix comes from the determinant equation $|\mathrm{det}(H)|^2=n^n$, where the left hand side is the norm of an Eisenstein integer and therefore is of the form $A^2-AB+B^2$ for some integers $A$ and $B$ [@cBRO], [@cWIN]. Consequently $\mathrm{BH}(n,6)$ matrices for $n\in\{5,11,15,17\}$ do not exist.
: The $\mathrm{BH}(14,7)$ matrices come from a doubling construction [@cBUT], [@cKYO] while $\mathrm{BH}(21,7)$ matrices do not exist by [@cWIN Theorem 5].
: Here $n=1$, or $n\geq2$ is necessarily even by Theorem \[LLMAIN\]. Existence follows from the existence of $\mathrm{BH}(n,4)$ matrices. A particular example of $\mathrm{BH}(6,8)$ matrix played an important role in disproving the “Spectral Set Conjecture” in $\mathbb{R}^3$, see [@cKMat]. This is one notable example of contemporary applications of complex Hadamard matrices.
: A $\mathrm{BH}(15,9)$ does not exist by [@cWIN Theorem 5].
: Nonexistence of $\mathrm{BH}(n,10)$ for $n\in\{6,7\}$ was proved in [@cBAN]. The discovery of a $\mathrm{BH}(9,10)$ matrix by Beauchamp and Nicoară (found also independently in [@cKA]) was rather unexpected [@cKarolweb]. There are no $\mathrm{BH}(11,10)$ or $\mathrm{BH}(13,10)$ matrices (see Theorem \[nonex11\] and \[nonex13\]). To the best of our knowledge $\mathrm{BH}(14,10)$ matrices were not known prior to this work, and Example \[ex1\] shows a new discovery.
: The Fourier matrix $F_{11}$ is unique [@cHIR].
: A $\mathrm{BH}(5,12)$ does not exist since all $5\times 5$ complex Hadamard were shown to be equivalent to $F_5$ in [@cHaa]. A $\mathrm{BH}(11,12)$ does not exist by Theorem \[nonex11\].
: The Fourier matrix $F_{13}$ is unique [@cHIR].
: Several nonexistence results are known. The matrices $\mathrm{BH}(n,14)$ for $n\in\{6,9,10\}$ were shown to be nonexistent in [@cBAN]. The matrices $\mathrm{BH}(11,14)$ do not exist by Theorem \[nonex11\]. Finally, there are no $\mathrm{BH}(21,14)$ matrices by [@cWIN Theorem 5].
: There are no $\mathrm{BH}(n,15)$ matrices for $n\in\{8,11\}$, see Theorem \[nonex8\] and Theorem \[nonex11\] respectively.
: Here $n=1$ or $n\geq 2$ is necessarily even. Existence follows from the existence of $\mathrm{BH}(n,4)$ matrices.
: The Fourier matrix $F_{17}$ was shown to be unique in [@cHIR] by computers.
Examples of matrices corresponding to the cases marked by “E” in Table \[tableBE\] can be obtained from either by viewing a matrix $H\in\mathrm{BH}(n,q)$ as a member of $\mathrm{BH}(n,r)$ with some $r$ which is a multiple of $q$; or by considering the Kronecker product of two smaller matrices [@cHOR Lemma 4.2]. In particular, if $H\in\mathrm{BH}(n_1,q_1)$ and $K\in\mathrm{BH}(n_2,q_2)$ then $H\otimes K\in\mathrm{BH}(n_1n_2,\mathrm{LCM}(q_1,q_2))$, where $\mathrm{LCM}(a,b)$ is the least commmon multiple of the positive integers $a$ and $b$. This construction shows that Butson matrices of composite orders are abundant. In contrast, very little is known about the prime order case [@cPET].
\[HADEQ\] Several authors, see e.g. [@cHOR Definition 4.12], [@cLOS], consider two $\mathrm{BH}(n,q)$ matrices Hadamard equivalent if either can be obtained from the other by performing a finite sequence of monomial equivalence preserving operations, and by replacing every entry by its image under a fixed automorphism of $\mathbb{Z}_q$. Given the classification of Butson matrices up to monomial equivalence it is a routine task to determine their number up to Hadamard equivalence. Indeed, let $\mathcal{X}$ be a complete set of representatives of $\mathrm{BH}(n,q)$ matrices up to monomial equivalence. Let $\varphi(.)$ denote the Euler’s totient function. Then for each $H\in\mathcal{X}$ let us denote by $c(\Psi(H))$ the number of matrices in $\Psi(H):=\{\psi(H)\colon\psi\in\mathrm{Aut}(\mathbb{Z}_q)\}$ up to monomial equivalence. For each $i\in\{1,\dots,\varphi(q)\}$ let us denote by $k_i$ the frequency distribution of the number $i$ occurring as the value of $c(\Psi(H))$ while it runs through $\mathcal{X}$. Then the number of Hadamard equivalence classes is $\sum_{i=1}^{\varphi(q)}k_i/i$, see Table \[tableHEQ\].
Classification of the BH(16,4) matrices {#seccase1}
---------------------------------------
Classification of the quaternary complex Hadamard matrices is motivated by their intrinsic connection to real Hadamard matrices, which is best illustrated by the following classical result.
\[turync\] Let $n\geq 1$. If $A$ and $B$ are $n\times n$ $\{-1,0,1\}$-matrices such that $A+\mathbf{i}B\in\mathrm{BH}(n,4)$ then $A\otimes\left[\begin{smallmatrix}1 &\hfill 1\\ 1 &\hfill -1\end{smallmatrix}\right]+B\otimes\left[\begin{smallmatrix}\hfill-1 & 1\\\hfill 1 & 1\end{smallmatrix}\right]\in\mathrm{BH}(2n,2)$.
It is conjectured [@cHOR p. 68] that $\mathrm{BH}(n,4)$ matrices exist for all even $n$. The resolution of this “Complex Hadamard Conjecture” would imply by Theorem \[turync\] the celebrated Hadamard Conjecture.
The classification of $\mathrm{BH}(16,4)$ matrices involved several steps. First we generated the set $\mathcal{O}(16,4)$. We note that $|\mathcal{O}(16,4)|=8$ by Lemma \[minor1\], and these elements can be obtained from Lemma \[l1\] by simple hand calculations. Then, we broke up the task of classification into $5$ smaller subproblems of increasing difficulty based on the presence of certain substructures. This allowed us to experiment with the simpler cases and to develop and test algorithms used for the more involved ones. In the following we introduce the type of a $\mathrm{BH}(n,4)$ matrix, a concept which is invariant up to monomial equivalence. A similar idea was used during the classification of $\mathrm{BH}(32,2)$ matrices [@cKha].
Let $n,r\geq 2$, let $R$ be an $r\times n$ orthogonal matrix with $4$th root entries, and let $r_1$ and $r_2$ be distinct rows of $L(R)$. Let $m$ denote the number of $0$ entries in the difference vector $r_1-r_2\in\mathbb{Z}_4^n$, and let $k:=\min\{m,n/2-m\}$. Then the subset of rows $\{r_1,r_2\}$ is said to be of type-$k$. The matrix $R$ is said to be of type-$k$, if $L(R)$ has no two rows which are of type-$\ell$ for any $\ell<k$.
Secondly, we fixed $k\in\{0,\dots,4\}$ and generated the $5\times 16$ canonical (see Section \[subsorder\]) type-$k$ matrices surviving the second column pruning strategy. Thirdly, we augmented each of these with three additional rows to obtain all $8\times 16$ matrices, but during this process a depth-first-search approach was employed, and the $r\times 16$ submatrices were not kept for $r\in\{6,7\}$. Finally, we finished the search by using breadth-first-search to generate all $r\times 16$ matrices step-by-step for each $r\in\{9,\dots,16\}$. The reader is invited to compare the size of the search trees involved with the $\mathrm{BH}(16,2)$ case displayed in Table \[table2\] and with the $\mathrm{BH}(14,4)$ case displayed in Table \[tabletreecomp\].
The search, which relied on only the standard [[C]{}]{} libraries and an army of 896 computing cores took more than $30$ CPU years, and yielded the following classification result.
The number of $\mathrm{BH}(16,4)$ matrices is $1786763$ up to monomial equivalence.
In Table \[tableautgrps\] we exhibit the automorphism group sizes along with their frequencies.
The total number of $\mathrm{BH}(16,4)$ matrices $($not considering equivalence$)$ is exactly $1882031756845055238646027031522819126506763059200000$.
Let $\mathcal{X}$ be a a complete set of representatives of $\mathrm{BH}(16,4)$ matrices up to monomial equivalence. Then the size of the set $\mathrm{BH}(16,4)$ can be inferred from an application of the Orbit-stabilizer theorem [@cKAS Theorem 3.20]. We have $|\mathrm{BH}(16,4)|=|G|\sum_{X\in\mathcal{X}}1/|\mathrm{Aut}(X)|$. Combining $|G|=(16!)^2\cdot4^{32}$ with the numbers shown in Table \[tableautgrps\] yields the result.
There are two main reasons for the existence of such a huge number of equivalence classes. First, Kronecker-like constructions can lift up the $\mathrm{BH}(8,4)$ matrices resulting in multi-parametric families of complex Hadamard matrices [@cDIT], [@cKarol]. The second reason is the presence of type-$0$ (that is: real) pair of rows. It is known that such a substructure can be “switched” [@cORR] in a continuous way [@cSZFPAR] thus escaping the monomial equivalence class of the matrices is possible. In contrast, matrices which cannot lead to continuous parametric families of complex Hadamard matrices are called isolated [@cKarol]. A notion to measure the number of free parameters which can be introduced into a given matrix is the defect [@cKarol], which serves as an upper bound. We remark that when $q\in\{2,3,4,6\}$ then computing the defect boils down to a rank computation of integer matrices which can be performed efficiently using exact integer arithmetic.
There are at least $7978$ isolated $\mathrm{BH}(16,4)$ matrices.
This is established by counting the number of $\mathrm{BH}(16,4)$ matrices with defect $0$. There are no isolated $\mathrm{BH}(16,4)$ matrices of type-$0$, because they contain a real pair of rows as a substructure. It is easy to see that such matrices cannot be isolated once the size of the matrices $n>2$, see [@cSZFPAR]. Computation reveals that there are no type-$k$ matrices with vanishing defect for $k\in\{1,3,4\}$, and there are exactly $7978$ type-$2$ matrices with defect $0$. Since the defect is an upper bound on the number of smooth parameters which can be introduced [@cKarol], these matrices are isolated.
Finally, we note a result connecting $\mathrm{BH}(2n,4)$ matrices with unreal $\mathrm{BH}(n,6)$ matrices.
\[newkron\] If $A$ and $B$ are $n\times n$ $\{-1,0,1\}$-matrices such that $A_{ij}B_{ij}=0$ for $i,j\in\{1,\hdots,n\}$, and $H:=A\omega+B\omega^2\in \mathrm{BH}(n,6)$ with $\omega=\mathrm{exp}(2\pi\mathbf{i}/3)$, then $K:=A\otimes\left[\begin{smallmatrix}1 &\hfill 1\\ 1 &\hfill -1\end{smallmatrix}\right]+B\otimes\left[\begin{smallmatrix}\hfill\mathbf{i}&\hfill-1\\\hfill-1&\hfill\mathbf{i}\end{smallmatrix}\right]\in\mathrm{BH}(2n,4)$.
Let $X:=\left[\begin{smallmatrix}1 &\hfill 1\\ 1 &\hfill -1\end{smallmatrix}\right]$ and $Y:=\left[\begin{smallmatrix}\hfill\mathbf{i}&\hfill-1\\\hfill-1&\hfill\mathbf{i}\end{smallmatrix}\right]$. We have $XX^\ast=YY^\ast=-(XY^\ast+YX^\ast)=2I_2$. Since $(A\omega+B\omega^2)(A^T\omega^2+B^T\omega)=n I_n$, we have $AB^T=BA^T$. Every entry of $K$ is some $4$th root of unity, and $KK^\ast=(AA^T+BB^T)\otimes (2I_2)+AB^T\otimes(XY^\ast+YX^\ast)=2nI_{2n}$.
The significance of this observation is that it implies the following recent result.
Let $n\geq 1$ be an integer. If there exists a $\mathrm{BH}(n,6)$ matrix with no $\pm1$ entries, then there exists a $\mathrm{BH}(4n,2)$.
Combine Theorem \[turync\] with Theorem \[newkron\].
Classification of BH(21,3) matrices {#bh213x}
-----------------------------------
In this section we briefly report on our computational results regarding the $\mathrm{BH}(21,3)$ matrices. The classification of $\mathrm{BH}(18,3)$ matrices was reported earlier in [@cHAR2] and independently in [@cLAM], while several examples of $\mathrm{BH}(21,3)$ matrices were reported in [@cAKI].
The major difference between this case and the case of $\mathrm{BH}(16,4)$ matrices discussed in Section \[seccase1\] is that due to the lack of building blocks (such as a $\mathrm{BH}(7,3)$) for Kronecker-like constructions here one does not expect many solutions to be found, and therefore one may try to approach this problem by employing slightly different techniques.
First, we classified all $r\times 21$ orderly-generated rectangular orthogonal $3$rd root matrices with the second column pruning technique, and found exactly $1$, $1$, $12$, $145$, and $74013$ such matrices up to monomial equivalence for $r\in\{1,2,\dots,5\}$. After this, we considered each of these $5\times 21$ starting-point matrices, say $R$, one-by-one, and generated a set $V$ containing those row vectors which are lexicographically larger than the $5$th row of $R$, and which are orthogonal to each $5$ rows of $R$. Then, following ideas used in [@cSPE], we created the compatibility graph $\Gamma(R)$ on $|V|$ vertices, where two vertices, say $x$ and $y$, indexed by elements of $V$, are adjacent if and only if the rows $x\in V$ and $y\in V$ are pairwise orthogonal. With this terminology the task was then to decide if $\Gamma(R)$ contains a clique of size $16$. It turned out that in most cases it does not, and therefore we could reject the matrix $R$. The Cliquer software [@cNIS], based on [@cOS], was used in the current work to prune inextendible matrices in this way.
It was estimated that around $500$ CPU years is required to solve this case [@cHAR2]. However, we have completed this task in just over $18$ CPU days.
The number of $\mathrm{BH}(21,3)$ matrices is $72$ up to monomial equivalence.
In Table \[tableautgrps2\] we display the automorphism group sizes along with their frequencies.
Nonexistence results
--------------------
Nonexistence results for Butson matrices were obtained in [@cBAN], [@cBRO], [@delaun1], [@cLL], [@cWIN]. To the best of our knowledge the results presented in this section are not covered by any of these previous theoretical considerations.
In this section we briefly report on several exhaustive computational searches which did not yield any Butson matrices. Most of these computations were done in two different ways. First, we established nonexistence by using Cliquer [@cNIS], which heavily pruned the search tree, that is reduced the number of cases to be considered. This was very efficient due to the lack of complete matrices. Once nonexistence was established, we verified it during a second run, but this time without relying on Cliquer. This was done in order to be able to prudently document the search, and to avoid the use of external libraries.
\[nonex8\] There does not exist a $\mathrm{BH}(8,15)$ matrix.
The proof is computational. We have generated the $r\times 8$ orthogonal matrices with $15$th root of unity entries with the orderly algorithm using the second column pruning strategy, and we found $1$, $1$, $6$, and $0$ such matrices for $r\in\{1,2,3,4\}$, respectively. Therefore there exist no $\mathrm{BH}(8,15)$ matrices.
\[nonex11\] There does not exist a $\mathrm{BH}(11,q)$ matrix for $q\in\{10,12,14,15\}$.
The proof is along the lines of the Proof of Theorem \[nonex8\]. Refer to Table \[tablenonex11\] for the number of orderly-generated, rectangular orthogonal $r\times 11$ matrices with $q$th roots of unity (where $q\in\{10,12,14,15\}$) surviving the second column pruning strategy. In each of the four cases no such matrices were found for some $r\in\{1,\dots,11\}$, hence $\mathrm{BH}(11,q)$ matrices do not exist. For comparison, the case $\mathrm{BH}(11,6)$ is also presented.
\[nonex13\] There does not exist a $\mathrm{BH}(13,10)$ matrix.
First, we classified the $r\times 13$ orthogonal $10$th root matrices surviving the second column pruning strategy for $r\in\{1,2,3\}$, and found $1$, $10$, and $127556$ such matrices, respectively. As a second step, we used Cliquer [@cNIS] to see if any of these $3\times 13$ starting-point matrices can be completed to a $\mathrm{BH}(13,10)$. This task took 250 CPU days, but unfortunately no complete matrices turned up during the search. We note that the number of $4\times 13$ matrices with the relevant properties is exactly $45536950$, and millions of $5\times 13$ and hundreds of $6\times 13$ matrices were found during an incomplete search.
Open problems {#sect99}
=============
We conclude the paper with the following problems.
Extend Table \[tableBE\] further by classifying some of the remaining cases of $\mathrm{BH}(n,q)$ matrices in the range $n\leq 21$ and $q\leq 17$, and possibly beyond.
Continue the classification of real Hadamard matrices by extending the work [@cORR].
Classify all $\mathrm{BH}(36,2)$ matrices. Is it true that every $H\in\mathrm{BH}(36,2)$ has an equivalent form with constant row sum?
For context regarding Problem \[sscj\] we refer the reader to [@cKMat].
\[sscj\] Let $n$ and $q$ be positive integers, such that $n\nmid q^2$. Are there rectangular matrices $A$ and $B$ with elements in $\mathbb{Z}_q$ of size $n\times 2$ and $2\times n$, respectively, such that $L(H)=AB$ (modulo $q$) for some $H\in\mathrm{BH}(n,q)$?
For context regarding Problem \[asym\] we refer the reader to [@cLL] (see also Remark \[magicsumq\]).
\[asym\] Let $n,q\geq 2$, let $H\in\mathrm{BH}(n,q)$, and let $r_1,r_2\in\mathbb{Z}_q^n$ be distinct rows of $L(H)$. Can $r_1-r_2\in\mathbb{Z}_q^n$ represent an “asymmetric” minimal $n$-term vanishing sum of $q$th roots of unity? In other words, is it possible that $\mathrm{Sort}(r_1-r_2)$ is minimal in the sense that it has no constituent of $m$-term vanishing subsums for $m<n$, yet it is not of the form $[0,1,\dots, p-1]\in\mathbb{Z}_q^n$ where $p$ is some prime divisor of $q$?
Several $\mathrm{BH}(n,q)$ matrices with large $n$ and $q$ were constructed in [@cPET], leading to infinite, parametric families of complex Hadamard matrices of prime orders for $n\equiv 1\ (\mathrm{mod}\ 6)$.
Find new examples of $\mathrm{BH}(n,q)$ matrices of prime orders $n\equiv 5\ (\mathrm{mod}\ 6)$.
Problem \[nextproblem\] asks if a non-Desarguesian projective plane of prime order $p$ exists [@cHIR].
\[nextproblem\] Let $p$ be a prime number. Decide the uniqueness of $F_p\in\mathrm{BH}(p,p)$.
The next problem asks for the classification of $q$th root mutually unbiased bases [@cKA].
Let $n,q\geq 2$, and let $H,K\in\mathrm{BH}(n,q)$. Classify all pairs $(H,K)$ for which $(HK^{\ast})/\sqrt{n}\in\mathrm{BH}(n,q)$.
[10]{}
<span style="font-variant:small-caps;">S.S. Agaian</span>: Hadamard matrices and their applications, Springer-Verlag Berlin (1980).
<span style="font-variant:small-caps;">K. Akiyama, M. Ogawa, C. Suetake</span>: On $\mathrm{STD}_6[18, 3]$’s and $\mathrm{STD}_7 [21, 3]$’s admitting a semiregular automorphism group of order 9, [*Elec. J. Combin.*]{}, [**16**]{} \#R148 21 pp. (2009).
<span style="font-variant:small-caps;">T. Banica, J. Bichon, J.-M. Schlenker</span>: Representation of quantum permutation algebras, [*J. Funct. Anal.*]{}, [**257**]{} 2864–2910 (2009).
<span style="font-variant:small-caps;">B.W. Brock</span>: Hermitian congruence and the existence and completion of generalized Hadamard matrices, [*J. Combin. Theory A*]{}, [**49**]{} 233–261 (1988).
<span style="font-variant:small-caps;">W. Bruzda, W. Tadej, K. Życzkowski</span>: Web page for complex Hadamard matrices, <http://chaos.if.uj.edu.pl/~karol/hadamard/>
<span style="font-variant:small-caps;">A.T. Butson</span>: Generalized Hadamard matrices, [*Proc. Amer. Math. Soc.*]{}, [**13**]{} 894–898 (1962).
<span style="font-variant:small-caps;">B. Compton, R. Craigen, W. de Launey</span>: Unreal $\mathrm{BH}(n,6)$’s and Hadamard matrices, [*Des. Codes. Crypt.*]{}, [**79**]{} 219–229 (2016).
<span style="font-variant:small-caps;">R. Craigen, W. Holzmann, H. Kharaghani</span>: Complex Golay sequences: structure and applications, [*Discrete Mathematics*]{}, [**252**]{} 73–89 (2002).
<span style="font-variant:small-caps;">W. de Launey</span>: Generalised Hadamard matrices which are developed modulo a group, [*Discr. Math.*]{}, [**104**]{} 49–65 (1992).
<span style="font-variant:small-caps;">P. Diţă</span>: Some results on the parametrization of complex Hadamard matrices, [*J. Phys. A: Math. Gen.*]{}, [**37**]{} 5355 (2004).
<span style="font-variant:small-caps;">D.Ž. oković</span>: Good Matrices of Orders $33$, $35$ and $127$, [*JCMCC*]{}, [**14**]{} 145–152 (1993).
<span style="font-variant:small-caps;">R. Egan, D. Flannery, P. Ó Catháin</span>: Classifying Cocyclic Butson Hadamard Matrices. In: Colbourn C. (eds) Algebraic Design Theory and Hadamard Matrices. Springer Proceedings in Mathematics & Statistics, [**133**]{} 93–106 (2015).
<span style="font-variant:small-caps;">P.B. Gibbons, R. Mathon</span>: Enumeration of Generalized Hadamard Matrices of Order 16 and Related Designs, [*J. Combin. Des.*]{}, [**17**]{} 119–135 (2009).
<span style="font-variant:small-caps;">U. Haagerup</span>: Orthogonal maximal abelian $\ast$-subalgebras of the $n\times n$ matrices and cyclic $n$-roots, in: S. Doplicher (Ed.), et al., Operator Algebras and Quantum Field Theory, International Press, 296–322 (1997).
<span style="font-variant:small-caps;">M. Harada, C. Lam, A. Munemasa, V.D. Tonchev</span>: Classification of Generalized Hadamard Matrices $H(6,3)$ and Quaternary Hermitian Self-Dual Codes of Length $18$, [*Electronic J. Combinatorics*]{}, [**17**]{} \#R171 (2010).
<span style="font-variant:small-caps;">M. Harada, C. Lam, V.D. Tonchev</span>: Symmetric (4,4)-nets and generalized Hadamard matrices over groups of order 4, [*Des. Codes Crpyt.*]{}, [**34**]{} 71–87 (2005).
<span style="font-variant:small-caps;">A.S. Hedayat, N.J.A. Sloane, J. Stufken</span>: Orthogonal Arrays, Springer (1999).
<span style="font-variant:small-caps;">M. Hirasaka, K.-T. Kim, Y. Mizoguchi</span>: Uniqueness of Butson Hadamard matrices of small degrees, [*J. Discrete Algorithms*]{}, [**34**]{} 70–77 (2015).
<span style="font-variant:small-caps;">K. Horadam</span>: Hadamard matrices and their applications, Princeton University Press (2006).
<span style="font-variant:small-caps;">B. Karlsson</span>: BCCB complex Hadamard matrices of order $9$, and MUBs, [*Linear Algebra Appl.*]{}, [**504**]{} 309–324 (2016).
<span style="font-variant:small-caps;">P. Kaski, P.R.J. Östergård</span>: Classification algorithms for codes and designs, Springer Berlin, (2006).
<span style="font-variant:small-caps;">H. Kharaghani, B. Tayfeh-Rezaie</span>: Hadamard matrices of order 32, [*J. Combin. Des.*]{}, [**21:5**]{} 212–221 (2013).
<span style="font-variant:small-caps;">H. Kimura</span>: Classification of Hadamard matrices of order 28, [*Discrete Mathematics*]{}, [**133**]{} 171–180 (1994).
<span style="font-variant:small-caps;">D.E. Knuth</span>: The Art of Computer Programming: Generating All Tuples and Permutations, [**4**]{}:2 Addison–Wesley, 2010.
<span style="font-variant:small-caps;">M.N. Kolountzakis, M. Matolcsi</span>: Complex Hadamard matrices and the spectral set conjecture, [*Collectanea Mathematica*]{}, Vol. Extra. 281–291 (2006).
<span style="font-variant:small-caps;">C. Lam, S. Lam, V. Tonchev</span>: Bounds on the number of affine, symmetric, and Hadamard designs and matrices, [*Journal of Combinatorial Theory A*]{}, [**92**]{} 186–196 (2000).
<span style="font-variant:small-caps;">T.Y. Lam, K.H. Leung</span>: On vanishing sums of roots of unity, [*J. Algebra*]{}, [**224**]{} 91–109 (2000).
<span style="font-variant:small-caps;">P.H.J. Lampio</span>: Classification of difference matrices and complex Hadamard matrices, PhD Thesis, Aalto University, (2015).
<span style="font-variant:small-caps;">P.H.J. Lampio, P.R.J. Östergård</span>: Classification of difference matrices over cyclic groups, [*J. Stat. Plan. Inference*]{}, [**141**]{} 1194–1207 (2011).
<span style="font-variant:small-caps;">P.H.J. Lampio, F. Szöllősi, P.R.J. Östergård</span>: The quaternary complex Hadamard matrices of order 10, 12, and 14, [*Discrete Math.*]{}, [**313**]{} 189–206 (2013).
<span style="font-variant:small-caps;">D. McNulty, S. Weigert</span>: Isolated Hadamard matrices from mutually unbiased product bases, [*Journal of Mathematical Physics*]{}, [**53**]{} 122202 (2012).
<span style="font-variant:small-caps;">R. Nicoară</span>: Subfactors and Hadamard matrices, [*Journal of Operator Theory*]{}, [**64:2**]{} 453–468 (2010).
<span style="font-variant:small-caps;">S. Niskanen, P.R.J. Östergård</span>: Cliquer user’s guide, version 1.0, [*Technical Report T48*]{}, Communications Laboratory, Helsinki University of Technology, Espoo, (2003).
<span style="font-variant:small-caps;">W.P. Orrick</span>: Switching operations for Hadamard matrices, [*SIAM J. Discrete Math.*]{}, [**22**]{} 31–50 (2008).
<span style="font-variant:small-caps;">P.R.J. Östergård</span>: A fast algorithm for the maximum clique problem, [*Discrete Appl. Math.*]{}, [**120**]{} 197–207 (2002).
<span style="font-variant:small-caps;">M. Petrescu</span>: Existence of continuous families of complex Hadamard matrices of prime dimensions, [*PhD Thesis*]{}, UCLA (1997).
<span style="font-variant:small-caps;">R.C. Read</span>: Every one a winner, or how to avoid isomorphism search when cataloguing combinatorial configurations, [*Annals of Discrete Math.*]{}, [**2**]{} 107–120 (1978).
<span style="font-variant:small-caps;">J. Seberry</span>: A construction for generalized Hadamard matrices, [*J. Statistical Planning and Inference*]{}, [**4**]{} 365–368 (1980).
<span style="font-variant:small-caps;">J. Seberry</span>: Complex Hadamard matrices, [*Linear and multilinear algebra*]{}, [**1**]{} 257–272 (1973).
<span style="font-variant:small-caps;">E. Spence</span>: Classification of Hadamard matrices of order $24$ and $28$, [*Discrete Mathematics*]{}, [**140**]{} 185–243 (1995).
<span style="font-variant:small-caps;">F. Szöllősi</span>: A note on the existence of $\mathrm{BH}(19,6)$ matrices, [*Australasian J. Combin.*]{}, [**55**]{} 31–34 (2013).
<span style="font-variant:small-caps;">F. Szöllősi</span>: Mutually Unbiased Bases, Gauss sums, and the asymptotic existence of Butson Hadamard matrices, [*RIMS Kokyuroku*]{}, [**1872**]{} 39–48 (2014).
<span style="font-variant:small-caps;">F. Szöllősi</span>: On quaternary complex Hadamard matrices of small orders, [*Advances in Mathematics of Communications*]{}, [**5**]{} 309–315 (2011).
<span style="font-variant:small-caps;">F. Szöllősi</span>: Parametrizing complex Hadamard matrices, [*European J. Combin.*]{}, [**29**]{} 1219–1234 (2008).
<span style="font-variant:small-caps;">W. Tadej, K. Życzkowski</span>: A concise guide to complex Hadamard matrices, [*Open Systems & Information Dynamics*]{}, [**13**]{} 133–177 (2006).
<span style="font-variant:small-caps;">R.J. Turyn</span>: Complex Hadamard matrices. In: R. Guy (Ed.), Combinatorial Structures and their Applications, Gordon and Breach, New York, 435–437 (1970).
<span style="font-variant:small-caps;">R.F. Werner</span>: All teleportation and dense coding schemes, [*J. Phys. A: Math. Gen.*]{}, [**34**]{} 7081 (2001).
<span style="font-variant:small-caps;">A. Winterhof</span>: On the non-existence of generalized Hadamard matrices, [*J. Statistical Planning and Inference*]{}, [**84**]{} 337–342 (2000).
Butson matrices up to Hadamard equivalence
==========================================
Compare Table \[tableBE\] with Table \[tableHEQ\] and see Remark \[HADEQ\].
[^1]: See <https://wiki.aalto.fi/display/Butson>.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we will analyze the status of gauge freedom in quantum mechanics (QM) and quantum field theory (QFT). Along with this analysis comparison with ordinary QFT will be given. We will show how the gauge freedom problem is connected with the spacetime coordinates status — the very point at which the difficulties of QM begin. A natural solution of the above mentioned problem will be proposed in which we give a slightly more general form of QM and QFT (in comparison to the ordinary QFT) with noncommutative structure of spacetime playing fundamental role in it. We achieve it by reinterpretation of the Bohr’s complementarity principle on the one hand and by incorporation of our gauge freedom analysis on the other. We will present a generalization of the Bargmann’s theory of exponents of ray representations. It will be given an example involving time dependent gauge freedom describing non-relativistic quantum particle in nonrelarivistic gravitational field. In this example we infer the most general Schrödinger equation and prove equality of the (passive) inertial and the gravitational masses of quantum particle.'
author:
- |
Jaros[ł]{}aw Wawrzyckii[^1]\
Institute of Nuclear Physics, ul. Radzikowskiego 152,\
31-342 Kraków, Poland
title: NONCOMMUTATIVE SPACETIME AND QUANTUM MECHANICS
---
General Introduction
====================
Probably it will be helpful to give a brief outline of this paper providing the general line of reasoning. For details, however, the reader must consult the foregoing sections.
Of late there has been proposed a reformulation of the standard QM ([*J. Wawrzycki, math-ph/0301005; Comm. Math. Phys., to appear*]{}), which is slightly more general in comparison to the ordinary form of the theory. This reformulation has emerged in a natural way in description of a quantum particle in the non-relativistic gravitational field with time dependent gauge freedom. Remember, please, that the states of a physical system do not correspond bi-uniquely to unit vectors $\phi$ of the respective Hilbert space $\mathcal{H}$ but to the rays, say $\boldsymbol{\phi}
= \{e^{i\xi} \phi\}$ in the QM and QFT, where $\xi$ is an arbitrary real number.[^2] Observe now, please, that in ordinary (non-relativistic) QM, when using Schrödinger picture, one can go considerably further with this observation. Namely, two Schrödinger waves $\psi$ and $e^{i\xi(t)}\psi$ are indistinguishable even when $\xi$ depends on time, but one have to assume simultaneously that Schrödinger wave equation possess a time dependent gauge freedom. Let us recall that the integral defining the scalar product is over the space coordinates and one can take a time dependent factor over the integral sign. After this, however, the Schrödinger wave functions should constitute the appropriate cross sections of a Hilbert bundle $\mathcal{\mathbb{R}}\triangle \mathcal{H}$ over time $\mathbb{R}$. The representations $T_{r}$ of a covariance as well as a symmetry groups act in $\mathcal{\mathbb{R}}\triangle \mathcal{H}$ and their exponents $\xi$ in the formula $$\label{rayrep}
T_{r}T_{s} = e^{i\xi(r,s,t)}T_{rs},$$ do depend on time $t \in \mathbb{R}$. Thus, at first sight the natural assumption that two Schrödinger waves differing by a time dependent phase are equivalent leads to a rather strange construction, namely, the Hilbert bundle — an object much more involved then the Hilbert space itself. One can prove, however, that in the non-relativistic Galilean invariant theory, this more general assumption leads exactly to the ordinary QM. The whole structure degenerates due to the specific structure of the Galilean group. Moreover, in the less trivial case of a quantum particle in non-relativistic gravitational field, when the time dependent gauge freedom is indispensable, the results are quite interesting. Namely, one can infer the most general wave equation for that particle and prove equality of the inertial and gravitational mass, the results confirmed by experiments! In the last case the Milne group plays the role of the Galilean group.
This non-relativistic generalization possesses also a very natural relativistic extension which can be incorporated within QFT rather then QM. In QFT one can still go a step further along with this line of generalizing the quantum mechanical principles. Remember, please, that in the Fock construction the Fourier components of the classical field constitute the arguments of the wave function. Anyway the arguments have nothing to do with ordinary spacetime coordinates. In other words the spacetime coordinates are mere parameters or the so-called c-numbers in Heisenberg canonical field quantization — just like the time in ordinary non-relativistic QM. One should, thus, assume the two wave functions to be equivalent whenever they differ by a spacetime dependent wave function. But, when treating this assumption seriously, the wave functions should constitute the appropriate cross sections of a Hilbert bundle $\mathcal{M}\triangle \mathcal{H}$ over spacetime $\mathcal{M}$. Accordingly the representations $T_{r}$ of covariance or symmetry groups possess spacetime dependent exponents $\xi = \xi(r,s,p)$: $$T_{r}T_{s} = e^{i\xi(r,s,p)}T_{rs},$$ with $p\in \mathcal{M}$ ([*J. Wawrzycki, math-ph/0301005; Comm. Math. Phys., to appear*]{}).
In particular one is forced to extend the ordinary classification theory of exponents $\xi$ of representations acting in ordinary Hilbert space so as to embrace the above case of representation acting in a Hilbert bundle with spacetime dependent $\xi$. It can be viewed as a generalization of the Bargmann’s theory ([*V. Bargmann, Ann. Math [**54**]{}, 1, 1954*]{}) of exponents of ray representations acting in ordinary Hilbert space with spacetime-independent $\xi$.
The fact that the simpler form of the theory with time dependent gauge freedom gives the correct form of the Schrödinger equation thought the author to treat seriously also the generalization with spacetime dependent gauge freedom. The natural question emerges if one can find a simple connection of our generalization (those with spacetime dependent gauge freedom) to the ordinary QFT, for example to QED and if the connection is so transparent as in the case of time dependent gauge freedom. At this place one have to note that in particular any representation $T_{r}$ with spacetime dependent exponent $\xi = \xi(r,s,p)$ in (\[rayrep\]) makes any sense if the representation of the algebra of Canonical Commutation Relations (CCR) is [**reducible**]{} and does possesses a [**nontrivial**]{} center. The diagonal algebra over which the above representation of CCR algebra can be decomposed into irreducible components corresponds to the classical spacetime. How one can interpret this strange result along with the ordinary QFT in which, according to “Wightman’s axioms” ([*A. S. Wightman and L. G[å]{}rding, “Fields as Operator Valued Distributions in Relativistic Quantum” Theory, Ark. Fys. [**23**]{}, No. 13, 1964*]{}), the algebra generated by quantum field operators (“smeared with appropriate test functions”) is [**irreducible**]{}?
Answer to this question is by no means trivial. In particular the situation is much less trivial then in the case of the theory with time dependent gauge freedom in non-relativistic QM. In our opinion one is forced to reconsider the fundamental principles of QM and QFT in answering the question. Of late the long-lasting dispute concerning the interpretation of QM is at its renaissance again. We will not go into detail of this dispute but we feel that something is missing in the standard picture and at least some points of the criticism are justified. We assume that the standard QM describes correctly the situation in which we have an atomic system measured with a macroscopic apparatus. In the standard QM the Bohr-Heisenberg cut between the system and the apparatus may range between them, and it is irrelevant how big is the system. It is important only that the cut is “placed” somewhere between the system and the apparatus (in this sense the Bohr-Heisenberg cut does not form any “normal” real physical boundary). This unavoidably implies that a macroscopic body can be in a (quantum) superposition of states with macroscopically different parameters like the center-of-mass position. On the other hand, so long as “big” and “small” are merely relative concepts, it is no help to one who wishes to account for the ultimate structure of matter. We agree with Dirac, that QM should be endowed with ideas in such a way as to give an absolute meaning to the words “small” and “big”. In the common opinion QM is the theory which incorporates the idea of absolute scale of action — the Planck’s quantum of action $\hslash$. Paradoxically, in ordinary QM the only idea, namely that of Bohr-Heisenberg cut, which separates the “small” from the “big” is purely relative such that we lose the possibility of introducing the absolute scale mentioned above. This is the element which in our opinion is missing in the standard interpretation of QM. Thus, we assume that QM with the standard interpretation works good but only within some limits, where the observed system involves only a few quantum particles whereas the apparatus constitute a macroscopic body consisting of an enormous number of quantum particles. But then the QFT is the appropriate scheme within which the systems consisting of many particles are naturally incorporated. Therefore, one has to be careful in applying the QM principles in QFT when the number of particles is too big. The application is justified only if one works within the neighborhood of the vacuum state — which does actually take place in practice e.g. when evaluating the [*Lamb shift*]{} or the [ *anomalous magnetic moment*]{}. For the states far removed from the vacuum one has to preserve an open mind on the ordinary QM principles. We maintain that our generalization of QFT is applicable if the number of particles in the system is large, just in opposite to the ordinary QFT applicable when the number is small. Indeed. The “Fock space” $\mathcal{H}_{F}$ is the direct sum of all $N$-particle spaces $\mathcal{H}_{N} =\mathcal{H}_{1}^{\otimes_{S} N}$ (resp. $\mathcal{H}_{N} =
\mathcal{H}_{1}^{\otimes_{A} N}$) i.e. the symmetrized $N$-fold products of the one-particle Hilbert space $\mathcal{H}_{1}$ (resp. anti-symmetrized $N$-fold products): $$\mathcal{H}_{F} = {\bigoplus}_{N=0}^{\infty} \mathcal{H}_{N}.$$ A state vector $\psi$ in $\mathcal{H}_{F}$ is an infinite hierarchy of symmetric (resp. anti-symmetric) wave functions $$\psi = \left( \begin{array}{ll}
c & \\
\psi_{1}(x) & \\
\psi_{2}(x_{1},x_{2}) & \\
\ldots &
\end{array} \right)$$ where we have written the argument $x$ to denote both position and spin component. $\psi_{N}$ is the probability amplitude for finding just $N$ particles and those in specified configuration. The complex number $c$ is the probability amplitude to find the vacuum. It is important to note that the argument $x$ in one particle state $\psi_{1}$ contains the space coordinates whereas $x_{1}, \ldots , x_{N}$ in the $N$-particle state have nothing to do with spacetime coordinates being rather the configuration coordinates. In the case of the free field when the number of particles $N$ is conserved our argument that $\psi$ and $e^{i\xi(p)}\psi$ are equivalent is valid (provided $N>1$). Remember, please, that $p$ stands for spacetime point. In this case the probability amplitude to find a number of particles different from $N$ is zero (in particular $\psi_{1} = 0$).[^3] However, relativistic QFT with interaction is a theory describing the phenomenon of creation and annihilation of particles as indispensable effects so as the total number of particles cannot be conserved. Thus, one cannot exclude the amplitude $\psi_{1}$ when working around the vacuum state. This one-particle amplitude is negligible only if the total number of particles is large. In this case, therefore $\vert(e^{i\xi(p)}\psi,\psi')\vert = \vert e^{i\xi(p)}(\psi,\psi')\vert =
\vert(\psi,\psi')\vert $ so that $\psi$ and $e^{i\xi(p)}\psi$ are equivalent; please, compare the first two footnotes — one can take the factor $e^{\xi(p)}$ over the argument of the inner product in this case. This shows that our generalization becomes adequate when the system consists of an appropriately large number of particles, just in the situation when the ordinary QFT is expected to be somewhat misleading.
Yet the (more) ultimate general theory, the scientist should intend to find, is expected to be adequate to account for systems of a few particles as well as systems consisting of an enormous number of them. We are certain that the ordinary QFT laws[^4] works perfectly in these cases of small values $N$ of particles in the system. Should it be possible at all to obtain a theory which can give correct answers also in the cases of small $N$-values? In the cases of small $N$ the ordinary QFT could act as the code-book — namely, one should pick up those general laws or theorems of QFT which can be formulated in terms independent of the actual number of particles. As such theorem which should guide us in further research we take the miracle that: [**1) Canonical quantization of a free scalar field leads to Fock space and an interpretation of states in terms of particle configurations**]{}, so it explains the wave-particle duality lying at the roots of QM. This theorem contains the general kinematical information that the states of quantized field serve as configuration space for quantum particles which may occupy them. As such this fact does not depend directly on the actual value of $N$. On the other hand, when the number $N$ of particles is large the classical theory is correct and may serve as a guide in this regime exactly as QFT does when $N$ is small, provided the laws we pick up to guide us do not depend on the number of degrees of freedom[^5], especially they should not depend on the fact if there is infinite number of them as in classical field or finite as e.g. for a rigid classical body. As the second guiding theorem we take the following theorem: [**2) The configuration space for the classical body consists of various space[^6] positions the body or its parts may occupy**]{} — a fact which seems to be independent of the actual number of degrees of freedom involved, i.e. independent of the number of independent parts. Next, we should observe that the algebraic formulation of QM and QFT is an approach in which the Hilbert space plays a secondary role and by this fact the number of particles $N$ as well as the number of degrees of freedom plays a very indirect role in it. Thus, the approach is the one we should work with in our research. Comparing now our guiding theorems 1) and 2) one can infer that the (noncommutative) quantum algebra of quantum field operators[^7] plays the role of noncommutative space for quantum particles, just like the classical space (i.e. space-like surface of spacetime) does for a classical body. In this way we arrive at the result that the space the quantum particles live in is a noncommutative space (in the sense of [*A. Connes, Noncommutative Geometry, Acad. Press 1994*]{}) with the noncommutative algebra of quantum field operators corresponding to the noncommutative space.
Still, it would be much better if we were able to find the counterpart of the whole classical spacetime and not only the counterpart of space.[^8] For this purpose let us note that the points of classical spacetime go into play when considering the time evolution of space position of the classical body. [*Per analogiam*]{}, if we are to have any hope to find the quantum counterpart of spacetime we must use the quantum substitute for the classical space position evolution. The state in the Fock space $\mathcal{H}_{F}$ in the Heisenberg picture (which we have used above) does not contain any information about time evolution. One could suppose that the simple replacement of the Heisenberg picture with the Schrödinger picture immediately resolves the problem. Unfortunately this is by no means the case. One has to recall that we still intend to find some general principles valid in general irrespectively if the number $N$ of particles is large or not. Unfortunately the standard connection between the two pictures Schrödinger’s and Heisenberg’s fails down when $N$ is large in general. Remember, please, that when $N$ is large the gauge freedom problem mentioned above goes into play: the two waves $\psi$ and $e^{i\xi(p)}\psi$ are to be considered equivalent. One can prove that instead of normal Schrödinger picture with ordinary Schrödinger waves (Hilbert space vectors parametrically dependent on time) one has to deal with cross sections of an appropriate Hilbert bundle $\mathcal{M}\triangle\mathcal{H}$ over whole spacetime $\mathcal{M}$. But, when using the Hilbert bundle $\mathcal{M}\triangle\mathcal{H}$ instead of $\mathcal{H}$, the respective algebra of quantum operators acting in $\mathcal{M}\triangle\mathcal{H}$ has to be reducible, compare a subsequent section where we give arguments for this fact. The spectrum of the center of this algebra is just equal to the classical spacetime $\mathcal{M}$. According to the above discussion this algebra corresponds to the noncommutative spacetime — an object plying the same role for quantum particles as the classical spacetime does for classical bodies.
In this way we have revealed the wave-particle duality as a manifestation of the noncommutative structure of spacetime. The noncommutative algebra $\mathcal{A}$ corresponding to the noncommutative spacetime is obtained as the smallest von Neumann algebra $\mathcal{A}^{CCR} \vee \mathcal{A}^{CCR}_{1}$ containing the von Neumann canonical commutation algebra $\mathcal{A}^{CCR}$ of field operators and the appropriate maximally Abelian subalgebra $\mathcal{A}^{CCR}_{1}$ in the commutant of $\mathcal{A}^{CCR}$. The center $\mathcal{A}^{CCR}_{1}$ (diagonal algebra) of $\mathcal{A}$ should not be trivial and corresponds to the classical spacetime.
Apparently, one can formulate a serious objection against our conclusion that quantum particle lives in a spacetime with a noncommutative structure, with the structure closely related to the algebra of quantum operators of the field corresponding to the particle. Apparently one can say that each type of particle lives in its own spacetime related to the corresponding type of field — which is a very strange idea rather. We maintain that the strange effect of many coexisting spacetimes is apparent. Indeed. One should note at this place that the representation of the algebra of quantum field operators $\mathcal{A}^{CCR}$ we are interested in should be reducible[^9]. As such the algebra cannot act in the ordinary Fock space. The space is no longer a direct sum of symmetrized (resp. anti-symmetrized) tensor products of one-particle Hilbert spaces, in which the tensor product of $N$ factors occurs once and only once for each natural $N$. This structure of Hilbert space is disturbed in our case. Let us recall that the symmetrization (resp. anti-symmetrization) is deeply connected with the [**irreducibility**]{} of the quantum algebra ([*H. Weyl, Gruppentheorie und Quantenmechanik, Leipzig, Verlag von S. Hirzel, 1931*]{}). In our case the representation is [**reducible**]{} so that the symmetrized and anti-symmetrized products may appear within the same representation space! In the standard theory if the system is in symmetrized state it remains symmetrized forever, regardless of what influences may act upon it — bosons and fermions do not mix each other. In our case (in the limit of large $N$) the situation is substantially different — the symmetric and anti-symmetric states do mix each other. Therefore, the algebra $\mathcal{A}$ is much more universal object in comparison to any particular operator algebra of any specified kind of field in standard QFT. It, therefore, seems to be capable as to account for different kinds of particles. Moreover, it is an advantage that in the space of the representation of our algebra the symmetrization (resp. anti-symmetrization) is disturbed. Remember, please, that the symmetrization (resp. anti-symmetrization) reflects the the Bose-Einstein (resp. Fermi-Dirac) correlations or the so-called entanglement of quantum states. Thus, in our case the entanglement of states is disturbed which may correspond to the fact that in practice for macroscopic bodies (when $N$ is large) the entanglement is negligible. It should be stressed that this fact is very difficult to explain within the ordinary QM and QFT.
Some other arguments concerning the apparent problem of many coexisting spacetimes corresponding to the various kinds of fields and particles seem advisable. The generalization of the wave-particle duality to a field-particle duality has dominated thinking in quantum theory for decades and has been heuristically useful in the development of elementary particle theory. Yet the belief in field-particle duality as a general principle, the idea that to each particle there is a corresponding field and to each field a corresponding particle has also been misleading. The number and the nature of different basic fields is related to the charge structure, not to the empirical spectrum of particles. For example, in the presently favored gauge theories the fields are the carriers of charges like colour and flavour but are not directly associated to observed particles like electrons. The biunique field-particle correspondence is therefore broken. Moreover, the spectacular issue of Connes and Lott ([*Nucl. Phys. [**18B**]{}, 29, 1990*]{}, see also [*A. Connes, loc. cit.*]{}) that the so much a long list of fields in the effective standard model can be considerably reduced with the cost of some noncommutative structure of spacetime involved. They considered the effective standard model, i.e. their analysis was confined to the classical context. The Lagrangean of this effective theory is a combination of a five terms representing to independent contributions. Connes and Lott showed that this artificially complicated Lagrangean is a natural generalization of the Maxwell-Dirac Lagrangean providing the appropriate noncommutative structure of spacetime.
Yet one has to be careful, however, and treat the specific structure of $\mathcal{A}$ mentioned above as a prototype rather then as the ultimate word one can say on this subject.
The fact that the concepts of “spacetime coincidence” and “observation” do require a thorough revision in QM was immediately noticed by the very founders of QM. It was especially evident in Bohr’s writings. In the year 1925 he described the situation in the words: “From these results it seems to follow that, in the general problem of the quantum theory, one is faced not with a modification of the mechanical and electrodynamical theories describable in terms of the usual physical concepts, but with an essential failure of the pictures in space and time on which the description of natural phenomena has hithero been based.” (*Nature*, [**116**]{}, p. 535.) Three years later he formulated the complementarity principle. As emphasized by Einstein every observation or measurement ultimately rests on the coincidence of two independent events at the same spacetime point. Now the quantum postulate implies that any observation of atomic phenomena will involve an interaction with the agency of observation not to be neglected. On one hand – Bohr stresses – the definition of the state of a physical system, as ordinary understood, claims the elimination of all external disturbances. But in that case, according to the quantum postulate, any observation will be impossible, and, above all, the concepts of space and time lose their immediate sense. On the other hand – he concludes – if in order to make observation possible we permit certain interactions with suitable agencies of measurement, not belonging to the system, an unambiguous definition of the state of the system is naturally no longer possible, and there can be no question of causality in the ordinary sense of the word (Heisenberg uncertainty principle). Concluding, we have the complementarity alternative: *Either* quantum particle is describable in terms of space and time but the state of particle is not well defined *or* the state of quantum particle is well defined but its description in terms of space and time impossible. We propose to interpret this complementarity principle as indicating that the spacetime event, in the classical sense of the word, cannot be ascribed to quantum particle according to our analysis presented above. The quantum particle lives in a noncommutative spacetime. The event in (classical) spacetime can be ascribed to a body which we would like to call classical – made of many quantum particles.
Gauge Symmetry
==============
Gauge Freedom and Heisenberg Commutation Relations
--------------------------------------------------
Let us back to the ordinary QM applied to a system with finite number of degrees of freedom. To be consequent we have to restrict our consideration to non-relativistic case for the moment — please, remember that any relativistic quantum system has to constitute a quantum field system with infinite number of degrees of freedom. The quantum states of our system constitute a Hilbert space $\mathcal{H}$. The inner product $(\centerdot,\centerdot)$ of the space is an relevant part of structure plying important role in the physical interpretation. Namely, the only contribution of the the vector $\varphi$ of the Hilbert space $\mathcal{H}$ to any measurable effects is inscribed into the set of numbers $$\frac{\vert(\phi, \varphi)\vert^{2}}{(\phi,\phi)(\varphi, \varphi)},
\, \, \, \phi \in \mathcal{H}.$$ Suppose we have an ideal source which prepares an ensemble in apure state, described by an element of $\mathcal{H}$. In particular if in addition we have an ideal detector, giving a yes-answer in a pure state $\varphi$ and the answer no in the orthogonal complement to $\phi$, then the probability of detecting an event in this set up of source and detector is given by $$\frac{\vert(\phi, \varphi)\vert^{2}}{(\phi,\phi)(\varphi, \varphi)}.$$ Therefore, two vectors $\varphi$ and $e^{i\xi}\varphi$ differing by a mere constant phase are indistinguishable. This is very important for the whole structure of quantum theory. For example, any kind of algebra of some quantities, like the algebra of classical observables (functions $f(p,q)$ on the phase space with ordinary point wise operations) — which is commutative — will not necessary be commutative if unitary represented as operator algebra acting in the Hilbert space of states. Let us explain how it follows from the constant-phase-equivalence of the two vectors $\varphi$ and $e^{i\xi}\varphi$, $\xi$ being a constant real number. Consider the set of invertible elements of the algebra in question, which constitutes a commutative group in our case. Because of the constant-phase-equivalence the two operators $A$ and $e^{i\xi}A$ are equivalent giving the same average values and having the same spectra. Therefore in general the relation $$AB = e^{i\xi(A,B)}BA$$ holds for representations acting in $\mathcal{H}$ instead of $$AB = BA.$$ Such a representation is mostly called [*ray representation*]{}. Let us consider representation $U_{i}$ and $V_{i}$ of the $2n$-dimensional Abelian group of phase coordinates $(p_{1}, \ldots p_{n},q_{1},
\ldots q_{n})$ in the classical phase space. The algebra of classical phase coordinates is generated by the one-parameter groups corresponding to $p_{i}$ and $q_{i}$ which should hold also for the quantum representation. If one assumes the unitary representation to be [**irreducible**]{} then one gets the relations $$\begin{aligned}
U_{i}V_{k} = & e^{i\xi(p_{i},q_{k})}V_{k}U_{i}, \nonumber \\
U_{i}U_{k} = & e^{i\xi(p_{i},p_{k})}U_{k}U_{i}, \nonumber \\
V_{i}V_{k} = & e^{i\xi(q_{i},q_{k})}V_{k}V_{i}, \nonumber \end{aligned}$$ equivalent to the Heisenberg commutation relations and moreover the representation is unique up to unitary equivalence. That is, denoting the generators of $U_{i}$ and $V_{k}$ by $P_{i}$ and $Q_{k}$ we get from the above relations the following result $$\begin{aligned}
Q_{i}P_{k} - P_{k}Q_{i} = & i\delta_{ik}, \nonumber \\
Q_{i}Q_{k} - Q_{k}Q_{i} = & 0, \nonumber \\
P_{i}P_{k} - P_{k}P_{i} = & 0, \nonumber \end{aligned}$$ i. e. the Heisenberg commutation relations. This is a well known result noticed by Weyl ([*loc. cit.*]{}) investigated further by v. Neumann, Rellich and Stone. The part of the statement concerning the uniqueness is mostly called von Neumann’s uniqueness theorem.
From the mathematical point of view this Weyl’s result is nothing else but a special case of Bargmann’s theory of ray representations ([*loc. cit.*]{}) applied to the Abelian group of canonical coordinates (i.e. to the translation group in the phase space).
The requirement that the set of $2n$ operators $P_{1} \ldots P_{n},Q_{1},
\ldots ,Q_{n}$ should be irreducible is very important in the proof of the above statement of Weyl but also in the whole of QM laws! This postulate is to be added to the Heisenberg commutation rules as an essential supplement. For example in QM there is a standard rule for description of systems composed of several individual parts (i. e. subsystems). Suppose a system $A$ to be composed of two parts $B$ and $C$. The states of the system and the subsystems are to be represented by vectors in the Hilbert spaces $\mathcal{H}_{A}$, $\mathcal{H}_{B}$ and $\mathcal{H}_{C}$ respectively. Then, the general rule says first that $\mathcal{H}_{A} =
\mathcal{H}_{B}\otimes\mathcal{H}_{C}$ is the ordinary tensor product of $\mathcal{H}_{B}$ and $\mathcal{H}_{C}$. Second, in accord to the rule, the only hermitian operators which have physical significance depend symmetrically on the two subsystems. Let us consider the very special situation of this kind when both $A$ and $B$ are quantum particles of the same kind. We, thus, have 12 matrices $P_{1} \ldots P_{6},Q_{1},
\ldots ,Q_{6}$ fulfilling Heisenberg commutation rules. In general they are reducible in accordance with the decomposition $$\mathcal{H}_{A} =
\mathcal{H}_{B}\otimes\mathcal{H}_{B} = \mathcal{H}_{B}\otimes_{S}\mathcal{H}_{B}
+ \mathcal{H}_{B}\otimes_{A}\mathcal{H}_{B}$$ of the Hilbert space of the system $A$ into the symmetric and anti-symmetric tensor product of the Hilbert spaces corresponding to the particles $B$ and $C=B$ (remember, please, that all $P_{i}$ and $Q_{k}$ depend symmetrically on subsystems $B$ and $C=B$). Experimental evidence tells us that there are only two kind of particles those with states either in $\mathcal{H}_{B}\otimes_{S}\mathcal{H}_{B}$ or $\mathcal{H}_{B}\otimes_{A}\mathcal{H}_{B}$. Thus the reducibility of the set of operators $P_{1} \ldots P_{n},Q_{1}, \ldots ,Q_{n}$ would contradict experiment.
Having given this comment on irreducibility let us back to the constant-phase equivalence and the Weyl’s scheme within which we have obtained the Heisenberg commutation rules from this equivalence. It is remarkable that the quantization of the problem of several particles also falls within this general scheme even for fermions. In dealing with it we are interested in that Abelian group whose basic elements $p_{\alpha}, q_{\alpha}$ are all of order 2 in the fermionic case. Such a group consists of the totality of monomials $$p_{1}^{n_{1}}q_{1}^{n_{2}}p_{2}^{n_{3}}q_{2}^{n_{4}}\centerdot \ldots,$$ where $n_{k} = 1$ or 0 and $n_{1} + n_{2} + \ldots \leq N$ (total number of particles), compare ([*Weyl, loc. cit.*]{}). In virtue that the two representations $P_{\alpha}, Q_{\alpha}$ and $P'_{\alpha} = e^{\beta(p_{\alpha})}P_{\alpha}, Q'_{\alpha}
= e^{\beta(q_{\alpha})}Q_{\alpha}$ are equivalent the gauge (exponent $\beta$) can be so chosen that the corresponding operators of the irreducible ray representation satisfy the anti-commutation rules $$\begin{aligned}
Q_{i}P_{k} + P_{k}Q_{i} = & i\delta_{ik}, \nonumber \\
Q_{i}Q_{k} + Q_{k}Q_{i} = & 0, \nonumber \\
P_{i}P_{k} + P_{k}P_{i} = & 0. \nonumber \end{aligned}$$
In passing to the relativistic QM we have to account for systems with infinite degrees of freedom. To obtain the Heisenberg rules within the same general scheme presents a very sophisticated mathematical problem. In the case of a free field, however, it is quite possible. But now the von Neumann uniqueness theorem no longer holds. There is a vast of inequivalent irreducible representations of Heisenberg commutation rules. We have to impose additional condition that the representation possesses a cyclic vector with some peculiar properties, namely the vacuum.
Because of this results one can see that the constant-phase equivalence of Hilbert space vectors lies at the very heart of QM. On the other hand there is a very natural way of generalizing this Weyl’s scheme. For example in the non-relativistic QM we can use the Schrödinger picture. But then the two wave functions $\psi$ and $e^{i\xi(t)}\psi$ are equivalent. Of course one has to assume that the Schrödinger equation is endowed with the appropriate time dependent gauge freedom, but this is very realistic and one cannot exclude such situation, compare the next subsection. For example when considering wave equation in gravitational field (of course non-relativistic field in this case) the time dependent gauge freedom is unavoidable. Because the constant-phase equivalence is so important one cannot ignore it and the consequences of this generalization should be investigated. Even more, one can go still a step further if starting from QFT instead of QM, as we argued in the first section of this Chapter.
It is,thus, justified to think of this phase-equivalence as of a special kind of gauge symmetry. In this way the Weyl’s constant-phase equivalence is a special kind of gauge symmetry, namely the constant-phase symmetry —the simplest possible one.
Time Dependent Gauge Freedom {#Time}
----------------------------
In this subsection we carry out general analysis of the representation $T_{r}$ of a covariance group, and compare it with the representation of a symmetry group. We also describe correspondence between the space of wave functions $\psi(\vec{x},t)$ and the Hilbert space. Here the analysis is performed in the non-relativistic case.
Before we give the general description, it will be instructive to investigate the problem for a free particle in the flat Galilean spacetime. The set of solutions $\psi$ of the Schrödinger equation, which are admissible in Quantum Mechanics, is precisely given by $$\psi(\vec x,t)=(2\pi)^{-3/2} \int \varphi(\vec k)e^{-i\frac{t}{2m}\vec{k}
\centerdot \vec k+i\vec k\centerdot\vec x} \, {{\mathrm{d}}}^{3}{k},$$ where $p=\hslash k$ is linear momentum and $\varphi(\vec{k})$ is any square integrable function. The functions $\varphi$ (wave functions in the “Heisenberg picture”) form a Hilbert space $\mathcal{H}$ with the inner product $$(\varphi_{1}, \varphi_{2})=\int \varphi_{1}^{*}(\vec{k})\varphi_{2}(\vec{k}) \,
{{\mathrm{d}}}^{3}{k}.$$ The correspondence between $\psi$ and $\varphi$ is one-to-one.
In general, however, the construction fails if the Schrödinger equation possesses nontrivial gauge freedom. Let us explain it. For example, the above construction fails for the non-relativistic quantum particle in the curved Newton-Cartan spacetime. Besides, in this spacetime we do not have any plane wave, see ([*J. Wawrzycki, Int. Jour. of Theor. Phys. [**40**]{}, 1595 (2001)*]{}). Thus, there does not exist any natural counterpart for the Fourier transform. However, we do not need to use the Fourier transform. What is the role of the Schrödinger equation in the above construction of $\mathcal{H}$? Please note that in general $$\Vert \psi \Vert^{2} \equiv
\int \psi^{*}(\vec{x},0)\psi(\vec{x},0) \, {{\mathrm{d}}}^{3}{x} = (\varphi,\varphi)$$
$$=\int \psi^{*}(\vec{x},t)\psi(\vec{x},t) \, {{\mathrm{d}}}^{3}{x}.$$
This is in accordance with the Born interpretation of $\psi$. Namely, if $\psi^{*}\psi(\vec{x},t)$ is the probability density, then $$\int \psi^{*}\psi \, {{\mathrm{d}}}^{3}{x}$$ has to be preserved over time. In the above construction, the Hilbert space $\mathcal{H}$ is isomorphic to the space of square integrable functions $\varphi(\vec{x})\equiv \psi(\vec{x},0)$, namely the set of square integrable initial data for the Schrödinger equation, cf. e.g. ([*D. Giulini, [*States, Symmetries and Superselection*]{}, in: [*Decoherence: Theoretical, Experimental and Conceptual Problems*]{}, (Lecture Notes in Physics, Springer Verlag 2000), page 87.*]{}). The connection between $\psi$ and $\varphi$ is given by the time evolution operator $U(0,t)$ (equivalently by the Schrödinger equation): $$U(0,t)\varphi=\psi.$$ The correspondence between $\varphi$ and $\psi$ has all formal properties, such as in the Fourier construction above. Of course, the initial data for the Schrödinger equation do not cover the whole Hilbert space $\mathcal{H}$ of square integrable functions, but the time evolution given by the Schrödinger equation can be uniquely extended over the whole Hilbert space $\mathcal{H}$ by the unitary evolution operator $U$.
The construction can be applied to the particle in the Newton-Cartan spacetime. As we implicitly assumed, the wave equation is such that the set of its admissible initial data is dense in the space of square integrable functions (we need this for the uniqueness of the extension). Because of the Born interpretation, the integral $$\int \psi^{*}\psi \, {{\mathrm{d}}}^{3}{x}$$ has to be preserved over time. Let us denote the space of the square-integrable initial data $\varphi$ on the simultaneity hyperplane $t(\vec{x},t)=t$ by ${\mathcal{H}}_{t}$. Then, the evolution is an isometry between ${\mathcal{H}}_{0}$ and ${\mathcal{H}}_{t}$. But such an isometry has to be a unitary operator, and the construction is well defined, *i.e.* the inner product of two states corresponding to the wave functions $\psi_{1}$ and $\psi_{2}$ does not depend on the choice of ${\mathcal{H}}_{t}$. Let us mention that the wave equation has to be linear in accordance with the Born interpretation of $\psi$ (since any unitary operator is linear the time evolution operator is linear as well). The space of wave functions $\psi(\vec{x},t) = U(0,t)\varphi(\vec{x})$ isomorphic to the Hilbert space ${\mathcal{H}}_{0}$ of $\varphi$’s is commonly called the “Schrödinger picture”.
However in general, the connection between $\varphi(\vec{x})$ and $\psi(\vec{x},t)$ is not unique if the wave equation possesses a gauge freedom. Namely, let us consider two states $\varphi_{1}$ and $\varphi_{2}$ and ask when these two states are equivalent, and indistinguishable. The answer is that they are equivalent if $$\vert(\varphi_{1},\varphi)\vert \equiv \Big\vert\int \psi_{1}^{*}(\vec{x},t)
\psi(\vec{x},t) \, {{\mathrm{d}}}^{3}{x}\Big\vert = \vert(\varphi_{2},\varphi)\vert$$
$$\label{row}
\equiv \Big\vert\int \psi_{2}^{*}(\vec{x},t)\psi(\vec{x},t) \,
{{\mathrm{d}}}^{3}{x}\Big\vert,$$
for any state $\varphi$ from $\mathcal{H}$, or for all $\psi=U\varphi$ ($\psi_{i}$ are defined to be = $U(0,t)\varphi_{i}$). By substituting $\varphi_{1}$ and then $\varphi_{2}$ for $\varphi$ and making use of the Schwarz’s inequality, one gets: $\varphi_{2}=e^{i\alpha}\varphi_{1}$, where $\alpha$ is any constant[^10]. The situation for $\psi_{1}$ and $\psi_{2}$ is however different. In general, condition (\[row\]) is fulfilled if $$\psi_{2}=e^{i\xi(t)}\psi_{1}$$ and the phase factor can depend on time. Of course, this has to be consistent with the wave equation, that is, together with a solution $\psi$ of the wave equation, the wave function $e^{i\xi(t)}\psi$ is also a solution of the appropriately gauged wave equation. *A priori* one cannot exclude the existence of such a consistent time evolution. This is not a new observation, as it was noticed by John von Neumann[^11], but it seems that it has never been deeply investigated (probably because the ordinary non-relativistic Schrödinger equation has gauge symmetry with constant $\xi$). The space of waves $\psi$ describing the system cannot be reduced in the above way to any fixed Hilbert space $\mathcal{H}_{t}$ with a fixed $t$. So, the existence of the nontrivial gauge freedom leads to the following
[**Hypothesis**]{}. *The two waves* $\psi$ *and* $e^{i\xi(t)}\psi$ *are quantum-mechanically indistinguishable*.
Moreover, we are obliged to use the whole Hilbert bundle $\mathcal{R}\triangle
\mathcal{H}: t \to \mathcal{H}_{t}$ over the time instead of a fixed Hilbert space $\mathcal{H}_{t}$, with the appropriate cross-sections $\psi$ as the waves (see the next section for details).
Let us consider now an action $T_{r}$ of a group $G$ in the space of waves $\psi$. Before we infer some consequences of the assumption that $G$ is a symmetry group, we need to state a:
[**Classical-like postulate**]{}. *Group $G$ is a symmetry group if and only if the wave equation is invariant under the transformation $x' = rx, r\in G$ of independent variables and the transformation $\psi'=
T_{r}\psi$ of the wave function*.
The above postulate is indeed commonly accepted in Quantum Mechanics even when the gauge freedom is not excluded. But it is a mere application of the symmetry definition for a classical field equation applied to the wave equation without any change. The wave $\psi$ is not a classical quantity, such as e.g. electromagnetic intensity. The above [**Hypothesis**]{} is not true for classical fields, and we have to be careful in forming the appropriate postulate for the wave equation compatible with the [**Hypothesis**]{}. Namely, the two wave equations differing by a mere gauge are indistinguishable. We call them *gauge-equivalent*. It is therefore natural to assume the
[**Quantum postulate**]{}. *Group $G$ is a symmetry group if and only if the transformation $x' = rx, r\in G$ of independent variables and the transformation $\psi'= T_{r}\psi$ of the wave function transform the wave equation into a gauge-equivalent one*.
Please note that not all possibilities admitted by the [**Hypothesis**]{} are included in the [**Classical-like postulate**]{}.
From the [**Classical-like postulate**]{} it follows that $\psi$ as well as $T_{r}\psi$ are solutions of exactly the same wave equation, in view of the invariance of the equation. Therefore, $\psi$ and $T_{r}\psi$ belong to the same “Schrödinger picture”, so that $$T_{r}T_{s}\psi = e^{i\xi(r,s)}T_{rs}\psi,$$ with $\xi = \xi(r,s)$ independent of time $t$! This is in accordance with the known theorem that
If $G$ is a symmetry group, then the phase factor $\xi$ should be time-independent.
But if we start from the [**Quantum postulate**]{}, we obtain instead $$\label{rowt}
T_{r}T_{s}\psi=e^{i\xi(r,s,t)}T_{rs}\psi$$ and get
[**Theorem 1’**]{} *If $G$ is a symmetry group, then the phase factor $\xi = \xi(r,s,t)$ is time-dependent in general*.
In this paper we propose to accept the [**Quantum postulate**]{}, which is compatible with the [**Hypothesis**]{}, and is more in spirit of Quantum Mechanics than the [**Classical-like postulate**]{}. It should be noted that in the special case when gauge freedom degenerates to the case of constant phase, the [**Quantum postulate**]{} is equivalent to to the [**Classical-like postulate**]{}.
We shall resolve the following paradox. Namely, a natural question arises why the phase factor $e^{i\xi}$ in (\[rowt\]) is time-independent for the Galilean group (even when the Galilean group is considered as a covariance group). The explanation of the paradox is as follows. The Galilean covariance group $G$ induces the representation $T_{r}$ in the space $\mathcal{R}\triangle \mathcal{H}$ and fulfills (\[rowt\]). But, as we will show later on, the structure of $G$ is such that there always exists a function $\zeta(r,t)$ continuous in $r$ and differentiable in $t$, with the help of which one can define a new equivalent representation $T'_{r}=e^{i\zeta(r,t)}T_{r}$ fulfilling $$T'_{r}T'_{s}=e^{i\xi(r,s)}T'_{rs}$$ with a time-independent $\xi$. The representations $T_{r}$ and $T'_{r}$ are equivalent because $T'_{r}\psi$ and $T_{r}\psi$ are equivalent for all $r$ and $\psi$. However, this is not the case in general, when the exponent $\xi$ depends on time, and this time dependence cannot be eliminated in the same way as for the Galilean group. We have a similar situation when we try to find the most general wave equation for a non-relativistic quantum particle in the Newton-Cartan spacetime. The relevant covariance group in this case is the Milne group which possesses representations with time-dependent $\xi$ not equivalent to any representations with a time-independent $\xi$. Moreover, the only physical representations of the Milne group are those with time-dependent $\xi$.
Spacetime Dependent Gauge Freedom
---------------------------------
There is a physical motivation to investigate representations $T_{r}$ fulfilling (\[rowt\]) with $\xi$ depending on spacetime point $p$: $$\label{rowX}
T_{r}T_{s}=e^{i\xi(r,s,p)}T_{rs}.$$ We have sketched the motivation in the first section. We have argued there, that the two wave functions $\psi$ and $\psi' =
e^{i\xi(p)}\psi$ are indistinguishable in the sense that they give the same transition probabilities: $\vert(\psi,\phi)\vert^{2} =
\vert(\psi',\phi)\vert^{2}$ for any $\phi$. One should provide, however, that we are sufficiently fare away from the vacuum. This additional assumption is an immediate consequence of the structure of states in the Fock space as well the form of the inner product in that space: $$\psi = \left( \begin{array}{ll}
c & \\
\psi_{1}(x) & \\
\psi_{2}(x_{1},x_{2}) & \\
\ldots &
\end{array} \right)$$ and $$(\psi,\psi') = c^{*}c' + \int_{\mathbb{R}^{3}}\psi_{1}^{*}\psi_{1}' \, d^{3}x +
\int_{\mathbb{R}^{2\times 3}}\psi_{2}^{*}\psi_{2}' \, d^{3}x_{1}d^{3}x_{2} + \ldots$$ in which the argument $x$ of the one-particle state $\psi_{1}$ contains the ordinary space coordinates. If one uses the Schrödinger picture the spacetime-dependent-phase equivalence effect is of primary importance, as one can expect by comparison with the previous subsection. Probably it would be superfluous to present in detail that in this case one is forced to use the whole Hilbert bundle $\mathcal{M}\triangle \mathcal{H}$ over spacetime $\mathcal{M}$ and respective cross-sections as the wave functions $\psi$ (see the next section for definitions). We do not present details as the reasoning is a simple analogue of that performed in the previous subsection.
Rather we concentrate on the heart of the whole problem, that is, on the spacetime-dependent-phase equivalence which seems more advisable. As we have said application of that equivalence is justified if working far from the vacuum state, when there is quite a number of particles present. But then the justification of this equivalence principle is the same as that of the fact that spacetime coordinates are $c$-numbers commuting with “everything”, so that the greater the number of particles the more commuting are the spacetime coordinates. This is natural and agrees with the well established knowledge that when dealing with one particle (within QM) the spacetime coordinates are not mere parameters and do not commute with “everything”; but, in passing to quantum field by canonical quantization (appropriate for many particles) the spacetime coordinates are ordinary $c$-numbers commuting with each other and all other quantities. Yet the state of affairs is not quite satisfactory and there is a problem which calls for a further analysis. As we have argued in first section the QFT is expected to work perfectly along with QM laws applied to infinite number of degrees of freedom when there is a very few particles present, i. e. near the vacuum state. This is the case in practice, for example, when computing both the Lamb shift and the anomalous magnetic moment — when we are dealing with one-electron problem. It seems, therefore, that application of our equivalence is not justified, but at the same time the commutativity of spacetime coordinates or their $c$-number character is not justified too! This contradict the canonical field quantization rule in which the coordinates do form a $c$-numbers! In this way we arrive at the puzzle of spacetime coordinate status which we shall try to resolve now.
First of all we should note that the spacetime-coordinates-problem does not exist when the quantum field is free — compare the first section. It goes into play when interaction is taken into account. We confine ourselves to QED in order to be more specific. One should like to work within the ordinary QM perturbation theory considered as causing transitions between the stationary states say of the free field. The ordinary QED in Schrödinger picture, however, presents so much a departure from logic in applying the QM perturbation theory, that it is even impossible to work within this formulation of QED. In general when a realistic interaction is present, so violent in the high frequencies, the ordinary picture of perturbation as causing transitions between ordinary stationary states of the free field (as in the anomalous magnetic moment) is destroyed and does require some extra caution. The Schödinder picture is unsuited for dealing with QED, because the vacuum fluctuations play such a dominant role in it. Still we shall try to reformulate the QED so as to be as much compatible with ordinary QM laws as possible. Such a reformulation was proposed by Dirac in his excellent book ([*[*Lectures of Quantum Field Theory*]{}, Academic Press, New York, 1966*]{}; see also the last Chap. of the Fourth revised 1981 ed. of [*The Principles of Quantum Mechanics*]{}). We have no room here to present the reformulation but we should quote some Dirac’s statements at least, which are of importance in our discussion. Suppose the ket $|Q\rangle$ represents a state for which there are no photons, electrons, or positrons present. One would be inclined to suppose this state to be the perfect vacuum, but it cannot be, because it is not stationary. For it to be stationary we should need to have $$H|Q\rangle = \lambda |Q\rangle$$ with $\lambda$ a number and $H$ the Hamiltonian of QED. Now $H$ contains the terms (we use the standard notation) $$\label{terms}
-e \int \bar{\psi}\alpha_{r}\mathcal{A}^{r}\psi \, d^{3}x
+\frac{1}{2} \int\!\!\!\int \frac{j_{0x}j_{0x'}}{|x-x'|} \, d^{3}xd^{3}x',$$ which do not give numerical factors when applied to $|Q\rangle$. If we start with the no-particle state it does not remain the no-particle state. Particles get created where none previously existed, their energy coming from the interaction part of the Hamiltonian. Let us call the no-particle state at a certain time by $|Q\rangle$. In order to study this spontaneous creation of particles, one takes the ket $|Q\rangle$ as initial in the Schrödinger picture and treat the terms (\[terms\]) as a perturbation giving rise to a probability of the state $|Q\rangle$ jumping into another state, in accordance with the ordinary perturbation theory of QM. The first term resolved into its Fourier components — the photon, electron and positron creation operators — contains a part $$\label{term1}
-e(\alpha_{r})^{ab} \int\!\!\!\int \mathcal{A}^{r}_{k} \bar{\xi}_{ap}
\zeta_{b p+k\hslash} \, d^{3}kd^{3}p,$$ causing transitions and corresponding to emission of a photon (creation operator $\mathcal{A}^{r}_{k}$) and simultaneously to creation of an electron-positron pair (creation operators $\bar{\xi}_{ap}$ and $\zeta_{b p+k\hslash}$). After a short time the transition probability is proportional to the squared length of the ket formed by multiplying (\[term1\]) into the initial ket $|Q\rangle$. But this length is infinite, so the transition probability is infinite. The second term of (\[terms\]) contains contributions with two electron-positron pairs created simultaneously. Again the transition probability due to this term is also infinite. One can conclude that the state $|Q\rangle$ is is not even approximately stationary. Even with a cutoff the no-particle state $|Q\rangle$ is not approximately stationary. This is why the above procedure presents so much a departure from the ordinary perturbation theory of QM, and seems to be not logically justified. Dirac proposes another way of dealing with QED, which is a less of departure from ordinary QM. From the the above calculations — says Dirac — it follows that the no-particle state $|Q\rangle$ differs very much from the vacuum state. The “vacuum” state must contain many particles, which may be pictured as a state of transient existence with violent fluctuations. Let us introduce the ket $|V\rangle$ to represent the “vacuum” state. It is the eigenket of $H$ belonging to the lowest eigenvalue. Here and subsequently $H$ denotes the Hamiltonian modified by the cutoff. One might try to calculate $|V\rangle$ as a perturbation of of the ket $|Q\rangle$, but such a method would be of doubtful validity, because the difference between $|V\rangle$ and $|Q\rangle$ is not small. No satisfactory way of calculating $|V\rangle$ is known. In any case the result would depend strongly on the cutoff, and since the cutoff is unspecified the result would not be a definite one. It follows that we must develop the theory without knowing $|V\rangle$. This is not a great hardship — argues further Dirac — because we are not manly interested in the “vacuum”. We are mainly interested in states which differ from the “vacuum” $|V\rangle$ through having a few particles present in addition to those associated with the vacuum fluctuations, and we want to know how this extra particles behave. For this purpose we focus our attention on an operator $K$ representing the creation of the extra particles, so that the state we are interested in appears $K|V\rangle$. We do not now how the ket $|V\rangle$ varies with time in the Schrödinger picture, since we do not now the lowest eigenvalue of $H$. To avoid this difficulty we work in the Heisenberg picture in which $|V\rangle$ is constant. We then require $K|V\rangle$ to represent another state in the Heisenberg picture and thus to be another constant ket. This leads to $$\label{evolution}
\frac{dK}{dt} \,\,\, \textrm{or} \,\,\,
i\hslash \frac{\partial K}{\partial t} + KH - HK =0.$$ We now have each physical state determined by a solution $K$ of (\[evolution\]). Dirac ([*loc. cit.*]{}) proceeded along these lines and built a theory more compatible with the standard QM. He was able to calculate the Lamb shift as well as the anomalous magnetic moment within this theory. Thus it is an open question if we are close to the “true” vacuum $|0\rangle$ when evaluating the anomalous magnetic moment. From the above calculations we expect rather that the state $|V\rangle$ is considerably far removed from the true vacuum $|0\rangle$ (no-particles present). But if we are sufficiently far from the vacuum to ensure the commutativity of sapacetime coordinates (as in the canonical field quantization) we will at the same time ensure the justification for our spacetime-dependent-equivalence of states! Is it therefore possible that we should use the Hilbert bundle $\mathcal{M}\triangle\mathcal{H}$ with appropriate cross sections instead of ordinary Hilbert space $\mathcal{H}$ with vectors of $\mathcal{H}$? It depends if the difference between $|V\rangle$ and the true vacuum is large or small — the fact on which we may speculate only. Anyway one has the following alternative: $$\begin{array}{rccc}
& \textrm{{\scriptsize Small difference between $|0\rangle$ and $|V\rangle$}} & & \textrm{{\scriptsize Large difference between $|0\rangle$ and $|V\rangle$}} \\
\textrm{{\bf Either}} & \textrm{\scriptsize $\mathcal{H}$} & \textrm{{\bf or}} & \textrm{\scriptsize $\mathcal{M}\triangle\mathcal{H}$} \\
& \textrm{{\scriptsize noncommuting spacetime coordinates}} & & \textrm{\scriptsize commuting spacetime coordinates}.
\end{array}$$ On the left hand side we have a case in which the spacetime-dependent equivalence is not always justified. In that case one actually works near the true vacuum $|0\rangle$ when considering one-particle problems like the anomalous magnetic moment. Ordinary QM with states as vectors (or Weyl’s rays) in ordinary Hilbert space is justified then, but the spacetime coordinates are not a mere $c$-numbers in it. On the right hand side we have a theory in which the space-time-dependent equivalence is in general justified. States are the appropriate cross sections $\psi: \, \mathcal{M}\ni p \to \psi_{p} \in \mathcal{H}_{p}$ in a Hilbert bundle $\mathcal{M}\triangle\mathcal{H}: \,\mathcal{M} \ni p \to \mathcal{H}_{p}$ over spacetime in which the respective Hilbert spaces $\mathcal{H}_{p}$ play a role rather. Now, one can see what a peculiar object is the QED. Namely, it is partly on the left hand side of the alternative as it use the basic methods of ordinary QM with $\mathcal{H}$ and at the same time it is at the right hand side of the alternative as a canonically quantized version of the classical field with [**commuting spacetime coordinates**]{}. Strictly speaking there are some troubles in constructing ordinary Hilbert space $\mathcal{H}$ for QED, as was pointed by Dirac ([*loc. cit.*]{}) even within his method of treatment mentioned above. Moreover, in the canonical quantization one is trying at the outset to implement the ordinary QM to a system with infinite number of degrees of freedom. We believe that the problem just mentioned mirrors an important physical truth. [**It strongly suggest that the spacetime coordinates in QED should not be perfectly commuting quantities, and that the less commuting they are the stronger is the cutoff effect**]{}. Indeed, if they are noncommuting, then we are on the left hand side of the alternative, where the difference between $|0\rangle$ and $|V\rangle$ is small. But this is possible only if the cutoff is large. On the other hand if they are commuting, then we are on the right hand side of the alternative where the difference between $|0\rangle$ and $|V\rangle$ is large, i. e. when the cutoff is small and not so violent. As is well known, the infinities in QED originate from the fact that we pass to the limit zero for the space and time intervals involved. But if the spacetime coordinates were noncommuting then the structure of spacetime would be more elaborate and the limit process would be meaning less. [**This suggest that the cutoff process reflects some important physical phenomena**]{}.[^12]
It is therefore advisable to construct a theory which like QED lies somewhat on the both sides of the above alternative. Namely, it this theory the states as cross sections (looking at the right hand side) should compose at the same time a Hilbert space (looking at the left hand side). There is a natural Hilbert space composed of appropriate cross sections of a Hilbert bundle in mathematics, namely the von Neumann direct integral $$\mathcal{H} =\int_{\mathcal{M}} \mathcal{H}_{p} \, d\mu(p)$$ of Hilbert spaces $\mathcal{H}_{p}$, compare the next section for definition. But this is possible if the relevant quantum algebra say $\mathcal{A}$ acting in the direct integral $\mathcal{H}$ is decomposable over a diagonal (commutative) algebra say $\mathcal{A}^{CCR}_{1}$, such that the spectrum of the diagonal algebra is the classical spacetime $\mathcal{M}$. Thus we arrive at the following interpretation of the canonical field quantization:
.
But the interpretation of the inner products of the Hilbert spaces $\mathcal{H}_{p}$, $p\in \mathcal{M}$ and of their direct integral $\mathcal{H}$ is not clear now. One has to preserve an open mind on this subject. Now we know only that these Hilbert spaces are of importance but their role depends on the respective regime (remember, please, our alternative). The problem should be investigated in a subsequent research.
The Dirac’s critique ([*loc. cit.*]{}) inspired Piron in his research who formulated some related ideas cf. ([*C. Piron, physics/0204083*]{}). He brings about within his analysis of Dirac’s work at the conclusion that some spectral families of Hilbert spaces instead of of a mere Hilbert space are indispensable. Piron’s reasoning was, however, completely different.
[**Appendix.**]{} It should be mentioned an independent argument which shows that the generalized ray representations may play a role in QED. Paradoxically, there should be no zero mass vector particles with helicity = 1, as a consequence of the theory of unitary representations of the Poincaré group, as shown by [Ł]{}opuszański ([*Fortschritte der Physik [**26**]{}, 261, (1978); Rachunek spinorów, PWN, Warszawa 1985 (in Polish)*]{}). This is apparently in contradiction to the experiment, because the photon is a vector particle with helicity = 1. What is the solution of this paradox? First, let us describe the solution on the grounds of the existing theory, which constitutes at the same time an orthodox view. We observe that we can build a zero mass vector state with $h=1$ but we must admit finite-dimensional irreducible and thus *non-unitary* representations of the small group, that is the two-dimensional non-compact Euclidean group. Next, please note that the representation of $G$ induced by the non-unitary representation of the small group remains “unitary” if we admit the inner product in the “Hilbert” space to be *not positively* defined, cf. ([*S. N. Gupta, Proc. Phys. Soc. [**63**]{}, 681, (1950); K. Bleuler, Helv. Phys. Acta [**23**]{}, 567, (1950)*]{}), or ([*S. Weinberg, The Quantum Theory of Fields, volume II, Univ. Press, Cambridge 1996*]{}). However, even with the most favorable attitude toward the orthodox view, this solution is rather obscure. We propose to proceed in another way. First, let us observe that the case of the free quantum vector field[^13] $A_{\mu}(x)$ with zero mass is exactly the same. As long as the inner product in the Hilbert space is positively defined, we are not able to introduce any vector potential which transforms as a vector field. However, we can introduce a local real electromagnetic field $F_{\mu\nu}(x)= -F_{\nu\mu}(x)$ which is a linear combination of a self-dual and an antiself-dual field with helicity $+1$ and $-1$ respectively. If we introduce a vector potential $A_{\mu}$ in some Lorentz frame such that $$F_{\mu\nu}=\partial_{\mu}A_{\nu} - \partial_{\nu}A_{\mu}$$ then in another Lorentz frame we will have ($\Lambda_{\mu}^{\nu}$ is the Lorentz transformation matrix corresponding to the Poincaré transformation $r$) $$\label{A}
A_{\mu}(x) \to U_{r}A_{\mu}(x){U_{r}}^{-1}
= {(\Lambda^{-1})}_{\mu}^{\nu}A_{\nu}(r^{-1}x)
+ \partial_{\mu}R(r, x), \,\,\, \partial_{\mu}R \neq 0,$$ while $F_{\mu\nu}$ transforms as a tensor field. We infer that gauge transformation of the second kind has to accompany the Poincaré transformation, or that gauge freedom is indispensable in the construction of the vector potential in the quantum field theory, cf. ([*J. [Ł]{}opuszański, Fortschritte der Physik [**26**]{}, 261, (1978); [*Rachunek spinorów*]{}, PWN, Warszawa 1985 (in Polish).*]{}) or ([*S. Weinberg, [*The Quantum Theory of Fields*]{}, volume II, Univ. Press, Cambridge 1996*]{}) , vol. I. Let $\Omega$ denote the vacuum state. According to QFT, we should define a photon state $\psi_{\mu}(x)$ in the following way $$\psi_{\mu}(x)=A_{\mu}(x)\Omega.$$ It immediately follows from Eq. (\[A\]) that $$\label{Uphot}
U_{r}\psi_{\mu}(x)= {(\Lambda^{-1})}_{\mu}^{\nu}\psi_{\nu}(r^{-1}x)
+ \partial_{\mu}\Theta(x),$$ where $\partial_{\mu}\Theta(x)$ denotes the vector-valued distribution $\partial_{\mu}R(x)\Omega$. The above representation spanned by the generalized vectors $\psi_{\mu}(x)$ induces a representation in the appropriate Hilbert space. Indeed, let us write $\varphi_{\mu}(x)$ and $\theta(x)$ for test functions which “smear” the distributions $\psi_{\mu}(x)$ and $\Theta(x)$ respectively. From formula (\[Uphot\]) we get the transformation law for $\varphi_{\mu}$ $$\label{Tphot}
T_{r}\varphi_{\mu}(x) =
{(\Lambda^{-1})}_{\mu}^{\nu}\varphi_{\nu}(r^{-1}x)
+ \partial_{\mu}\theta(x).$$ By construction, the space of test functions is dense in the corresponding Hilbert space and the above representation $T_{r}$ can be uniquely extended. As we are dealing with a gauge-invariant theory, the two quantum vector potentials $A_{\mu}(x)$ and $A_{\mu} + \partial_{\mu}\Phi(x)$ are unitary equivalent. Accordingly, two photon states differing by a gradient, as well as their two corresponding vectors $\varphi_{\mu}(x)$ and $\varphi_{\mu}(x) + \partial_{\mu}\phi(x)$ should be unitary equivalent. This means that it is more adequate to consider all $\varphi_{\mu}(x) + \partial_{\mu}\phi(x)$ instead of the respective $\varphi_{\mu}(x)$ alone. We write $\varphi_{\mu}(x)+\partial_{\mu}\phi(x)$ as a pair $\{\phi(x), \varphi_{\mu}(x)\}$. The action of our representation $T_{r}$ in the space of pairs $\{\phi(x),\varphi_{\mu}(x)\}$ is as follows $$T_{r}\{\phi(x),\varphi_{\mu}(x)\} = \{\phi(r^{-1}x)+\theta(r,x),
U_{r}\varphi_{\mu}(x)\},$$ where $U_{r}$ acts as an ordinary vector transformation: $$U_{r}\varphi_{\mu}(x) = {(\Lambda^{-1})}_{\mu}^{\nu}\varphi_{\nu}(r^{-1}x).$$ Moreover, we have $$\label{QED}
T_{r}T_{s}\{\phi(x),\varphi_{\mu}(x)\} = T_{rs}\{\phi(x),\varphi_{\mu}(x)\}
+\{\xi(r,s,x),0\},$$ where $$\xi(r,s,x) = \theta(r,x) + \theta(s,r^{-1}x) - \theta(rs,x).$$ But this seems to be a kind of a (generalized) ray representation of the Poincaré group $G$ fulfilling (\[rowX\]) with the spacetime-dependent exponent $\xi(r,s,x)$! Summing up the discussion, we have just seen that the generalized representation in the sense of Eq. (\[rowX\]) seems to be indispensable if we wish to work with ordinary Hilbert spaces with positive norms while having a theory which describes photons.
Generalization of Bargmann’s Theory
===================================
We have shown that we are forced to generalize the Bargmann’s theory of factors to embrace the spacetime-dependent factors of representations acting in a Hilbert bundle over space-time (time) and apply the theory in the simplest non-relativistic case. We have already done it, the results will be presented in the subsequent part of this Chapter.
Generalized Wave Rays and Operator Rays {#generalization}
---------------------------------------
In this section we give strict mathematical definitions of the notions of the preceding section, and formulate the problem stated there in an exact way. From the pure mathematical point of view, the analysis of spacetime-dependent $\xi(r,s,p)$ is more general, so at the outset we confine ourselves to this case [^14].
Let us recall some definitions, cf. e.g. ([*G. W. Mackey, Unitary Group Representations in Physics, Probability, and Number Theory. Addison-Wesley Publishing Company, INC. New York, Amsterdam, Wokingham-UK (1989)*]{}). Let $\mathcal{M}$ be a set endowed with an analytic Borel structure.
By a *Hilbert bundle over* $\mathcal{M}$ or a *Hilbert bundle with base* $\mathcal{M}$ we shall mean an assignment $\mathcal{H}: p \to\mathcal{H}_{p}$ of a Hilbert space $\mathcal{H}_{p}$ to each $p \in \mathcal{M}$. The set of all pairs $(p,\psi)$ with $\psi \in \mathcal{H}_{p}$ will be denoted by $\mathcal{M}\triangle\mathcal{H}$ and called the *space of the bundle*. By a *cross section* of our bundle we shall mean an assignment $\psi: p \to \psi_{p}$ of a member of $\mathcal{H}_{p}$ to each $p\in\mathcal{M}$. If $\psi$ is a cross section and $(p_{0},\phi_{0})$ a point of $\mathcal{M}\triangle\mathcal{H}$, we may form a scalar product $(\phi_{0},\psi_{p_{0}})$. In this way, every cross-section $\psi$ defines a complex-valued function $f_{\psi}$ on $\mathcal{M}\triangle\mathcal{H}$. By a *Borel Hilbert bundle* we shall mean a Hilbert bundle together with an analytic Borel structure in $\mathcal{M}\triangle\mathcal{H}$ such that the following conditions are fulfilled
1. Let $\pi(p,\psi)=p$. Then $E \subseteq \mathcal{M}$ is a Borel set if and only if $\pi^{-1}(E)$ is a Borel set in $\mathcal{M}\triangle\mathcal{H}$.
2. There exist countably many cross-sections $\psi^{1}, \psi^{2}, \ldots$ such that
- the corresponding complex-valued functions on $\mathcal{M}\triangle\mathcal{H}$ are Borel functions,
- these Borel functions separate points in the sense that no two distinct points $(p_{i},\phi_{i})$ of $\mathcal{M}\triangle\mathcal{H}$ assign the same values to all $\psi^{j}$ unless $\phi_{1}=\phi_{2}=0$, and
- $p \to (\psi^{i}(p),\psi^{j}(p))$ is a Borel function for all $i$ and $j$.
A cross-section is said to be a *Borel cross-section* if the function on $\mathcal{M}\triangle\mathcal{H}$ defined by the cross-section is a Borel function. All Borel cross-sections compose a linear space under the obvious operations, cf. Mackey ([*loc. cit*]{}). Now let $\mu$ be a measure on $\mathcal{M}$. The cross-section $p \to \varphi_{p}$ is said to be *square summable* with respect to $\mu$ if $$\int_{\mathcal{M}} (\varphi_{p}, \varphi_{p}) \, d\mu(p) < \infty.$$ The space $\mathcal{L}^{2}(\mathcal{M}, \mu, \mathcal{H})$ of all equivalence classes of square-summable cross-sections, where two cross-sections $\varphi$ and $\varphi'$ are in the same equivalence class if $\varphi_{p}=\varphi'_{p}$ for almost all $p \in \mathcal{M}$, forms a separable Hilbert space with the inner product given by $$(\varphi, \theta) = \int_{\mathcal{M}} (\varphi_{p}, \theta_{p}) \, d\mu(p),$$ cf. Mackey ([*loc. cit.*]{}). It is called the direct integral of the $\mathcal{H}_{p}$ with respect to $\mu$ and is denoted by $\int_{\mathcal{M}} \mathcal{H}_{p} \, d\mu(p)$.
Identification with the previous section is partially suggested by the notation itself. We shall make this identification more explicit. The set $\mathcal{M}$ plays the role of spacetime or real line $\mathbb{R}$ of time $t$ respectively. The wave functions $\psi$ of the preceding section are the Borel cross-sections of $\mathcal{M}\triangle\mathcal{H}$ but if they do belong to the subset $\mathcal{L}^{2}(\mathcal{M},\mu,\mathcal{H})$ of cross-sections which are square integrable presents an open question. The separate Hilbert spaces $\mathcal{H}_{p}$ with their inner products play some role in experiments as well as the inner product in their direct integral product. But would be better to leave unspecified the precise role they play in experiments for now. We have also used $\psi(p)$ and $\psi_{p}$ as well as $(\psi_{p}, \theta_{p})$ and $(\psi, \theta)_{p}$ interchangeably.
By an *isomorphism* of the Hilbert bundle $\mathcal{M}\triangle\mathcal{H}$ with the Hilbert bundle $\mathcal{M}'\triangle\mathcal{H}'$ we shall mean a Borel isomorphism $T$ of $\mathcal{M}\triangle\mathcal{H}$ on $\mathcal{M}'\triangle\mathcal{H}'$ such that for each $p \in \mathcal{M}$ the restriction of $T$ to $p \times \mathcal{H}_{p}$ has some $q \times \mathcal{H}_{q}'$ for its range and is unitary when regarded as a map of $\mathcal{H}_{p}$ on $\mathcal{H}_{q}'$. The induced map carrying $p$ into $q$ is clearly a Borel isomorphism of $\mathcal{M}$ with $\mathcal{M}'$ and we denote it by $T^{\pi}$. The above-defined $T$ is said to be an *automorphism* if $\mathcal{M}\triangle\mathcal{H}=\mathcal{M}'\triangle\mathcal{H}'$. Please note that for any automorphism $T$ we have $(T\psi,T\phi)_{T^{\pi}p}=(\psi,\phi)_{p}$, but in general $(T\psi,T\phi)_{p}\neq(\psi,\phi)_{p}$. By this token, any automorphism $T$ is what is frequently called *bundle isometry*.
The function $r \to T_{r}$ from group $G$ into the set of automorphisms (bundle isometry) of $\mathcal{M}\triangle\mathcal{H}$ is said to be a *general factor representation* of $G$ associated to the action $G \times \mathcal{M} \ni r,p \to r^{-1}p \in \mathcal{M}$ of $G$ on $\mathcal{M}$ if $T_{r}^{\pi}(p) \equiv r^{-1}p$ for all $r \in G$, and $T_{r}$ satisfy condition (\[rowX\]).
Of course, $T_{r}$ is to be identified with that of the preceding section. Our further specializing assumptions partly following from the above interpretation are as follows. We assume $\mathcal{M}$ to be endowed with the manifold structure inducing a topology associated with the above-assumed Borel structure. We confine ourselves to a finite dimensional Lie group $G$ which acts smoothly and transitively on spacetime $\mathcal{M}$, such that a $G$-invariant measure $\mu$ exists on $\mathcal{M}$.
By a *factor representation* of a Lie group we mean a general factor representation with the exponent $\xi(r,s,p)$ differentiable in $p \in \mathcal{M}$.
Now we define the *operator ray* $\boldsymbol{T}$ corresponding to a given bundle isometry operator $T$ to be the set of operators $${\boldsymbol{T}}=\{\tau T , p \to \tau(p) \in {\mathcal{D}} \, \,
\textrm{and} \, \, \vert\tau \vert = 1\},$$ where $\mathcal{D}$ denotes the set of all differentiable real functions on $\mathcal{M}$. Any $T\in {\boldsymbol{T}}$ will be called a *representative* of the ray ${\boldsymbol{T}}$. The product ${\boldsymbol{TV}}$ is defined as the set of all products $TV$ such that $T\in {\boldsymbol{T}}$ and $V \in {\boldsymbol{V}}$.
Please note that not all Borel sections are physically realizable. By interpreting the discussion of the preceding section in the Hilbert bundle language, we see that the role of the Schrödinger equation is essentially to establish all the physical sections. Any two sections $\psi(p)$ and $\psi'(p) =e^{i\zeta(p)}\psi(p)$ are indistinguishable giving the same probabilities $\vert f_{\psi}\vert^{2}
=\vert f_{\psi'}\vert^{2}$. After this, any group $G$ acting in $\mathcal{M}$ induces a *ray representation* of $G$, i.e. a mapping $r \to \boldsymbol{T}_{r}$ of $G$ into the space of rays of bundle automorphisms (bundle isometrics) of $\mathcal{M}\triangle\mathcal{H}$, fulfilling the condition $$\boldsymbol{T}_{r}\boldsymbol{T}_{s}=\boldsymbol{T}_{rs}.$$ For any cross-section $\psi$ we define its corresponding *ray* $\boldsymbol{\psi} =\{e^{i\zeta(p)}\psi(p),
\zeta \in \mathcal{D}\}$. If $\psi$ is a physical cross-section, then we get the *physical ray* of the preceding section. Selecting a representative $T_{r}$ for each $\boldsymbol{T}_{r}$, we get a factor representation fulfilling (\[rowX\]). Please note that $T_{r}$ transforms rays into rays, and we have $T_{r}(e^{i\xi(p)}\psi)=e^{i\xi_{r}(p)}T_{r}\psi$. Further on we assume that that operators $T_{r}$ are such that $\xi_{r}(p)=\xi(r^{-1}p)$, where $r^{-1}p$ denotes the action of $r^{-1} \in G$ on the spacetime point $p \in \mathcal{M}$. This is a natural assumption which does actually take place in practice.
Now we shall make the last assumption, namely that all transition probabilities vary continuously with the continuous variation of the coordinate transformation $s \in G$:
1. For any element $r$ in $G$, any ray ${\boldsymbol{\psi}}$ and any positive $\epsilon$, there exists a neighborhood $\mathfrak{N}$ of $r$ on $G$ such that $d_{p}(\boldsymbol{T}_{s}\boldsymbol{\psi},
\boldsymbol{T}_{r}\boldsymbol{\psi}) < \epsilon$ if $s \in \mathfrak{N}$ and $p \in \mathcal{M}$,
where $$d_{p}(\boldsymbol{\psi_{1}},\boldsymbol{\psi_{2}}) =
\inf_{\psi_{i} \in \boldsymbol{\psi}_{i}}
\Vert \psi_{1} - \psi_{2} \Vert_{p}
= \sqrt{2\vert1-\vert(\psi_{1},\psi_{2})_{p}\vert \, \vert}.$$
Basing on the continuity assumption, one can prove the following
Let ${\boldsymbol{T}}_{r}$ be a continuous ray representation of a group $G$. For all $r$ in a suitably chosen neighborhood $\mathfrak{N}_{0}$ of the unit element $e$ of $G$ one may select a strongly continuous set of representatives $T_{r}\in {\boldsymbol{T}}_{r}$. That is, for any compact set $\mathcal{C} \subset \mathcal{M}$, any wave function $\psi$, any $r \in {\mathfrak{N}}_{0}$ and any positive $\epsilon$ there exists a neighborhood $\mathfrak{N}$ of $r$ such that $\Vert T_{s}\psi
- T_{r}\psi \Vert_{p} < \epsilon$ if $s \in \mathfrak{N}$ and $p \in \mathcal{C}$.
There are numerous possible selections of such factor representations. But many among them merely differ by a differentiable phase factor and are physically indistinguishable. We call them equivalent. Our task then is to classify all possible factor representations with respect to this equivalence.
Local Exponents {#exponents}
---------------
The representatives $T_{r} \in \boldsymbol{T}_{r}$ selected as in Theorem 2 will be called *admissible*, with the representation $T_{r}$ obtained in this way referred to as an *admissible* representation. There are infinitely many possibilities of such a selection of admissible representations $T_{r}$. We confine ourselves to the local *admissible* representations defined on a fixed neighborhood $\mathfrak{N}_{o}$ of $e \in G$, as in Theorem 2.
Let $T_{r}$ be an *admissible* representation. With the help of the phase $e^{i\zeta(r,p)}$ with a real function $\zeta(r,p)$ differentiable in $p$ and continuous in $r$, we can define $$\label{4}
T'_{r} = e^{i\zeta(r,p)}T_{r},$$ which is a new *admissible* representation. This is trivial if one defines the continuity of $\zeta(r,p)$ in $r$ appropriately. Namely, from Theorem 2 it follows that the continuity has to be defined in the following way. The function $\zeta(r,p)$ *will be called strongly continuous in r at $r_{0}$ if and only if for any compact set $\mathcal{C}
\subset \mathcal{M}$ and any positive $\epsilon$ there exists a neighborhood ${\mathfrak{N}}_{0}$ of $r_{0}$ such that* $$\vert \zeta(r_{0},p) - \zeta(r,p) \vert < \epsilon,$$ *for all* $r \in {\mathfrak{N}}_{0}$ *and for all* $p \in \mathcal{C}$. But the converse is also true. Indeed, if $T'_{r}$ is also an *admissible* representation, then (\[4\]) has to be fulfilled for a real function $\zeta(r,p)$ differentiable in $p$ because $T'_{r}$ and $T_{r}$ belong to the same ray. Moreover, because both $T'_{r}\psi$ and $T_{r}\psi$ are strongly continuous (in $r$ for any $\psi$), then $\zeta(r,p)$ has to be *strongly continuous* (in $r$).
Let $T_{r}$ be an *admissible* representation, and thus continuous in the sense indicated in Theorem 2. One can always choose the above $\zeta$ in such a way that $T_{e} = 1$ as will be assumed from now on.
Because $T_{r}T_{s}$ and $T_{rs}$ belong to the same ray, one has $$\label{5}
T_{r}T_{s} = e^{i\xi(r,s,p)} T_{rs}$$ with a real function $\xi(r,s,p)$ differentiable in $p$. From the fact that $T_{e} = 1$, we have $$\label{9}
\xi(e,e,p) = 0.$$ From the associative law $(T_{r}T_{s})T_{g} = T_{r}(T_{s}T_{g})$ one gets $$\label{10}
\xi(r,s,p) + \xi(rs,g,p) = \xi(s,g,r^{-1}p) + \xi(r,sg,p).$$ Formula (\[10\]) is very important and our analysis largely rests on this relation. From the fact that the representation $T_{r}$ is *admissible* follows that the exponent $\xi(r,s,p)$ is continuous in $r$ and $s$. Indeed, let us take a $\psi$ belonging to a unit ray $\boldsymbol{\psi}$. Then, making use of (\[5\]), we get $$e^{i\xi(r,s,p)}(T_{rs} - T_{r's'})\psi + (T_{r'}(T_{s'} - T_{s})\psi +
(T_{r'} - T_{r})T_{s}\psi$$
$$= (e^{i\xi(r',s',p)} - e^{i\xi(r,s,p)}) T_{r's'}\psi.$$
Taking norms $\Vert \centerdot \Vert_{p}$ of both sides, we get $$\vert e^{i\xi(r',s',p)} - e^{i\xi(r,s,p)} \vert \leq
\Vert (T_{r's'} - T_{rs})\psi \Vert_{p} +$$
$$+ \Vert T_{r'}(T_{s'} - T_{s})\psi \Vert_{p} + \Vert (T_{r'} -
T_{r})T_{s}\psi \Vert_{p}.$$
From this inequality and the continuity of $T_{r}\psi$, the continuity of $\xi(r,s,p)$ in $r$ and $s$ follows. Moreover, from Theorem 2 and the above inequality follows the *strong continuity* of $\xi(r,s,p)$ in $r$ and $s$.
Formula (\[4\]) suggests the following definition. Two *admissible* representations $T_{r}$ and $T'_{r}$ are called *equivalent* if and only if $T'_{r} = e^{i\zeta(r,p)}T_{r}$ for some real function $\zeta(r,p)$ differentiable in $p$ and *strongly continuous* in $r$. Thus, making use of (\[5\]), we get $T'_{r}T'_{s} = e^{i\xi'(r,s,p)}T'_{rs}$, where $$\label{13}
\xi'(r,s,p) = \xi(r,s,p) + \zeta(r,p) + \zeta(s,r^{-1}p) - \zeta(rs,p).$$ Then the two exponents $\xi$ and $\xi'$ are equivalent if and only if (\[13\]) is fulfilled with $\zeta(r,p)$ *strongly continuous* in $r$ and differentiable in $p$.
From (\[9\]) and (\[10\]) it immediately follows that $$\label{11}
\xi(r,e,p) = 0 \, \, \, and \, \, \, \xi(e,g,p) = 0,$$
$$\label{12}
\xi(r,r^{-1},p) = \xi(r^{-1},r,r^{-1}p).$$
Relation (\[13\]) between $\xi$ and $\xi'$ will be written in short by $$\label{13'}
\xi' = \xi + \Delta[\zeta].$$ The relation (\[13\]) between exponents $\xi$ and $\xi'$ defines an equivalence relation, which preserves the linear structure.
We introduce now group $H$, a very important notion for our further investigations. It is evident that all operators $T_{r}$ contained in all rays $\boldsymbol{T}_{r}$ form a group under multiplication. Indeed, let us consider an *admissible* representation $T_{r}$ with a well-defined $\xi(r,s,p)$ in formula (\[5\]). Because any $T_{r} \in \boldsymbol{T}_{r}$ has the form $e^{i\theta(p)}T_{r}$ (with a real and differentiable $\theta$), one has $$\label{H-action}
\Big( e^{i\theta(p)}T_{r}\Big) \Big( e^{i\theta'(p)}T_{s}\Big) =
e^{i\{\theta(p) + \theta'(r^{-1}p) + \xi(r,s,p)\}}T_{rs}.$$ This important relation suggests the following definition of the local group $H$ connected with the *admissible* representation or with the exponent $\xi(r,s,p)$. Namely, $H$ consists of the pairs $\{\theta(p), r\}$ where $\theta(p)$ is a differentiable real function and $r \in G$. The multiplication rule, suggested by the above relation, is defined as follows $$\label{15}
\{\theta(p),r\} \centerdot \{\theta'(p),r'\} = \{\theta(p) +
\theta'(r^{-1}p) + \xi(r,r',p), \, rr' \}.$$ The associative law for this multiplication rule is equivalent to (\[10\]) (in complete analogy with the classical Bargmann’s theory). The pair $\check{e} =\{0,e\}$ plays the role of the unit element in $H$. For any element $\{\theta(p), r\} \in H$ there exists the inverse $\{\theta(p),r\}^{-1} = \{-\theta(rp) - \xi(r,r^{-1},rp), \, r^{-1} \}$. Indeed, from (\[12\]) it follows that $\{\theta, r\}^{-1} \centerdot
\{\theta,r \} = \{\theta, r\} \centerdot \{\theta, r\}^{-1} = \check{e}$. The elements $\{\theta(p), e\}$ form an Abelian subgroup $N$ of $H$. Any $\{\theta,r\} \in H$ can be uniquely written as $\{\theta(p),r\} =
\{\theta(p),e\} \centerdot \{ 0, r\}$. The same element can be also uniquely expressed in the form $\{\theta(p),r\} = \{0,r\} \centerdot \{\theta(rp),e\}$. Thus, we have $H = N \centerdot G = G \centerdot N$. The Abelian subgroup $N$ is a normal factor subgroup of $H$. But this time, $G$ does not form any normal factor subgroup of $H$ (contrary to the classical case investigated by Bargmann, when the exponents do not depend on $p$). So, this time $H$ is not direct product $N \otimes G$, but a semidirect product $N \circledS G$. In this case, however, the theorem that $G$ is locally isomorphic to the factor group $H/N$ is still valid. Then group $H$ composes a *semicentral extension* of $G$, and not a central extension of $G$ as in the Bargmann’s theory.
The rest of this paper is based on the following reasoning (the author being largely inspired by Bargmann’s work ([*Ann. Math. [**59**]{}, 1, 1954*]{}). If the two exponents $\xi$ and $\xi'$ are *equivalent*, that is $\xi' = \xi + \Delta[\zeta]$, then the *semicentral extensions* $H$ and $H'$ connected with $\xi$ and $\xi'$ are homomorphic. The homomorphism $h: \{\theta,r\}
\mapsto \{\theta', r'\}$ is given by $$\label{izo}
\theta'(p) = \theta(p) - \zeta(r,p), \, \, r' = r.$$ Using an *Iwasawa-type construction* we show that any exponent $\xi(r,s,p)$ is equivalent to a differentiable one (in $r$ and $s$). We can then confine ourselves to the differentiable $\xi$ and $\xi'$. We show that $\zeta(r,p)$ is also a differentiable function of $(r,p)$. Moreover, we show that any $\xi$ is equivalent to the canonical one, that is such $\xi$ which is differentiable and for which $\xi(r,s,p) = 0$ whenever $r$ and $s$ belong to the same one-parameter subgroup. Then we can restrict our investigation to the canonical $\xi$ considering the subgroup of all elements $\{\theta(p),r\} \in H$ with differentiable $\theta(p)$. For simplicity let us denote the subgroup by the same symbol $H$ . We embed the subgroup in an infinite dimensional Lie group $D$ with manifold structure modeled on a Banach space. Then we consider the subgroup $\overline{H}$ which is the closure of $H$ in $D$. After this, $\overline{H}$ turns into a Lie group and the homomorphism (\[izo\]) becomes an isomorphism of the two Lie groups. Thus, the group $\overline{H}$ has the Banach Lie algebra $\overline{\mathfrak{H}}$. We apply the general theory of analytic groups developed in ([*G. Birkhoff, Continuous Groups and Linear Spaces, Recueil Mathématique (Moscow) [**1**]{}(5), 635, (1935); Analytical Groups, Trans. Am. Math. Soc. [**43**]{}, 61, (1938)*]{}) and ([*E. Dynkin, Uspekhi Mat. Nauk [**5**]{}, (1950), 135; Amer. Math. Soc. Transl. [**9**]{}(1), (1950), 470*]{}). From this theory it follows that the correspondence between the local $\overline{H}$ and $\overline{\mathfrak{H}}$ is bi-unique and one can construct uniquely the local group $\overline{H}$ from the algebra $\overline{\mathfrak{H}}$ as well. As we will see, the algebra defines a spacetime-dependent anti-linear form $\Xi$ on the Lie algebra $\mathfrak{G}$ of $G$, the so-called *infinitesimal exponent* $\Xi$. By this we reduce the classification of local $\xi$’s which define $\overline{H}$’s to the classification of $\Xi$’s which define $\overline{\mathfrak{H}}$’s. So, we will simplify the problem of the classification of local $\xi$’s to a largely linear problem. Here are the details.
[**Iwasawa construction**]{}. Let us denote by ${{\mathrm{d}}}r$ and ${{\mathrm{d}}}^{*}r$ the left and right invariant Haar measure on $G$. Let $\nu(r)$ and $\nu^{*}(r)$ be two infinitely differentiable functions on $G$ with compact supports contained in the fixed neighborhood $\mathfrak{N}_{0}$ of $e$. By multiplying them by the appropriate constants, we can always obtain: $\int_{G} \nu(r) \, {{\mathrm{d}}}r =
\int_{G} \nu^{*}(r) \, {{\mathrm{d}}}^{*}r =1$. Let $\xi(r,s,p)$ be any *admissible* local exponent defined on $\mathfrak{N}_{0}$. We will construct a differentiable (in $r$ and $s$) exponent $\xi''(r,s,p)$ *equivalent* to $\xi(r,s,p)$ and defined on $\mathfrak{N}_{0}$, in the following two steps: $\xi' = \xi + \Delta[\zeta]$ and $\xi'' = \xi' + \Delta[\zeta']$, where $\zeta(r,p)$ is the left invariant integral of $l\to -\xi(r,l,p)\nu(l)$, while $\zeta'(r,p)$ is the right invariant integral of $u\to -\xi'(u,r,up)\nu^{*}(u)$. A rather simple computation in which we use (\[13\]) and (\[10\]) and the invariance property of the Haar measures shows that $\xi''(r,s,p)$ is a differentiable (up to any order)exponent in all variables. Next we shall show that *if two differentiable exponents $\xi$ and $\xi'$ are equivalent, that is, if $\xi' = \xi + \Delta[\zeta]$, then $\zeta(r,p)$ is differentiable in $r$*. Clearly, the difference $\xi'-\xi$ is differentiable. Similarly, both $(\xi'-\xi)\nu$ ( with $\nu$ defined as above), as well as its left invariant integral $\eta$ are differentiable. It is easy to show that $\zeta' = \eta - \zeta$ is also differentiable. In this way we arrive at differentiability of $\zeta = \eta - \zeta'$ is differentiable. A slightly more complicated argumentation shows that *every (local) exponent of one-parameter group is equivalent to zero*. However, the argumentation is quite analogous to that of Bargmann. We can treat such a group as the additive group of real numbers, so that the first two arguments of $\xi$ are real numbers. Let us set $\vartheta(\tau,\sigma,p)$ as the derivative of $\xi$ with respect to the second argument. It is not hard to show that $\xi + \Delta[\zeta]=0$, where $\zeta(\tau,p)$ is the ordinary Riemann integral of $\mu\to \tau\vartheta(\mu\tau,0,p)$ over the unit interval $[0,1]$. But it means that $\xi$ is equivalent to zero.
Let us recall that the continuous curve $r(\tau)$ in a Lie group $G$ is a one-parameter subgroup if and only if $r(\tau_{1})r(\tau_{2}) =
r(\tau_{1} + \tau_{2})$, *i.e.* $r(\tau) = (r_{0})^{\tau}$ for some element $r_{0} \in G$. (Please note that the real power $r^{\tau}$ is well defined on a Lie group, at least on some neighborhood of $e$). The coordinates $\rho^{k}$ in $G$ are *canonical* if and only if any curve of the form $r(\tau) = \tau \rho^{k}$ (where the coordinates $\rho^{k}$ are fixed) is a one-parameter subgroup (the curve $r(\tau) = \tau \rho^{k}$ will be denoted in short by $\tau a$, with the coordinates of $a$ equal to $\rho^{k}$). The “vector” $a$ is called by physicists the *generator* of the one-parameter subgroup $\tau a$.
A local exponent $\xi$ of a Lie group $G$ is called *canonical* if $\xi(r,s,p)$ is differentiable in all variables, and $\xi(r,s,p) = 0$ if $r$ and $s$ are elements of the same one-parameter subgroup.
Almost the same argument used to show that every $\xi$ on a one-parameter group is equivalent to zero also shows that *every local exponent $\xi$ of a Lie group is equivalent to a canonical local exponent*. In order to prove this, we shall apply the argument to the exponent $\xi_{0}(\tau, \sigma,p)
:=\xi(\tau a, \sigma a,p)$, cf. ([*J. Wawrzycki, math-ph/0301005*]{}). Up to now, the argumentation has been more or less analogous to that of Bargmann. From now on, the argumentation becomes entirely different. *Let $\xi$ and $\xi'$ be two differentiable and equivalent local exponents of a Lie group $G$, assuming $\xi$ to be canonical. Then $\xi'$ is canonical if and only if $\xi' = \xi
+ \Delta[\Lambda]$, where $\Lambda(r,p)$ is a linear form in the canonical coordinates of $r$ fulfilling the condition that $\Lambda(a,(\tau a)p)$ is constant as a function of $\tau$, i.e. it follows that*[^15] $$\label{condLem}
\boldsymbol{a}\Lambda(a,p) = \frac{d\Lambda(a, (\tau a)p)}{d\tau}
= \lim_{\epsilon \to 0}\frac{\Lambda(a,(\epsilon a)p)
- \Lambda(a,p)}{\epsilon} = 0.$$ While sufficiency of the condition in the above statement is almost evident, proving its necessity is quite nontrivial. Hereafter we outline the argumentation. Because the exponents are equivalent we have $\xi'(r,s,p) = \xi(r,s,p) +\Delta[\zeta]$. Since both $\xi$ and $\xi'$ are differentiable then $\zeta(r,p)$ is also a differentiable function, which follows from what has been said above. Let us suppose that $r=\tau a$ and $s= \tau' a$. Because both $\xi$ and $\xi'$ are *canonical*, we have $\xi(\tau a, \tau' a,p) = \xi'(\tau a, \tau' a,p) = 0$, so that $\Delta[\zeta](\tau a, \tau' a,p) = 0$. Applying the last formula recurrently one gets $$\zeta(\tau a,p) = \sum_{k=0}^{n-1} \zeta(\frac{\tau}{n}a,
(-\frac{k}{n}\tau a)p).$$ Then we use the Taylor Theorem to each summand in the above expression, and pass to the limit $n \to + \infty$. In this way, we obtain $$\label{18}
\zeta(\tau a,p) = \int_{0}^{\tau} \varsigma(a,(-\sigma a)p) \, {{\mathrm{d}}} \sigma,$$ where $\varsigma = \varsigma(r,p)$ is a differentiable function, cf. ([*J. Wawrzycki, math-ph/0301005*]{}). If we differentiate now expression (\[18\]) with respect to $\tau$ at $\tau = 0$, we will immediately see that the function $\varsigma(a,p)$ is linear with respect to $a$. Let us suppose that the spacetime coordinates are chosen in such a way that the integral curves $p(x) = (xa)p_{0}$ are coordinate lines, which is possible for appropriately small $x$. There are of course three remaining families of coordinate lines besides $p(x)$, which can be chosen in an arbitrary way, with their parameters denoted by $y_{i}$. After this, $$\zeta(a,x,y_{i}) = \frac{1}{\tau} \int_{0}^{\tau} \varsigma(a,x
- \sigma, y_{i}) \, {{\mathrm{d}}} \sigma
= \frac{1}{\tau} \int_{x - \tau}^{x} \varsigma(a,z,y_{i}) \, {{\mathrm{d}}}z,$$ for any $\tau$ (of course, with appropriately small $\vert \tau \vert$, in our case $\vert \tau \vert \leq 1$) and for any (appropriately small) $x$. But this is possible for the function $\varsigma(a,x,y_{k})$ continuous in $x$ (in our case, differentiable in $x$) if and only if $\varsigma(a,x,y_{k})$ does not depend on $x$. This means that $\zeta(a,x,y_{k})$ does not depend on $x$ and the condition of the statement is hereby proved.
[**Infinitesimal exponents and embedding of $H$ in a Lie group $D$**]{}. According to what has been shown already, we can assume that the exponent is canonical. We also confine ourselves to the subgroup of $\{\theta(p),r\} \in H$ with differentiable $\theta$, and denote this subgroup by the same letter $H$. We embed this subgroup $H$ in an infinite dimensional Lie group with the manifold structure modeled on a Banach space. We will extensively use the theory developed by Birkhoff ([*loc. cit.*]{}) and Dynkin ([*loc. cit.*]{}). For the systematic treatment of manifolds modeled on Banach spaces, see e.g. ([*S. Lang, Differential Manifolds, Springer-Verlag, Berlin, Heidelberg, New York (1985)*]{}). By this embedding we ascribe bi-uniquely a Lie algebra to the group $H$ with the convergent Baker-Hausdorff series.
Please note first that the formula $$H \times \mathcal{L}^{2}(\mathcal{M},\mu,\mathcal{H})\ni
(\{\theta(p),r \}, \phi) \to e^{i\theta(p)}T_{r}\phi$$ (together with (\[H-action\])) can be viewed as a rule giving the action of $H$ in the direct integral Hilbert space $\int_{\mathcal{M}} \mathcal{H}_{p} \, d\mu(p)$ defined in Section [**3**]{}. Moreover, this is a unitary action, provided $\mu$ is $G$-invariant. In accordance to Birkhoff ([*loc. cit.*]{}), the group $D$ of *all* unitary operators of a Hilbert space is an infinite dimensional Lie group. Hence, $H = N \circledS G$ can be viewed as a subgroup of a Lie group.
We consider now the closure $\overline{H}$ of $H$ in the sense of the topology in $D$. It is remarkable that *the subgroup $\overline{H}$ has locally the structure of the semi-direct product $\overline{N} \circledS G$ as well*. This is a consequence of the following four facts. (1) $\overline{N}$ is a normal subgroup of $\overline{H}=\overline{N\centerdot G}$. (2) $G$ is finite dimensional, so $\overline{G} = G$. (3) Locally (in a neighborhood $\mathcal{O}$), the multiplication in $D$ is given by the Baker-Hausdorff formula in the Banach algebra of $D$. Because $\overline{N}$ is normal in $\overline{H}$, then the above exponential mapping converts locally the multiplication $\overline{N}\centerdot S$ of $\overline{N}$ by any subset $S$ of $\overline{H}$ into the sum $\overline{N} + S$. Because $G$ is finite-dimensional, and hence locally compact, the neighborhood $\mathcal{O}$ can be chosen in such a way that locally (in the closure of $\mathcal{O} + \mathcal{O}$) the following holds: $$\overline{N}+\overline{G}
= \overline{N+G} = \overline{H}.$$ (4) The local $\overline{N}$ (intersected with $\overline{\mathcal{O}}$) has a finite co-dimension in local $\overline{N+G}$ (intersected with $\overline{\mathcal{O} + \mathcal{O}}$) and thus it splits locally $\overline{N+G}$. So, we have locally *i.e.* in $\overline{\mathcal{O} + \mathcal{O}}$: $$\overline{N} + \overline{G}
= \overline{H} = \overline{N}\oplus G',$$ where $\overline{G'} = G'$ and $\oplus$ stands for a direct sum. From this it follows that $G' =\overline{G}$ locally. This shows that $\overline{H}=\overline{N}\circledS G$.
Because $\overline{H}= \overline{N}\circledS G$, every $h\in \overline{H}$ is uniquely representable in the form $ng$, where $n \in \overline{N}$ and $g \in G$. Please note now that $$(n_{1}g_{1})(n_{2}g_{2}) = n_{1}g_{1}n_{2}g_{1}^{-1}g_{1}g_{2} =
[n_{1}(g_{1}n_{2}g_{1}^{-1})](g_{1}g_{2})$$ and that $g_{1}n_{2}g_{1}^{-1} \in \overline{N}$ because $\overline{N}$ is normal in $\overline{H}$. Let us denote the automorphism $n \to gng^{-1}$ of $\overline{N}$ by $R_{g}$. The group $\overline{H}$ can be locally viewed as a topological product of Banach spaces $\overline{\mathfrak{N}}\times
\mathfrak{G}$, one of which (namely $\mathfrak{G}$) is finite-dimensional and isomorphic tho the Lie algebra of $G$. The multiplication in $\overline{H}$ can be written as $(n_{1}, g_{1})(n_{2},g_{2}) = (n_{1}R_{g_{1}}(n_{2}), g_{1}g_{2})$. Moreover, $\overline{N}$ can be viewed locally as the Banach space $\overline{\mathfrak{N}}$ with the multiplication law given by the vector addition in $\overline{\mathfrak{N}}$.
Our task now is to reconstruct the Lie algebra $\overline{\mathfrak{H}}$ corresponding to the subgroup $\overline{H}$. Let $\lambda \to \lambda a$ be a one-parameter subgroup of $G$. The mapping $(\lambda, n) \to (R_{\lambda a}n, \lambda a)$ of the Banach space $\mathcal{R} \times \overline{\mathfrak{N}}$ into the Banach space $\overline{\mathfrak{N}} \times \mathfrak{G}$ is continuous. In consequence, $\mathcal{R}\ni \lambda \to R_{\lambda a}n \in
\overline{\mathfrak{N}}$ as well as $\overline{\mathfrak{N}}
\ni n \to R_{\lambda a}n$ are continuous. Therefore, the function $\lambda \to R_{\lambda a}n$ can be integrated over any compact interval and $$\tau \to (n_{\tau a}, \tau a)
:= \Big(\int_{0}^{\tau}R_{\sigma a}n \, {\mathrm{d}}\sigma, \tau a\Big)$$ is a one-parameter subgroup of $\overline{H}$ with generator[^16] $(n, a)$, cf. Birkhoff ([*loc. cit.*]{}), Dynkin ([*loc. cit*]{}). Having obtained this, we are able to reconstruct the algebra. The elements of $H \subset \overline{H}$ are representable in the ordinary form $\{\alpha,r\}$ with differentiable $\alpha = \alpha(p), p\in \mathcal{M}$, and $r \in G$. Let us consider the above-defined operator $R_{\lambda a}$. Its restriction to $H \subset \overline{H}$ is given by (please remember that $\xi$ is canonical) $$\alpha(p) \to (R_{\lambda a}\alpha)(p) = \alpha((\lambda a)^{-1}p).$$ We can now compute explicitly the Lie bracket and the Jacobi identity for all the elements $\{\alpha(p), a\}$ of the subalgebra $\mathfrak{H}
\subset \overline{\mathfrak{H}}$ corresponding to the subgroup $H$. The result is as follows[^17] $$\label{20'}
[\check{a}, \check{b}] = \{\boldsymbol{a}\beta - \boldsymbol{b}\alpha
+ \Xi(a,b, p), [a,b] \},$$
$$\Xi(a,b,p) = \lim_{\tau \to 0} \tau^{-2}\{\xi((\tau a)(\tau b),
(\tau a)^{-1}(\tau b)^{-1},p) +$$
$$\label{20''}
+ \xi(\tau a, \tau b, p) + \xi((\tau a)^{-1},(\tau b)^{-1},
(\tau b)^{-1}(\tau a)^{-1}p) \},$$
From the associative law in $H$ one gets $$\Xi([a,a'],a'',p) + \Xi([a',a''],a,p) + \Xi([a'',a],a',p) =$$
$$\label{21}
= \boldsymbol{a}\Xi(a',a'',p) + \boldsymbol{a'}\Xi(a'',a,p)
+ \boldsymbol{a''}\Xi(a,a',p),$$
which can be shown to be equivalent to the Jacobi identity $$\label{21'}
[[\check{a}, \check{a}'], \check{a}''] + [[\check{a}',\check{a}''],\check{a}]
+ [[\check{a}'',\check{a}], \check{a}'] = 0.$$ Thus, in this way we have reconstructed the Lie algebra $\overline{\mathfrak{H}}$ giving explicitly $[\check{a}, \check{b}]$ for all $\check{a},\check{b} \in \mathfrak{H} \subseteq \overline{\mathfrak{H}}$. Because $\mathfrak{H}$ is dense in $\overline{\mathfrak{H}}$, the local exponent $\Xi$ determines the algebra $\overline{\mathfrak{H}}$ uniquely. But from the theory of Lie groups the correspondence between the algebras $\overline{\mathfrak{H}}$ and local Lie groups $\overline{H}$ is bi-unique, at least locally, cf. e.g. Birkhoff ([*loc. cit.*]{}) and Dynkin ([*loc. cit.*]{}). Therefore, *the correspondence* $\overline{H} \to
\overline{\mathfrak{H}}$ *between the local group* $\overline{H}$ *and the algebra* $\overline{\mathfrak{H}}$ *is one-to-one*. Because the exponent $\xi$ determines the multiplication rule in $H$ and vice-versa, then it follows that *the correspondence $\xi \to \Xi$ between the local $\xi$ and the infinitesimal exponent $\Xi$ is one-to-one*. Please note that the term ’local $\xi = \xi(r,s,p)$’ means that $\xi(r,s,p)$ is defined for $r$ and $s$ belonging to a fixed neighborhood ${\mathfrak{N}}_{0} \subset G$ of $e \in G$, but *in our case it is defined globally as a function of the spacetime variable $p \in \mathcal{M}$*.
[**Infinitesimal exponents and local exponents**]{}. Now, let us move on to describing the relation between the infinitesimal exponents $\Xi$ and the local exponents $\xi$. First, let us compute the infinitesimal exponents $\Xi$ and $\Xi'$ given by (\[20”\]), which correspond to the two equivalent canonical local exponents $\xi$ and $\xi' = \xi + \Delta[\Lambda]$. Inserting $\xi' = \xi + \Delta[\Lambda]$ into formula (\[20”\]), one gets $$\label{24}
\Xi'(a,b,p) = \Xi(a,b,p) + \boldsymbol{a}\Lambda(b,p)
- \boldsymbol{b}\Lambda(a,p) - \Lambda([a,b],p).$$ According to what has been said, we can confine ourselves to the canonical exponents. Then, as one of our previous statements said, $\Lambda = \Lambda(a, (\tau b)p)$ is a constant function of $\tau$ if $a = b$, and $\Lambda(a,p)$ is linear with respect to $a$ (we use the canonical coordinates on $G$). Hence $\Xi'(a,b,p)$ is antisymmetric in $a$ and $b$ and fulfills (\[21\]) only if $\Xi(a,b,p)$ is antisymmetric in $a$ and $b$ and fulfills (\[21\]). This suggests the definition: *two infinitesimal exponents $\Xi$ and $\Xi'$ will be called equivalent if and only if relation* (\[24\]) *holds*. For brevity, we write relation (\[24\]) as follows: $$\Xi' = \Xi + d[\Lambda].$$ Finally, we maintain that *two canonical local exponents $\xi$ and $\xi'$ are equivalent if and only if the corresponding infinitesimal exponents $\Xi$ and $\Xi'$ are equivalent*. Indeed. (1) Assume $\xi$ and $\xi'$ to be equivalent. Then, by the definition of equivalence of infinitesimal exponents: $\Xi' = \Xi + d[\Lambda]$. (2) Assume $\Xi$ and $\Xi'$ to be equivalent: $\Xi' = \Xi + d[\Lambda]$ for some linear form $\Lambda(a,t)$ such that $\Lambda(a,(\tau a)p)$ does not depend on $\tau$. Then $\xi + \Delta[\Lambda] \to \Xi'$, and by the uniqueness of the correspondence $\xi \to \Xi$ we have $\xi' = \xi + \Delta[\Lambda]$, *i.e.* $\xi$ and $\xi'$ are equivalent. In this way, we arrive at the following:
\(1) On a Lie group G, every local exponent $\xi(r,s,p)$ is equivalent to a canonical local exponent $\xi'(r,s,p)$ which, on some canonical neighborhood ${\mathfrak{N}}_{0}$, is analytic in canonical coordinates of r and s, and vanishes if r and s belong to the same one-parameter subgroup. Two canonical local exponents $\xi,\xi'$ are equivalent if and only if $\xi' = \xi + \Delta[\Lambda]$ on some canonical neighborhood, where $\Lambda(r,p)$ is a linear form in the canonical coordinates of $r$ such that $\Lambda(r,sp)$ does not depend on $s$ if $s$ belongs to the same one-parameter subgroup as $r$. (2) To every canonical local exponent of $G$ there corresponds uniquely an infinitesimal exponent $\Xi(a,b,p)$ on the Lie algebra $\mathfrak{G}$ of $G$, i.e. a bilinear antisymmetric form which satisfies the identity $\Xi([a,a'],a'',p) +\Xi([a',a''],a,p)+ \Xi(a'',a],a',p)
= \boldsymbol{a}\Xi(a',a'',p) + \boldsymbol{a'}\Xi(a'',a,p) +
\boldsymbol{a''}\Xi(a,a',p)$. The correspondence is linear. (3) Two canonical local exponents $\xi,\xi'$ are equivalent if and only if the corresponding $\Xi,\Xi'$ are equivalent, i.e. $\Xi'(a,b,p) = \Xi(a,b,p)
+ \boldsymbol{a}\Lambda(b,p) - \boldsymbol{b}\Lambda(a,p) - \Lambda([a,b],p)$ where $\Lambda(a,p)$ is a linear form in $a$ on $\mathfrak{G}$ such that $\tau \to \Lambda(a,(\tau b)p)$ is constant if $a = b$. (4) There exist a one-to-one correspondence between the equivalence classes of local exponents $\xi$ (global in $p \in \mathcal{M}$) of $G$ and the equivalence classes of infinitesimal exponents $\Xi$ of $\mathfrak{G}$.
Global Extensions of Local Exponents
------------------------------------
Theorem 3 provides full classification of exponents $\xi(r,s,p)$ local in $r$ and $s$, defined for all $p \in \mathcal{M}$. But if $G$ is both connected and simply connected, then we have the following theorems. (1) If an extension $\xi'$ of a given local (in $r$ and $s$) exponent $\xi$ does exist, then it is uniquely determined (up to the equivalence transformation (\[13\])) (Theorem 4). (2) There exists such an extension $\xi'$ (Theorem 5), proved for $G$, which possess finite-dimensional extension $\mathfrak{H}'$ only.
We are not able to prove that the (*global*) homomorphism (\[izo\]) is continuous when $\xi$ is not canonical. Please note that any $\xi$ is equivalent to its canonical counterpart, but only *locally*! This is why the topology of $H$ induced from $D$ is not applicable in the global analysis. We introduce another topology. Because of the semidirect structure We introduce another topology. Because of the semidirect structure of $H = N\circledS G$, it is sufficient to introduce it into $N$ and $G$ separately in such a manner that $G$ acts continuously on $N$, cf. e.g. Mackey ([*loc. cit.*]{}). From the discussion of Section \[exponents\] it is sufficient to introduce the Fréchet topology of almost uniform convergence in the function space $N$. Indeed, from the strong continuity of $\xi$ and $\zeta$ in (\[izo\]) it follows that the multiplication rule as well as the homomorphism (\[izo\]) are continuous.
Let $\xi$ and $\xi'$ be two equivalent local exponents of a connected and simply connected group $G$, so that $\xi' = \xi
+ \Delta[\zeta]$ on some neighborhood, assuming the exponents $\xi_{1}$ and $\xi_{1}'$ of $G$ to be extensions of $\xi$ and $\xi'$ respectively. Then, for all $r,s \in G$: $\xi_{1}'(r,s,p) = \xi_{1}(r,s,p) + \Delta[\zeta_{1}]$, where $\zeta_{1}(r,p)$ is strongly continuous in $r$ and differentiable in $p$, and $\zeta_{1}(r,p)
= \zeta(r,p)$, for all $p \in \mathcal{M}$ and for all $r$ belonging to some neighborhood of $e \in G$.
Here is the proof outline. The two exponents $\xi_{1}$ and $\xi_{1}'$ being *strongly continuous* (by assumption) define two semicentral extensions $H_{1} = N_{1} \circledS G$ and $H_{1}'= N'_{1}\circledS G$, which are continuous groups. Please note that the linear groups $N_{1},N'_{1}$ are connected and simply connected. Because both $H_{1}$ and $H_{1}'$ are semi-direct products of two connected and simply connected groups they are both connected and simply connected. Eq. (\[izo\]) defines a local isomorphism mapping $h: \check{r} \to
\check{r}' = h(\check{r})$ of $H_{1}$ into $H_{1}'$. Because $H_{1}$ and $H_{1}'$ are connected and simply connected, the isomorphism $h$ given by (\[izo\]) can be uniquely extended to an isomorphism $h_{1}$ of the entire groups $H_{1}$ and $H_{1}'$ such that $h_{1}(\check{r}) = h(\check{r})$ on some neighborhood of $H_{1}$, cf. ([*L. Pontrjagin, [*Topological groups*]{}, Moscow (1984) (in Russian)*]{}), Theorem 80. The isomorphism $h_{1}$ defines an isomorphism of the two Abelian subgroups $N_{1}$ and $h_{1}(N_{1})$. By (\[izo\]), $h_{1}(\theta,e) = \{\theta, e\}$ locally in $H_{1}$, that is for $\theta$ lying appropriately close to 0 (in the metric sense defined previously). Both $N_{1}$ and $h_{1}(N_{1})$ are connected, and $N_{1}$ is in addition simply connected, so applying once again Theorem 80 of Pontrjagin ([*loc. cit.*]{}), one can see that $h_{1}(\theta,e) = \{\theta,e\}$ for all $\theta$. A rather simple computation shows that $\zeta_{1}$ defined by the equality $h_{1}(0,r)
= \{ - \zeta_{1}(p),g(r)\}$ fulfills the conditions of our theorem.
The following theorem is proved for the group $G$ with a finite-dimensional extended algebra $\mathfrak{H}'$.
Let $G$ be a connected and simply connected Lie group. Then to every exponent $\xi(r,s,X)$ of $G$ defined locally in $(r,s)$ there exists an exponent $\xi_{0}$ of $G$ defined on the whole group $G$ which is an extension of $\xi$. If $\xi$ is differentiable, $\xi_{0}$ may be chosen differentiable.
Because the proof of Theorem 5 is almost identical to that of Theorem 5.1 in Bargmann ([*loc. cit.*]{}), we do not present it explicitly[^18]. Please note that the proof rests largely on the global theory of classical (finite-dimensional) Lie groups. Namely, it rests on the theorem that to any finite dimensional Lie group there always exists a universal covering group . We can use those methods because of the existence of a finite-dimensional extension $H'$ of $G$.
We have obtained the full classification of time-dependent $\xi$ defined on the whole group $G$ for Lie groups $G$ which are connected and simply connected in non-relativistic theory. But for any Lie group $G$ there exists a universal covering group $G^{*}$ which is connected and simply connected. Thus, for $G^{*}$ the correspondence $\xi \to \Xi$ is one-to-one, that is, to every $\xi$ there exists a unique $\Xi$ and vice versa, to every $\Xi$ corresponds a unique $\xi$ defined on the whole group $G^{*}$, and the correspondence preserves the equivalence relation. Because $G$ and $G^{*}$ are locally isomorphic, the infinitesimal exponents $\Xi$’s are exactly the same for $G$ and for $G^{*}$. Since to every $\Xi$ there does exist exactly one $\xi$ on $G^{*}$, so, if to a given $\Xi$ there exists the corresponding $\xi$ on the whole $G$, then such a $\xi$ is unique. In this way, we have obtained the full classification of $\xi$ defined on a whole Lie group $G$ for any Lie group $G$, in the sense that no $\xi$ can be omitted in the classification. The set of equivalence classes of $\xi$ is considerably smaller than that for $\Xi$; it may happen that to some local $\xi$ there does not exist any global extension.
Examples
--------
[**Example 1: The Galilean Group**]{}. According to the conclusions of subsection \[Time\] one should *a priori* investigate such representations of the Galilean group $G$ which fulfill Eq. (\[rowt\]), with $\xi$ depending on time. Then, the following paradox arises. Why has the transformation law $T_{r}$ under the Galilean group a time-independent $\xi$ in (\[rowt\]), regardless of whether it is a covariance group or a symmetry group? We will solve the paradox in this subsection. Namely, we will show that any representation of the Galilean group fulfilling (\[rowt\]) is equivalent to a representation fulfilling (\[rowt\]) with time-independent $\xi$. This is a rather peculiar property of the Galilean group, not valid in general. For example, this is not true for the group of Milne transformations.
In non-relativistic theory $\xi=\xi(r,s,t)$ depends on the time. In this case, according to our assumption about $G$, any $r \in G$ transforms simultaneity hyperplanes into simultaneity hyperplanes. Thus, there are two possibilities for any $r \in G$. First, when $r$ does not change time: $t(rp) = t(p)$, and the second in which time is changed, but in such a way that $t(rp) - t(p) = f(t)$. *We assume in addition that the base generators $a_{k} \in \mathfrak{G}$ can be chosen in such a way that only one acts on time as translation and the remaining ones do not act on time*. We can assume that the operators $\boldsymbol{a}$ are ordinary differential operators. Hence, the Jacobi identity (\[21\]) reads $$\label{Jac1}
\Xi([a,a'],a'') + \Xi([a', a''],a) + \Xi([a'',a],a')
= \partial_{t} \Xi(a',a''),$$ if one and only one among $a,a',a''$ is the time-translation generator, namely $a$, and $$\begin{aligned}
\label{Jac2}
\Xi([a,a'],a'') + \Xi([a',a''], a) + \Xi([a'',a],a') = 0,\end{aligned}$$ in all of the remaining cases. The Jacobi identities (\[Jac1\]) and (\[Jac2\]) can be treated as a system of ordinary differential linear equations for the finite set of unknown functions $\Xi_{ij}(t) = \Xi(a_{i}, a_{j},t)$, where $a_{i}$ is the base in the Lie algebra of $G$.
According to Section \[generalization\], in order to classify all $\xi$ of $G$ we shall determine all equivalence classes of infinitesimal exponents $\Xi$ of the Lie algebra $\mathfrak{G}$ of $G$. The commutation relations for the Galilean group are as follows $$\label{26a}
[a_{ij},a_{kl}] = \delta_{jk}a_{il} - \delta_{ik}a_{jl}+\delta_{il}a_{jk}-
\delta_{jl}a_{ik},$$
$$\label{26b}
[a_{ij},b_{k}] = \delta_{jk}b_{i} - \delta_{ik}b_{j}, \, [b_{i},b_{j}] = 0,$$
$$\label{26c}
[a_{ij.d_{k}}] = \delta_{jk}d_{i} - \delta_{ik}d_{j}, \, [d_{i},d_{j}]
= 0, [b_{i},d_{j}] = 0,$$
$$\label{26d}
[a_{ij},\tau] = 0, [b_{k},\tau] = 0, [d_{k},\tau] = b_{k},$$
where $b_{i},d_{i}$ and $\tau$ stand for the generators of space translations, with the proper Galilean transformations and time translation respectively and $a_{ij} = - a_{ji}$ being rotation generators. Please note that the Jacobi identity (\[Jac2\]) is identical to that in the ordinary Bargmann’s theory of time-independent exponents see Bargmann ([*loc. cit.*]{}), Eqs (4.24) and (4.24a)). Thus, using (\[26a\]) – (\[26c\]) we can proceed exactly after Bargmann ([*loc. cit.*]{}), pages 39, 40) and show that any infinitesimal exponent defined on the subgroup generated by $b_{i}, d_{i}, a_{ij}$ is equivalent to an exponent equal to zero, with the possible exception of $\Xi(b_{i},d_{k},t) = \gamma \delta_{ik}$, where $\gamma = \gamma(t)$. Hence, the only components of $\Xi$ defined on the whole algebra $\mathfrak{G}$ which can be a priori not equal to zero are: $\Xi(b_{i},d_{k},t) = \gamma \delta_{ik}, \, \Xi(a_{ij},
\tau,t), \,
\Xi(b_{i}, \tau,t)$ and $\Xi(d_{k},\tau,t)$. First, we compute function $\gamma(t)$. Substituting $a= \tau, \, a' = b_{i}, a'' = d_{k}$ to (\[Jac1\]), we get ${{\mathrm{d}}}\gamma/{{\mathrm{d}}}t = 0$, so that $\gamma$ is a constant, denoting the constant value of $\gamma$ by $m$. By inserting $a = \tau, \, a' = a_{i}^{s}, \, a'' = a_{sj}$ to (\[Jac1\]) and summing up with respect to $s$, we get $\Xi(a_{ij},\tau,t) = 0$. In the same way, but with the substitution $a = \tau, a' = a_{i}^{s}, a''
= b_{s}$, one shows that $\Xi(b_{i},\tau, t) = 0$. At last, the substitution $a=\tau, a' = a_{i}^{s}, a''=d_{s}$ to (\[Jac1\]) and summation with respect to $s$ gives $\Xi(d_{i},\tau,t) = 0$. In this way, we have proved that any time-dependent $\Xi$ on $\mathfrak{G}$ is equivalent to a time-independent one. In other words, we get a one-parameter family of possible $\Xi$, with the parameter equal to the inertial mass $m$ of the system in question. Any infinitesimal time-dependent exponent of the Galilean group is equivalent to the above time-independent exponent $\Xi$ with some value of the parameter $m$; and any two infinitesimal exponents with different values of $m$ are nonequivalent. As was argued in subsection \[generalization\] (Theorems 3 $\div$ 5), the classification of $\Xi$ gives a full classification of $\xi$. Moreover, it can be shown that the classification of $\xi$ is equivalent to the classification of possible $\theta$-s in the transformation law $$\label{trapsi}
T_{r}\psi(p) = e^{i\theta(r,p)}\psi(r^{-1}p)$$ for the spinless non-relativistic particle. On the other hand, the exponent $\xi(r,s,t)$ of the representation $T_{r}$ given by (\[trapsi\]) can be easily computed to be equal to $\theta(rs,p) - \theta(r,p) - \theta(s,r^{-1}p)$, and the infinitesimal exponent belonging to $\theta$ defined as $\theta(r,p) = -m\vec{v}\centerdot \vec{x} + \frac{m}{2}\vec{v}^{2}t$, covers the whole one-parameter family of the classification (its infinitesimal exponent is equal to that infinitesimal exponent $\Xi$, which has been found above). Thus, the standard $\theta(r,p) = - m\vec{v}
\centerdot \vec{x} + \frac{m}{2}\vec{v}^{2}t$, covers the full classification of possible $\theta$-s in (\[trapsi\]) for the Galilean group. Inserting the standard form for $\theta$ we see that $\xi$ does not depend on time but only on $r$ and $s$. By this, any time-dependent $\xi$ on $G$ is equivalent to its time-independent counterpart.
In this way, we have reconstructed the standard result. Using now the formula[^19] $$A_{i}\psi(p) = \lim_{\sigma \to 0} \frac{(T_{\sigma a_{i})}
- \boldsymbol{1})\psi(p)}{\sigma},$$ for the generator $A_{i}$ corresponding to $a_{i}$, we get the standard commutation relations for the *ray* representation $T_{r}$ of the Galilean group $$[A_{ij},A_{kl}] = \delta_{jk}A_{il} - \delta_{ik}A_{jl} - \delta_{jl}A_{ik},$$
$$[A_{ij},B_{k}] = \delta_{jk}B_{i} - \delta_{ik}B_{j}, \, [B_{i},B_{j}] = 0,$$
$$[A_{ij}, D_{k}] = \delta_{jk}D_{i} - \delta_{ik}D_{j},$$
$$[D_{i},D_{j}] =0, \, [B_{i},D_{j}] = m\delta_{ij},$$
$$[A_{ij}, T] = 0, \, [B_{k},T] = 0, \, [D_{k}, T] = B_{k}.$$
Please note that to any $\xi$ (or $\Xi$) there exists a corresponding $\theta$ (and such a $\theta$ is unique up to a trivial equivalence relation). As we will see, this is not the case for the Milne group, where some $\Xi$’s do exist which cannot be realized by any $\theta$.
[**Example 2: Milne group and equality of inertial and gravitational masses**]{}. In here we apply the theory of Section \[generalization\] to the Milne transformations group. We proceed like with the Galilean group in the preceding section. The Milne group $G$ does not form any Lie group, but in the physical application it is sufficient for us to consider some Lie subgroups $G(m)$ of the Milne group. We will go on according to the following plan. First, we compute the infinitesimal exponents and exponents for each $G(m)$, $m = 1,2, \ldots $, and then the $\theta$ in (\[trapsi\]) for $G(m)$. Please compare ([*J. Wawrzycki, math-ph/0301005*]{}), where the result is extended on the whole group.
The Milne transformation is defined as follows $$\label{tra}
(\vec{x},t) \to (R\vec{x} + \vec{A}(t), t + b),$$ where $R$ is an orthogonal matrix, and $b$ is constant. The extent of arbitrariness of function $\vec{A}(t)$ in (\[tra\]) will be left undetermined for now. It is convenient to rewrite the Milne transformations (\[tra\]) in the following form $$\vec{x'} = R\vec{x} + A(t) \vec{v}, \, \, \, t' = t + b,$$ where $\vec{v}$ is a constant vector which does not depend on time $t$. We define the subgroup $G(m)$ of $G$ as the group of the following transformations $$\vec{x'} = R\vec{x} + \vec{v}_{(0)} + t\vec{v}_{(1)}
+ \frac{t^{2}}{2!}\vec{v}_{(2)} + \ldots + \frac{t^{m}}{m!}\vec{v}_{(m)},
\, \, \, t' = t + b,$$ where $R = (R_{a}^{b}), v_{(n)}^{k}$ are the group parameters; in particular, the group $G(m)$ has the dimension equal to $3m + 7$.
Now Let us investigate the group $G(m)$, that is, classify its infinitesimal exponents. The commutation relations of $G(m)$ are as follows $$\label{Milne1}
[a_{ij},a_{kl}] = \delta_{jk}a_{il} - \delta_{ik}a_{jl}
+ \delta_{il}a_{jk} - \delta_{il}a_{ik},$$
$$\label{Milne2}
[a_{ij},d_{k}^{(n)}] = \delta_{jk}d_{i}^{(n)}
- \delta_{ik}d_{j}^{(n)}, \, [d_{i}^{(n)},d_{j}^{(k)}] = 0,$$
$$\label{Milne3}
[a_{ij},\tau] = 0, \, [d_{i}^{(0)},\tau]=0, \, [d_{i}^{(n)},\tau]
= d_{i}^{(n-1)},$$
where $d_{i}^{(n)}$ is the generator of the transformation $r(v_{(n)}^{i})$: $${x'}^{i} = x^{i} + \frac{t^{n}}{n!}v_{(n)}^{i},$$ which will be called the $n$-acceleration, and 0-acceleration in the particular case of the ordinary space translation. All relations (\[Milne1\]) and (\[Milne2\]) are identical to (\[26a\]) $\div$ (\[26c\]) with the $n$-acceleration instead of the Galilean transformation. Thus, the same argumentation as that used for the Galilean group gives: $\Xi(a_{ij}, a_{kl})= 0$, $\Xi(a_{ij},d_{k}^{(n)}) = 0$, and $\Xi(d_{i}^{(n)}, d_{j}^{(n)}) = 0$. Substituting $a_{i}^{h}, a_{hi},
\tau$ for $a,a',a''$ into Eq. (\[Jac1\]),making use of the commutation relations and summing up with respect to $h$, we get $\Xi(a_{ij},\tau) = 0$. In an analogous way, substituting $a_{i}^{h}, d_{h}^{(l)}, d_{k}^{(n)}$ for $a,a',a''$ into Eq. (\[Jac2\]), we get $\Xi(d_{i}^{(l)}, d_{k}^{(n)})
= \frac{1}{3}\Xi(d^{(l)h},d_{h}^{(n)}) \, \delta_{ik}$. Substituting $a_{i}^{h}, d_{h}^{(n)}, \tau$ for $a,a',a''$ into Eq. (\[Jac1\]), making use of the commutation relations, and summing up with respect to $h$, we get $\Xi(d_{i}^{(n)}, \tau) = 0$. Now, we substitute $d_{k}^{(n)}, d_{i}^{(0)}, \tau$ for $a,a',a''$ in (\[Jac1\]), and proceed recurrently with respect to $n$, obtaining in this way $\Xi(d_{i}^{(0)},d_{k}^{(n)})
= P^{(0,n)}(t)\delta_{ik}$, where $P^{(0,n)}(t)$ is a polynomial of degree $n-1$, and the time derivation of $P^{(0,n)}(t)$ has to be equal to $P^{(0,n-1)}(t)$, and $P^{(0,0)}(t) = 0$. Substituting $d_{k}^{(n)}, d_{i}^{(l)}, \tau$ into (\[Jac1\]), in the same way we get $\Xi(d_{k}^{(l)}, d_{i}^{(n)}) = P^{(l,n)}(t)\delta_{ki}$, where $\frac{{{\mathrm{d}}}}{{{\mathrm{d}}}t}P^{(l,n)} = P^{(l-1,n)} + P^{(l,n-1)}$. This allows us to determine all $P^{(l,n)}$ by the recurrent integration process. Please note that $P^{(0,0)} = 0$, and $P^{(l,n)} = - P^{(n,l)}$, so given the $P^{(0,n)}$ we can compute all $P^{(1,n)}$. Indeed, we have $P^{(1,0)} = - P^{(0,1)}, P^{(1,1)} = 0, {{\mathrm{d}}}P^{(1,2)}/{{\mathrm{d}}}t = P^{(0,2)}
+ P^{(1,1)}, {{\mathrm{d}}}P^{(1,3)}/{{\mathrm{d}}}t = P^{(0,3)} + P^{(1,2)},
\ldots$, and after $m-1$ integrations we compute all $P^{(1,n)}$. Each elementary integration introduces a new independent parameter (the arbitrary additive integration constant). Exactly in the same way, given all $P^{(1,n)}$ we can compute all $P^{(2,n)}$ after $m-2$ elementary integration processes. In general, the $P^{(l-1,n)}$ allows us to compute all $P^{(l,n)}$ after $m-l$ integrations. Thus, $P^{(l,n)}(t)$ are $l+n - 1$-degree polynomial functions of $t$, and all are determined by $m(m+1)/2$ integration constants. Because $d[\Lambda](d_{i}^{(n)}, d_{k}^{(l)}) = 0$, the exponents $\Xi$ defined by different polynomials $P^{(l,n)}$ are inequivalent. Therefore, the space of nonequivalent classes of $\Xi$ is $m(m+1)/2$-dimensional.
However, not all $\Xi$ can be realized by the transformation $T_{r}$ of the form (\[trapsi\]). It can seen that any integration constant $\gamma_{(l,q)}$ of the polynomial $P^{(l,q)}(t)$ has to be equal to zero if $l,q \neq 0$, provided the exponent $\Xi$ belongs to the representation $T_{r}$ of the form (\[trapsi\]). By this, all exponents of $G(m)$ which can be realized by the transformations $T_{r}$ of the form (\[trapsi\]) are determined by the polynomial $P^{(0,m)}$, that is, by $m$ constants. We omit the proof of this fact, and refer the reader to ([*J. Wawrzycki, math-ph/0301005*]{}), Example 2. In the proof we compute the exponent $\Xi$ directly for the transformation $T_{r}$ of the form (\[trapsi\]) and compare it with the classification results above.
Consider the $\theta$, given by the formula $$\label{thetaG(m)}
\theta(r,p) = \gamma_{1}\frac{{{\mathrm{d}}}\vec{A}}{{{\mathrm{d}}}t}
+ \gamma_{2}\frac{{{\mathrm{d}}}^{2}\vec{A}}{{{\mathrm{d}}}t^{2}} + \ldots +
\gamma_{m}\frac{{{\mathrm{d}}}^{m}\vec{A}}{{{\mathrm{d}}}t^{m}} + \widetilde{\theta}(t),$$ for $r \in G(m)$, where $\gamma_{i}$ are the integration constants which define the polynomial $P^{(0,m)} = \gamma_{1}\frac{t^{m-1}}{(m-1)!}
+ \gamma_{2}\frac{t^{(m-2)}}{(m-2)!} + \ldots + \gamma_{m}$, and $\widetilde{\theta}(t)$ is any function of time $t$, and eventually of the group parameters. A rather simple computation shows that this $\theta$ covers all possible $\Xi$ which can be realized by (\[trapsi\]). That is, the infinitesimal exponents corresponding to the $\theta$ given by (\[thetaG(m)\]) yield all possible $\Xi$ with all integration constants $\gamma_{(k,n)} = 0$, for $k,n \neq 0$. Thus, the most general $\theta(r,p)$ defined for $r \in G(m)$ is given by (\[thetaG(m)\]).
At this point we make use of the assumption that the wave equation is local. It can be shown then (we leave this without proof) that the $\theta(r,p)$ can be a function of a finite order derivatives of $\vec{A}(t)$, say $k$-th at most, while the higher derivatives cannot enter into $\theta$. Therefore, the most general $\theta(r,p)$ defined for $r \in G(m)$ has the following form $$\label{thetaG}
\theta(r,X) = \gamma_{1}\frac{{{\mathrm{d}}}\vec{A}}{{{\mathrm{d}}}t} + \ldots
+ \gamma_{k}\frac{{{\mathrm{d}}}^{k}\vec{A}}{{{\mathrm{d}}}t^{k}}
+ \widetilde{\theta}(t).$$
Having obtained this we can infer the most general Schrödinger equation for a spinless particle in Newton-Cartan spacetime, cf. ([*J. Wawrzycki, Acta Phys. Polon. [**B 35**]{}, 613 (2004); gr-qc/0301102.*]{}). The inertial and the gravitational masses are always equal in this equation. Our assumptions are, more precisely, as follows: (i) The quantum particle, when its kinetic energy is small in comparison to its rest energy $mc^{2}$, does not exert any influence on the space-time structure. (ii) The Born interpretation for the wave function is valid, and the transition probabilities in the Newton-Cartan space-time which describes geometrically Newtonian gravity, are equal to the ordinary integral over a simultaneity hyperplane and are preserved under the coordinate transformations. (iii) The wave equation is linear, of second order at most, generally covariant, and can be built in a local way with the help of the geometrical objects describing the space-time structure. (iv) The probability density $\rho = \psi^{*}\psi$ is a scalar field (with the scalar transformation rule). In fact the conditions (i), (ii), (iii) and (iv) are somewhat interrelated. The coefficients $a, b^{i}, \ldots$ in the wave equation $$\Big[ a\partial_{t}^{2} + b^{i}\partial_{i}\partial_{t} +
c^{ij}\partial_{i}\partial_{j} + f^{i}\partial_{i} +
d\partial_{t} + g \Big]\psi = 0,$$ are local functions of the potential and therefore cannot depend on arbitrary high order derivatives of the potential. Within a rather standard analysis one gets the Schrödinger equation from (i), (ii), (iii), and (iv), which after the ordinary notation of constants has the form $$\Big[ \frac{\hslash^{2}}{2m}\delta^{ij}\partial_{i}\partial_{j}
+i\hslash\partial_{t} -m\phi+\Lambda\Big]\psi = 0,$$ with the $\theta$ in $T_{r}$ given by $$\theta=\frac{m}{2\hslash} \int_{0}^{t} \dot{\vec{A}}^{2}(\tau) \, {{\mathrm{d}}{\tau}}
+\frac{m}{\hslash}\dot{A}_{i}x^{i}.$$ $\phi$ is the gravitational potential and $\Lambda$ is one of the Kronecker’s invariants of the matrix $(\partial_{a}\partial_{b}\phi)$ in the above equation.
Note that the the inertial mass $m$ in the equation is equal to the parameter at the gravitational potential. That is, the gravitational mass must be equal to the inertial mass.
[^1]: Electronic address: [email protected]
[^2]: Let us recall that all relevant information carried by $\phi$ is contained in the set of numbers $$\frac{\vert(\phi, \varphi)\vert^{2}}{(\phi,\phi)(\varphi, \varphi)},$$ where $(\phi, \varphi)$ is the scalar product of the Hilbert space $\mathcal{H}$. As such $\phi$ and $e^{i\xi} \phi$ are equivalent containing exactly the same information.
[^3]: Probably it will be helpful to recall the formula for the scalar product in $\mathcal{H}_{F}$ $$(\psi,\psi') = c^{*}c' + \int_{\mathbb{R}^{3}}\psi_{1}^{*}\psi_{1}' \, d^{3}x +
\int_{\mathbb{R}^{2\times 3}}\psi_{2}^{*}\psi_{2}' \, d^{3}x_{1}d^{3}x_{2} + \ldots$$ to see the equivalence of $\psi$ and $e^{i\xi(p)}\psi$, compare the preceding footnote.
[^4]: The term “ordinary” means here the ordinary QM applied to system with infinite number of degrees of freedom.
[^5]: Remember, please, that the number of particles $N$ in quantum theory corresponds to to the number of degrees of freedom in classical theory, where the degree of freedom is in the sense of the Lagrange-Hamilton theory. The laws serving us as guiding receipt should not depend on $N$ as we have mentioned earlier. Therefore, those laws — if taken from classical theory — should not depend on the number of Lagrange degrees of freedom.
[^6]: Here “space” stands for an appropriate space-like Cauchy section of (classical) spacetime.
[^7]: More precisely we have to construct first the specific representation of the algebra with the vacuum plying the role of the cyclic vector of the representation. Then, the the states of the Hilbert space of the representation constitute the analogue of classical space positions.
[^8]: The notion of a noncommutative space at a particular classical time is rather a strange mixture of classical and noncommutative properties which cannot serve as a generally valid universal concept.
[^9]: Strictly speaking the matter is even more involved but we do not go into detail now in order to avoid excessive mathematical complexities. We have assumed implicitly that the spacetime is compact so that the wave functions — the respective cross sections of the Hilbert bundle $\mathcal{M}\triangle\mathcal{H}$ — do belong to the so-called direct integral Hilbert space $\int_{M} \mathcal{H}_{p} \, d\mu(p)$, $p\in\mathcal{M}$; compare the respective section of this Chapter. After this one can think of $\mathcal{A}^{CCR}$ and $\mathcal{A}$ as of acting in this direct integral Hilbert space.
[^10]: This gives the conception of the ray, introduced to Quantum Mechanics by Hermann Weyl \[H. Weyl, [*Gruppentheorie und Quantenmechanik*]{}, Verlag von S. Hirzel in Leipzig (1928)\]: a physical state does not correspond uniquely to a normed state $\varphi \in \mathcal{H}$, but it is uniquely described by a *ray*; two states belong to the same ray if they differ by a constant phase factor.
[^11]: J. v. Neumann, [*Mathematical Principles of Quantum Mechanics*]{}, University Press, Princeton (1955). He did not mention gauge freedom on that occasion. However, gauge freedom is necessary for the equivalence of $\psi_{1}$ and $\psi_{2}
= e^{i\xi(t)}\psi_{1}$.
[^12]: People usually think that the cutoff is a man-made process which one has to perform due to the fact that the theory at our disposal is incapable of the high energy processes. This opinion was justified in the early seventies, when the strong interactions were out of our scope. One could hope that the strong interactions will cancel the infinities. But now it seems that this hope turned out to be vain. Strong interactions do not cancel the infinities. Paradoxically QED serves now as the paradigm for a successful quantum field theory.
[^13]: This time $A_{\mu}(x)$ is an operator-valued distribution
[^14]: It becomes clear in further analysis that the group $G$ in question has to fulfill the consistency condition requiring that for any $r\in G$, $rt$ is a function of time only in the case of non-relativistic theory with (\[rowt\])
[^15]: The limit in the expression can be understood in the ordinary point-wise sense with respect to $p$, but also in any linear topology in the function linear space (with obvious addition) of $\theta(p)$, providing that $p \to \Lambda(a, p)$ is differentiable in the sense of this linear topology. Further on, the simple notation $$\boldsymbol{a} f(p) = \frac{df((\tau a)p)}{d\tau}\Big\vert_{\tau = 0}
= \lim_{\epsilon \to 0}\frac{f((\epsilon a)p) - f(p)}{\epsilon},$$ will be used.
[^16]: The limit process with the help of which the generator is computed refers to the topology in $D$, of course.
[^17]: Let us stress once more that $$\label{strong}
\boldsymbol{a}\theta(p) = \lim_{\epsilon \to 0}\frac{\theta((\epsilon a)p)
- \theta(p)}{\epsilon},$$ and the limit is in the sense of topology induced from the Lie group $D$.
[^18]: In this proof we consider the finite-dimensional extension $H'$ of $G$ instead of the Lie group $H$ in the proof presented in Bargmann ([*loc. cit.*]{}). The remaining replacements are rather trivial, but we mark them here explicitly to simplify the reading (1) Instead of the formula $\bar{r}' = \bar{t}(\theta)\bar{r}
= \bar{r}\bar{t}(\theta)$ of (5.3) in Bargmann ([*loc. cit.*]{}), we have $\check{r}' = \check{t}(\theta(r^{-1}p))\check{r}
= \check{r}\check{t}(\theta(p))$. Thus, from the formula $(\check{h}_{1}(r)\check{h}_{1}(s))\check{h}_{1}(g)$ $= \check{h}_{1}(r)(\check{h}_{1}(s)\check{h}_{1}(g))$ see [*V. Bargmann, loc. cit.*]{}) it follows that $\xi(r,s,p) + \xi(rs,g,p)$ $= \xi(s,g,r^{-1}p) + \xi(r,sg,p)$ instead of (5.8) in Bargmann ([*loc. cit.*]{}). (2) Instead of (4.9), (4.10) and (4.11) we use the Iwasawa-type construction presented in this paper.
[^19]: The transformation $T_{r}$ does not act in the ordinary Hilbert space but in the Hilbert bundle space $\mathcal{R}\triangle\mathcal{H}$, hence we cannot immediately appeal to the Stone and G[å]{}rding Theorems. Nonetheless, $T_{r}$ induces a unique unitary representation acting in the Hilbert space $\int_{\mathcal{R}} \mathcal{H}_{t} d\mu(t)$ and it can be shown that it is meaningful to talk about the generators $A$ of $T_{r}$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a new spectroscopic classification for OB stars based on $H-$band (1.5 to 1.8 ) observations of a sample of stars with optical spectral types. Our initial sample of nine stars demonstrates that the combination of 1.7002 and H Brackett series absorption can be used to determine spectral types for stars between $\sim$ O4 and B7 (to within $\sim$ $\pm$ 2 sub–types). We find that the Brackett series exhibits luminosity effects similar to the Balmer series for the B stars. This classification scheme will be useful in studies of optically obscured high mass star forming regions. In addition, we present spectra for the OB stars near 1.1 and 1.3 which may be of use in analyzing their atmospheres and winds.'
author:
- 'R. D. Blum, T. M. Ramond, P. S. Conti'
- 'D. F. Figer'
- 'K. Sellgren'
title: '$H-$Band Spectroscopic Classification of OB Stars'
---
2o[H$_2$O]{}
[*accepted for publication in the*]{} AJ
INTRODUCTION
============
OB stars are massive and thus short lived. Because they have short lives, they will be confined to regions relatively close to their birthplaces and will be found close to the Galactic plane. The youngest massive stars will also still be in or near their dusty star forming environment. Due to the strong extinction from interstellar dust at optical wavelengths, OB stars at large distances from the sun, or in young star forming regions, are more easily studied at near infrared (1 $-$ 5 ) or longer wavelengths. Note that the extinction at $V$ is about 10 times greater in $magnitudes$ than at 2.2 , and for 3.5 $\lambda$ 0.9 , the extinction can be represented by a power–law dependence: $A_{\lambda}$ / $A_{2.2 \mu m}$ = $(\lambda / 2.2)^{-1.7}$ (Mathis 1990).
The advent of sensitive infrared array detectors has made near infrared classification schemes commonplace for a wide range of stellar spectral types. For example, Kleinmann & Hall (1986), Greene & Meyer (1995), Ali et al. (1995), and Ramírez et al. (1997) have presented $K-$band spectra of late–type stars useful for classification at 2 . Lançon & Rocca–Volmerange (1992) have presented $H$ and $K$ spectra of optically classified stars ranging in spectral type from O6 to M7 (only three OB stars were included). Eenens et al. (1991), Blum et al. (1995$a$), Tamblyn et al. (1996), Morris et al. (1996), and Figer et al. (1996, 1997 ) have presented $K-$band spectra of known Wolf–Rayet and other emission–line stars. A detailed classification scheme in the $K-$band for early–type (OB) stars has been presented by Hanson & Conti (1994) and Hanson et al. (1996, hereafter HCR96). All these infrared classification spectra are now routinely applied to the study of obscured stars, typically in star forming regions and/or along sight lines with large interstellar obscuration by dust (e.g. Greene & Meyer 1995; Blum et al. 1995$a$,$b$; Tamblyn et al. 1996; Figer et al. 1996; Ali 1996; Blum et al. 1996).
Classification of hot stars at near infrared wavelengths is challenging because the photospheric absorption features used in the classification (primarily H, , and ) are generally weaker than the molecular and atomic (metal) lines used for late–type stars. In addition, there can also be strong circumstellar emission which dilutes the photospheric lines. It is this latter problem that motivates the present work. Since we expect circumstellar dust emission to dominate the $K-$band spectra of some stars in star forming regions (e.g., in M17, Hanson & Conti 1995), a more robust classification scheme for high mass star forming regions will also utilize shorter wavelengths which may be less affected by circumstellar excess emission. Conversely, we wish to observe at the longest possible wavelengths shortward of $K$ to overcome the extinction of intervening dust. In this paper, we present a preliminary classification of OB stars based on spectra between 1.5 and 1.8 (i.e. in the $H-$band). We also present spectra of the same stars between 1 and 1.3 which may prove useful in the analysis of massive stars atmospheres.
OBSERVATIONS and DATA REDUCTION
===============================
The OB stars were observed on the nights of 9, 10, and 13 June 1996 using the Ohio State Infrared Imager and Spectrometer (OSIRIS) on the Perkins 1.8–m telescope of the Ohio Wesleyan and Ohio State Universities at the Lowell Observatory. The Perkins telescope is located on Anderson Mesa near Flagstaff, Arizona. OSIRIS is more fully described by DePoy et al. (1993). The target stars were taken from the catalog of HCR96 and are listed, along with their optical spectral types, in Table \[stars\]. In addition to the target OB stars, we observed A–type and G–type stars for use in canceling telluric absorption features. We will refer to these stars as “atmospheric standards.”
\[stars\]
\[eqw\]
OSIRIS was used in cross–dispersed mode which gives $\lambda/\Delta\lambda \approx 570$ (2 pixels) while covering the $J$ ($\lambda _{\circ} \approx$ 1.25 , $\Delta\lambda \approx$ 0.30 ), $H$ ($\lambda_\circ \approx$ 1.65 , $\Delta\lambda \approx$ 0.37 ), and $K$ ($\lambda_\circ \approx$ 2.20 , $\Delta\lambda \approx$ 0.50 ) bands simultaneously. One star (HD183143) was also observed in long slit mode at $J$ and $H$. Long slit mode is similar to the cross–dispersed mode (same spectral resolution and spatial scale) except only one wavelength band is acquired at a time and the long slit fills the entire spatial dimension on the array. Similar results were obtained from both sets of spectra; however, the long slit mode spectra are slightly higher S/N and are therefore presented here. In addition, five of the stars observed in cross–dispersed mode were also observed at $\lambda$ $\approx$ 1.10 ($\Delta\lambda$ $\approx$ 0.25 ), a bandpass which we call $I'$, in long slit mode. Analysis of $H-$band OH sky lines results in linear dispersions of 9.70, 11.60, 14.53, and 19.37 Å at $I'$, $J$, $H$, and $K$, respectively. The sky lines were identified using the list of Oliva & Origlia (1992). $I'$, $J$, and $K$ band dispersions follow from the $H-$band dispersion and the appropriate order number, m (m $=$ 6, 5, 4, 3 for $I'$, $J$, $H$, $K$). In cross–dispersed mode, OSIRIS has a $\sim$ 60$''$ $\times$ 5$''$ slit (1.5$''$ ). The target stars were observed at 12 uniformly spaced positions along the slit. Combining spectra over a uniform grid along the slit greatly reduces systematic errors which may be introduced, for example, by scattered light in the dome flats and fringing (see Blum et al. 1995$a$ for a discussion of these problems in OSIRIS). The seeing varied throughout each night from 2$''$ to 4$''$. None of the nights were photometric; we observed through thin clouds. This will not affect our results since we are interested in the relative intensities and line strengths, not absolute fluxes.
All basic data reduction was accomplished using IRAF[^1]. Each image was flat–fielded using dome flats and then sky subtracted using another image from the grid with the star displaced by several positions along the slit (cross–dispersed mode) or with a median combination sky image (long slit mode). Individual spectra were extracted from each program star image and atmospheric standard image using IRAF “APEXTRACT.” Synthesized apertures $\pm$ 3 pixels wide were used. The entire grid of 1–d spectra for each star was then combined (after scaling). In the case of the long slit spectra, the individual 1–d spectra were first shifted (0 to $\sim$ 4 pixels) to account for anamorphic demagnification along the slit spatial dimension.
The final spectra were obtained by ratioing the program stars with a atmospheric standard star which had first been corrected for intrinsic H absorption lines (P$\beta$ 1.281 , and Brackett series 1.51 to 2.165 ). The atmospheric standards were observed in pairs of A–type and G–type stars. The H lines in the A stars were corrected in the following way. The majority (seven) of our program stars were corrected using an A0 V and a G2 V. In this case, the G2 star (BS 6847) was first corrected for H line absorption using the NOAO solar atlas (Livingston & Wallace 1991). Next a ratio of the A star (BS 7734) to the corrected G star was made. Since these two were taken at nearly the same airmass, the resulting ratio contains essentially only the H line spectrum of the A star and perhaps some “emission” lines due to metal lines in the G star. The H lines in this ratio were then fit with Gaussian profiles. The resultant line fits were used to correct the H absorption in the A star. The two remaining stars (HD183143 and BD+24 3866) were corrected using an A3 V (BS 7958) and a G9 III (BS 7760). We have no intrinsic spectrum to correct BS 7760 with, so the $J$ and $K-$band lines were corrected by eye. No correction was made for the $H-$band Brackett lines in BS 7760. Since BS 7760 was not corrected using a matching intrinsic H spectrum, we are less confident in the precise H line absorption resulting in HD183143 and BD+24 3866; however, none of our primary results is sensitive to the accuracy of our measurement of H absorption in the program stars but only the appearance or absence of a definitive Brackett series. No telluric correction was made for the $I'$ spectra as a cursory inspection of the spectra shows no strong telluric lines. We did obtain A and G star spectra for use in making a similar telluric correction as described above. Making this correction shows no difference in the final $I'$ spectra of the OB stars at the few percent level and serves only to introduce more noise.
We have also included two $H-$band spectra (HD 157955 and HD 169033) which were obtained as part of a program concentrating on $K-$band spectra of stars in an inner Galaxy star forming region called the Quintuplet (Figer et al. 1996). These spectra have similar spectral (16.75 Å) resolution as the OSIRIS data and spatial scale somewhat smaller (0.7$''$ ); see Figer et al. (1995). The telluric correction was made by ratioing the stars to a Quintuplet star (q3 in Moneti et al. 1994) and a dusty Wolf-Rayet star (WR118), neither of which have spectral lines near the or Brackett lines. These stars were observed at higher airmass ($\sim$ 2) than the other stars presented here ($\sim$ 1) and may not be corrected as well (the higher Brackett lines appear to suffer from incomplete correction).
RESULTS and DISCUSSION
======================
The $H-$band Spectra
--------------------
The final $H-$band spectra are shown in Figure \[spect\]. To highlight the absorption features, we have divided each spectrum by a low order fit to the continuum. We do not present our $K-$band spectra here since spectra for these stars have been presented by HCR96 (at higher spectral resolution). We note that our lower resolution $K-$band spectra reproduce the basic features of the HCR96 spectra. Line identifications used throughout this paper were taken from Weise et al. (1966, H, ), and Garcia & Mack (1965, ). All wavelengths used and/or quoted are in air. Line equivalent widths (Table \[eqw\]) were measured by fitting Gaussian profiles using the LINER program (Pogge 1997). Uncertainties are the formal one sigma errors derived from the rms scatter in the continuum corresponding to a bandpass equal to the full width at zero intensity of the lines. The errors do not reflect the line fit.
### The Brackett Series
The H lines are sensitive to the excitation in the stellar atmosphere. For example, the strength of the well–known Balmer lines increases with as more atoms are excited to the n$=$2 level. The Balmer lines reach peak strength at $\sim$ 9000 K (Gray 1992) as Hydrogen becomes appreciably ionized at higher temperatures. The Brackett series excitation is similar to that for the Balmer series (10.8 eV compared to 10.2 eV), so we expect the lines to behave in a similar manner. This basic behavior with (spectral type) is seen in Figure \[spect\].
Inspection of Figure \[spect\] suggests the Brackett lines behave similarly to the Balmer lines in early–type stars as first noted by Adams & Joy (1922, 1923): the width of the H lines decreases with stellar luminosity for the B stars; further, the last line visible in the series increases with luminosity. Br19 is confidently detected in the B supergiants while Br15 or Br16 is the last of the series for the B dwarfs. The measured line widths (FWHM $\sim$ 60 Å for Br11–Br15) for the two B dwarf stars are roughly two times the instrumental resolution (FWHM $=$ 29 Å). This can be compared to a maximum of 44 Å in the B supergiants for the same lines. These “luminosity” effects are well understood as resulting from the linear Stark effect (Hulbert 1924; Struve 1929; Gray 1992) which is sensitive to the electron pressure.
In Figure \[spect\] it can be noted that the upper Brackett lines are missing in the earliest O-type stars, although Br$\gamma$ is present in these same stars (HCR96). There are, as yet, no published model predictions for the upper Brackett series lines, but we suspect their absence in the earliest O-type stars, while Br$\gamma$ is still present, is analogous to the case for the Balmer lines. LTE model predictions of Auer & Mihalas (1972) clearly underestimate H$\gamma$ absorption compared to observations, while their non–LTE predictions provide a much better fit (Conti 1973). Peterson & Scholz (1971) demonstrate that the difference between the observations and LTE predictions is much less for the higher series member H8 than for H$\gamma$. The upper series members of the Brackett lines probably are still close to LTE (similar to H8) and thus, quite weak due to the advanced ionization state of hydrogen. The lower series members, such as Br$\gamma$, exhibit non–LTE behavior (similar to H$\gamma$) and are thus still present.
### and
A relatively strong line of (4–3) is visible at 1.7002 . To be useful for classification purposes, the wavelength range of interest must contain various lines sensitive to different excitation in the stellar atmosphere. In the present case, the Brackett series lines (as discussed above) and 1.7002 line are sensitive to lower and higher , respectively, in the atmospheres of OB stars. In Figure \[he1\], we plot the equivalent width of the (4–3) 1.7002 line versus the spectral type (Table \[stars\]). This line is at maximum strength in the late O/early B types. In the earliest O stars, the Brackett series will be weak or absent (Figure \[spect\]), whereas in the late B stars, the Brackett lines are stronger. These features, therefore, provide a rough spectral class in the $H-$band for stars in the range $\sim$ O4 to B7 (Figure \[he1\] suggests a spectral type could be determined to $\sim$ $\pm$ 2 sub–types). Note that 1.7002 is detected in HD183143 (B7 Ia) even though the telluric correction appears worse for this star: compare the region just to the red wavelength side of the 1.7 line. The dip in the spectrum is due to a relatively poor air–mass match between object and standard, but the He I 1.7002 line is visible in the uncorrected spectrum.
Figure \[heii\] shows a more detailed look at the $H-$band spectra of the four O stars. The figure suggests a tentative detection of (13-7) at 1.5719 and perhaps (12–7) at 1.6918 . The former is clearly in a region of higher signal–to–noise; the latter sits in a region with larger telluric features (the strongest of which are to the blue of the 1.7002 line). The 1.5719 line appears in both of the O star spectra of Lançon & Rocca–Volmerange (1992) as well, including their spectrum of HD 199579, one of the stars observed in our sample. Detection of is difficult since the 1.5719 line is near the Brackett 15–4 line (1.5701 ); and hence, the line could depend on how well we have corrected the Brackett series lines in the A star which was used as a atmospheric standard (see §2). The lines may prove more useful when observed at higher spectral resolution.
$I'$ and $J-$band Spectra
-------------------------
Figure \[spect2\] shows spectra of the OB stars near 1.1 and 1.3 . The primary lines which fall within our $I'$ and $J-$band spectra are due to H and , as in the $H-$band spectra. However the H lines, Pa$\gamma$ (1.0938 ) and Pa$\beta$ (1.2818 ) are blended with lines of and at our resolution (see Figure \[spect2\]), so they are of limited use for determining spectral types. We note that 1.0830 is present in several of the stars (HD190429, P Cygni profile; HD183143). We anticipate that the behavior of this triplet line (2$p ^3P^{\circ}$–2$s ^3S$) will $not$ be well correlated with spectral type, following the example of the analogous singlet feature at 2.0581 (HCR96). We have only five $I'$ spectra in our initial sample, so we do not attempt to make any firm conclusions about this wavelength region. In any case, the $I'$ and $J-$ bands should be useful for comparison to model atmosphere predictions, especially when they are obtained for “free,” i.e. in cross–dispersed mode.
SUMMARY
=======
We have presented $H-$band spectra of a preliminary set of OB stars and identified absorption features that can be used to classify young, massive stars in obscured H II regions and/or along sight lines with large optical extinction due to interstellar dust. The absorption strength of the (4–3) 1.7002 line, in conjunction with the presence or absence of the $H-$band lines of the Brackett series, is well correlated with optical spectral types and can be used as a coarse spectral classification. The behavior of the infrared lines is analogous to well–studied lines in the optical wavelength region including basic effects due to excitation (temperature) and gravity (pressure). In particular, we see the effects of linear Stark broadening, resulting from changes in pressure in B dwarfs and supergiants, on the Brackett lines which will be useful as a luminosity indicator in the B stars. We have also presented spectra of the OB stars between 1.1 and 1.3 . Features in these wavelength regions may prove useful in cases where excess emission is still strong at $H$ or for the purposes of testing atmosphere models of hot stars.
Support for this work was provided by NASA through grant number HF 01067.01 – 94A from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5–26555. OSIRIS was built with support from NSF grants AST 90–16112 and AST 9218449. This work was also supported by NSF grant AST 93–14808.
Adams, W. S. & Joy, A. H. 1922, , 56, 242
Adams, W. S. & Joy, A. H. 1923, , 57, 294
Ali, B., Carr, J. S., DePoy, D. L., Frogel, J. A., & Sellgren, K. 1995, , 110, 2415
Ali, B. 1996, PhD Dissertation, The Ohio State University, Columbus, OH.
Auer, L. & Mihalas, D. 1972, , 24, 193
Blum, R. D., DePoy, D. L., & Sellgren, K. 1995$a$, , 441, 603
Blum, R. D., Sellgren, K., & DePoy, D. L. 1995$b$, , 440, L17
Blum, R. D., Sellgren, K., & DePoy, D. L. 1996, , 112, 1988
Conti, P. S. 1973, , 179, 161
DePoy, D. L., Atwood, B., Byard, P., Frogel, J. A., & O’Brien, T. 1993, in SPIE 1946, “Infrared Detectors and Instrumentation,” pg 667
Eenens, P. R. J., Williams, P. M., & Wade, R. 1991, , 252, 300
Figer, D. F., Morris, M., & McLean, I. S. 1995, , 447, L29
Figer, D. F., Morris, M., & McLean, I. S. 1996, in The Galactic Center 4th ESO/CTIO Workshop, Ed. R. Gredel, PASP Conf. Ser., Vol. 102, p 263
Figer, D. F., McLean, I. S., & Najarro, F. 1997, , submitted
Garcia, J. D. & Mack, J. E. 1965, J. Opt. Soc. Am., 55, 654
Gray, D. F. 1992, The Observation and Analysis of Stellar Photospheres, Cambridge University Press, Cambridge, p219–225, 291
Greene, T. P. & Meyer, M. R. 1995, , 450, 233
Hanson, M. M. & Conti, P. S. 1994, , 423, L129
Hanson, M. M. & Conti, P. S. 1995, , 448, L45
Hanson, M. M., Conti, P. S., & Rieke, M. J. 1996, , 107 281 (HCR96)
Hoffleit, D., 1982, The Bright Star Catalog, Yale University Press, New Haven
Hulburt, E. O. 1924, , 59, 177
Kleinmann, S.G. & Hall, D.N.B. 1986, , 62, 501
Lançon, A. & Rocca–Volmerange, B. 1992, , 96, 593
Lennon, D. J., Dufton, P. L., & Fitzsimmons, A. 1992, , 94, 569
Livingston, W. & Wallace, L. 1991, N.S.O. Technical Report $\#$91–001, 1991 July
Mathis, J.S. 1990, , 28, 37
Moneti, A., Glass, I. S., & Moorwood, A. F. M. 1994, , 268, 194
Morris, P. M, Hanson, M. M., Conti, P. S., Eenens, P. R. J., & Blum, R. D. 1996, , 470, 597
Oliva, E. & Origlia, L. 1992, , 254, 466
Peterson, D. M. & Scholz, M. 1971, , 163, 51
Pogge, R. 1997, private communication
Ramírez, S. V., DePoy, D. L., Frogel, J. A., Sellgren, K., & Blum, R. D. 1997, , in press
Struve, O. 1929, , 69, 173
Tamblyn, P., Rieke, G. H., Hanson, M. M., Close, L. M., McCarthy, D. W., JR., & Rieke, M. J. 1996, , 456, 206
Wiese, W. L., Smith, M. W., Glennon, B. M. 1966, Atomic Transition Probabilities, United States Dept. of Commerce, report no. NSRDS–NBS 4
[^1]: IRAF is distributed by the National Optical Astronomy Observatories.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'While end-to-end neural conversation models have led to promising advances in reducing hand-crafted features and errors induced by the traditional complex system architecture, they typically require an enormous amount of data due to the lack of modularity. Previous studies adopted a hybrid approach with knowledge-based components either to abstract out domain-specific information or to augment data to cover more diverse patterns. On the contrary, we propose to directly address the problem using recent developments in the space of continual learning for neural models. Specifically, we adopt a domain-independent neural conversational model and introduce a novel neural continual learning algorithm that allows a conversational agent to accumulate skills across different tasks in a data-efficient way. To the best of our knowledge, this is the first work that applies continual learning to conversation systems. We verified the efficacy of our method through a conversational skill transfer from either synthetic dialogs or human-human dialogs to human-computer conversations in a customer support domain.'
author:
- |
Sungjin Lee\
Microsoft Research / Redmond, WA, USA\
[[email protected]]{}
bibliography:
- 'acl2017.bib'
title: Toward Continual Learning for Conversational Agents
---
Introduction {#sec:intro}
============
Conversational bots become increasingly popular in a wide range of business areas. In order to support the rapid development of bots, a number of bot building platforms have been launched, for example, from Microsoft, Amazon and so on. Despite this progress, the development of a business-critical bot still requires a serious effort from the design to actual implementation of several components such as language understanding, state tracking, action selection, and language generation. Not only does this complexity prevent casual developers from building quality bots but also introduces an unavoidable degradation in performance due to some non-trivial problems including unclear state representation design, insufficient labeled data and error propagation down the pipeline.
Recently, end-to-end (E2E) approaches using Deep Neural Networks (DNNs) have shown the potential to solve such problems – the DNNs induce a latent representation in the course of the joint optimization of all components without requiring any labeling on internal state. Despite such appealing aspects, the neural E2E approaches also have major challenges to overcome. The state-of-the-art systems require numerous dialogs only to learn simple behaviors. In general, it is expensive to collect a sufficient amount of dialogs from a target task. A possibility to the data-intensiveness problem would be to repurpose already built models through fine-tuning with a few additional dialogs from the target task. For example, one can add a payment handling capability to a new bot by repurposing any model that is already trained on payment-related conversations. It has been shown, however, that neural models tend to forget what it previously learned when it continuously trains on a new task, which is what is called [*Catastrophic Forgetting*]{} [@french1999catastrophic]. Another possible approach would be to compose only relevant parts from each of the pretrained models for the target task. Unfortunately, it is also unclear how to compose neural models due to the lack of modularity.
There have been prior studies that partly address the data-intensiveness problem outside the neural model [@wen2016network; @williams2017hybrid; @zhao2017generative; @eshghi2017bootstrapping; @wenmultinlg16; @zhao2017generative]. In this work, we instead propose to directly address the problem with recent developments in the space of continual learning for neural models. Specifically, we adopt a domain-independent neural conversational model and introduce a novel [*Adaptive Elastic Weight Consolidation*]{} (AEWC) algorithm to continuously learn a new task without forgetting valuable skills that are already learned. To test our method, we first continuously train a model on synthetic data that only consists of general conversation patterns like opening/closing turns and then on a corpus of human-computer (H-C) dialogs in a customer support domain that barely has general conversations. Then, we show that the resulting model is able to handle both general and task-specific conversations without forgetting general conversational skills. Second, we continuously train a model on a large amount of human-human (H-H) dialogs and then on a small number of H-C dialogs in the same customer support domain. As H-H conversations typically cover various out-of-domain topics including general conversations, this allows us to show that the resulting model can carry out the target task while handling general conversations that do not occur in the H-C dialogs.
The rest of this paper is organized as follows. In Section \[sec:related\] we present a brief summary of related work. In Section \[sec:method\] we describe our approach for continual learning for conversational agents. In Section \[sec:experiments\] we discuss our experiments. We finish with conclusions and future work in Section \[sec:conclusion\].
Related Work {#sec:related}
============
#### Conversational Systems
Traditionally, a dialog system has a pipeline architecture, typically consisting of language understanding, dialog state tracking, dialog control policy, and language generation [@jokinen2009spoken; @young2013pomdp]. With this architecture, developing dialog systems requires designing the input and output representation of multiple components. It also, oftentimes, involves writing a large number of handcrafted rules and laboriously labeling dialog datasets to train statistical components. In order to avoid this costly manual endeavor, a line of research has emerged to introduce end-to-end trainable neural models [@sordoni2015neural; @vinyals2015neural; @serban2016building; @wen2016network; @bordes2016learning]. But the E2E neural approach is data-intensive requiring over thousands of dialogs to learn even simple behaviors. Broadly there are two lines of work addressing the data-intensiveness problem. The first makes use of domain-specific information and linguistic knowledge to abstract out the data [@wen2016network; @williams2017hybrid; @zhao2017generative; @eshghi2017bootstrapping]. The second line of work adopts data recombination approaches to generate counterfeit data that mimics target domain dialogs [@wenmultinlg16; @zhao2017generative]. The prior approaches, however, partly bring us back to the downside of traditional approaches: The difficulty in maintenance increases as the number of rules grows; The system quality depends on external expertise; It is hard to scale out over different domains. Furthermore, the increased size of training data, as a result of data recombination approaches, would lead to a significant increase in training time. Unlike prior work, we approach the problem from the standpoint of continual learning where a single neural network accumulates task-relevant knowledge over time and exploits this knowledge to rapidly learn a new task from a small number of examples.
#### Continual Learning for Neural Networks
Previous studies that address the catastrophic forgetting problem are broadly partitioned into three groups. First, architectural approaches reduce interference between tasks by altering the architecture of the network. The simplest form is to copy the entire network for the previous task and add new features for a new task. [@rusu2016progressive; @lee2016dual]. Though this prevents forgetting on earlier tasks, the architectural complexity grows with the number of tasks. proposes a single network where a subset of different modules gets picked for each task based on an evolutionary idea to alleviate the complexity issue. Second, functional approaches encourage similar predictions for the previous and new tasks. applies the old network to the training data of the new task and uses the output as pseudo-labels. performs a similar regularization on the distance between the final hidden activations. But the additional computation using the old network makes functional approaches computationally expensive. Lastly, implicitly distributed knowledge approaches use neural networks of a large capacity to distribute knowledge for each task using dropout, maxout, or local winner-take-all [@goodfellow2013empirical; @srivastava2013compete]. But these earlier approaches had limited success and failed to preserve performance on the old task when an extreme change to the environment occurred. Recently, proposed elastic weight consolidation (EWC) which makes use of a point estimate for the Fisher information metric as a weighting factor for a distance penalty between the parameters of the new and old tasks. To alleviate the cost of exactly computing the diagonal of the Fisher metric, presented an online method that accumulates the importance of individual weights over the entire parameter trajectory during training. This method, however, could yield an inaccurate importance measure when the loss function is not convex. Thus, we propose an adaptive version of the online method that applies exponential decay to cumulative quantities.
Continual Learning for Conversational Agents {#sec:method}
============================================
In this section, we describe a task-independent conversation model and an adaptive online algorithm for continual learning which together allow us to sequentially train a conversation model over multiple tasks without forgetting earlier tasks.
Task-Independent Conversation Model {#conv_model}
-----------------------------------
As we need to use the same model structure across different tasks, including open-domain dialogs as well as task-oriented dialogs, we adopt a task-independent model. Thus, our model should be able to induce a meaningful representation from raw observations without access to hand-crafted task-specific features. [^1] In order to achieve this goal, we employ a neural encoder for state tracking and a neural ranking model for action selection.
#### State Tracking {#state_tracking}
Due to the sequential nature of the conversation, variants of Recurrent Neural Networks (RNNs) [@medsker1999recurrent] have been widely adopted for modeling conversation. State tracking models with RNN architecture, however, usually tend to become less effective as the length of a sequence increases due to the gradual information loss along the recurring RNN units. To address this problem, we use a hierarchical recurrent encoder architecture which has been recently adopted for generative conversation models [@serban2016building]. Specifically, our model mimics the natural structure in language — a conversation consists of a sequence of utterances, an utterance is a sequence of words, a word, in turn, is composed of characters. In order to capture character-level patterns in word embeddings, we concatenate the word embeddings with the output of the bidirectional character-level RNN encoder. We use Long Short-Term Memory (LSTM) [@hochreiter1997long] as the RNN unit that takes an input vector $x$ and a state vector $h$ to output a new state vector $h' = \phi(x, h)$. Let $\mathcal{C}$ denote the set of characters and $\mathcal{W}$ the set of words. $w \in \mathcal{W}$ is a sequence of characters $(c_1 \ldots c_m) \in \mathcal{C}^m$. We compute the embedding of a word, $v$, as follows: $$\begin{aligned}
f^{\mathcal{C}}_i &= \phi^{\mathcal{C}}_f{\left(e_{c_i}, f^{\mathcal{C}}_{i-1}\right)} &&\forall i = 1 \ldots m \\
b^{\mathcal{C}}_i &= \phi^{\mathcal{C}}_b{\left(e_{c_i}, b^{\mathcal{C}}_{i+1}\right)} &&\forall i = m \ldots 1\\
v &= f^{\mathcal{C}}_m \oplus b^{\mathcal{C}}_1 \oplus e_w &&\end{aligned}$$ where $\oplus$ denote the vector concatenation operation. $e_c \in {\ensuremath{\mathbb{R}}}^{d_c}$ and $e_w \in {\ensuremath{\mathbb{R}}}^{d_w}$ are character embeddings and word embeddings, respectively.
Next, we have another bidirectional LSTM-RNN layer that takes the word embeddings $(v_1 \ldots v_n)$ of an utterance $(w_1 \ldots w_n) \in \mathcal{W}^n$ to generate an utterance embedding $u \in {\ensuremath{\mathbb{R}}}^{d_u}$: $$\begin{aligned}
f^{\mathcal{W}}_i &= \phi^{\mathcal{W}}_f{\left(v_i, f^{\mathcal{W}}_{i-1}\right)} &&\forall i = 1 \ldots n \\
b^{\mathcal{W}}_i &= \phi^{\mathcal{W}}_b{\left(v_i, b^{\mathcal{W}}_{i+1}\right)} &&\forall i = n \ldots 1 \\
u &= f^{\mathcal{W}}_n \oplus b^{\mathcal{W}}_1 \label{utt_emb}\end{aligned}$$
After that, we have a unidirectional LSTM-RNN layer that takes a sequence of pairs of user utterance embedding, $(u_1 \ldots u_t)$, and previous system action embeddings [^2], $(a_1 \ldots a_t)$, as input to induce state embeddings $(s_1 \ldots s_t)$: $$\begin{aligned}
\label{state_emb}
s_i &= \phi_s{\left(u_i \oplus a_i, s_{i-1}\right)} &&\forall i = 1 \ldots t\end{aligned}$$
#### Action Ranking
In a continual learning setting, there is no predefined action set from which a classifier selects an action across different tasks. Thus, we cast the action selection problem as a ranking problem where we consider the affinity strength between a state embedding $s$ (Eq. \[state\_emb\]) and a set of candidate system action embeddings $\{a_i\}$ (Eq. \[utt\_emb\]): [^3] $$\rho(a|s) = s^T M a$$ where $M \in {\ensuremath{\mathbb{R}}}^{d_s} \times {\ensuremath{\mathbb{R}}}^{d_u}$ projects the state embedding onto the action space. In order to optimize the projection matrix, $M$, we adopt the [*Plackett-Luce*]{} model [@plackett1975analysis]. The Plackett-Luce model normalizes ranking score by transforming real-valued scores into a probability distribution: $$p(a|s) = \frac{\exp(s^T M a)}{\sum_i \exp(s^T M a_i)}
$$ With the transformed probability distribution, we minimize the cross-entropy loss against the true label distribution. Thanks to the normalization, our ranking model performs a more effective penalization to negative candidates.
Adaptive Elastic Weight Consolidation
-------------------------------------
In order to achieve continual learning, we need to minimize the total loss function summed over all tasks, $\mathcal{L} = \Sigma_\mu L_\mu$, without access to the true loss functions of prior tasks. A catastrophic forgetting arises when minimizing $L_\mu$ leads to an undesirable increase of the loss on prior tasks $L_\nu$ with $\nu < \mu$. Variants of the EWC algorithm tackle this problem by optimizing a modified loss function: $$\label{surrogate_loss}
\tilde{L}_\mu = L_\mu + \underbrace{c\sum_k\Omega_k^\mu(\bar{\theta}_k-\theta_k)^2}_{\text{surrogate loss}}$$ where $c$ represents an weighting factor between prior and current tasks, $\theta$ all model parameters introduced in Section \[conv\_model\], $\bar{\theta}_k$ the parameters at the end of the previous task and $\Omega_k^\mu$ regularization strength per parameter $k$. The bigger $\Omega_k^\mu$, the more influential is the parameter. EWC defines $\Omega^\mu$, for example, to be a point estimate which is equal to the diagonal entries of the [*Fisher information matrix*]{} at the final parameter values. Since EWC relies on a point estimate, we empirically noticed that sometimes $\Omega^\mu$ fails to capture the parameter importance when the loss surface is relatively flat around the final parameter values as $\Omega^\mu$ essentially decreases to zero.
In contrast to EWC, @zenke2017continual ([-@zenke2017continual]) computes an importance measure online by taking the path integral of the change in loss along the entire trajectory through parameter space. Specifically, the per-parameter contribution $\omega_k^\mu$ to changes in the total loss is defined as follows: $$\omega_k^\mu = -\int_{t^{\mu-1}}^{t^\mu}g_k(\theta(t))\theta^\prime_k(t)dt$$ where $\theta(t)$ is the parameter trajectory as a function of time $t$, $g(\theta) = \frac{\partial L}{\partial \theta}$ and $\theta^\prime_k(t) = \frac{\partial \theta}{\partial t}$. Note that the minus sign indicates that we are interested in decreasing the loss. In practice, we can approximate $\omega_k^\mu$ as the sum of the product of the gradient $g_k(t)$ with the parameter update $\Delta_k(t)$. Having defined $\omega_k^\mu$, $\Omega_k^\mu$ is defined such that the regularization term carries the same units as the loss by dividing $\omega_k^\mu$ by the total displacement in parameter space: $$\Omega_k^\mu = \sum_{\nu < \mu}\frac{\omega_k^\nu}{(\Delta_k^\nu)^2 + \zeta}$$ where $\Delta_k^\nu$ quantifies how far the parameter moved during the training process for task $\nu$. $\zeta$ is introduced to keep the expression from exploding in cases where the denominator gets close to zero. Note that, with this definition, the quadratic surrogate loss in (\[surrogate\_loss\]) yields the same change in loss over the parameter displacement $\Delta_k$ as the loss function of the previous tasks.
The path integral over the entire trajectory of parameters during training, however, can yield an inaccurate importance measure when the loss function is not convex. As the loss function of neural networks is generally not convex, we propose to apply exponential decay to $\omega_k$ and $\Delta_k$: $$\label{path_int}
\omega_k = \lambda \cdot \omega_k - g_k(\theta(t))\Delta_k(t)$$ $$\Delta_k = \lambda \cdot \Delta_k + \Delta_k(t)$$ where $0 < \lambda < 1$ is a decay factor.
Experiments {#sec:experiments}
===========
#### Data
\[1\][>p[\#1]{}]{}
Train Data \# Dialogs Avg. Dialog Len Avg. User Len Avg. System Len
---------------------- ------------ ----------------- --------------- -----------------
`open_close` 10 2 1.75 5.25
`HH_reset_password` 746 12.84 9.35 21.06
`HC_reset_password` 520 1.93 8.14 12.57
Test Data \# Dialogs Avg. Dialog Len Avg. User Len Avg. System Len
`HC_reset_password` 184 1.99 8.33 12.22
`HC_reset_password+` 184 3.99 5.23 8.86
In order to test our method, we used four dialog datasets — `open_close`, `HH_reset_password`, `HC_reset_password` and `HC_reset_password+`. [^4] Basic statistics of the datasets are shown in Table \[tab:data\]. Example dialogs are provided in the Appendix.
#### `open_close`:
A synthetic corpus that we created in order to clearly demonstrate the phenomena of catastrophic forgetting and the impact of our method. In `open_close`, every conversation has only opening/closing utterances without any task-related exchanges.
#### `HH_reset_password`:
A real H-H conversation dataset that we obtained from our company’s text-based customer support chat system. All conversation logs are anonymized and filtered by a particular problem which is “reset password" for this study. Data from this system is interesting in that H-H conversations typically cover not only task-related topics but also various out-of-domain topics including general conversations. If we can transfer such conversational skills to an H-C conversation model, the H-C model will be able to handle many situations even without seeing relevant examples in its training data. But using H-H dialogs to bootstrap a task-oriented dialog system has been shown to be difficult even with serious annotation effort [@bangalore2008learning].
#### `HC_reset_password`:
A real corpus of H-C conversations that we obtained from our company’s text-based customer support dialog system. This data is distinctive since it is used by real users and the system was developed by customer support professionals at our company with sophisticated rules. As the data was obtained from a dialog system, however, it includes mistakes that system made. Thus, we modified the dialog data to correct system’s mistakes by replacing it with the most appropriate system action given the context and then discarding the rest of the dialog, since we do not know how the user would have replied to this new system action. The resulting dataset is a mixture of complete and partial conversations, containing only correct system actions.
#### `HC_reset_password+`:
The `HC_reset_password` dataset barely has general conversations since the dialog system for a particular problem only starts after customers are routed to the system from somewhere else according to the brief problem description. But for a standalone system, it is natural to have opening and closing utterances. Thus, we extended the `HC_reset_password` dataset with opening and closing utterances which were randomly sampled from the `open_close` dataset. This allows us to test if our method can keep the conversation skills that it learned from either `open_close` or `HH_reset_password` after we further train the model on `HC_reset_password`.
#### Training Details
To implement the state tracker, we use three LSTM-RNNs with – 25 hidden units for each direction of the RNNs for word encoding, 128 hidden units for each direction of the RNNs for utterance encoding and 256 hidden units for the RNN for state encoding. We shared the RNNs for word and utterance encoding for both user utterances and (candidate) system actions. We initialized all LSTM-RNNs using orthogonal weights [@saxe2013exact]. We initialized the character embeddings with 8-dimensional vectors randomly drawn from the uniform distribution $U(−0.01, 0.01)$ while the word embedding weight matrix is initialized with the GloVe embeddings with 100 dimension [@pennington2014glove]. We use the Adam optimizer [@kingma2014adam], with gradients computed on mini-batches of size 1 and clipped with norm value 5. The learning rate was set to $1 \times 10^{-3}$ throughout the training and all the other hyper parameters were left as suggested in . To create train sets for the Plackett-Luce model, we performed negative sampling for each dataset to generate 9 distractors for each truth action.
#### Continual Learning Details
As a simple transfer mechanism, we initialized all weight parameters with prior weight parameters when there is a prior model. Given that our focus is few-shot learning, we don’t assume the existence of development data with which we can decide when to stop training. Instead, training terminates after 100 epochs which is long enough to reconstruct all training examples. The $\omega_k$ and $\Delta_k$ are updated continuously during training, whereas the importance measure $\Omega_k$ and the prior weight $\tilde{\theta}$ are only updated at the end of each task. We set the trade-off parameter $c$ to $0.01$. If the path integral (Eq. \[path\_int\]) is exact, $c = 1$ would mean an equal weighting among tasks. However, the evaluation of the integral typically involves various noises, leading to an overestimate of $\omega_k$. To compensate the overestimation, $c$ generally has to be chosen smaller than one.
#### Results on Synthetic Dialog to H-C Dialog
[|c||C[1.2cm]{}|C[1.2cm]{}|C[1.2cm]{}||C[1.2cm]{}|C[1.2cm]{}|C[1.2cm]{}|]{} Test Data & &\
& NT & WT & AEWC & NT & WT & AEWC\
1 & 47.32 & 49.85 & 50.73 & 18.02 & 24.01 & **[27.17]{}\
2 & 54.45 & 55.52 & 55.38 & 22.28 & 22.33 & **[35.79]{}\
3 & 58.91 & 60.41 & 58.86 & 25.70 & 23.48 & **[47.65]{}\
4 & 61.88 & 61.38 & 60.13 & 27.03 & 23.99 & **[49.71]{}\
5 & 63.03 & 62.51 & 60.32 & 27.93 & 22.84 & **[51.14]{}\
**********
In order to clearly demonstrate the catastrophic forgetting problem, we compare three models trained by different training schemes: 1) [*No Transfer*]{} (NT) – we train a model on `HC_reset_password` from scratch, 2) [*Weight Transfer*]{} (WT) – we train a model on `open_close`, and continued to train the model on `HC_reset_password`, 3) AEWC – the same as 2) except for AEWC being applied. We compared accuracy on both the evaluation data from `HC_reset_password` and `HC_reset_password+`. The result is shown in Table \[tab:res\_syn\]. All the models perform well on `HC_reset_password` due to the similarity between the training and evaluation data. But the performances of NT and WT on `HC_reset_password+` significantly drop down. This surprisingly poor result confirms which found that neural conversation models can be badly affected by systemic noise. In this case, we systemically introduced unseen turns into dialogs. On the contrary, AWEC shows a higher performance than the others by trying to find optimal parameters not only for the previous task but also for the new task. One of key observations is that the performance of WT on `HC_reset_password+` starts strong but keeps decreasing as more training examples are given. This indicates that weight transfer alone cannot help carry prior knowledge to a new task, rather it might lead to poor local optima if the prior knowlege is not general enough.
#### Results on H-H Dialog to H-C Dialog
[|c||C[1.2cm]{}|C[1.2cm]{}|C[1.2cm]{}||C[1.2cm]{}|C[1.2cm]{}|C[1.2cm]{}|]{} Test Data & &\
& NT & WT & AEWC & NT & WT & AEWC\
1 & 45.67 & 47.87 & 48.07 & 16.70 & 44.63 & **[50.10]{}\
2 & 54.09 & 56.15 & 56.05 & 22.79 & 51.93 & **[63.31]{}\
3 & 58.87 & 60.88 & 60.60 & 25.62 & 54.89 & **[68.88]{}\
4 & 62.86 & 64.08 & 63.18 & 27.20 & 57.68 & **[71.53]{}\
5 & 63.84 & 65.08 & 64.07 & 27.39 & 57.81 & **[72.99]{}\
**********
Now we turn to a more challenging problem — we bootstrap a conversation model by first learning from H-H dialogs without any manual labeling. We again compare three models trained by different training schemes: 1) [*No Transfer*]{} – we train a model on `HC_reset_password` from scratch, 2) [*Weight Transfer*]{} – we train a model on `HH_reset_password`, and continued to train the model on `HC_reset_password` 3) AEWC – the same as 2) except for AEWC being applied. The characteristics of H-H dialogs are vastly different from H-C dialogs. For example, the average utterance length and conversation length are way longer than H-C dialogs, introducing complex long dependencies. Also, H-H dialogs cover much broader topics that do not exactly match the conversation goal introducing many new vocabularies that do not occur in the corresponding H-C dialogs. In order to make the knowledge transfer process more robust to such non-trivial differences, the dropout regularization [@srivastava2014dropout] was applied to both utterance and state embeddings with the ratio of 0.4. We also limited the maximum utterance length to 20 words. The result is shown in Table \[tab:res\_hh\_hc\], which generally agrees on the previous experimental results obtained with synthetic data. All three models work well on `HC_reset_password`. NT performs poorly on `HC_reset_password+` as expected whereas WT interestingly shows a much higher performance than NT. This can be attributed to the fact that `HH_reset_password` contains a broader array of conversation skills including some knowledge on the target task compared to `open_close`, which leads to a more general and transferable model that would not necessarily be forgotten when faced with a new task. But the increasing performance gap between WT and AEWC on `HC_reset_password+` clearly shows that WT gradually forgets the prior knowledge as it trains on more examples from a new task. Another noteworthy observation is that, when just one training example is given, the performances of WT and AEWC are much higher than that of NT. This result demonstrates that our method can capture meaningful knowledge from prior tasks in such a way that it partly works on a related new task.
Conclusion {#sec:conclusion}
==========
We have presented a novel approach to the data-intensiveness problem due to the lack of compositionality of end-to-end neural conversation models. We tackled the problem with a continual learning technique for neural networks. Specifically, we proposed a universal conversational model that can be used across different tasks and a novel [*Adaptive Elastic Weight Consolidation*]{} method which together allows us to continuously train on a sequence of tasks without forgetting earlier tasks. We tested our method with two experimental configurations – conversation skill transfer from synthetic dialogs/H-H dialogs to H-C dialogs in a customer support domain. Future work includes in-depth analysis on how our algorithm distributes different pieces of knowledge across the network. Also, it would be interesting to apply the continual learning method for composing multiple tasks to handle a more complex task.
[**Appendices**]{}
`open_close` example dialog
===========================
Hello.\
Hello. What can I help you?\
Okay. Thank you.\
Sure thing! Have a great day.
`HC_reset_password` example dialog
==================================
I forgot my password and now I have locked my account for 30 days\
Okay, you don’t need to remember your password, we can reset it. Would you like to try that?\
Yes please\
SOLUTION: To reset your password, go to xx\_url\_xx. Was that helpful?\
Okay that’s also another problem I have. I added extra security & the number I have on fine is no longer active. I chnsged my number and because I can’t access it I can’t long into my account & it won’t let me reset my password\
Let’s connect you to a person who can help you.
`HC_reset_password+` example dialog
===================================
Good morning.\
Hi. What can I help you?\
forgot password\
Okay, you don’t need to remember your password, we can reset it. Would you like to try that?\
already tried that\
Let’s connect you to a person who can help you.\
Okay thank you.\
Sure thing! Have a great day.
`HH_reset_password` example dialog
==================================
hi, thanks for visiting answer desk! i’m xx\_firstname\_xx q.\
hello i am having trouble accessing my laptop\
hi, how may i help you today?\
i forgot the password i changed the pw on my account, but the computer is still not able to be accessed\
oh that’s bad , that might be very important to you but no worries i will help you out with your issue. to start with may i have your complete name please?\
my name is xx\_firstname\_xx xx\_firstname\_xx, i am contacting you on behalf of xx\_firstname\_xx mine\
that’s okay xx\_firstname\_xx. may i also know your email address and phone number please?\
my email, xx\_email\_xx, xx\_phonenumber\_xx\
thank you for the information. xx\_firstname\_xx what is your current operating system?\
windows 10\
may i know what is the error message you received upon trying to unlock your computer?\
the password is not working i forgot the pw, i tried to reset the pw from the account\
is it a local account or microsoft account?\
i am not sure i thought it was a microsoft account but the pw didnt change\
can you send me the email so that i can check if it is microsoft account?\
xx\_email\_xx though i think i may have created one by accident setting up the computer it might be xx\_email\_xx or some variation of that i chnaged the pw on the xx\_email\_xx but it had no effect\
since we’re unable to know what exactly happening to your computer. i will provide you our technical phone support so that you will be well instructed on what you are going to do to get your computer work again. would that be okay with you?\
fine\
one moment please.
[^1]: Inspired by @williams2017hybrid ([-@williams2017hybrid]), one can still make use of action masks without hurting the task independence.
[^2]: We treat system actions the same as user utterances and encode both with a common LSTM-RNN encoder
[^3]: We employ the same utterance-level encoder as in the state tracking to yield the embeddings of the candidate system actions.
[^4]: The datasets are not publicly available, but we are unaware of suitable H-H and H-C paired data in the public domain.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, stochastic inertial manifold for damped wave equations subjected to additive white noise is constructed by the Lyapunov-Perron method. It is proved that when the intensity of noise tends to zero the stochastic inertial manifold converges to its deterministic counterpart almost surely.\
[*Keywords:*]{} Stochastic inertial manifold; Wave equation; Random dynamical system
author:
- |
Zhenxin Liu\
[College of Mathematics, Jilin University, Changchun 130012, People’s Republic of China]{}\
[[email protected]]{}
title: '**Stochastic inertial manifolds for damped wave equations[^1]**'
---
Introduction
============
The inertial manifold (IM) introduced by Constantin, Foias, Nicolaenko, Sell, and Temam [@Fo; @Foi; @FNST0; @FNST; @Con0] is a finite dimensional Lipschitz invariant manifold attracting solutions exponentially, which goes back to the works of Mañé, Henry and Mora [@Man; @Hen; @Mor]. Global attractor is an invariant compact set attracting solutions which often has a finite (fractal) dimension and, therefore, it is an important object for the study of long time behavior of evolution equations. At the present level of understanding of dynamical systems, global attractors are expected to be very complicated objects (fractals) and their practical utilization, for instance for numerical simulations, may be difficult. The IMs, when they exist, are more convenient objects which are able to describe the large-time behavior of dynamical systems. One of the important properties of inertial manifolds is that they contain global attractors, so the study of dynamics of infinite dimensional nonlinear systems can be reduced to the study of dynamics of flows on the inertial manifold, which, in turn, is described by the dynamics of an ordinary differential equation. There are extensive works on IMs. See, for example, Chow and Lu [@Cho], Chow et al [@Cho1], Constantin et al [@Con0], Constantin et al [@Con], Foias et al [@FNST0; @FNST], Foias et al [@Fo; @Foi], Foias et al [@Foi1], Mallet-Paret and Sell [@Mal], Sell and You [@Sel], Temam [@Tem], among others.
Stochastic partial differential equations (SPDE) have been drawing more and more attention for their importance in describing many natural phenomenon under random influences. With the rapid development of random dynamical systems (RDS) [@Ar1], many SPDEs are studied in the framework of RDS. On many occasions, the development of SPDE and RDS mimics the deterministic case and many efforts are devoted to establish the results for SPDE and RDS corresponding to that for the deterministic case. This is true for IM: there have been some works on Stochastic IMs, see, for example, Bensoussan and Flandoli [@Ben], Chueshov and Girya [@Chu], Chueshov and Scheutzow [@Chu1], Da Prato and Debussche [@Dap], Duan et al [@Dua; @Dua1]. These works mainly deal with stochastic parabolic equations. In present paper, we aim to obtain the existence of stochastic IM for damped wave equations subjected to additive white noise. Moreover, we will show that the stochastic IM converges to its deterministic counterpart [*almost surely*]{} when the intensity of noise tends to zero. As in the deterministic case, the usual methods to obtain the existence of stochastic IMs are Hadamard’s graph transform method [@Had] and Lyapunov-Perron’s method [@Lia; @Per]. In this paper, we adopt the latter one. In Section 2 we introduce some preliminaries and prove the existence theorem of stochastic IM for abstract evolution equations with random coefficients; in Section 3 we apply the result of Section 2 to damped wave equations subjected to additive white noise and study the property of its IM.
Existence of Stochastic IM for abstract equations
=================================================
Let $X$ be a metric space with a metric $d_X$. A [*random dynamical system (RDS)*]{}, shortly denoted by $\varphi$, consists of two ingredients:\
(i) A model of the noise, namely a metric dynamical system $(\Omega,
\mathscr F, \mathbb P, (\theta_t)_{t\in \mathbb R})$, where $(\Omega, \mathscr F, \mathbb P)$ is a probability space and $(t,
\omega)\mapsto \theta_t\omega$ is a measurable flow which leaves $\mathbb P$ invariant, i.e. $\theta_t\mathbb P=\mathbb P$ for all $t\in \mathbb R$.\
(ii) A model of the system perturbed by noise, namely a cocycle $\varphi$ over $\theta$, i.e. a measurable mapping $\varphi: \mathbb
R^+\times \Omega\times X \rightarrow X,
(t,\omega,x)\mapsto\varphi(t,\omega,x)$, such that: $$\label{phi}
\varphi(0,\omega,\cdot)={\rm id}_X,
\varphi(t+s,\omega,\cdot)=\varphi(t,\theta_s
\omega,\varphi(s,\omega,\cdot))\quad {\rm for ~all}\quad
t,s\in\mathbb R^+,\omega\in\Omega.$$
Although it is well known that a large class of partial differential equations with stationary random coefficients and Itô stochastic ordinary differential equations generate RDS (for details see Chapter 1 of [@Ar1]), this problem is still unsolved for SPDE with general noise terms, see [@Dua] for the reason. Indeed, the existence of RDS generated by SPDE has been proved in relatively narrow generality. In fact, only cases in which the SPDE can be reduced to a deterministic one with random coefficients can be treated in the framework of RDS. See, for example, [@Car; @Cra; @Cra1; @Dua1].
For later use, assume $z$ is an Ornstein-Uhlenbeck process which satisfies the following equation $$\label{z}
{\rm d}z+\lambda z{\rm d}t=\delta{\rm d}W$$ for some $\lambda>0$ and $\delta>0$. The process $z$ has the following properties, see [@Car; @Dua] for the proof.
\[zpr\] (i) There exists a $\{\theta_t\}_{t\in\mathbb R}$-invariant set $\Omega\in\mathcal B(C_0(\mathbb R,\mathbb R))$ of full measure with sublinear growth: $$\lim_{t\rightarrow\pm\infty}\frac{|\omega(t)|}{|t|}=0,~\omega\in\Omega.$$ (ii) For $\omega\in\Omega$ the random variable $$z(\omega)=-\lambda\delta\int_{-\infty}^0e^{\lambda\tau}\omega(\tau){\rm
d}\tau$$ exists and generates a unique stationary solution of (\[z\]) given by $$(t,\omega)\rightarrow
z(\theta_t\omega)=-\lambda\delta\int_{-\infty}^0e^{\lambda\tau}\theta_t\omega(\tau){\rm
d}\tau=-\lambda\delta\int_{-\infty}^0e^{\lambda\tau}\omega(\tau+t){\rm
d}\tau+\delta\omega(t).$$ The map $t\rightarrow z(\theta_t\omega)$ is continuous.\
(iii) In particular, we have $$\lim_{t\rightarrow\pm\infty}\frac{|z(\theta_t\omega)|}{|t|}=0~{\rm
for~}\omega\in\Omega.$$ (iv) In addition, $$\lim_{t\rightarrow\pm\infty}\frac1t\int_0^tz(\theta_\tau\omega){\rm
d}\tau=0~{\rm
for~}\omega\in\Omega,~\lim_{t\rightarrow\pm\infty}\frac1t\int_0^t|z(\theta_\tau\omega)|{\rm
d}\tau=\mathbb E|z|<\infty.$$
Let $H$ be a separable Hilbert space with norm $|\cdot|$ and inner product $\langle\cdot,\cdot\rangle$. Consider the Stratonovich SPDE on $H$ $$\label{u}
\frac{{\rm d}u}{{\rm d}t}=Au+F(u)+u\dot W,$$ where $u\in H$, $W(t)$ is the standard real-valued two-sided Wiener process and the generalized time-derivative $\dot W$ formally describes a white noise. Here we assume that $F$ is globally Lipschitz continuous on $H$ with Lipschitz constant ${\rm Lip}F$. For the existence and uniqueness theory of (\[u\]) we can first write it into its equivalent Itô equation and then refer to [@Dap1] for details. Under the transformation $T(\omega,u)=ue^{-z(\omega)}$, (\[u\]) is conjugated to the following equation with random coefficients $$\label{main}
\frac{{\rm d}u}{{\rm
d}t}=Au+z(\theta_t\omega)u+G(\theta_t\omega,u),~~u(0)=x\in H,$$ where $z$ satisfies $${\rm d}z+z{\rm d}t={\rm d}W,$$ and $G(\omega,u)=e^{-z(\omega)}F(ue^{z(\omega)})$. It is clear that ${\rm Lip}_uG={\rm Lip}F$.
Assume $A: D(A)\rightarrow H$ is a linear operator which generates a strongly continuous semigroup $e^{At}$ on $H$, which satisfies the pseudo exponent dichotomy condition with exponents $0>\alpha>\beta$ and bound $K>0$, i.e. there exists a continuous projection $P$ on $H$ such that\
(i) $Pe^{At}=e^{At}P$;\
(ii) the restriction $e^{At}|_{R(P)}, t\ge 0$, is an isomorphism of the range $R(P)$ of $P$ onto itself, and we denote $e^{At}$ for $t<0$ the inverse map;\
(iii) $$\label{expo}
\left.
\begin{array}{ll}
|e^{At}Px|\le Ke^{\alpha t}|x|, & t\le 0, \\
|e^{At}Qx|\le Ke^{\beta t}|x|, & t\ge 0,
\end{array}
\right.$$ where $Q=I-P$.
\[im\] A random set is called [*invariant*]{} for RDS $\varphi$ if $$\varphi(t,\omega,M(\omega))\subset M(\theta_t\omega), ~{\rm
for~any~} t\ge 0.$$ If an invariant set $M(\omega)$ can be represented by a Lipschitz or $C^k$ mapping $$h(\cdot,\omega):PH\rightarrow QH$$ such that $$M(\omega)=\{\xi+h(\xi,\omega)|\xi\in PH\},$$ then we call $M(\omega)$ a [*Lipschitz or $C^k$ invariant manifold*]{}. Furthermore, if $PH$ is finite dimensional and $M(\omega)$ attracts exponentially all the orbits of $\varphi$, then we call $M(\omega)$ a [*stochastic inertial manifold of $\varphi$*]{}.
\[the\] ([@Dua1]) If $$\label{gap}
K{\rm
Lip}F\left(\frac{1}{\alpha-\eta}+\frac{1}{\eta-\beta}\right)<1,$$ then there exists a Lipschitz invariant manifold for the random evolutionary Equation (\[main\]), which is given by $$\label{M}
M(\omega)=\{\xi+h(\xi,\omega)|\xi\in PH\},$$ where $h:PH\rightarrow QH$ is a Lipschitz continuous mapping given by $$\label{h}
h(\xi,\omega)=\int_{-\infty}^0e^{-As+\int_s^0z(\theta_r\omega){\rm
d}r}QG(\theta_s\omega,u(s;\xi,\omega)).$$
It is easy to see that if $F$ is $C^1$, then the stochastic invariant manifold obtained in Theorem \[the\] is $C^1$ by Theorem 5.3 of [@Dua1].
Theorem \[the\] says that (\[main\]) has a Lipschitz manifold if the spectral gap condition (\[gap\]) holds. To show that the manifold is an inertial manifold for (\[main\]), we should verify that it attracts exponentially all the orbits of $\varphi$. A stronger reduction property is the exponential tracking property [@Foi1], also called [*asymptotical completeness property*]{} [@Rob]: each trajectory of the evolution equation tends exponentially to a trajectory on the inertial manifold. To be more specific, we states it as follows:
\[asymp\] Let $M(\omega)$ be an invariant manifold for RDS $\varphi$. If for $\forall x\in H$, there exists an $\bar x\in M(\omega)$ such that $$|\varphi(t,\omega,x)-\varphi(t,\omega,\bar x)|\le
c_1e^{-c_2t}|x-\bar x|, ~\forall t\ge 0,$$ where $c_1>0$ is a constant dependent on $\omega$, $x$ and $\bar x$, while $c_2$ is a constant independent of these variables, then $M(\omega)$ is said to have the [*asymptotic completeness property*]{}.
If $M(\omega)$ has the asymptotic completeness property, then the asymptotic behavior of $\varphi$ on $H$ can be reduced to $M(\omega)$. Hence the the original infinite dimensional SPDE problem on $H$ is reduced to a finite dimensional stochastic ODE problem on $M(\omega)$.
Denote $$C_\eta^+:=\{\phi:[0,\infty)\rightarrow H|\phi~{\rm
continuous},~\sup_{t\ge 0}e^{-\eta t-\int_0^tz(\theta_r\omega){\rm
d}r}|\phi(t)|<\infty\},$$ then $C_\eta^+$ is a Banach space with norm $|\phi|_{C_\eta^+}:=\sup_{t\ge 0}e^{-\eta
t-\int_0^tz(\theta_r\omega){\rm d}r}|\phi(t)|$.
\[th1\] If we have the spectral gap condition $$\label{gap1}
K{\rm
Lip}F\left(\frac{1}{\alpha-\eta}+\frac{1}{\eta-\beta}\right)+K^2{\rm
Lip}h\cdot{\rm Lip}F\frac{1}{\alpha-\eta}<1,$$ then the Lipschitz invariant manifold for (\[main\]) obtained in Theorem \[the\] has the asymptotic completeness property.
[**Proof.**]{} Assume $u,\bar u$ are two solutions of (\[main\]) and let $w=\bar u-u$, then $w$ satisfies the following equation: $$\label{w}
\frac{{\rm d}w}{{\rm d}t}=Aw+z(\theta_t\omega)w+\tilde
F(\theta_t\omega,w),$$ where $$\tilde
F(\theta_t\omega,w):=G(\theta_t\omega,u+w)-G(\theta_t\omega,u).$$ It is clear that $$\label{tildeF}
\tilde F(\theta_t\omega,0)=0,~{\rm Lip}_w\tilde F={\rm Lip}_uG={\rm
Lip}F.$$
First if $w\in C_\eta^+$ is a solution of (\[w\]), then $w$ can be expressed by $$\begin{aligned}
\label{w1}
w(t)=&e^{At+\int_0^tz(\theta_r\omega){\rm
d}r}Qw(0)+\int_0^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}Q\tilde F(\theta_s\omega,w(s)){\rm
d}s\nonumber\\
&~~+\int_{\infty}^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}P\tilde F(\theta_s\omega,w(s)){\rm d}s.\end{aligned}$$ In fact, since $w$ is a solution of (\[w\]), we have $$w(t)=e^{A(t-t_0)+\int_{t_0}^tz(\theta_r\omega){\rm
d}r}w(t_0)+\int_{t_0}^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}\tilde F(\theta_s\omega,w(s)){\rm d}s.$$ This implies $$Pw(t)=e^{A(t-t_0)+\int_{t_0}^tz(\theta_r\omega){\rm
d}r}Pw(t_0)+\int_{t_0}^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}P\tilde F(\theta_s\omega,w(s)){\rm d}s.$$ When $t_0>t$, by (\[expo\]) we have $$\begin{aligned}
|e^{A(t-t_0)+\int_{t_0}^tz(\theta_r\omega){\rm d}r}Pw(t_0)|&\le
Ke^{\alpha(t-t_0)+\int_{t_0}^tz(\theta_r\omega){\rm d}r}|w(t_0)|\\
&\le Ke^{-(\alpha-\eta)t_0+\alpha t+\int_{0}^tz(\theta_r\omega){\rm
d}r}|w|_{C_\eta^+}.\end{aligned}$$ By the property of $z(\omega)$ we obtain $$e^{A(t-t_0)+\int_{t_0}^tz(\theta_r\omega){\rm d}r}Pw(t_0)\rightarrow
0{\rm ~as~}t_0\rightarrow\infty.$$ Therefore, $$Pw(t)=\int_{\infty}^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}P\tilde F(\theta_s\omega,w(s)){\rm d}s.$$ Thus (\[w1\]) holds.
We then show that (\[w1\]) has solutions on $C_\eta^+$ and $\bar
u(0)=u(0)+w(0)\in M(\omega)$. From [@Dua1] we know that the solution $\bar u$ lies on $M$ if and only if $Q\bar u(0)=h(P\bar
u(0),\omega)$, recalling that $M(\omega)=\{\xi+h(\xi,\omega)|\xi\in
PH\}$. That is $$\label{qw0}
Qw(0)=-Qu(0)+h(Pu(0)+Pw(0),\omega).$$ Let $$\begin{aligned}
\tilde Tw(t)&=e^{At+\int_0^tz(\theta_r\omega){\rm d}r}Qw(0),\\
Tw(t)&=\int_0^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm d}r}Q\tilde
F(\theta_s\omega,w(s)){\rm d}s
+\int_{\infty}^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm d}r}P\tilde
F(\theta_s\omega,w(s)){\rm d}s,\end{aligned}$$ then (\[w1\]) reads as $$w(t)=\tilde Tw(t)+Tw(t).$$ We assert that $\tilde T$ and $T$ map $C_\eta^+$ to $C_\eta^+$. In fact, $$\begin{aligned}
e^{-\eta t-\int_0^tz(\theta_r\omega){\rm d}r}|\tilde Tw(t)|& \le
Ke^{-(\eta-\beta)t}|Qw(0)|\\
&\le K|Qw(0)|\\
&\le^{(\ref{qw0})}K\bigl(|-Qu(0)+h(Pu(0),\omega)|\\
&\qquad\qquad +|h(Pu(0)+Pw(0),\omega)-h(Pu(0),\omega)|\bigr)\\
&\le K\bigl(|-Qu(0)+h(Pu(0),\omega)|+{\rm Lip}h|Pw(0)|\bigr)\\
&\le^{(\ref{w1})} K\biggl(|-Qu(0)+h(Pu(0),\omega)|\\
&\qquad\qquad+{\rm
Lip}h\left|\int_\infty^0e^{-As+\int_s^0z(\theta_r\omega){\rm
d}r}P\tilde
F(\theta_s\omega,w(s)){\rm d}s\right|\biggr)\\
&\le^{(\ref{tildeF})} K\bigl(|-Qu(0)+h(Pu(0),\omega)|+K{\rm
Lip}h\cdot{\rm Lip}F\frac1{\alpha-\eta}|w|_{C_\eta^+}\bigr)\end{aligned}$$ and $$\begin{aligned}
e^{-\eta t-\int_0^tz(\theta_r\omega){\rm d}r}|Tw(t)|& \le Ke^{-\eta
t-\int_0^tz(\theta_r\omega){\rm
d}r}\biggl(\int_0^te^{\beta(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}|\tilde
F(\theta_s\omega,w(s))|{\rm d}s\\
&\qquad
+\left|\int_\infty^te^{\alpha(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}\tilde F(\theta_s\omega,w(s)){\rm d}s\right|\biggr)\\
&\le^{(\ref{tildeF})} K{\rm Lip}Fe^{-\eta
t-\int_0^tz(\theta_r\omega){\rm
d}r}\biggl(\int_0^te^{\beta(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}|w(s)|{\rm d}s\\
&\qquad +\int_t^\infty e^{\alpha(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}|w(s)|{\rm d}s\biggr)\\
&\le K{\rm Lip}F\biggl(\int_0^te^{-(\eta-\beta)(t-s)}{\rm
d}s+\int_t^\infty e^{(\alpha-\eta)(t-s)}{\rm d}s\biggr)|w|_{C_\eta^+}\\
&\le K{\rm
Lip}F\left(\frac1{\eta-\beta}+\frac1{\alpha-\eta}\right)|w|_{C_\eta^+}.\end{aligned}$$
Next we show that under the spectral gap condition (\[gap1\]), the map $\tilde T+T:C_\eta^+\rightarrow C_\eta^+$ is contractive. To this end, assume $w,\bar w\in C_\eta^+$, then we have $$\begin{aligned}
e^{-\eta t-\int_0^tz(\theta_r\omega){\rm d}r}|\tilde Tw(t)-\tilde
T\bar w(t)|& \le^{(\ref{qw0})}
Ke^{-(\eta-\beta)t}|h(Pu(0)+Pw(0),\omega)\\
&\qquad\qquad -h(Pu(0)+P\bar w(0),\omega)|\\
&\le K{\rm Lip}h|Pw(0)-P\bar w(0)|\\
&\le K{\rm
Lip}h\left|\int_\infty^0e^{-As+\int_s^0z(\theta_r\omega){\rm
d}r}P{\rm Lip}\tilde F|w(s)-\bar w(s)|{\rm d}s\right|\\
&\le K^2{\rm Lip}h\cdot{\rm
Lip}F\int_0^\infty e^{-(\alpha-\eta)s}|w-\bar w|_{C_\eta^+}{\rm d}s\\
&\le K^2{\rm Lip}h\cdot{\rm Lip}F\frac1{\alpha-\eta}|w-\bar
w|_{C_\eta^+},\end{aligned}$$ and $$\begin{aligned}
&e^{-\eta t-\int_0^tz(\theta_r\omega){\rm d}r}|Tw(t)-T\bar w(t)|\\
\le & e^{-\eta t-\int_0^tz(\theta_r\omega){\rm
d}r}\biggl(\int_0^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}Q|\tilde F(\theta_s\omega,w(s))-\tilde F(\theta_s\omega,\bar
w(s))|{\rm
d}s\\&+\left|\int_\infty^te^{A(t-s)+\int_s^tz(\theta_r\omega){\rm
d}r}P|\tilde F(\theta_s\omega,w(s))-\tilde F(\theta_s\omega,\bar
w(s))|{\rm d}s\right|\biggr)\\
\le & K\biggl(\int_0^te^{\beta(t-s)-\eta
t+\int_s^0z(\theta_r\omega){\rm d}r}{\rm Lip}F|w(s)-\bar w(s)|{\rm
d}s\\
&\quad +\left|\int_\infty^te^{\alpha(t-s)-\eta
t+\int_s^0z(\theta_r\omega){\rm d}r}{\rm Lip}F|w(s)-\bar w(s)|{\rm
d}s\right|\biggr)\\
\le & K{\rm Lip}F\biggl(\int_0^te^{-(\eta-\beta)(t-s)}|w-\bar
w|_{C_\eta^+}{\rm d}s+\int_t^\infty e^{(\alpha-\eta)(t-s)}|w-\bar
w|_{C_\eta^+}{\rm d}s\biggr)\\
\le & K{\rm
Lip}F\left(\frac1{\eta-\beta}+\frac1{\alpha-\eta}\right)|w-\bar
w|_{C_\eta^+}.\end{aligned}$$ That is $$\begin{aligned}
&|\tilde Tw-\tilde T\bar w|_{C_\eta^+}\le K^2{\rm Lip}h\cdot{\rm
Lip}F\frac1{\alpha-\eta}|w-\bar w|_{C_\eta^+},\\
&|Tw-T\bar w|_{C_\eta^+}\le K{\rm
Lip}F\left(\frac1{\eta-\beta}+\frac1{\alpha-\eta}\right)|w-\bar
w|_{C_\eta^+}.\end{aligned}$$ Therefore, $$|(\tilde T+T)w-(\tilde T+T)\bar w|_{C_\eta^+}\le \left[K^2{\rm
Lip}h\cdot{\rm Lip}F\frac1{\alpha-\eta}+K{\rm
Lip}F\left(\frac1{\eta-\beta}+\frac1{\alpha-\eta}\right)\right]|w-\bar
w|_{C_\eta^+}.$$ Then by the spectral gap condition (\[gap1\]) we obtain that $\tilde T+T$ has a unique fixed point $w^*$ on $C_\eta^+$, which satisfies $$\bar u(0)=u(0)+w^*(0)\in M(\omega)$$ as desired. Hence $$\begin{aligned}
|\bar u(t,\omega,\bar u_0)-u(t,\omega,u_0)|& \le e^{\eta
t+\int_0^tz(\theta_r\omega){\rm d}r}|\bar u_0-u_0|\\
&\le c(\omega)e^{\eta t}|\bar u_0-u_0|, ~t\ge 0\end{aligned}$$ for some $c(\omega)>0$ by the property of $z(\omega)$. $\Box$
Stochastic IM for wave equations
================================
Consider the following wave equation in $[0,\pi]$ perturbed by additive white noise: $$\label{wa}
\epsilon^2{\rm d}u_t+(u_t-\Delta u){\rm d}t=f(u){\rm
d}t+\delta\phi{\rm d}W$$ with $$u(0,x)=u_0(x),~u_t(0,x)=u_1(x),~u(t,0)=u(t,\pi)=0,$$ where $\phi\in H_0^1(0,\pi)$, $u_t:=\dfrac{{\rm d}u}{{\rm d}t}$. We assume that the nonlinear term $f$ is globally Lipschitz continuous on $L^2(0,\pi)$ with Lipschitz constant ${\rm Lip}f$.
Rewrite (\[wa\]) as $$\left\{
\begin{array}{l}
u_t=v, \\
\epsilon^2v_t+v+\tilde Au=f(u)+\delta\phi\dfrac{{\rm d}W}{{\rm d}t},
\end{array}
\right.$$ where $\tilde Au:=-\Delta u$, and $(u,v)\in E:=H_0^1(0,\pi)\times
L^2(0,\pi)$. Let $\bar u=u$, $\bar v=v-\delta\phi z$. Here $z$ satisfies $$\label{zp}
\epsilon^2{\rm d}z+z{\rm d}t={\rm d}W.$$
Let $U=(\bar u,\bar v)\in E$, then $U$ satisfies $$\label{V}
\dot{U}=AU+F(\theta_t\omega,U),$$ where $$\label{af}
A:=\left(
\begin{array}{cc}
0 & {\rm id}_{L^2} \\
-\epsilon^{-2}\tilde A & -\epsilon^{-2}{\rm id}_{L^2} \\
\end{array}
\right),~~
F(\omega,U):=\left(
\begin{array}{c}
\delta\phi z \\
\epsilon^{-2}f(\bar u)\\
\end{array}
\right).$$ Noting that (\[V\]) is a particular form of (\[main\]) with $z=0$. It is easy to verify that $A$ is the infinitesimal generator of a $C^0$-semigroup $e^{At}$ on Hilbert space $E$. Since $F$ is Lipschitz continuous with respect to $U$ (see (\[lip\])), by the classical semigroup theory concerning the local existence and uniqueness of the solutions of evolution differential equations in [@Paz], we obtain the existence and uniqueness of (\[V\]) and hence (\[wa\]).
Since the eigenvalues of $\tilde A$ are $\tilde\lambda_k=k^2$ with corresponding eigenvectors $\tilde e_k=\sin kx$, $k=1,2,\cdots$, the eigenvalues of the operator $A$ are $$\lambda_k^\pm=\frac{-1\pm\sqrt{1-4\epsilon^2k^2}}{2\epsilon^2}$$ with corresponding eigenvectors $$e_k^\pm=\left(
\begin{array}{c}
\sin kx\\
\lambda_k^\pm\sin kx\\
\end{array}
\right),~k=1,2,\cdots.$$ It is clear that $$\label{lambda}
\lambda_k^+\rightarrow -k^2 ~{\rm as}~\epsilon\rightarrow 0.$$
Denote $$E_1:={\rm Span}\{e_k^+|1\le k\le N\}, ~E_{-1}:={\rm
Span}\{e_k^-|1\le k\le N\},$$ $$E_{11}:=E_1\oplus E_{-1},~E_{22}:={\rm Span}\{e_k^\pm|k\ge N+1\},
~E_2=E_{-1}\oplus E_{22}.$$ By the orthogonality of $\sin kx$, we have $$E_1\bot E_{22},~E_{-1}\bot E_{22},$$ while $E_1$ is not orthogonal to $E_{-1}$.
Following [@Mor1], we define an equivalent new inner product on $E$. In this section, we use $\langle\cdot,\cdot\rangle$, $\|\cdot\|$ to denote the usual inner product and norm on $L^2(0,\pi)$, respectively. Let $U_1=(u_1,v_1)$, $U_2=(u_2,v_2)$ are two vectors in $E$ or $E_{11}$, $E_{22}$. Recalling that the usual inner product on $E$ defined by $$\langle U_1,U_2\rangle=\langle u_1,u_2\rangle+\langle \tilde
A^{\frac12}u_1,\tilde A^{\frac12}u_2\rangle+\langle v_1,v_2\rangle.$$
Assume $\dfrac1{2\epsilon}>N+1$, define the new inner product as follows: $$\begin{aligned}
\langle U_1,U_2\rangle_{E_{11}}:=& \frac1{4\epsilon^2}\langle
u_1,u_2\rangle-\langle\tilde A^{\frac12}u_1,\tilde
A^{\frac12}u_2\rangle+\langle\frac1{2\epsilon}u_1+\epsilon
v_1,\frac1{2\epsilon}u_2+\epsilon v_2\rangle,\\
\langle U_1,U_2\rangle_{E_{22}}:=& \langle\tilde
A^{\frac12}u_1,\tilde
A^{\frac12}u_2\rangle+\big(\frac1{4\epsilon^2}-2(N+1)^2\big)\langle
u_1,u_2\rangle+\langle\frac1{2\epsilon}u_1+\epsilon
v_1,\frac1{2\epsilon}u_2+\epsilon v_2\rangle.\end{aligned}$$ For $U=U_{11}+U_{22}$, $V=V_{11}+V_{22}$, define $$\langle U,V\rangle_{E}:=\langle
U_{11},V_{11}\rangle_{E_{11}}+\langle U_{22},V_{22}\rangle_{E_{22}}.$$
Since $\dfrac1{2\epsilon}>N+1$, it is clear that $\langle
\cdot,\cdot\rangle_{E_{11}}$ is equivalent to the usual inner product on $E_{11}$, and $\langle\cdot,\cdot\rangle_{E_{22}}$ is equivalent to the usual inner product on $E_{22}$. Hence the new inner product $\langle\cdot,\cdot\rangle_{E}$ is equivalent to the usual product on $E$, see [@Mor1] for details.
By the definition of new inner product, it is clear that for $U=(u,v)$ with $u=0$ we have $$\label{inp}
\|U\|_E=\epsilon\|v\|,$$ and for any $U=(u,v)\in E$ we have $$\label{inpr}
\|U\|_E\ge\sqrt{\dfrac1{4\epsilon^2}-(N+1)^2}~\|u\|.$$
Under the new inner product $\langle\cdot,\cdot\rangle_E$, by the orthogonality of $\sin kx$ it is easy to verify that we have $$E_1\bot E_{22},~E_{-1}\bot E_{22}.$$ Moreover, we have $E_1\bot E_{-1}$ and hence $E_1\bot E_2$. In fact, by the definition of $\langle\cdot,\cdot\rangle_{E_{11}}$ it follows that $$\left\{
\begin{array}{l}
\langle e_k^+,e_l^-\rangle_{E_{11}}=0, ~{\rm when}~1\le k,l\le N,~ k\neq l,\\
\langle
e_k^+,e_k^-\rangle_{E_{11}}=\dfrac1{4\epsilon^2}-k^2
+(\dfrac1{2\epsilon}+\epsilon\lambda_k^+)(\dfrac1{2\epsilon}+\epsilon\lambda_k^-)=0,~{\rm
for}~1\le k\le N,
\end{array}
\right.$$ which verifies $E_1\bot E_{-1}$.
We use $A_1$, $A_2$, $A_{-1}$, $A_{22}$ to denote $A|_{E_1}$, $A|_{E_2}$, $A|_{E_{-1}}$, $A|_{E_{22}}$, respectively. Then similar to [@Mor1], we have $$\begin{aligned}
&\|e^{A_1t}\|=e^{\lambda_N^+t}, ~{\rm for}~t\le 0,\label{ex1}\\
&\|e^{A_{-1}t}\|=e^{\lambda_N^-t}, ~{\rm for}~t\ge 0,\label{ex2}\\
&\|e^{A_{22}t}\|=e^{\lambda_{N+1}^+t}, ~{\rm for}~t\ge 0,\label{ex3}\end{aligned}$$ where $\|\cdot\|$ denotes the operator norm in Hilbert space $(E,\langle\cdot,\cdot,\rangle_E)$. By (\[ex2\]), (\[ex3\]) we have $$\|e^{A_2t}\|=e^{\lambda_{N+1}^+t}, ~{\rm for}~t\ge 0.\label{ex4}$$
Next we show that $F$ is Lipschitz with respect to $U$ under the norm $\|\cdot\|_E$ and the Lipschitz constant is independent of $\epsilon$ when $\epsilon$ is small. In fact, $$\begin{aligned}
\|F(\omega,U_1)-F(\omega,U_2)\|_E & \le\left\|\epsilon^{-2}\left(
\begin{array}{c}
0\\
f(u_1)-f(u_2)\nonumber\\
\end{array}
\right)
\right\|_E\nonumber\\
& \le^{(\ref{inp})}\epsilon^{-1}\|f(u_1)-f(u_2)\|\nonumber\\
& \le \epsilon^{-1}{\rm Lip}f\|u_1-u_2\|\nonumber\\
& \le^{(\ref{inpr})}\epsilon^{-1} \dfrac{{\rm
Lip}f}{\sqrt{\dfrac1{4\epsilon^2}-(N+1)^2}}\|U_1-U_2\|_E\nonumber\\
&\le \dfrac{{\rm
Lip}f}{\sqrt{\dfrac14-\epsilon^2(N+1)^2}}\|U_1-U_2\|_E\nonumber\\
&\le 3{\rm Lip}f\|U_1-U_2\|_E,\label{lip}\end{aligned}$$ where the last “=" holds when $\epsilon$ is appropriately small.
Consider stochastic wave equation (\[wa\]). There exists some $\epsilon_0>0$ such that for any $\epsilon\in(0,\epsilon_0)$, the equation (\[wa\]) has a stochastic IM.
[**Proof.**]{} Consider (\[V\]) and let $H=(E,\langle\cdot,\cdot\rangle_E)$, $A$ be as in (\[af\]), $\alpha=\lambda_N^+$, $\beta=\lambda_{N+1}^+$ and $\eta=\dfrac{\alpha+\beta}{2}$. By (\[ex1\]) and (\[ex4\]), the pseudo exponent dichotomy condition (\[expo\]) holds with $PH=E_1$, $QH=E_2$ and $K=1$. According to (\[lambda\]) and (\[lip\]), there exists $\epsilon_0>0$ such that for any $\epsilon\in(0,\epsilon_0)$ the spectral gap condition (\[gap1\]) holds when $N$ is appropriately large. Hence Theorem \[th1\] holds for (\[V\]), i.e. there exists a stochastic IM $M(\omega)$ for (\[V\]).
For $U\in E$, define the transform $$T(\omega,U)=U-(0,\delta\phi z),~T^{-1}(\omega,U)=U+(0,\delta\phi z).$$ If $(t,\omega,U_0)\rightarrow \varphi(t,\omega,U_0)$ is the RDS generated by (\[V\]), then it is easy to verify that $$(t,\omega,U_0)\rightarrow\tilde\varphi:=T^{-1}(\theta_t\omega,\varphi(t,\omega,T(\omega,U_0)))$$ is the RDS generated by (\[wa\]).
Let $$\tilde
M(\omega):=T^{-1}(\omega,M(\omega))=\{\xi+h(\xi,\omega)+(0,\delta\phi
z)|\xi\in PE\},$$ then $\tilde M(\omega)$ is a stochastic IM for (\[wa\]). In fact, $$\begin{aligned}
\tilde\varphi(t,\omega,\tilde
M(\omega))&=T^{-1}(\theta_t\omega,\varphi(t,\omega,T(\omega,\tilde
M(\omega)))) \\
&=T^{-1}(\theta_t\omega,\varphi(t,\omega,M(\omega)))\\
&\subset
T^{-1}(\theta_t\omega,M(\theta_t\omega))=\tilde M(\theta_t\omega),\end{aligned}$$ i.e. $\tilde M(\omega)$ is an invariant manifold for (\[wa\]).
Assume $\tilde U_1$ is a solution of (\[wa\]), then it is easy to verify that $$U_1:=T(\theta_t\omega,\tilde U_1(t,\omega,T^{-1}(\omega,\tilde
U_1(0))))$$ is a solution of (\[V\]). By the asymptotic complete property of $M(\omega)$, there exists a solution $U_2$ of (\[V\]) lying on $M(\omega)$ such that $$\|\varphi(t,\omega,U_1(0))-\varphi(t,\omega,U_2(0))\|_E\le
c(\omega)e^{\eta t}\|U_1(0)-U_2(0)\|_E,~\forall t\ge 0.$$ Let $\tilde U_2:=T^{-1}(\theta_t\omega,U_2(t,\omega,U_2(0)))$, then it is easy to verify that $\tilde U_2$ is a solution of (\[wa\]) and $\tilde U_2$ lies on $\tilde M(\omega)$. Furthermore, $$\begin{aligned}
\|\tilde\varphi(t,\omega,\tilde
U_1(0))-\tilde\varphi(t,\omega,\tilde
U_2(0))\|_E&=\|\varphi(t,\omega,U_1(0))-\varphi(t,\omega,U_2(0))\|_E\\
&\le c(\omega)e^{\eta t}\|U_1(0)-U_2(0)\|_E\\
& \le c(\omega)e^{\eta t}\|\tilde U_1(0)-\tilde U_2(0)\|_E,~\forall
t\ge 0.\end{aligned}$$ Therefore, $\tilde M(\omega)$ has asymptotic completeness property and hence it is a stochastic IM for (\[wa\]). The proof is complete. $\Box$
In above theorem we obtain the existence of stochastic IM when $\epsilon$ is small. In fact, when $\epsilon$ is large, counterexample has shown that the attractor of (\[wa\]) in the deterministic case (i.e. $\delta=0$) is not contained in any finite dimensional manifold, see [@Mor3] for details. It seems that the corresponding result holds for stochastic case, i.e. we would not obtain the existence of stochastic IM for (\[wa\]) when $\epsilon$ is large.
Denote $$C_\eta^-:=\{\phi:(-\infty,0]\rightarrow E|\phi~{\rm
continuous},~\sup_{t\le 0}e^{-\eta t}\|\phi(t)\|_E<\infty\},$$ then $C_\eta^-$ is a Banach space with norm $\|\phi\|_{E,C_\eta^-}:=\sup_{t\le 0}e^{-\eta t}\|\phi(t)\|_E$. Assume $R>0$ and $M_\delta(\omega)$ is a stochastic IM of (\[wa\]). Let $$M_\delta^R(\omega):=\{\xi+h(\xi,\omega)|\xi\in PE,\|\xi\|_E\le R\},$$ where the graph of $h$ gives the IM $M_\delta(\omega)$. The following theorem states that the stochastic IM of (\[wa\]) converges to its deterministic counterpart almost surely when the intensity of noise tends to zero.
Assume $M_\delta(\omega)$ is a stochastic IM of (\[wa\]) and $M_0$ is the IM of (\[wa\]) when $\delta=0$ with the same dimension as that of $M_\delta(\omega)$, then, for any $R>0$, we have $$\lim_{\delta\rightarrow 0}\sup_{U\in M_\delta^R(\omega)}\inf_{V_\in
M_0}\|U-V\|_E=0$$ almost surely.
[**Proof.**]{} Assume $\bar u$, $u$ satisfy $$\epsilon^2{\rm d}\bar u_t+(\bar u_t+\tilde A\bar u){\rm d}t=f(\bar
u){\rm d}t+\delta\phi{\rm d}W$$ and $$\epsilon^2{\rm d}u_t+(u_t+\tilde A u){\rm d}t=f(u){\rm d}t,$$ respectively. We also assume that $(\bar u,\bar u_t)$, $(u,u_t)$ lie on $M_\delta(\omega)$, $M_0$, respectively. Let $w=\bar u-u$, then $w$ satisfies $$\label{wor}
\epsilon^2w_{tt}+w_t+\tilde Aw=f(u+w)-f(u)+\delta\phi\frac{{\rm
d}W}{{\rm d}t}.$$ Let $\bar W=(w,w_t-\delta\phi z)$, where $z$ satisfies $\epsilon^2{\rm d}z+z{\rm d}t={\rm d}W$, then $\bar W$ satisfies $$\label{W}
\dot{\bar W}=A\bar W+F(\theta_t\omega,\bar W),$$ where $$A=\left(
\begin{array}{cc}
0 & {\rm id}_{L^2} \\
-\epsilon^{-2}\tilde A & -\epsilon^{-2}{\rm id}_{L^2} \\
\end{array}
\right),~~ F=\left(\begin{array}{c}
\delta\phi z \\
\epsilon^{-2}[f(u+w)-f(u)]
\end{array}\right).$$ It is clear that the form of (\[W\]) is the same as that of (\[V\]) except that the nonlinear term $F$ is not the same. But it is easy to verify that the nonlinear term $F$ in (\[W\]) is globally Lipschitz continuous with respect to $\bar W$, so (\[W\]) has a stochastic IM and by similar argument to that of Theorem \[th1\] (see also (27) in [@Dua1]) we have $\bar W\in
C_\eta^-$ and $\bar W$ satisfies $$\begin{aligned}
\bar W(t)=& e^{At}P\bar
W(0)+\int_0^te^{A(t-s)}PF(\theta_s\omega,\bar W(s)){\rm
d}s\\
&\quad +\int_{-\infty}^te^{A(t-s)}QF(\theta_s\omega,\bar W(s)){\rm
d}s.\end{aligned}$$ Since $PE=E_1$ is of finite dimension, we can choose $(u(0),
u_t(0))$ such that $P\bar W(0)=P(\bar u(0)-u(0),\bar
u_t(0)-u_t(0)-\delta\phi z)=0$. Therefore, $$\begin{aligned}
e^{-\eta t}\|\bar W(t)\|_E
& \le e^{-\eta
t}\int_t^0e^{\lambda_N^+(t-s)}\left\|\left(
\begin{array}{c}
\delta\phi z(\theta_s\omega) \\
\epsilon^{-2}[f(u+w)-f(u)]\\
\end{array}
\right)
\right\|_E{\rm d}s\\
&\qquad +e^{-\eta
t}\int_{-\infty}^te^{\lambda_{N+1}^+(t-s)}\left\|\left(
\begin{array}{c}
\delta\phi z(\theta_s\omega) \\
\epsilon^{-2}[f(u+w)-f(u)]\\
\end{array}
\right)
\right\|_E{\rm d}s \\
& \le e^{-\eta t}\int_t^0e^{\lambda_N^+(t-s)}\left\|\left(
\begin{array}{c}
0\\
\epsilon^{-2}[f(u+w)-f(u)]\\
\end{array}
\right)
\right\|_E{\rm d}s\\
&\qquad +e^{-\eta t}\int_{t}^0e^{\lambda_{N}^+(t-s)}\left\|\left(
\begin{array}{c}
\delta\phi z(\theta_s\omega) \\
0\\
\end{array}
\right)
\right\|_E{\rm d}s\\
&\qquad +e^{-\eta
t}\int_{-\infty}^te^{\lambda_{N+1}^+(t-s)}\left\|\left(
\begin{array}{c}
0\\
\epsilon^{-2}[f(u+w)-f(u)]\\
\end{array}
\right)
\right\|_E{\rm d}s \\
&\qquad +e^{-\eta
t}\int_{-\infty}^te^{\lambda_{N+1}^+(t-s)}\left\|\left(
\begin{array}{c}
\delta\phi z(\theta_s\omega) \\
0\\
\end{array}
\right)
\right\|_E{\rm d}s\\
&\le 3{\rm Lip}f\int_t^0e^{(\lambda_N^+-\eta)(t-s)}\left\| \left(
\begin{array}{c}
w \\
0\\
\end{array}
\right) \right\|_{E,C_\eta^-}{\rm d}s\\
&\qquad +3{\rm
Lip}f\int_{-\infty}^te^{(\lambda_{N+1}^+-\eta)(t-s)}\left\| \left(
\begin{array}{c}
w \\
0\\
\end{array}
\right) \right\|_{E,C_\eta^-}{\rm d}s\\
&\qquad +c_1(\omega)\left(\int_t^0e^{(\lambda_N^+-\eta)(t-s)}{\rm
d}s+\int_{-\infty}^te^{(\lambda_{N+1}^+-\eta)(t-s)}{\rm
d}s\right)\left\|\left(
\begin{array}{c}
\delta\phi \\
0\\
\end{array}
\right)
\right\|_E\\
&\le 3{\rm Lip}f\left\| \left(
\begin{array}{c}
w \\
0\\
\end{array}
\right) \right\|_{E,C_\eta^-}\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\\
&\qquad +c_1(\omega)\left\|\left(
\begin{array}{c}
\delta\phi \\
0\\
\end{array}
\right)
\right\|_E\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\\
&\le 3{\rm Lip}f\|\bar
W\|_{E,C_\eta^-}\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\\
&\qquad +c_1(\omega)\left\|\left(
\begin{array}{c}
\delta\phi \\
0\\
\end{array}
\right)
\right\|_E\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right),\end{aligned}$$ where the third “$\le$" holds for some $c_1(\omega)$ due to (\[lip\]) and the sublinear growth of $z(\theta_s\omega)$ with respect to $s$. Hence we have $$\left[1-3{\rm Lip}f\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\right]\|\bar
W\|_{E,C_\eta^-} \le c_1(\omega)\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\left\|\left(
\begin{array}{c}
\delta\phi \\
0\\
\end{array}
\right)
\right\|_E.$$ When $N$ is appropriately large and $\epsilon$ is appropriately small we have $$3{\rm Lip}f\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\le\frac12,~\left(\dfrac{1}{\lambda_N^+-\eta}
+\dfrac{1}{\eta-\lambda_{N+1}^+}\right)\le 1,$$ which implies that $$\|\bar W\|_{E,C_\eta^-} \le 2c_1(\omega)\left\|\left(
\begin{array}{c}
\delta\phi \\
0\\
\end{array}
\right)
\right\|_E.$$ Returning back to (\[wor\]), we let $\hat W=\bar W+(0,\delta\phi
z)=(w,w_t)$, then $$\begin{aligned}
\|\hat W\|_{E,C_\eta^-}& \le \|\bar W\|_{E,C_\eta^-}+\left\|\left(
\begin{array}{c}
0\\
\delta\phi z\\
\end{array}
\right)
\right\|_{E,C_\eta^-}\\
&\le \|\bar W\|_{E,C_\eta^-}+\left\|\left(
\begin{array}{c}
0\\
\delta\phi\\
\end{array}
\right)
\right\|_{E}\sup_{t\le 0}e^{-\eta t}|z(\theta_t\omega)|\\
& \le 2c_1(\omega)\left\|\left(
\begin{array}{c}
\delta\phi\\
0\\
\end{array}
\right)
\right\|_{E}+c_2(\omega)\left\|\left(
\begin{array}{c}
0 \\
\delta\phi\\
\end{array}
\right)
\right\|_{E}\\
&\le 2\delta c_1(\omega)\left\|\left(
\begin{array}{c}
\phi\\
0\\
\end{array}
\right)
\right\|_{E}+\delta c_2(\omega)\left\|\left(
\begin{array}{c}
0 \\
\phi\\
\end{array}
\right)
\right\|_{E}\end{aligned}$$ for some $c_2(\omega)$. Thus $$\|\hat W(0)\|_E\le \|\hat W\|_{E,C_\eta^-}\le\delta c_3(\omega),$$ where $$c_3(\omega):=2c_1(\omega)\left\|\left(
\begin{array}{c}
\phi\\
0\\
\end{array}
\right)
\right\|_{E}+c_2(\omega)\left\|\left(
\begin{array}{c}
0 \\
\phi\\
\end{array}
\right)
\right\|_{E}.$$ The proof is complete. $\Box$
Acknowledgement {#acknowledgement .unnumbered}
===============
I am most indebted to my advisor, Professor Yong Li, not only for his direct helpful suggestions but primarily for his continual instruction, encouragement and support over all these years.
Appendix {#appendix .unnumbered}
========
It is well-known that, for deterministic evolution equations, the inertial manifolds contain the corresponding global attractors when they both exist. Like deterministic case, we have the same result for stochastic evolution equations: stochastic IM contains the corresponding random attractor when they both exist. Here we give a simple proof of this result.
First let us recall the definition of (global) random attractor.
([@Cra1])\[attractor\] Assume $\varphi$ is an RDS on a Polish space $X$, then a random compact set $A(\omega)$ is called a [*(global) random attractor*]{} for the RDS $\varphi$ if
- $A(\omega)$ is invariant, i.e. $$\label{inva}
\varphi(t,\omega,A(\omega))=A(\theta_t\omega),~\forall t\ge 0$$ for almost all $\omega\in\Omega$;
- $A(\omega)$ pull-back attracts every bounded deterministic set, i.e. for any bounded deterministic set $B\subset X$, we have $$\label{att}
\lim_{t\rightarrow\infty}d(\varphi(t,\theta_{-t}\omega,B),A(\omega))=0$$ almost surely.
In (\[att\]), $d(D_1,D_2)$ denotes the Hausdorff semi-metric between $D_1$ and $D_2$, i.e. $$d(D_1,D_2):=\sup_{x\in D_1}\inf_{y\in D_2}d_X(x,y)$$ for any two closed sets $D_1$, $D_2$ in $X$.
The global random attractor for RDS $\varphi$ is the [*minimal*]{} random closed set which attracts all the bounded deterministic sets and it is the [*largest*]{} random compact set which is invariant in the sense of (\[inva\]), see [@Cra] for details. The random attractor defined above is unique and it is uniquely determined by attracting [*deterministic compact*]{} sets, see [@Cr] for details.
Assume an SPDE has a stochastic IM $M(\omega)$ and a random attractor $A(\omega)$. Then we have $A(\omega)\subset M(\omega)$ almost surely.
[**Proof.**]{} If the assertion is false, then $$\mathbb P\{\omega|A(\omega)\not\subset M(\omega)\}>0.$$ Let $\tilde A(\omega)=A(\omega)\cap M(\omega)$. Since $A(\omega)$ is “minimal", there exists a deterministic compact set $D$ and $\epsilon_1,\epsilon_2>0$ such that $$\label{con}
\mathbb
P\{\omega|\lim_{t\rightarrow\infty}d(\varphi(t,\theta_{-t}\omega,
D),\tilde A(\omega))\ge\epsilon_1\}=\epsilon_2>0.$$ Since $A(\omega)$ is the random attractor, we have $$\label{con1}
\mathbb
P\{\omega|\lim_{t\rightarrow\infty}d(\varphi(t,\theta_{-t}\omega,
D),A(\omega))=0\}=1.$$ On the other hand we have $$\begin{aligned}
& \mathbb
P\{\omega|\lim_{t\rightarrow\infty}d(\varphi(t,\theta_{-t}\omega,
D),M(\omega))=0\}\nonumber\\
=& \mathbb P\{\omega|\lim_{t\rightarrow\infty}d(\varphi(t,\omega,
D),M(\theta_t\omega))=0\}=1\label{con2}\end{aligned}$$ by the measure preserving of $\{\theta_t\}_{t\in\mathbb R}$ and the fact that $M(\omega)$ is a stochastic IM for $\varphi$. According to (\[con1\]), (\[con2\]) and the definition of Hausdorff semi-metric, we have $$\mathbb
P\{\omega|\lim_{t\rightarrow\infty}\varphi(t,\theta_{-t}\omega,D)\subset
A(\omega)\cap M(\omega)=\tilde A(\omega)\}=1,$$ a contradiction to (\[con\]). The proof is complete. $\Box$
[99]{}
L. Arnold, Random Dynamical Systems, Springer-Verlag, Berlin Heidelberg New York, 1998.
A. Bensoussan, F. Flandoli, Stochastic inertial manifold, Stochast. Stoch. Rep. 53 (1995) 13-39.
T. Caraballo, P.E. Kloeden, B. Schmalfuss, Exponentially stable stationary solutions for stochastic evolution equations and their perturbation, Appl. Math. Optim. 50 (2004) 183-207.
S.N. Chow, K. Lu, Invariant manifolds for flows in Banach spaces, J. Diff. Eqs. 74 (1988) 285-317.
S.N. Chow, K. Lu, G.R. Sell, Smoothness of inertial manifolds, J. Math. Anal. Appl. 169 (1992) 283-312.
I.D. Chueshov, T.V. Girya, Inertial manifolds for stochastic dissipative dynamical systems, Doklady Acad. Sci. Ukraine 7 (1994) 42-45.
I.D. Chueshov, M. Scheutzow, Inertial manifolds and forms for stochastically perturbed retarded semilinear parabolic equations, J. Dynam. Differential Equations 13 (2001) 355-380.
P. Constantin, C. Foias, B. Nicolaenko, R. Temam, Nouveaux résultats sur les variétés inertielles pour les équations différentielles dissipatives (New results on the inertial manifolds for dissipative differential equations), C. R. Acad. Sci., Paris, Sér. I. 302 (1986) 375-378.
P. Constantin, C. Foias, B. Nicolaenko, R. Temam, Integral Manifolds and Inertial Manifolds for Dissipative Partial Differential Equations, Applied Mathematical Sciences, 70. Springer-Verlag, New York, 1989.
H. Crauel, Global random attractors are uniquely determined by attracting deterministic compact sets, Ann. Mat. Pura Appl. 176 (1999) 57-72.
H. Crauel, A. Debussche, F. Flandoli, Random attractors, J. Dynam. Differential Equations 9 (1997) 307-341.
H. Crauel, F. Flandoli, Attractors for random dynamical systems, Prob. Theory Rel. Fields 100 (1994) 365-393.
G. Da Prato, A. Debussche, Construction of stochastic inertial manifolds using backward integration, Stochast. Stoch. Rep. 59 (1996) 305-324.
G. Da Prato, J. Zabczyk, Stochastic Equations in Infinite Dimensions, Cambridge University Press, Cambridge, 1992.
J. Duan, K. Lu, B. Schmalfuss, Invariant manifolds for stochastic partial differential equations, Ann. Probab. 31 (2003) 2109-2135.
J. Duan, K. Lu, B. Schmalfuss, Smooth stable and unstable manifolds for stochastic evolutionary equations, J. Dynam. Differential Equations 16 (2004) 949-972.
C. Foias, B. Nicolaenko, G.R. Sell, R. Temam, Variétés inertielles pour l¡¯¨¦quation de Kuramoto-Sivashinsky (Inertial manifolds for the Kuramoto-Sivashinsky equation), C. R. Acad. Sci., Paris, S¨¦r. I 301 (1985) 285-288.
C. Foias, B. Nicolaenko, G.R. Sell, R. Temam, Inertial manifolds for the Kuramoto-Sivashinsky equation and an estimate of their lowest dimension. J. Math. Pures Appl. 67 (1988) 197-226.
C. Foias, G.R. Sell, R. Temam, Variétés inertielles des équations différentielles dissipatives (Inertial manifolds for dissipative differential equations), C. R. Acad. Sci., Paris, Sér. I 301 (1985) 139-141.
C. Foias, G.R. Sell, R. Temam, Inertial manifolds for non-linear evolutionary equations, J. Differential Equations 73 (1988) 309-353.
C. Foias, G.R. Sell, E.S. Titi, Exponential tracking and approximation of inertial manifolds for dissipative nonlinear equations, J. Dynam. Differential Equations 1 (1989) 199-244.
J. Hadamard, Sur l’itération et les solutions asymptotiques des équations différentielles, Bull. Soc. Math. France 29 (1901) 224-228.
D. Henry, Geometric Theory of Semilinear Parabolic Equations, Lecture Notes in Mathematics, Vol. 840, Springer, Berlin, 1981.
M.A. Liapounoff, Problème général de la stabilité du mouvement, Ann. Fac. Sci. Toulouse 9, 1907. \[Translation of the Russian edition, Kharkov 1892, reprinted by Princeton University Press, Princeton, NJ, 1949 and 1952.\]
J. Mallet-Paret, G.R. Sell, Inertial manifolds for reaction diffusion equations in higher space dimensions, J. Amer. Math. Soc. 1 (1988) 804-866.
R. Mañé, Reduction of semilinear parabolic equations to finite dimensional $C^1$-flows. Geometry and topology (Proc. III Latin Amer. School of Math., Inst. Mat. Pura Aplicada CNPq, Rio de Janeiro, 1976), pp. 361-378. Lecture Notes in Math., Vol. 597, Springer, Berlin, 1977.
X. Mora, Finite-dimensional attracting manifolds in reaction-diffusion equations, Nonlinear partial differential equations (Durham, N.H., 1982), 353-360, Contemp. Math., 17, Amer. Math. Soc., Providence, R.I., 1983.
X. Mora, Finite-dimensional attracting invariant manifolds for damped semilinear wave equations, Contributions to nonlinear partial differential equations, Vol. II (Paris, 1985), 172-183, Pitman Res. Notes Math. Ser., 155, Longman Sci. Tech., Harlow, 1987.
X. Mora, J. Solà-Morales, Existence and nonexistence of finite-dimensional globally attracting invariant manifolds in semilinear damped wave equations, Dynamics of infinite-dimensional systems (Lisbon, 1986), 187-210, NATO Adv. Sci. Inst. Ser. F Comput. Systems Sci., 37, Springer, Berlin, 1987.
A. Pazy, Semigroups of Linear Operators and Application to Partial Differential Equations, Springer, New York, 1983.
O. Perron, Über Stabilität und asymptotisches Verhalten der Integrale von Differentialgleichungssystemen, Math. Z. 29 (1928) 129-160.
J.C. Robinson, The asymptotic completeness of inertial manifolds, Nonlinearity 9 (1996) 1325-1340.
G.R. Sell, Y. You, Inertial manifolds: the non-self-adjoint case, J. Differential Equations 96 (1992) 203-255.
R. Temam, Infinite-Dimensional Dynamical Systems in Mechanics and Physics, 2nd Ed., Applied Mathematical Sciences, Vol. 68, Springer, New York, 1997.
[^1]: This work is partially supported by the 985 project of Jilin University.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study quantum bipartite systems in a random pure state, where von Neumann entropy is considered as a measure of the entanglement. Expressions of the first and second exact cumulants of von Neumann entropy, relevant respectively to the average and fluctuation behavior, are known in the literature. The focus of this paper is on its skewness that specifies the degree of asymmetry of the distribution. Computing the skewness requires additionally the third cumulant, an exact formula of which is the main result of this work. In proving the main result, we obtain as a byproduct various summation identities involving polygamma and related functions. The derived third cumulant also leads to an improved approximation to the distribution of von Neumann entropy.'
address: 'Department of Electrical and Computer Engineering, University of Michigan - Dearborn, Michigan 48128, USA'
author:
- Lu Wei
title: Skewness of von Neumann entanglement entropy
---
[*Keywords*]{}: quantum entanglement, von Neumann entropy, skewness, random matrix theory, polygamma functions
Introduction and the main result
================================
Classical information theory is the theory behind modern development of computing, communications, and other fields. As its classical counterpart, quantum information theory aims at understanding the theoretical underpinnings of quantum science and technology. One of the most fundamental features of quantum mechanics is the phenomenon of entanglement, which is the resource and medium that enable quantum technologies.
In this work, we consider the quantum bipartite model proposed in the seminal work of Page [@Page93] in the year $1993$, which becomes a standard model in describing the interaction of a physical object and its environment. For such a model, we wish to understand the degree of entanglement as measured by the von Neumann entropy, the statistical behavior of which can be studied from its cumulants/moments. In principle, the knowledge of all moments determines uniquely the distribution of von Neumann entropy due to its compact support (a.k.a. Hausdorff’s moment problem). In practice, a finite number of cumulants can be utilized to construct approximations to the distribution of the entropy, where the higher cumulants describe the tail distribution. The higher cumulants also provide information such as whether the average entropy is typical [@Bianchi19]. In the literature, the mean and variance of von Neumann entropy have been investigated in [@Page93; @Foong94; @Ruiz95; @VPO16; @Wei17; @Bianchi19] among others. The focus of this paper is to study the skewness (involves the second and third cumulants) of von Neumann entropy that measures the degree of asymmetry of the distribution.
The bipartite model proposed by Page [@Page93] is formulated as follows. Consider a composite quantum system that consists of two subsystems $A$ and $B$ of Hilbert space dimensions $m$ and $n$, respectively. The Hilbert space $\mathcal{H}_{A+B}$ of the composite system is given by the tensor product of the Hilbert spaces of the subsystems, $\mathcal{H}_{A+B}=\mathcal{H}_{A}\otimes\mathcal{H}_{B}$. A random pure state of the composite system is written as a linear combination of the random coefficients $x_{i,j}$ and the complete basis $\left\{\Ket{i^{A}}\right\}$ and $\left\{\Ket{j^{B}}\right\}$ of $\mathcal{H}_{A}$ and $\mathcal{H}_{B}$, $$\Ket{\psi}=\sum_{i=1}^{m}\sum_{j=1}^{n}x_{i,j}\Ket{i^{A}}\otimes\Ket{j^{B}}.$$ The corresponding density matrix in the random pure state is $$\label{eq:rho}
\rho=\Ket{\psi}\Bra{\psi}=\sum_{i,k=1}^{m}\sum_{j,l=1}^{n}x_{i,j}x_{k,l}^{\dag}\Ket{i^{A}}\Bra{k^{A}}\otimes\Ket{j^{B}}\Bra{l^{B}},$$ which has the natural constraint $\tr(\rho)=1$ (or equivalently $\braket{\psi|\psi}=1$). This implies that the $m\times n$ random coefficient matrix $\mathbf{X}=(x_{i,j})$ satisfies $$\label{eq:pcv}
\tr\left(\mathbf{XX}^{\dag}\right)=1.$$ We assume without loss of generality that $m\leq n$. The reduced density matrix $\rho_{A}$ of the smaller subsystem $A$ is computed by partial tracing of the full density matrix (\[eq:rho\]) over the other subsystem $B$ (interpreted as the environment) as $$\rho_{A}=\tr_{B}(\rho)=\sum_{i,k=1}^{m}\sum_{j=1}^{n}x_{i,j}x_{k,j}^{\dag}\Ket{i^{A}}\Bra{k^{A}}=\sum_{i,k=1}^{m}w_{i,k}\Ket{i^{A}}\Bra{k^{A}},$$ where $w_{i,k}$ is the $(i,k)$-th entry of the $m\times m$ Hermitian matrix $\mathbf{W}=\mathbf{XX}^{\dag}$. The Schmidt decomposition of $\rho_{A}$ is given by $$\rho_{A}=\sum_{i=1}^{m}\lambda_{i}\Ket{\phi_{i}^{A}}\Bra{\phi_{i}^{A}},$$ where $0<\lambda_{m}<\dots<\lambda_{1}<1$ are the eigenvalues of $\mathbf{W}$, and the condition (\[eq:pcv\]) now implies the fixed-trace constraint $$\label{eq:lamcon}
\sum_{i=1}^{m}\lambda_{i}=1.$$ The probability measure of $\rho_{A}$ is the Haar measure satisfying the additional constraint (\[eq:lamcon\]). The corresponding eigenvalue density of $\mathbf{W}$ is well-known (see, e.g., [@Page93]) $$\label{eq:fte}
f\left(\bm{\lambda}\right)=\frac{\Gamma(mn)}{C}~\delta\left(1-\sum_{i=1}^{m}\lambda_{i}\right)\prod_{1\leq i<j\leq m}\left(\lambda_{i}-\lambda_{j}\right)^{2}\prod_{i=1}^{m}\lambda_{i}^{n-m},$$ where $\delta(\cdot)$ is the Dirac delta function and the constant $$\label{eq:con}
C=\prod_{i=1}^{m}\Gamma(n-i+1)\Gamma(i).$$ The random matrix ensemble (\[eq:fte\]) is also known as the (unitary) fixed-trace ensemble. The above discussed bipartite model is useful in describing the interaction of various real-world quantum systems. For example, in [@Page93] the subsystem $A$ is the black hole and the subsystem $B$ is the associated radiation field. In another example [@Majumdar], the subsystem $A$ is a set of spins and the subsystem $B$ represents the environment of a heat bath.
The degree of entanglement of subsystems can be measured by entanglement entropies, which are functions of eigenvalues of $\mathbf{W}$. We consider the standard measure of von Neumann entropy of the subsystem[^1] $$\label{eq:vN}
S=-\tr\left(\rho_{A}\ln\rho_{A}\right)=-\sum_{i=1}^{m}\lambda_{i}\ln\lambda_{i},~~~~~~S\in\left[0, \ln{m}\right],$$ which achieves the separable state ($S=0$) when $\lambda_{1}=1$, $\lambda_{2}=\dots=\lambda_{m}=0$ and the maximally-entangled state ($S=\ln{m}$) when $\lambda_{1}=\dots\lambda_{m}=1/m$. Statistical information of the von Neumann entropy is encoded through its moments $$\mathbb{E}_{f}\!\left[S^{k}\right],~~~~~~k=1,2,3,\dots,$$ where the expectation is taken over the density (\[eq:fte\]). In general, the moment sequence, $m_{1},m_{2},m_{3},\dots$, and the cumulant sequence, $\kappa_{1},\kappa_{2},\kappa_{3},\dots$, for a random variable are related: the $i$-th moment is an $i$-th degree polynomial in the first $i$ cumulants and vice versa. In particular, the relation pairs up to $i=3$ are $$\begin{aligned}
m_{1} &= \kappa_{1} \label{eq:1km} \\
m_{2} &= \kappa_{2}+\kappa_{1}^{2}~~~~~~~~~~~&\kappa_{2} = m_{2}-m_{1}^{2} \label{eq:2mk} \\
m_{3} &= \kappa_{3}+3\kappa_{2}\kappa_{1}+\kappa^{3}_{1}~~~~~~~~~~~&\kappa_{3} = m_{3}-3m_{2}m_{1}+2m_{1}^{3}. \label{eq:3mk}\end{aligned}$$ It turns out that the cumulants/moments of von Neumann entropy can be expressed through polygamma functions, the $i$-th order of which is defined as $$\label{eq:pgd}
\psi_{i}(z)=\frac{\partial^{i+1}\ln\Gamma(z)}{\partial z^{i+1}}=(-1)^{i+1}i!\sum_{k=0}^{\infty}\frac{1}{(k+z)^{i+1}}.$$ For positive integer arguments, the digamma function ($0$-th order polygamma function) is simplified to a finite sum as $$\label{eq:p0}
\psi_{0}(l)=-\gamma+\sum_{k=1}^{l-1}\frac{1}{k}$$ with $\gamma\approx0.5772$ being the Euler’s constant, and the polygamma functions of order $j\geq1$ can be also reduced to finite sums as $$\label{eq:pg}
\psi_{j}(l)=(-1)^{j+1}j!\left(\zeta(j+1)-\sum_{k=1}^{l-1}\frac{1}{k^{j+1}}\right),$$ where $$\label{eq:z}
\zeta(s)=\sum_{k=1}^{\infty}\frac{1}{k^{s}}$$ is the Riemann zeta function. In particular, the present paper involves finite sum form $$\label{eq:p1}
\psi_{1}(l)=\frac{\pi^{2}}{6}-\sum_{k=1}^{l-1}\frac{1}{k^{2}}$$ of the trigamma function, and finite sum form $$\label{eq:p2}
\psi_{2}(l)=-2\zeta(3)+2\sum_{k=1}^{l-1}\frac{1}{k^{3}}$$ of the second order polygamma function with $$\zeta(3)\approx1.20206$$ being the Apéry’s constant.
With the above definitions, we now discuss the state of the art in discovering the exact cumulants of von Neumann entropy. The mean value of von Neumann entropy (first cumulant) relevant to the typical behavior of entanglement was conjectured by Page, in the same work [@Page93] the bipartite model was proposed, as $$\label{eq:k1}
\kappa_{1}=\psi_{0}(mn+1)-\psi_{0}(n)-\frac{m+1}{2n}.$$ Page’s conjecture was proved shortly afterwards in [@Foong94; @Ruiz95] among others. The variance of von Neumann entropy (second cumulant) that describes the fluctuation of entanglement around the typical value was conjectured in [@VPO16] as $$\label{eq:k2}
\kappa_{2}=-\psi_{1}\left(mn+1\right)+\frac{m+n}{mn+1}\psi_{1}\left(n\right)-\frac{(m+1)(m+2n+1)}{4n^{2}(mn+1)}.$$ This variance formula was firstly proved in [@Wei17], and was independently proved in [@Bianchi19] recently, see also [@Wei19] for a discussion on the latter proof. In the present work, we focus on the skewness of von Neumann entropy defined as the third standardized moment $$\label{eq:skew}
\gamma_{1}=\mathbb{E}_{f}\!\left[\left(\frac{S-\kappa_{1}}{\sqrt{\kappa_{2}}}\right)^{3}\right]=\frac{\kappa_{3}}{\kappa_{2}^{3/2}},$$ where the notation $\gamma_{1}$ is due to Karl Pearson. The skewness quantifies the (lack of) symmetry of a probability distribution, where a symmetric distribution such as the Gaussian distribution has a skewness of zero. As seen from the definition (\[eq:skew\]), calculating the skewness requires the additional knowledge on the third cumulant of $S$, an expression of which is given by $$\begin{aligned}
\label{eq:k3}
\fl\kappa_{3}&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\psi_{2}\left(mn+1\right)-\frac{m^{2}+3mn+n^{2}+1}{(mn+1)(mn+2)}\psi_{2}(n)+\frac{(m^2-1)(mn-3n^2+1)}{n(mn+1)^{2}(mn+2)}\psi_{1}(n+1)-\nonumber\\
&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\frac{(m+1)\left(2m^{3}n+3m^{2}n^{2}+2m^{2}+4mn^{3}+15mn^{2}+12mn-2n^{2}+6n+6\right)}{4n^{3}(mn+1)^{2}(mn+2)}.\end{aligned}$$ Proving the above formula of the exact third cumulant of von Neumann entropy is the main contribution of this work. Note that the third moment formula (\[eq:k3\]) has been recently reported in [@Bianchi19], where the authors described the computations that led to the result (\[eq:k3\]) as “The computation is an herculean task but with the help of Wolfram’s Mathematica we are able to simplify the exact formula for $\mu_{3}$”. The rest of the paper is to decode the above description by providing a proof to the claimed result (\[eq:k3\]). Note also that despite having a different starting point of the calculation in [@Bianchi19 Eq. (S32)] than that in the current paper (\[eq:mST\]), the subsequent bulk of the calculations omitted in [@Bianchi19] necessarily involves the calculations performed here in Sections \[sec:IABC\] and \[sec:simp\]. To see the above statement in the second cumulant computation, we refer to [@Wei19]. Interestingly, as will be seen the proof relies crucially on new summation identities derived in this work that in fact have not been implemented in Mathematica [@KG].
Approximations to the distribution of von Neumann entropy can be constructed from the closed-form cumulant expressions. For convenience, we first standardize the von Neumann entropy as $$\label{eq:X}
X=\frac{S-\kappa_{1}}{\sqrt{\kappa_{2}}}$$ so that the first two cumulants of the random variable $X$ become $$\label{eq:cX12}
\kappa_{1}^{(X)}=0,~~~~~~\kappa_{2}^{(X)}=1.$$ The higher order cumulants of $S$ and $X$, beyond the first two in (\[eq:k1\]), (\[eq:k2\]), (\[eq:cX12\]), are related by $$\label{eq:SXc}
\kappa_{j}^{(X)}=\frac{\kappa_{j}}{\kappa_{2}^{j/2}},~~~~~~j\geq3.$$ In principle, the probability density function of the standardized variable $X$ can be represented as [@Cramer] $$\label{eq:app}
f_{X}(x)=\varphi_{X}(x)+r_{X}(x),$$ where the function $r_{X}(x)$ is the reminder term of the initial approximation $\varphi_{X}(x)$. Since the variable (\[eq:X\]) is supported in $X\in(-\infty,\infty)$ with the first two cumulants given by (\[eq:cX12\]), we consider a standard Gaussian distribution as the initial approximation, i.e., $$\label{eq:iappr}
f_{X}(x)\approx\varphi_{X}(x)=\frac{1}{\sqrt{2\pi}}\e^{-\frac{1}{2}x^{2}}.$$ The corresponding initial Gaussian approximation to $S$ incorporates its first two cumulants (\[eq:k1\]) and (\[eq:k2\]) through the affine transformation (\[eq:X\]). The reminder term $r_{X}(x)$ associated with $\varphi_{X}(x)$ in (\[eq:iappr\]) admits different types of expansions. We shall adopt an expansion based on orthogonal polynomials that are induced from the initial approximation $\varphi_{X}(x)$, where the reminder term is formally expanded as [@Cramer] $$\label{eq:expan}
r_{X}(x)=\sum_{k=1}^{\infty}\frac{d_{k}}{k!}\varphi_{X}^{(k)}(x).$$ The $k$-th derivative $\varphi_{X}^{(k)}(x)$ of the Gaussian distribution (\[eq:iappr\]) gives rise to the (probabilists’) Hermite polynomials $H_{k}(x)$ of degree $k$ as $$\label{eq:reH}
\varphi_{X}^{(k)}(x)=(-1)^{k}H_{k}(x)\varphi_{X}(x).$$ The Hermite polynomials satisfy the orthogonality relation [@Cramer; @Forrester] $$\label{eq:ocH}
\int_{-\infty}^{\infty}\!\!\varphi_{X}(x)H_{k}(x)H_{l}(x){\,\mathrm{d}}{x}=k!\delta_{kl}$$ with $\delta_{kl}$ being the Kronecker delta function, where the first a few of $H_{k}(x)$ are $$\begin{aligned}
H_{0}(x) &=1,~~~~~~~~~ &H_{1}(x)=x \label{eq:H01} \\
H_{2}(x) &=x^{2}-1,~~~~~~~~~ &H_{3}(x)=x^{3}-3x. \label{eq:H23}\end{aligned}$$ With the choices of $\varphi_{X}(x)$ in (\[eq:iappr\]) and $r_{X}(x)$ in (\[eq:expan\]), the expansion (\[eq:app\]) is also known as the type-A Gram-Charlier series [@Cramer]. Its $k$-th coefficient $d_{k}$ can be conveniently expressed as a polynomial in the first $k$ cumulants of the standardized random variable $X$ in (\[eq:X\]). In particular, it can be directly verified by the orthogonality relation (\[eq:ocH\]) and the results (\[eq:H01\]), (\[eq:H23\]) that $$d_{1}=d_{2}=0$$ and that $d_{3}$ equals the negative of $\kappa_{3}^{(X)}$, which also equals the negative of the skewness of $S$ (cf. (\[eq:SXc\]) and (\[eq:skew\])), as $$d_{3}=-\kappa_{3}^{(X)}=-\frac{\kappa_{3}}{\kappa_{2}^{3/2}}=-\gamma_{1}.$$ As a result, a refined approximation (cf. (\[eq:iappr\])) to the distribution of standardized von Neumann entropy (\[eq:X\]) is obtained as $$\begin{aligned}
f_{X}(x)&=&\varphi_{X}(x)+\sum_{k=1}^{\infty}\frac{d_{k}}{k!}\varphi_{X}^{(k)}(x)\\
&\approx&\varphi_{X}(x)+\frac{d_{3}}{3!}\varphi_{X}^{(3)}(x), \label{eq:k3app}\end{aligned}$$ where the correction term $$\label{eq:ct}
\frac{d_{3}}{3!}\varphi_{X}^{(3)}(x)=\frac{\kappa_{3}}{6\kappa_{2}^{3/2}}H_{3}(x)\varphi_{X}(x)$$ incorporates the derived third cumulant (\[eq:k3\]).
![Probability densities of standardized von Neumann entropy (\[eq:X\]) of subsystem dimensions $m=4$ and $n=8$: A comparison of Gaussian approximation (\[eq:iappr\]) (dashed line in blue) to the refined approximation (\[eq:k3app\]) (dash-dot line in red) with the $\kappa_{3}$ correction term (\[eq:ct\]). The solid line in black represents simulated true distribution.[]{data-label="fig:p1"}](p1.pdf){width="0.92\linewidth"}
In Figure \[fig:p1\], we numerically compare the refined approximation (\[eq:k3app\]) to the initial Gaussian approximation (\[eq:iappr\]), where the dimensions of the subsystems are $m=4$ and $n=8$. Comparing with the simulated true distribution, it is observed that the new approximation (\[eq:k3app\]) that incorporates the additional knowledge on the skewness via the correction term (\[eq:ct\]) is more accurate than the first two cumulants based Gaussian approximation (\[eq:iappr\]). As compared to the symmetric Gaussian distribution, we also see from Figure \[fig:p1\] that the true distribution of von Neumann entropy is indeed a skewed one. The distribution appears to be left-skewed (a.k.a. a negative skewness), where the left tail of the distribution is longer.
![Probability densities of standardized von Neumann entropy (\[eq:X\]) of subsystem dimensions $m=16$ and $n=32$: A numerical support to the conjectured asymptotic limit. The solid line in black, the dashed line in blue, and the dash-dot line in red represents the simulated true distribution, the standard Gaussian distribution (\[eq:iappr\]), and the refined approximation (\[eq:k3app\]), respectively.[]{data-label="fig:p2"}](p2.pdf){width="0.92\linewidth"}
In Figure \[fig:p2\], we perform another numerical study by simultaneously increasing the subsystem dimensions to $m=16$ and $n=32$ with their ratio $$c=\frac{m}{n}=\frac{1}{2}$$ kept the same as the value in Figure \[fig:p1\]. We observe from Figure \[fig:p2\] that both the simulated and the approximate (\[eq:k3app\]) distributions of $X$ approach the standard Gaussian distribution (\[eq:iappr\]), where the three curves almost overlap. As the author has learnt from Sean O’Rourke, the observed asymptotic Gaussian behavior is typical for a wide class of linear spectral statistics[^2] over different random matrix ensembles. In fact, we conjecture the following central limit theorem for the linear spectral statistics of von Neumann entropy (\[eq:vN\]) over the fixed-trace ensemble (\[eq:fte\]).
In the limit $$\label{eq:lim}
m\to\infty,~~~~n\to\infty,~~~~\frac{m}{n}=c\in(0,1],$$ the standardized von Neumann entropy defined in (\[eq:X\]) converges in distribution to a Gaussian random variable with zero mean and unit variance.
Note that the regime (\[eq:lim\]) is a typical asymptotic regime in random matrix theory (a.k.a. high-dimensional asymptotic regime), where both dimensions $m$ and $n$ approach infinity. This is different from the classical asymptotic regime as studied in, e.g., [@Page93; @Bianchi19], where the dimension $m$ is fixed as $n$ goes to infinity. Proving the above conjecture requires showing that all the higher cumulants (\[eq:SXc\]) vanish in the regime (\[eq:lim\]), i.e., $$\label{eq:ks}
\kappa_{j}^{(X)}=\frac{\kappa_{j}}{\kappa_{2}^{j/2}}\rightarrow0,~~~~~~j\geq3.$$ With the obtained main result (\[eq:k3\]) of this work, we can show that the third cumulant of $X$ vanishes in the regime (\[eq:lim\]) as follows. First, by the asymptotic behavior of polygamma functions $$\psi_{j}(x)=\Theta\left(\frac{1}{x^{j}}\right),~~~~~~x\to\infty,~~~~~~j\geq1,$$ the second cumulant (\[eq:k2\]) and third cumulant (\[eq:k3\]) of $S$ scale in the limit (\[eq:lim\]) as $$\label{eq:k2a}
\kappa_{2}=\Theta\left(\frac{1}{n^{2}}\right),$$ and $$\label{eq:k3a}
\kappa_{3}=\Theta\left(\frac{1}{n^{4}}\right),$$ respectively. Consequently, as claimed, the third cumulant of $X$ (or the skewness of $S$) $$\kappa_{3}^{(X)}=\frac{\kappa_{3}}{\kappa_{2}^{3/2}}=\frac{\Theta(1/n^{4})}{\Theta(1/n^{3})}=\Theta\left(\frac{1}{n}\right)$$ vanishes in the regime (\[eq:lim\]). In addition, based on the structure of the first three exact cumulants and the asymptotic results (\[eq:k2a\]), (\[eq:k3a\]), we further conjecture the limiting behavior of all higher cumulants as $$\kappa_{j}=\Theta\left(\frac{1}{n^{2j-2}}\right),~~~~~~j\geq4,$$ or equivalently, cf. (\[eq:ks\]) and (\[eq:k2a\]), $$\kappa_{j}^{(X)}=\Theta\left(\frac{1}{n^{j-2}}\right),~~~~~~j\geq4.$$
The rest of the paper is organized as follows. In Section \[sec:deri\], we derive the main result (\[eq:k3\]) on the third cumulant of von Neumann entropy. Specifically, in Section \[sec:rela\] the original problem is reduced to the task of computing three integrals (\[eq:IA\]), (\[eq:IB\]), and (\[eq:IC\]) by exploring the relation to a more convenient random matrix ensemble. The three integrals are calculated explicitly in Section \[sec:IABC\] to the form of finite summations involving polygamma functions. The remaining part of the derivation is performed in Section \[sec:simp\], where we evaluate and simplify the resulting sums with the help of two types of polygamma summation identities. The relevant summation identities are listed in the appendices, where we also discuss the strategies in finding them.
Derivation of the third cumulant {#sec:deri}
================================
Cumulant relation {#sec:rela}
-----------------
The task of this subsection is to convert the third cumulant of von Neumann entropy to that of a related random variable, the computation of which can then be conveniently performed.
We start by studying the relation between the third moments. By construction, the random coefficient matrix $\mathbf{X}$ has a natural relation with a Wishart matrix $\mathbf{YY}^{\dag}$ as $$\label{eq:wf}
\mathbf{XX}^{\dag}=\frac{\mathbf{YY}^{\dag}}{\tr\left(\mathbf{YY}^{\dag}\right)},$$ where $\mathbf{Y}$ is an $m\times n$ ($m\leq n$) matrix of independently and identically distributed complex Gaussian entries. The density of the eigenvalues $0<\theta_{m}<\dots<\theta_{1}<\infty$ of $\mathbf{YY}^{\dag}$ equals [@Forrester] $$\label{eq:we}
g\left(\bm{\theta}\right)=\frac{1}{C}\prod_{1\leq i<j\leq m}\left(\theta_{i}-\theta_{j}\right)^{2}\prod_{i=1}^{m}\theta_{i}^{n-m}\e^{-\theta_{i}},$$ where $C$ is given by (\[eq:con\]) and the above ensemble is known as the Wishart-Laguerre ensemble. The trace of the Wishart matrix $$r=\tr\left(\mathbf{YY}^{\dag}\right)=\sum_{i=1}^{m}\theta_{i}$$ follows a gamma distribution with the density $$\label{eq:r}
h_{mn}(r)=\frac{1}{\Gamma(mn)}\e^{-r}r^{mn-1},~~~~~~r\in[0,\infty).$$ The relation (\[eq:wf\]) induces the change of variables $$\label{eq:cv}
\lambda_{i}=\frac{\theta_{i}}{r},~~~~~~i=1,\ldots,m,$$ that leads to a well-known relation (see, e.g. [@Page93]) among the densities (\[eq:fte\]), (\[eq:we\]), and (\[eq:r\]) as $$\label{eq:relation}
f\left(\bm{\lambda}\right)h_{mn}(r){\,\mathrm{d}}r\prod_{i=1}^{m}{\,\mathrm{d}}\lambda_{i}=g\left(\bm{\theta}\right)\prod_{i=1}^{m}{\,\mathrm{d}}\theta_{i}.$$ The above relation implies that $r$ is independent of each $\lambda_{i}$, $i=1,\ldots,m$, since their densities factorize. We now define the random variable $$\label{eq:T}
T=\sum_{i=1}^{m}\theta_{i}\ln\theta_{i},$$ as the induced entropy[^3] over the Wishart-Laguerre ensemble (\[eq:we\]). The relation (\[eq:relation\]) has been utilized to convert the first and second moment of $S$ to $T$ in [@Page93] and [@Wei17], respectively. In a similar manner, the computation of the third moment of $S$ can be also converted to that of $T$ as follows. First, by the change of variables (\[eq:cv\]), one has $$\label{eq:ST}
S=-\sum_{i=1}^{m}\frac{\theta_{i}}{r}\ln\frac{\theta_{i}}{r}=r^{-1}\left(r\ln r-T\right),$$ and consequently $$\begin{aligned}
S^{3} &=& r^{-3}\left(-T^{3}+T^{2}3r\ln r-T3r^{2}\ln^{2}r+r^{3}\ln^{3}r\right)\\
&=& r^{-3}\left(-T^{3}+S^{2}3r^{3}\ln r-S3r^{3}\ln^{2}r+r^{3}\ln^{3}r\right),\label{eq:TS}\end{aligned}$$ where the second equality is obtained by replacing $T$ (except for the the highest power term $T^{3}$) by $S$ using the relation (\[eq:ST\]). As will be seen, the form (\[eq:TS\]) makes it possible to utilize the independence between $r$ and $\bm{\lambda}$ so as to perform the subsequent calculations. The third moment of $S$ can then be computed as $$\begin{aligned}
\fl\mathbb{E}_{f}\!\left[S^{3}\right]&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=&\int_{\bm{\lambda}}r^{-3}\left(-T^{3}+S^{2}3r^{3}\ln r-S3r^{3}\ln^{2}r+r^{3}\ln^{3}r\right)f\left(\bm{\lambda}\right)\prod_{i=1}^{m}{\,\mathrm{d}}\lambda_{i} \label{eq:S2T} \\
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=&\int_{\bm{\lambda}}r^{-3}\left(-T^{3}+S^{2}3r^{3}\ln r-S3r^{3}\ln^{2}r+r^{3}\ln^{3}r\right)f\left(\bm{\lambda}\right)\prod_{i=1}^{m}{\,\mathrm{d}}\lambda_{i}\int_{r}h_{mn+3}(r){\,\mathrm{d}}{r} \nonumber\\
&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!=&\frac{-\mathbb{E}_{g}\!\left[T^{3}\right]+3\mathbb{E}_{h}\!\left[r^{3}\ln r\right]\mathbb{E}_{f}\!\left[S^{2}\right]-3\mathbb{E}_{h}\!\left[r^{3}\ln^{2}r\right]\mathbb{E}_{f}\!\left[S\right]+\mathbb{E}_{h}\!\left[r^{3}\ln^{3}r\right]}{(mn)_{3}}\label{eq:mST},\end{aligned}$$ where in the second line we multiple an appropriate constant $$1=\int_{r}h_{mn+3}(r){\,\mathrm{d}}{r},$$ and the last equality (\[eq:mST\]) is obtained by the using identity (with $(a)_{n}$ denoting the Pochhammer’s symbol) $$r^{-3}h_{mn+3}(r)=\frac{h_{mn}(r)}{(mn)_{3}},$$ as well as the change of measures (\[eq:relation\]) for the $T^{3}$ term along with the fact that $r$ and $\bm{\lambda}$ are independent (so that the integrals involving $r$ and $S$ are evaluated separately). In (\[eq:mST\]), the expected values over the density $h_{mn}(r)$ in (\[eq:r\]) are computed by the identities $$\begin{aligned}
\int_{0}^{\infty}\!\e^{-r}r^{a-1}\ln{r}{\,\mathrm{d}}{r}&=&\Gamma(a)\psi_{0}(a) \\
\int_{0}^{\infty}\!\e^{-r}r^{a-1}\ln^{2}{r}{\,\mathrm{d}}{r}&=&\Gamma(a)\left(\psi_{0}^{2}(a)+\psi_{1}(a)\right) \\
\int_{0}^{\infty}\!\e^{-r}r^{a-1}\ln^{3}{r}{\,\mathrm{d}}{r}&=&\Gamma(a)\left(\psi_{0}^{3}(a)+3\psi_{0}(a)\psi_{1}(a)+\psi_{2}(a)\right),\end{aligned}$$ obtained by taking derivatives with respect to the parameter $a$ of gamma function $$\int_{0}^{\infty}\!\e^{-r}r^{a-1}{\,\mathrm{d}}{r}=\Gamma(a),$$ as $$\begin{aligned}
\fl\mathbb{E}_{h}\!\left[r^{3}\ln r\right] &=& (mn)_{3}\psi_{0}(mn+3) \\
\fl\mathbb{E}_{h}\!\left[r^{3}\ln^{2}r\right] &=& (mn)_{3}\left(\psi_{0}^{2}(mn+3)+\psi_{1}(mn+3)\right) \\
\fl\mathbb{E}_{h}\!\left[r^{3}\ln^{3}r\right] &=& (mn)_{3}\left(\psi_{0}^{3}(mn+3)+3\psi_{0}(mn+3)\psi_{1}(mn+3)+\psi_{2}(mn+3)\right).\end{aligned}$$ The first and second moment of $S$ in (\[eq:mST\]) are also known, cf. (\[eq:1km\]), (\[eq:2mk\]), (\[eq:k1\]), and (\[eq:k2\]), $$\begin{aligned}
\mathbb{E}_{f}\!\left[S\right] &=& \kappa_{1} \label{eq:S1} \\
\mathbb{E}_{f}\!\left[S^{2}\right] &=& \kappa_{2}+\kappa_{1}^{2}. \label{eq:S2}\end{aligned}$$ Therefore, to obtain $\mathbb{E}_{f}\!\left[S^{3}\right]$ the remaining term to compute in (\[eq:mST\]) is $\mathbb{E}_{g}\!\left[T^{3}\right]$. Since $$\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!T^{3}=\sum_{i=1}^{m}\theta_{i}^{3}\ln^{3}\theta_{i}+3\sum_{1\leq i\neq j\leq m}\theta_{i}^{2}\theta_{j}\ln^{2}\theta_{i}\ln\theta_{j}+6\sum_{1\leq i\neq j\neq k\leq m}\theta_{i}\theta_{j}\theta_{k}\ln\theta_{i}\ln\theta_{j}\ln\theta_{k},$$ the computation of $\mathbb{E}_{g}\!\left[T^{3}\right]$ involves one, two, and three arbitrary eigenvalue densities, denoted respectively by $g_{1}(x_{1})$, $g_{2}(x_{1},x_{2})$, and $g_{3}(x_{1},x_{2},x_{3})$, of the Wishart-Laguerre ensemble (\[eq:we\]) as $$\begin{aligned}
\fl\mathbb{E}_{g}\!\left[T^{3}\right] &=& {m\choose1}\int_{0}^{\infty}\!\!x_{1}^{3}\ln^{3}x_{1}~g_{1}(x_{1}){\,\mathrm{d}}x_{1}+3{m\choose1}{m-1\choose1}\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!x_{1}^{2}x_{2}\ln^{2}x_{1}\ln x_{2}\times \\
\fl&&\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!g_{2}\left(x_{1},x_{2}\right){\,\mathrm{d}}x_{1}{\,\mathrm{d}}x_{2}+6{m\choose3}\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!x_{1}x_{2}x_{3}\ln x_{1}\ln x_{2}\ln x_{3}~g_{3}\left(x_{1},x_{2},x_{2}\right){\,\mathrm{d}}x_{1}{\,\mathrm{d}}x_{2}{\,\mathrm{d}}x_{3}.\end{aligned}$$ It is a well-known result in random matrix theory that the joint density $g_{N}(x_{1},\dots,x_{N})$ of $N$ (out of $m$) arbitrary eigenvalues of various matrix models, including the Wishart-Laguerre ensemble, can be written in terms of a determinant of a correlation kernel $K(x_{i},x_{j})$ as [@Forrester] $$\label{eq:Nei}
g_{N}(x_{1},\dots,x_{N})=\frac{(m-N)!}{m!}\det\left(K\left(x_{i},x_{j}\right)\right)_{i,j=1}^{N}.$$ The determinant in (\[eq:Nei\]) is known as the $N$-point correlation function, where the symmetric correlation kernel $K(x_{i},x_{j})$ uniquely specifies the random matrix ensemble. As a result of (\[eq:Nei\]), the arbitrary eigenvalue densities needed to compute $\mathbb{E}_{g}\!\left[T^{3}\right]$ are $$\begin{aligned}
\fl g_{1}(x_{1}) &\!\!\!\!=& \frac{1}{m}K\left(x_{1},x_{1}\right) \label{eq:g1} \\
\fl g_{2}(x_{1},x_{2}) &\!\!\!\!=& \frac{1}{m(m-1)}\left(K(x_{1},x_{1})K(x_{2},x_{2})-K^{2}(x_{1},x_{2})\right) \\
\fl g_{3}(x_{1},x_{2},x_{3}) &\!\!\!\!=& \frac{1}{m(m-1)(m-2)}(K(x_{1},x_{1})K(x_{2},x_{2})K(x_{3},x_{3})+K(x_{1},x_{2})K(x_{2},x_{3})\times\nonumber\\ &&K(x_{3},x_{1})+K(x_{1},x_{3})K(x_{3},x_{2})K(x_{2},x_{1})-K(x_{1},x_{2})K(x_{2},x_{1})K(x_{3},x_{3})\nonumber \\
&&-K(x_{1},x_{3})K(x_{3},x_{1})K(x_{2},x_{2})-K(x_{2},x_{3})K(x_{3},x_{2})K(x_{1},x_{1})),\end{aligned}$$ and consequently the third moment of $T$ can be written as $$\begin{aligned}
\label{eq:T3}
\fl\mathbb{E}_{g}\!\left[T^{3}\right]&=&\left(\int_{0}^{\infty}\!\!x\ln{x}~K(x,x){\,\mathrm{d}}{x}\right)^{3}+3\int_{0}^{\infty}\!\!x\ln{x}~K(x,x){\,\mathrm{d}}{x}\int_{0}^{\infty}\!\!x^{2}\ln^{2}{x}~K(x,x){\,\mathrm{d}}{x}+ \nonumber \\
\fl&&\int_{0}^{\infty}\!\!x^{3}\ln^{3}{x}~K(x,x){\,\mathrm{d}}{x}-3\int_{0}^{\infty}\!\!x\ln{x}~K(x,x){\,\mathrm{d}}{x}\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!xy\ln x\ln y\times\nonumber \\
\fl&&K^{2}(x,y){\,\mathrm{d}}x{\,\mathrm{d}}y-3\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!x^{2}y\ln^{2}x\ln y~K^{2}(x,y){\,\mathrm{d}}x{\,\mathrm{d}}y+2\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!xyz\times \nonumber \\
\fl&& \ln x\ln y\ln z~K(x,y)K(y,z)K(z,x){\,\mathrm{d}}x{\,\mathrm{d}}y{\,\mathrm{d}}z.\end{aligned}$$
We now turn to the third cumulant of the induced entropy $T$, which will result in a more compact expression (\[eq:T3ABC\]) than that of the corresponding moment (\[eq:T3\]). Similarly as the third moment (\[eq:T3\]), the first two moments of $T$ can also be represented as integrals involving the correlation kernel as $$\begin{aligned}
\mathbb{E}_{g}\!\left[T\right]&=&\int_{0}^{\infty}\!\!x\ln{x}~K(x,x){\,\mathrm{d}}{x} \label{eq:T1} \\
\mathbb{E}_{g}\!\left[T^{2}\right]&=&\left(\int_{0}^{\infty}\!\!x\ln{x}~K(x,x){\,\mathrm{d}}{x}\right)^{2}+\int_{0}^{\infty}\!\!x^{2}\ln^{2}x~K(x,x){\,\mathrm{d}}{x}-\nonumber\\
&&\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!xy\ln x\ln y~K^{2}(x,y){\,\mathrm{d}}x{\,\mathrm{d}}y. \label{eq:T2}\end{aligned}$$ The above integrals have been computed in [@Ruiz95] and [@Wei17], respectively, as $$\begin{aligned}
\mathbb{E}_{g}\!\left[T\right]&=&mn\psi_{0}(n)+\frac{1}{2}m(m+1) \label{eq:T1R} \\
\mathbb{E}_{g}\!\left[T^{2}\right]&=&mn(m+n)\psi_{1}(n)+mn(mn+1)\psi_{0}^{2}(n)+m\big(m^{2}n+mn+ \nonumber\\
&&m+2n+1\big)\psi_{0}(n)+\frac{1}{4}m(m+1)\left(m^2+m+2\right). \label{eq:T2R}\end{aligned}$$ Inserting the first three moments (\[eq:T1\]), (\[eq:T2\]), and (\[eq:T3\]) into the moment-to-cumulant relation (\[eq:3mk\]), the third cumulant of $T$ is simplified to $$\begin{aligned}
\kappa_{3}^{T}&=&\mathbb{E}_{g}\!\left[T^{3}\right]-3\mathbb{E}_{g}\!\left[T^{2}\right]\mathbb{E}_{g}\!\left[T\right]+2\mathbb{E}_{g}^{3}\!\left[T\right]\\
&=&I_{A}-3I_{B}+2I_{C}, \label{eq:T3ABC}\end{aligned}$$ where we denote the integrals $$\begin{aligned}
I_{A} &=& \int_{0}^{\infty}\!\!x^{3}\ln^{3}{x}~K(x,x){\,\mathrm{d}}{x} \label{eq:IA} \\
I_{B} &=& \int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!x^{2}y\ln^{2}x\ln y~K^{2}(x,y){\,\mathrm{d}}x{\,\mathrm{d}}y \label{eq:IB} \\
I_{C} &=& \int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!\int_{0}^{\infty}\!\!xyz\ln x\ln y\ln z~K(x,y)K(y,z)K(z,x){\,\mathrm{d}}x{\,\mathrm{d}}y{\,\mathrm{d}}z. \label{eq:IC}\end{aligned}$$ By the relations (\[eq:1km\])-(\[eq:3mk\]) and the existing results (\[eq:T1R\]), (\[eq:T2R\]), it remains to compute the integrals $I_{A}$, $I_{B}$, and $I_{C}$ in order to obtain the final result (\[eq:k3\]). The prior results discussed so far also allow us to express the desired third cumulant of $S$ in terms of the first three cumulants of $T$ as\
$$\begin{aligned}
\label{eq:k3T3}
\kappa_{3}&=&\frac{1}{(mn)_{3}}\Bigg(\!-\kappa_{3}^{T}+\frac{6}{mn}\kappa_{1}^{T}\kappa_{2}^{T}+\frac{3(2mn+3)}{mn+1}\kappa_{2}^{T}-\frac{4}{m^{2}n^{2}}\left(\kappa_{1}^{T}\right)^{3}-\nonumber\\
&&\frac{3(3mn+4)}{mn(mn+1)}\left(\kappa_{1}^{T}\right)^{2}-\frac{6(mn+2)}{mn+1}\kappa_{1}^{T}\Bigg)+\psi_{2}(mn+1),\end{aligned}$$ where (\[eq:T1R\]) and (\[eq:T2R\]) directly lead to $$\begin{aligned}
\fl\kappa_{1}^{T}&=&mn\psi_{0}(n)+\frac{1}{2}m(m+1) \label{eq:kT1} \\
\fl\kappa_{2}^{T}&=&mn(m+n)\psi_{1}(n)+mn\psi_{0}^{2}(n)+m\left(m+2n+1\right)\psi_{0}(n)+\frac{1}{2}m(m+1), \label{eq:kT2}\end{aligned}$$ and we have also utilized the identities below, cf. (\[eq:pg\]), in simplifying (\[eq:k3T3\]), $$\begin{aligned}
\psi_{0}(l+n)&=&\psi_{0}(l)+\sum_{k=0}^{n-1}\frac{1}{l+k} \label{eq:0s} \\
\psi_{1}(l+n)&=&\psi_{1}(l)-\sum_{k=0}^{n-1}\frac{1}{(l+k)^2} \label{eq:1s} \\
\psi_{2}(l+n)&=&\psi_{2}(l)+2\sum_{k=0}^{n-1}\frac{1}{(l+k)^3}. \label{eq:2s}\end{aligned}$$ The rest of the paper is to compute the integrals $I_{A}$, $I_{B}$, $I_{C}$, where we will eventually show that $$\begin{aligned}
\kappa_{3}^{T}&=&I_{A}-3I_{B}+2I_{C} \label{eq:ABC} \\
&=&mn\left(m^2+3mn+n^2+1\right)\psi_{2}(n)+6mn(m+n)\psi_{0}(n)\psi_{1}(n)+\nonumber\\
&&m\left(2m^2+12mn+3m+6n^{2}+3n+1\right)\psi_{1}(n)+2mn\psi_{0}^{3}(n)+\nonumber\\
&&3m(m+3n+1)\psi_{0}^{2}(n)+6m(m+n+1)\psi_{0}(n)+m(m+1). \label{eq:T3R}\end{aligned}$$ Inserting the above expression as well as (\[eq:kT1\]) and (\[eq:kT2\]) into (\[eq:k3T3\]), the claimed main result (\[eq:k3\]) will then be established.
Calculations of integrals $I_{A}$, $I_{B}$, and $I_{C}$ {#sec:IABC}
-------------------------------------------------------
To compute the remaining integrals (\[eq:IA\]), (\[eq:IB\]), and (\[eq:IC\]), we need the following results on the Wishart-Laguerre ensemble. First, its correlation kernel can be explicitly written as [@Forrester] $$\label{eq:ker}
K(x_{i},x_{j})=\sqrt{\e^{-x_{i}-x_{j}}(x_{i}x_{j})^{n-m}}\sum_{k=0}^{m-1}\frac{C_{k}(x_{i})C_{k}(x_{j})}{k!(n-m+k)!},$$ where $$C_{k}(x)=(-1)^{k}k!L_{k}^{(n-m)}(x)$$ with $$\label{eq:Lar}
L_{k}^{(n-m)}(x)=\sum_{i=0}^{k}(-1)^{i}{n-m+k\choose k-i}\frac{x^i}{i!}$$ being the (generalized) Laguerre polynomial of degree $k$. Similarly as the Hermite polynomials (\[eq:ocH\]), the Laguerre polynomials also satisfy an orthogonality relation [@Forrester] $$\label{eq:oc}
\int_{0}^{\infty}\!\!x^{n-m}\e^{-x}L_{k}^{(n-m)}(x)L_{l}^{(n-m)}(x){\,\mathrm{d}}{x}=\frac{(n-m+k)!}{k!}\delta_{kl}.$$ Instead of the summation form (\[eq:ker\]), the one arbitrary eigenvalue density (\[eq:g1\]) admits a more convenient form by the Christoffel-Darboux formula [@Ruiz95; @Forrester] as $$\label{eq:one}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!g_{1}(x)=\frac{(m-1)!}{(n-1)!}x^{n-m}\e^{-x}\left(\left(L_{m-1}^{(n-m+1)}(x)\right)^{2}-L_{m-2}^{(n-m+1)}(x)L_{m}^{(n-m+1)}(x)\right).$$ We also need the following integral, due to Schr[ö]{}dinger [@Schrodinger1926], that generalizes the identity (\[eq:oc\]) to $$\label{eq:Swm}
\fl\int_{0}^{\infty}\!\!x^{q}\e^{-x}L_{s}^{(\alpha)}(x)L_{t}^{(\beta)}(x){\,\mathrm{d}}{x}=(-1)^{s+t}\sum_{k=0}^{\min(s,t)}{q-\alpha\choose s-k}{q-\beta\choose t-k}\frac{\Gamma(q+1+k)}{k!}.$$ By taking up to the third derivatives of the Schr[ö]{}dinger’s integral (\[eq:Swm\]) with respect to $q$, we obtain three more useful integrals shown in (\[eq:1d\]) (see also [@Ruiz95]), (\[eq:2d\]), and (\[eq:3d\]) below, denoted respectively by $A_{s,t}^{(\alpha,\beta)}(q)$, $B_{s,t}^{(\alpha,\beta)}(q)$, and $C_{s,t}^{(\alpha,\beta)}(q)$, where we have also denoted $$\begin{aligned}
\Psi_{j}&=&\psi_{j}(q+1+k)+\psi_{j}(q-\alpha+1)+\psi_{j}(q-\beta+1)- \nonumber \\
&&\psi_{j}(q-\alpha-s+1+k)-\psi_{j}(q-\beta-t+1+k).\end{aligned}$$ Note that the explicit kernel expression (\[eq:ker\]) together with the corresponding Schr[ö]{}dinger’s integral (\[eq:Swm\]) (and its higher order derivatives) of the Wishart-Laguerre ensemble, unavailable for the fixed-trace ensemble, makes the subsequent calculation possible. This fact is the motivation behind the approach of moments (cumulants) conversion (\[eq:mST\]) between the two ensembles.
$$\begin{aligned}
\fl A_{s,t}^{(\alpha,\beta)}(q)&=&\int_{0}^{\infty}\!\!x^{q}\e^{-x}\ln{x}~L_{s}^{(\alpha)}(x)L_{t}^{(\beta)}(x){\,\mathrm{d}}{x} \label{eq:i1d}\\
\fl&=&(-1)^{s+t}\sum_{k=0}^{\min(s,t)}{q-\alpha\choose s-k}{q-\beta\choose t-k}\frac{\Gamma(q+1+k)}{k!}\Psi_{0}.\label{eq:1d}\end{aligned}$$
$$\begin{aligned}
\fl B_{s,t}^{(\alpha,\beta)}(q)&=&\int_{0}^{\infty}\!\!x^{q}\e^{-x}\ln^{2}{x}~L_{s}^{(\alpha)}(x)L_{t}^{(\beta)}(x){\,\mathrm{d}}{x} \label{eq:i2d}\\
\fl&=&(-1)^{s+t}\sum_{k=0}^{\min(s,t)}{q-\alpha\choose s-k}{q-\beta\choose t-k}\frac{\Gamma(q+1+k)}{k!}\left(\Psi_{0}^{2}+\Psi_{1}\right).\label{eq:2d}\end{aligned}$$
$$\begin{aligned}
\fl C_{s,t}^{(\alpha,\beta)}(q)&=&\int_{0}^{\infty}\!\!x^{q}\e^{-x}\ln^{3}{x}~L_{s}^{(\alpha)}(x)L_{t}^{(\beta)}(x){\,\mathrm{d}}{x}\label{eq:i3d}\\
\fl&=&(-1)^{s+t}\sum_{k=0}^{\min(s,t)}{q-\alpha\choose s-k}{q-\beta\choose t-k}\frac{\Gamma(q+1+k)}{k!}\left(\Psi_{0}^{3}+3\Psi_{0}\Psi_{1}+\Psi_{2}\right).\label{eq:3d}\end{aligned}$$
### Computing $I_{A}$
Inserting the one arbitrary eigenvalue density (\[eq:one\]) (see also (\[eq:g1\])) into (\[eq:IA\]), $I_{A}$ is expressed via the integral (\[eq:i3d\]) as $$\label{eq:IAI}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!I_{A}=\frac{m!}{(n-1)!}\left(C_{m-1,m-1}^{(n-m+1,n-m+1)}(n-m+3)-C_{m-2,m}^{(n-m+1,n-m+1)}(n-m+3)\right),$$ where the integrals in (\[eq:IAI\]) are evaluated by invoking the identity (\[eq:3d\]) and then collecting the non-zero contributions. In particular, the non-zero contributions consist of indeterminate terms as a result of zero or negative arguments of gamma and polygamma functions. These indeterminacy can be resolved by interpreting the gamma and polygamma functions involved as the limits $\epsilon\to0$ of (with $l\geq0$) \[eq:pgna\]$$\begin{aligned}
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\Gamma(-l+\epsilon)&=&\frac{(-1)^{l}}{l!\epsilon}\left(1+\psi_{0}(l+1)\epsilon+o\left(\epsilon^2\right)\right)\label{eq:pgna1}\\
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\psi_{0}(-l+\epsilon)&=&-\frac{1}{\epsilon}+\psi_{0}(l+1)+\left(2\psi_{1}(1)-\psi_{1}(l+1)\right)\epsilon+\frac{1}{2}\psi_{2}(l+1)\epsilon^{2}+o\left(\epsilon^3\right)\label{eq:pgna2}\\
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\psi_{1}(-l+\epsilon)&=&\frac{1}{\epsilon^{2}}-\psi_{1}(l+1)+\psi_{1}(1)+\zeta(2)+o\left(\epsilon\right)\label{eq:pgna3}\\
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\psi_{2}(-l+\epsilon)&=&-\frac{2}{\epsilon^{3}}+\psi_{2}(l+1)+\psi_{2}(1)+2\zeta(3)+o\left(\epsilon\right).\label{eq:pgna4}\end{aligned}$$ By the above procedure, the integrals $C_{m-1,m-1}^{(n-m+1,n-m+1)}(n-m+3)$ and $C_{m-2,m}^{(n-m+1,n-m+1)}(n-m+3)$ are evaluated, respectively, as $$\begin{aligned}
\fl&&C_{m-1,m-1}^{(n-m+1,n-m+1)}(n-m+3)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+3}\e^{-x}\ln^{3}{x}~\left(L_{m-1}^{(n-m+1)}(x)\right)^{2}{\,\mathrm{d}}{x}\\
\fl=&&\frac{(n-1)!}{2(m-1)!}\Big(18m^{2}n+39m^2-30mn-57m+12n+30+3\big(13m^{2}n+12m^2+4mn^2-\nonumber\\
\fl&&3mn-4m-4n^2+2n+4\big)\psi_{0}(n)+6\left(3m^{2}n+m^2+4mn^2+3mn+m-n^2\right)\times\nonumber\\
\fl&&\left(\psi_{0}^{2}(n)+\psi_{1}(n)\right)+2n\left(m^2+4mn+m+n^2-n\right)\!\left(\psi_{0}^{3}(n)+3\psi_{0}(n)\psi_{1}(n)+\psi_{2}(n)\right)\!\Big)+\nonumber\\
\fl&&\sum_{k=1}^{m-3}\frac{6(n-k)!}{(m-3-k)!}\left(\frac{3}{k+2}-\frac{3}{k}+\frac{1}{k^2}+\frac{4}{(k+1)^2}+\frac{1}{(k+2)^2}\right)\big(\psi_{0}(n+1-k)-\nonumber\\
\fl&&2\psi_{0}(k)+2\psi_{0}(1)+3\big),\label{eq:IAS1}\end{aligned}$$ and $$\begin{aligned}
\fl&&C_{m-2,m}^{(n-m+1,n-m+1)}(n-m+3)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+3}\e^{-x}\ln^{3}{x}~L_{m-2}^{(n-m+1)}(x)L_{m}^{(n-m+1)}(x){\,\mathrm{d}}{x}\\
\fl=&&\frac{(n-1)!}{8(m-2)!}\Big(40mn-m^2+69m-32n-38+8\big(8mn-m^2+9m+3n^2+5n+1)\times\nonumber\\
\fl&&\psi_{0}(n)+2\left(8mn-m^2+5m+18n^2+26n+6\right)\left(\psi_{0}^{2}(n)+\psi_{1}(n)\right)+8n(n+1)\times\nonumber\\
\fl&&\left(\psi_{0}^{3}(n)+3\psi_{0}(n)\psi_{1}(n)+\psi_{2}(n)\right)+\sum_{k=1}^{m-4}\frac{(n-k-1)!}{(m-k-4)!}\Bigg(\frac{1}{2k}-\frac{4}{k+1}+\frac{4}{k+3}-\nonumber\\
\fl&&\frac{1}{2(k+4)}+\frac{6}{(k+2)^2}\Bigg)\left(\psi_{0}(n-k)-\psi_{0}(k+2)-\psi_{0}(k)+2\psi_{0}(1)+3\right).\label{eq:IAS2}\end{aligned}$$ Inserting (\[eq:IAS1\]) and (\[eq:IAS2\]) back into (\[eq:IAI\]), we arrive at a finite summation form of $I_{A}$.
### Computing $I_{B}$
Similarly, inserting the kernel expression (\[eq:ker\]) into (\[eq:IB\]), $I_{B}$ is expressed via the integrals (\[eq:i1d\]) and (\[eq:i2d\]) as $$\begin{aligned}
\label{eq:IBI}
\fl I_{B}&=&\sum_{k=0}^{m-1}\frac{k!^{2}A_{k,k}^{(n-m,n-m)}(n-m+1)B_{k,k}^{(n-m,n-m)}(n-m+2)}{(k+n-m)!^{2}}+\nonumber\\
\fl&&\sum_{k=0}^{m-2}\sum_{j=0}^{m-k-2}\frac{2j!(k+j+1)!A_{j,k+j+1}^{(n-m,n-m)}(n-m+1)B_{j,k+j+1}^{(n-m,n-m)}(n-m+2)}{(j+n-m)!(k+j+n-m+1)!}.\end{aligned}$$ By the identities (\[eq:1d\]), (\[eq:2d\]) and with the help of $(106)$, the integrals $A_{k,k}^{(n-m,n-m)}(n-m+1)$ and $B_{k,k}^{(n-m,n-m)}(n-m+2)$ in (\[eq:IBI\]) are calculated as $$\begin{aligned}
\fl&&A_{k,k}^{(n-m,n-m)}(n-m+1)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+1}\e^{-x}\ln{x}~\left(L_{k}^{(n-m)}(x)\right)^{2}{\,\mathrm{d}}{x} \\
\fl=&&\frac{(k+n-m)!}{k!}\left((2k-m+n+1)\psi_{0}(k+n-m+1)+2k+1\right),\label{eq:IBS1}\end{aligned}$$
$$\begin{aligned}
\fl&&B_{k,k}^{(n-m,n-m)}(n-m+2)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+2}\e^{-x}\ln^{2}{x}~\left(L_{k}^{(n-m)}(x)\right)^{2}{\,\mathrm{d}}{x} \\
\fl=&&\frac{(k+n-m)!}{k!}\Bigg(\frac{1}{2}\left(17k^2+k(4n-4m+7)+4\right)+2\big(7k^2+k(4n-4m+7)+2n- \nonumber\\
\fl&&2m+3\big)\psi_{0}(k+n-m+1)+\left(6k^2+6k(n-m+1)+(n-m+2)(n-m+1)\right)\times \nonumber\\
\fl&&\left(\psi_{0}^{2}(k+n-m+1)+\psi_{1}(k+n-m+1)\right)+\sum_{i=3}^{k}\frac{2(k-i+n-m+2)!}{(k-i)!}\times \nonumber\\
\fl&&\left(\frac{3}{i}-\frac{3}{i-2}+\frac{1}{i^2}+\frac{4}{(i-1)^2}+\frac{1}{(i-2)^2}\right)\!\Bigg),\label{eq:IBS2}\end{aligned}$$
and the integrals $A_{j,k+j+1}^{(n-m,n-m)}(n-m+1)$ and $B_{j,k+j+1}^{(n-m,n-m)}(n-m+2)$ admit different expressions for different ranges of $k$ as $$\begin{aligned}
\fl&&A_{j,j+1}^{(n-m,n-m)}(n-m+1)\\
\fl&=&\int_{0}^{\infty}\!\!x^{n-m+1}\e^{-x}\ln{x}~L_{j}^{(n-m)}(x)L_{j+1}^{(n-m)}(x){\,\mathrm{d}}{x} \\
\fl&=&-\frac{(j+n-m)!}{j!}\left((j+n-m+1)\psi_{0}(j+n-m+1)+\frac{3j}{2}+n-m+2\right),\label{eq:IBS3}\end{aligned}$$
$$\begin{aligned}
\fl&&A_{j,k+j+1}^{(n-m,n-m)}(n-m+1)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+1}\e^{-x}\ln{x}~L_{j}^{(n-m)}(x)L_{k+j+1}^{(n-m)}(x){\,\mathrm{d}}{x} \\
\fl=&&\frac{(j+n-m)!}{j!(k+1)}\left(\frac{j+n-m+1}{k}-\frac{j}{k+2}\right),~~~~~~k>0,\label{eq:IBS4}\end{aligned}$$
$$\begin{aligned}
\fl&&B_{j,j+1}^{(n-m,n-m)}(n-m+2)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+2}\e^{-x}\ln^{2}{x}~L_{j}^{(n-m)}(x)L_{j+1}^{(n-m)}(x){\,\mathrm{d}}{x} \\
\fl=&&-\frac{2(j+n-m)!}{j!}\Bigg(\frac{7j^{2}}{2}+\frac{5j}{2}(n-m+3)+2n-2m+5+\Bigg(\frac{16j^{2}}{3}+j\Bigg(6n-6m+\nonumber\\
\fl&&\frac{38}{3}\Bigg)+(n-m)^2+7(n-m)+8\Bigg)\psi_{0}(j+n-m+1)+(j+n-m+1)(2j+n-\nonumber\\
\fl&&m+2)\left(\psi_{0}^{2}(j+n-m+1)+\psi_{1}(j+n-m+1)\right)\!\Bigg)+\sum_{l=3}^{j}\frac{2(j-l+n-m+2)!}{(j-l)!}\times\nonumber\\
\fl&&\left(-\frac{1}{3(l+1)}-\frac{3}{l}+\frac{3}{l-1}+\frac{1}{3(l-2)}-\frac{2}{l^2}-\frac{2}{(l-1)^2}\right),\label{eq:IBS5}\end{aligned}$$
$$\begin{aligned}
\fl&&B_{j,j+2}^{(n-m,n-m)}(n-m+2)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+2}\e^{-x}\ln^{2}{x}~L_{j}^{(n-m)}(x)L_{j+2}^{(n-m)}(x){\,\mathrm{d}}{x} \\
\fl=&&\frac{(j+n-m)!}{j!}\Bigg(\frac{10j^2}{3}+\frac{2j}{3}(7n-7m+20)+(n-m)^2+9(n-m)+13+\frac{1}{6}\big(25j^2+\nonumber\\
\fl&&j(44n-44m+87)+6(n-m+3)(3n-3m+4)\big)\psi_{0}(j+n-m+1)+(j+n-\nonumber\\
\fl&&m+1)(j+n-m+2)\left(\psi_{0}^{2}(j+n-m+1)+\psi_{1}(j+n-m+1)\right)\!\Bigg)+\nonumber\\
\fl&&\sum_{l=3}^{j}\!\frac{4(j-l+n-m+2)!}{(j-l)!}\left(\frac{1}{3(l+1)}-\frac{1}{24(l+2)}-\frac{1}{3(l-1)}+\frac{1}{24(l-2)}+\frac{1}{2l^2}\right)\!,\label{eq:IBS6}\end{aligned}$$
$$\begin{aligned}
\fl&&B_{j,j+k+1}^{(n-m,n-m)}(n-m+2)\\
\fl=&&\int_{0}^{\infty}\!\!x^{n-m+2}\e^{-x}\ln^{2}{x}~L_{j}^{(n-m)}(x)L_{j+k+1}^{(n-m)}(x){\,\mathrm{d}}{x} \\
\fl=&&-\frac{2(j+n-m)!}{j!(k-1)_{5}}\Bigg(j^{2}\left(k^2-7k+24\psi_{0}(1)+56\right)+j\big(k^{2}(2m-2n-5)+k(18n-\nonumber\\
\fl&&18m+55+12(n-m+2)\psi_{0}(1))+4(3(3n-3m+4)\psi_{0}(1)+18n-18n+31)\big)+\nonumber\\
\fl&&(k+2)(k+3)\big(3m^2-6mn-13m+3n^2+13n+12+2(n-m+1)(n-m+2)\times\nonumber\\
\fl&&\psi_{0}(1)\big)+2\big(12j^2+6j(k(n-m+2)+3n-3m+4)+(k+2)(k+3)(n-m+1)\times\nonumber\\
\fl&&(n-m+2)\big)\left(\psi_{0}(j+n-m+1)-\psi_{0}(k-1)\right)\!\Bigg)+\sum_{l=3}^{j}\frac{8(j-l+n-m+2)!}{(j-l)!}\times\nonumber\\
\fl&&\frac{1}{(l-2)(l-1)l(k+l-1)(k+l)(k+l+1)},~~~~~~k>1.\label{eq:IBS7}\end{aligned}$$
Inserting (\[eq:IBS1\]), (\[eq:IBS2\]), (\[eq:IBS3\]), (\[eq:IBS4\]), (\[eq:IBS5\]), (\[eq:IBS6\]), and (\[eq:IBS7\]) back into (\[eq:IBI\]), we obtain a finite summation representation of $I_{B}$.
### Computing $I_{C}$
In the same manner, by inserting (\[eq:ker\]) into (\[eq:IC\]), $I_{C}$ can be expressed in terms of the integral (\[eq:i1d\]) as $$\begin{aligned}
\label{eq:ICI}
\fl
I_{C}&=&\sum_{k=0}^{m-1}\frac{k!^{3}\left(A_{k,k}^{(n-m,n-m)}(n-m+1)\right)^{3}}{(k+n-m)!^{3}}+\sum_{j=0}^{m-1}\sum_{i=j+1}^{m-1}\frac{3i!j!\left(A_{i,j}^{(n-m,n-m)}(n-m+1)\right)^{2}}{(i+n-m)!(j+n-m)!}\times\nonumber\\
\fl&&\left(\frac{i!A_{i,i}^{(n-m,n-m)}(n-m+1)}{(i+n-m)!}+\frac{j!A_{j,j}^{(n-m,n-m)}(n-m+1)}{(j+n-m)!}\right)+\sum_{k=0}^{m-1}\sum_{j=k+1}^{m-1}\sum_{i=j+1}^{m-1}6\times\nonumber\\
\fl&&\frac{i!j!k!A_{i,j}^{(n-m,n-m)}(n-m+1)A_{j,k}^{(n-m,n-m)}(n-m+1)A_{k,i}^{(n-m,n-m)}(n-m+1)}{(i+n-m)!(j+n-m)!(k+n-m)!}.\end{aligned}$$ The integrals in the above expression have been computed in (\[eq:IBS1\]), (\[eq:IBS3\]), and (\[eq:IBS4\]), inserting which back into (\[eq:ICI\]) leads to a finite summation form of $I_{C}$.
Evaluation of summations in $I_{A}$, $I_{B}$, and $I_{C}$ {#sec:simp}
---------------------------------------------------------
The rest of the task is to evaluate the summations in the finite summation forms of $I_{A}$, $I_{B}$, and $I_{C}$, after inserting back the computed integrals (\[eq:IAS1\]), (\[eq:IAS2\]), (\[eq:IBS1\]), (\[eq:IBS2\]), (\[eq:IBS3\]), (\[eq:IBS4\]), (\[eq:IBS5\]), (\[eq:IBS6\]), (\[eq:IBS7\]). The evaluation of the summations amounts to applying the closed-form identities listed in \[sec:ap1-c\] and \[sec:ap2-c\] as well as the semi closed-form identities listed in \[sec:ap1-sc\] and \[sec:ap2-sc\]. The derivation of these identities is discussed in \[sec:ap1-re\] and \[sec:ap2-re\].
Evaluating the summations is a tedious but straightforward procedure, which eventually leads to $I_{A}$, $I_{B}$, and $I_{C}$ being expressed respectively by (\[eq:IAf\]), (\[eq:IBf\]), and (\[eq:ICf\]) as shown below. The coefficients $a_{i}$, $b_{i}$, and $c_{i}$ in the expressions (\[eq:IAf\]), (\[eq:IBf\]), and (\[eq:ICf\]) can be found in Table \[t:IAc\], Table \[t:IBc\], and Table \[t:ICc\], respectively. As seen from (\[eq:IAf\]), (\[eq:IBf\]), and (\[eq:ICf\]), we have shifted the argument of each polygamma function in $I_{A}$, $I_{B}$, and $I_{C}$ to one of following: $1$, $m$, $n$, $n-m$, $n-m+k$, with the help of (89). This choice leads the coefficients in Table \[t:IAc\], Table \[t:IBc\], and Table \[t:ICc\], to become polynomials in $m$ and $n$.
$$\begin{aligned}
\fl I_{A}&=&a_{1}+a_{2}\psi_{0}(n)+a_{3}\psi_{0}(n-m)+a_{4}\psi_{0}(1)\psi_{0}(n-m)+a_{5}\psi_{0}(m)\psi_{0}(n-m)+\nonumber\\
\fl&&a_{6}\psi_{0}(n)\psi_{0}(n-m)+a_{7}\psi_{0}^{2}(n-m)+a_{8}\psi_{0}(1)\psi_{0}(n)\psi_{0}(n-m)+\nonumber\\
\fl&&a_{9}\psi_{0}(1)\psi_{0}^{2}(n-m)+a_{10}\psi_{0}(m)\psi_{0}(n)\psi_{0}(n-m)+a_{11}\psi_{0}(m)\psi_{0}^{2}(n-m)+\nonumber\\
\fl&&a_{12}\psi_{0}(n)\psi_{0}^{2}(n-m)+a_{13}\psi_{0}^{3}(n-m)+a_{14}\psi_{1}(n-m)+a_{15}\psi_{0}(n)\psi_{1}(n-m)+\nonumber\\
\fl&&a_{16}\psi_{2}(n-m)+a_{17}\sum_{k=1}^m\frac{\psi_{0}(k+n-m)}{k}+a_{18}\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+n-m)}{k}.\label{eq:IAf}\end{aligned}$$
[cl]{}
\
& $\frac{m}{288}\big(37 m^3+4012 m^2 n-30 m^2+4410 m n^2-330 m n-169 m+84 n^3-$\
& $30 n^2-250 n+162\big)$\
& $-\frac{n}{24}\big(12 m^3+414 m^2 n-6 m^2+364 m n^2+6 m n-94 m+7 n^3-6 n^2-$\
& $67 n+90\big)$\
& $-\frac{1}{24}\big(7 m^4+352 m^3 n+18 m^3+336 m^2 n+5 m^2-352 m n^3+360 m n^2+$\
& $216 m n+42 m-7 n^4+6 n^3+139 n^2+222 n+72\big)$\
$a_{4}=$ & $\frac{m}{2}\left(m^3+28 m^2 n+6 m^2+30 m n^2+18 m n+11 m+6 n^2+26 n+6\right)$\
$a_{5}=$ & $-\frac{m}{2}\left(m^3+28 m^2 n+6 m^2+30 m n^2+18 m n+11 m+6 n^2+26 n+6\right)$\
$a_{6}=$ & $\frac{n}{2}\left(30 m^2 n-18 m^2+28 m n^2-54 m n+26 m+n^3-18 n^2+11 n-18\right)$\
& $\frac{1}{4}\big(m^4+28 m^3 n+6 m^3-30 m^2 n^2+6 m^2 n+11 m^2-56 m n^3-$\
& $30 m n^2-26 m n+6 m-2 n^4-12 n^3-22 n^2-12 n\big)$\
$a_{8}=$ & $6 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{9}=$ & $6 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{10}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{11}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{12}=$ & $3 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{13}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{14}=$ & $\frac{m}{4}\left(m^3+28 m^2 n+6 m^2+30 m n^2+18 m n+11 m+6 n^2+26 n+6\right)$\
$a_{15}=$ & $3 m n \left(m^2+3 m n+n^2+1\right)$\
$a_{16}=$ & $m n \left(m^2+3 m n+n^2+1\right)$\
& $\frac{m}{2}\big(m^3+28 m^2 n+6 m^2+30 m n^2+18 m n+11 m+6 n^2+26 n+6+$\
& $12n\left(m^2+3 m n+n^2+1\right)\psi_{0}(n)+24n\left(m^2+3 m n+n^2+1\right)\psi_{0}(n-m)\big)$\
$a_{18}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
\
\[t:IAc\]
$$\begin{aligned}
\fl I_{B}&=&b_{1}+b_{2}\psi_{0}(n)+b_{3}\psi_{0}(n-m)+b_{4}\psi_{0}(1)\psi_{0}(n-m)+b_{5}\psi_{0}(m)\psi_{0}(n-m)+\nonumber\\
\fl&&b_{6}\psi_{0}^{2}(n)+b_{7}\psi_{0}(n)\psi_{0}(n-m)+b_{8}\psi_{0}^{2}(n-m)+b_{9}\psi_{0}(1)\psi_{0}(n)\psi_{0}(n-m)+\nonumber\\
\fl&&b_{10}\psi_{0}(1)\psi_{0}^{2}(n-m)+b_{11}\psi_{0}(m)\psi_{0}(n)\psi_{0}(n-m)+b_{12}\psi_{0}^{2}(n)\psi_{0}(n-m)+\nonumber\\
\fl&&b_{13}\psi_{0}(m)\psi_{0}^{2}(n-m)+b_{14}\psi_{0}(n)\psi_{0}^{2}(n-m)+b_{15}\psi_{0}^{3}(n-m)+b_{16}\psi_{1}(n)+\nonumber\\
\fl&&b_{17}\psi_{0}(n-m)\psi_{1}(n)+b_{18}\psi_{1}(n-m)+b_{19}\psi_{0}(1)\psi_{1}(n-m)+\nonumber\\
\fl&&b_{20}\psi_{0}(m)\psi_{1}(n-m)+b_{21}\psi_{0}(n)\psi_{1}(n-m)+b_{22}\psi_{0}(n-m)\psi_{1}(n-m)+\nonumber\\
\fl&&b_{23}\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k}+b_{24}\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+n-m)}{k}+b_{25}\sum_{k=1}^{m}\frac{\psi_{1}(k+n-m)}{k}.\label{eq:IBf}\end{aligned}$$
[cl]{}
\
& $\frac{m}{864}\big(111 m^3+12036 m^2 n+2198 m^2+13230 m n^2+4530 m n+1629 m+$\
& $252 n^3+150 n^2-414 n-194\big)$\
& $-\frac{1}{72}\big(36 m^3 n+1242 m^2 n^2+174 m^2 n+72 m^2+1092 m n^3+690 m n^2-$\
& $234 m n+168 m+21 n^4+14 n^3-321 n^2+358 n+24\big)$\
& $-\frac{1}{72}\big(21 m^4+1056 m^3 n+230 m^3+864 m^2 n+255 m^2-1056 m n^3+$\
& $360 m n^2+216 m n+94 m-21 n^4-14 n^3+249 n^2+434 n+144\big)$\
$b_{4}=$ & $\frac{m}{6}\left(3 m^3+84 m^2 n+10 m^2+90 m n^2+6 m n+21 m-6 n^2+66 n+14\right)$\
$b_{5}=$ & $-\frac{m}{6}\left(3 m^3+84 m^2 n+10 m^2+90 m n^2+6 m n+21 m-6 n^2+66 n+14\right)$\
$b_{6}=$ & $-\frac{2n}{3}\left(3 m^2+12 m n-3 m+n^2-3 n+2\right)$\
& $\frac{1}{6}\big(-12 m^3+90 m^2 n^2-138 m^2 n-24 m^2+84 m n^3-150 m n^2+66 m n-$\
& $12 m+3 n^4-50 n^3+45 n^2-46 n\big)$\
& $\frac{1}{12}\big(3 m^4+84 m^3 n+26 m^3-90 m^2 n^2+66 m^2 n+45 m^2-168 m n^3-$\
& $66 m n^2-66 m n+22 m-6 n^4-36 n^3-66 n^2-36 n\big)$\
$b_{9}=$ & $2 m n \left(3 m^2+9 m n-2 m+3 n^2-2 n+3\right)$\
$b_{10}=$ & $6 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{11}=$ & $-2 m n \left(3 m^2+9 m n-2 m+3 n^2-2 n+3\right)$\
$b_{12}=$ & $-4 m n (m+n)$\
$b_{13}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{14}=$ & $m n \left(3 m^2+9 m n+2 m+3 n^2+2 n+3\right)$\
$b_{15}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{16}=$ & $-\frac{n}{6}n\left(30 m^2 n-6 m^2+28 m n^2-18 m n+26 m+n^3-6 n^2+11 n-6\right)$\
$b_{17}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
& $\frac{1}{12}\big(m^4+28 m^3 n-2 m^3+90 m^2 n^2-18 m^2 n-m^2+56 m n^3+18 m n^2+$\
& $66 m n+2 m+2 n^4+12 n^3+22 n^2+12 n\big)$\
$b_{19}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{20}=$ & $2 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{21}=$ & $m n \left(m^2+3 m n-2 m+n^2-2 n+1\right)$\
$b_{22}=$ & $2 m n \left(m^2+3 m n+n^2+1\right)$\
& $\frac{m}{6}\big(3 m^3+84 m^2 n+10 m^2+90 m n^2+6 m n+21 m-6 n^2+66 n+14+$\
& $12n\left(3 m^2+9 m n-2 m+3 n^2-2 n+3\right)\psi_{0}(n)+72n(m^2+3 m n+$\
& $n^2+1)\psi_{0}(n-m)\big)$\
$b_{24}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$b_{25}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
\
\[t:IBc\]
$$\begin{aligned}
\fl I_{C}&=&c_{1}+c_{2}\psi_{0}(n)+c_{3}\psi_{0}(n-m)+c_{4}\psi_{0}(1)\psi_{0}(n-m)+c_{5}\psi_{0}(m)\psi_{0}(n-m)+\nonumber\\
\fl&&c_{6}\psi_{0}^{2}(n)+c_{7}\psi_{0}(n)\psi_{0}(n-m)+c_{8}\psi_{0}^{2}(n-m)+c_{9}\psi_{0}(1)\psi_{0}(n)\psi_{0}(n-m)+\nonumber\\
\fl&&c_{10}\psi_{0}(m)\psi_{0}(n)\psi_{0}(n-m)+c_{11}\psi_{0}^{3}(n)+c_{12}\psi_{0}^{2}(n)\psi_{0}(n-m)+\nonumber\\
\fl&&c_{13}\psi_{0}(1)\psi_{0}^{2}(n-m)+c_{14}\psi_{0}(m)\psi_{0}^{2}(n-m)+c_{15}\psi_{0}(n)\psi_{0}^{2}(n-m)+\nonumber\\
\fl&&c_{16}\psi_{0}^{3}(n-m)+c_{17}\psi_{1}(n)+c_{18}\psi_{0}(n)\psi_{1}(n)+c_{19}\psi_{0}(n-m)\psi_{1}(n)+c_{20}\psi_{1}(n-m)+\nonumber\\
\fl&&c_{21}\psi_{0}(1)\psi_{1}(n-m)+c_{22}\psi_{0}(m)\psi_{1}(n-m)+c_{23}\psi_{0}(n)\psi_{1}(n-m)+\nonumber\\
\fl&&c_{24}\psi_{0}(n-m)\psi_{1}(n-m)+c_{25}\psi_{2}(n)+c_{26}\psi_{2}(n-m)+c_{27}\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k}+\nonumber\\
\fl&&c_{28}\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+n-m)}{k}+c_{29}\sum_{k=1}^{m}\frac{\psi_{1}(k+n-m)}{k}.\label{eq:ICf}\end{aligned}$$
[cl]{}
\
& $\frac{m}{288}\big(37 m^3+4012 m^2 n+1114 m^2+4410 m n^2+2430 m n+1043 m+$\
& $84 n^3+90 n^2-82 n-34\big)$\
& $-\frac{1}{24}\big(12 m^3 n+414 m^2 n^2+90 m^2 n-36 m^2+364 m n^3+342 m n^2-$\
& $142 m n+12 m+7 n^4+10 n^3-127 n^2+134 n+12\big)$\
& $-\frac{1}{24}\big(7 m^4+352 m^3 n+106 m^3+264 m^2 n+125 m^2-352 m n^3+$\
& $26 m-7 n^4-10 n^3+55 n^2+106 n+36\big)$\
$c_{4}=$ & $\frac{m}{2}\left(m^3+28 m^2 n+2 m^2+30 m n^2-6 m n+5 m-6 n^2+20 n+4\right)$\
$c_{5}=$ & $-\frac{m}{2}\left(m^3+28 m^2 n+2 m^2+30 m n^2-6 m n+5 m-6 n^2+20 n+4\right)$\
$c_{6}=$ & $\frac{1}{2}\left(-6 m^2 n+3 m^2-24 m n^2+15 m n+3 m-2 n^3+6 n^2-4 n\right)$\
& $-\frac{1}{2}\big(6 m^3-30 m^2 n^2+60 m^2 n+12 m^2-28 m n^3+48 m n^2-20 m n+$\
& $6 m-n^4+16 n^3-17 n^2+14 n\big)$\
& $\frac{1}{4}\big(m^4+28 m^3 n+10 m^3-30 m^2 n^2+30 m^2 n+17 m^2-56 m n^3-$\
& $18 m n^2-20 m n+8 m-2 n^4-12 n^3-22 n^2-12 n\big)$\
$c_{9}=$ & $6 m n \left(m^2+3 m n-m+n^2-n+1\right)$\
$c_{10}=$ & $-6 m n \left(m^2+3 m n-m+n^2-n+1\right)$\
$c_{11}=$ & $mn$\
$c_{12}=$ & $-6 m n (m + n)$\
$c_{13}=$ & $6 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{14}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{15}=$ & $3 m n \left(m^2+3 m n+m+n^2+n+1\right)$\
$c_{16}=$ & $-2 m n \left(m^2+3 m n+n^2+1\right)$\
& $\frac{1}{4}\big(4 m^3-30 m^2 n^2+30 m^2 n+6 m^2-28 m n^3+30 m n^2-20 m n+$\
& $2 m-n^4+6 n^3-11 n^2+6 n\big)$\
$c_{18}=$ & $3 m n (m+n)$\
$c_{19}=$ & $-3 m n \left(m^2+3 m n+n^2+1\right)$\
& $\frac{1}{4}\big(\!-4 m^3+30 m^2 n^2-18 m^2 n-6 m^2+28 m n^3+6 m n^2+20 m n-$\
& $2 m+n^4+6 n^3+11 n^2+6 n\big)$\
$c_{21}=$ & $-3 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{22}=$ & $3 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{23}=$ & $-3 m n (m+n)$\
$c_{24}=$ & $3 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{25}=$ & $\frac{mn}{2}\left(m^2+3 m n+n^2+1\right)$\
$c_{26}=$ & $-\frac{mn}{2}\left(m^2+3 m n+n^2+1\right)$\
& $\frac{m}{2}\big(m^3+28 m^2 n+2 m^2+30 m n^2-6 m n+5 m-6 n^2+20 n+4+$\
& $12n\left(m^2+3 m n-m+n^2-n+1\right)\psi_{0}(n)+24n\big(m^2+3 m n+$\
& $n^2+1\big)\psi_{0}(n-m)\big)$\
$c_{28}=$ & $-6 m n \left(m^2+3 m n+n^2+1\right)$\
$c_{29}=$ & $-3 m n \left(m^2+3 m n+n^2+1\right)$\
Finally, inserting (\[eq:IAf\]), (\[eq:IBf\]), and (\[eq:ICf\]) into (\[eq:ABC\]), we observe substantial cancellation among the terms in $I_{A}-3I_{B}+2I_{C}$. In particular, polygamma functions of argument $n-m$ and the three types of unsimplifiable sums $$\begin{aligned}
&\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k} \label{eq:ub1} \\
&\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+n-m)}{k} \label{eq:ub2} \\
&\sum_{k=1}^{m}\frac{\psi_{1}(k+n-m)}{k}. \label{eq:ub3}\end{aligned}$$ cancel completely. The surviving terms give us $$\begin{aligned}
\fl I_{A}-3I_{B}+2I_{C}&=&2c_{25}\psi_{2}(n)+2c_{18}\psi_{0}(n)\psi_{1}(n)+\left(2c_{17}-3b_{16}\right)\psi_{1}(n)+2c_{11}\psi_{0}^{3}(n)+\nonumber\\
\fl&&\left(2c_{6}-3b_{6}\right)\psi_{0}^{2}(n)+\left(a_2-3b_2+2c_2\right)\psi_{0}(n)+a_{1}-3b_{1}+2c_{1} \label{eq:T3f}\\
\fl&=&mn\left(m^2+3mn+n^2+1\right)\psi_{2}(n)+6mn(m+n)\psi_{0}(n)\psi_{1}(n)+\nonumber\\
\fl&&m\left(2m^2+12mn+3m+6n^{2}+3n+1\right)\psi_{1}(n)+2mn\psi_{0}^{3}(n)+\nonumber\\
\fl&&3m(m+3n+1)\psi_{0}^{2}(n)+6m(m+n+1)\psi_{0}(n)+m(m+1),\end{aligned}$$ which is the desired identity (\[eq:T3R\]). This completes the proof of the main result (\[eq:k3\]).
Before the end of the paper, a few remarks are in order. Firstly, note that the results (\[eq:IAf\]), (\[eq:IBf\]), and (\[eq:ICf\]) become indeterminate when $m=n$ since some of the polygamma functions approach infinity. The resulting identity (\[eq:T3R\]) is still valid for $m=n$ due to the cancellation of all polygamma functions of argument $n-m$ as observed in (\[eq:T3f\]). On the other hand, the indeterminacy can be also resolved by taking appropriate limits using (106). Secondly, note that as a result of applying the semi closed-form formulas, we in fact ended up with four types of unsimplifiable summations. Namely, in addition to (\[eq:ub1\]), (\[eq:ub2\]), and (\[eq:ub3\]), we also have $$\label{eq:ub0}
\sum_{k=1}^{m}\frac{\psi_{0}(k)}{k+n-m}.$$ The sum (\[eq:ub0\]), however, can be expressed by the sum (\[eq:ub1\]) through the identity $$\begin{aligned}
\fl\sum_{k=1}^{m}\left(\frac{\psi_{0}(k+a)}{k}+\frac{\psi_{0}(k)}{k+a}\right)&=&\psi_{0}(m+1)\psi_{0}(a+m+1)-\psi_{0}(1)\psi_{0}(a+1)+\frac{1}{a}\times\nonumber\\
&&\left(\psi_{0}(a+m+1)-\psi_{0}(a+1)-\psi_{0}(m+1)+\psi_{0}(1)\right),\end{aligned}$$ which is a special case of the result [@Milgram Eq. (23)] $$\begin{aligned}
\fl\sum_{k=1}^{m}\left(\frac{\psi_{0}(k+a)}{k+b}+\frac{\psi_{0}(k+b)}{k+a}\right)&=&\psi_{0}(a+m+1)\psi_{0}(b+m+1)-\psi_{0}(a+1)\psi_{0}(b+1)+\nonumber\\
&&\frac{1}{a-b}(\psi_{0}(a+m+1)-\psi_{0}(b+m+1)-\psi_{0}(a+1)+\nonumber\\
&&\psi_{0}(b+1)).\end{aligned}$$ In the limit $a$ approaches $b$, the above identity reduces to [@Milgram Eq. (26)] $$\fl\sum_{k=1}^{m}\frac{\psi_{0}(k+a)}{k+a}=\frac{1}{2}\left(\psi_{1}(a+m+1)-\psi_{1}(a+1)+\psi_{0}^{2}(a+m+1)-\psi_{0}^{2}(a+1)\right),$$ which has also been utilized in simplifying the summations. Since no further relation seems to exist among (\[eq:ub1\]), (\[eq:ub2\]), and (\[eq:ub3\]), we call these sums bases in representing the unsimplifiable summations in $I_{A}$, $I_{B}$, and $I_{C}$. In the simplification, we also find that the unsimplifiable sum $$\label{eq:vb1}
\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+a)}{k+a}$$ always comes in pairs with the unsimplifiable sum $$\label{eq:vb2}
\sum_{k=1}^{m}\frac{\psi_{1}(k+a)}{k+a}.$$ It can be verified by the principle discussed in \[sec:ap1-re\] the following closed-form relation of the two unsimplifiable sums $$\begin{aligned}
\fl\sum_{k=1}^{m}\frac{\psi_{0}^{2}(k+a)+\psi_{1}(k+a)}{k+a}&=&\frac{1}{3}\Big(\psi_{2}(a+m+1)-\psi_{2}(a+1)+3\psi_{0}(a+m+1)\times\nonumber\\
&&\psi_1(a+m+1)-3\psi_{0}(a+1)\psi_{1}(a+1)+\psi_{0}^{3}(a+m+1)-\nonumber\\
&&\psi_{0}^{3}(a+1)\Big),\end{aligned}$$ which excludes (\[eq:vb1\]) or (\[eq:vb2\]) being considered as an unsimplifiable sum basis. Finally, as seen from this section, in computing the integrals in (\[eq:IAI\]), (\[eq:IBI\]), and (\[eq:ICI\]) an essential task that the formulation in [@Bianchi19] will also inevitably end up with is to capture the cancellation of each of the unsimplifiable sums (\[eq:ub1\]), (\[eq:ub2\]), (\[eq:ub3\]), (\[eq:ub0\]). This in particular requires deriving tailor-made semi closed-form formulas as listed in \[sec:ap1-sc\] and \[sec:ap2-sc\], which are unavailable in the computer algebra system Mathematica [@KG]. It is therefore unclear the statement in [@Bianchi19] that the calculations were performed by using Mathematica to yield the result (\[eq:k3\]).
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks Bjordis Collaku and Xhoendi Collaku for helping with the simplification task. The author also wishes to thank Sean O’Rourke and Nicholas Witte for the discussion.
Polygamma summation identities of the first type {#App1}
================================================
In this appendix, we list finite summation identities of polygamma functions of the type $$\label{eq:A1}
\sum_{k=1}^{n}k^{c}\psi_{j_{1}}^{b_{1}}(k+a_{1})\psi_{j_{2}}^{b_{2}}(k+a_{2})\cdots\psi_{j_{m}}^{b_{m}}(k+a_{m})$$ (hereinafter referred to as the first type) useful in the simplification process in Section \[sec:simp\], where $\{a_{l}\}_{l=1}^{m}$, $\{b_{l}\}_{l=1}^{m}$, $\{j_{l}\}_{l=1}^{m}$, and $c$, are non-negative integers. We list the corresponding closed-form identities in \[sec:ap1-c\] and semi closed-form identities (that contain an unsimplifiable term) in \[sec:ap1-sc\]. Some remarks on derivation and implementation of the listed formulas are provided in \[sec:ap1-re\].
Close-form expressions {#sec:ap1-c}
----------------------
$$\label{eq:A2}
\fl\sum_{k=1}^n\psi_{0}(k+a)=(a+n)\psi_{0}(a+n+1)-a\psi_{0}(a+1)-n$$
$$\begin{aligned}
\label{eq:A3}
\fl\sum_{k=1}^{n}k\psi_{0}(k+a)&=&\frac{1}{2}\left(-a^2+a+n^2+n\right)\psi_{0}(a+n+1)+\frac{1}{2}(a-1)a\psi_{0}(a+1)+\nonumber \\
&&\frac{1}{4}n(2a-n-3)\end{aligned}$$
$$\begin{aligned}
\label{eq:A4}
\fl\sum_{k=1}^{n}k^{2}\psi_{0}(k+a)&=&\frac{1}{6}\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{0}(a+n+1)-\frac{1}{6}a\Big(2a^2-\nonumber\\
&&3a+1\Big)\psi_{0}(a+1)-\frac{1}{36}n\left(12a^2-6an-24a+4n^2+15n+17\right)\end{aligned}$$
$$\begin{aligned}
\label{eq:A5}
\fl\sum_{k=1}^{n}k^3\psi_{0}(k+a)&=&\frac{1}{4} \left(a^4-2a^3+a^2-n^4-2n^3-n^2\right)\psi_{0}(a+n+1)+\frac{1}{4}(a-1)^{2}a^{2}\times\nonumber\\
&&\psi_{0}(a+1)-\frac{1}{48}n\Big(\!-12a^3+6a^{2}n+30a^2-4an^2-18an-26a+\nonumber\\
&&3n^3+14n^2+21n+10\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A6}
\fl\sum_{k=1}^{n}\psi_{0}^{2}(k+a)&=&(a+n)\psi_{0}^{2}(a+n+1)-(2a+2n+1)\psi_{0}(a+n+1)-a\psi_{0}^{2}(a+1)+\nonumber\\
&&(2a+1)\psi_{0}(a+1)+2n\end{aligned}$$
$$\begin{aligned}
\label{eq:A7}
\fl\sum_{k=1}^{n}k\psi_{0}^{2}(k+a)&=&\frac{1}{2}\left(-a^2+a+n^2+n\right)\psi_{0}^{2}(a+n+1)+\frac{1}{4}\Big(6a^2+4an-2a-2n^2-\nonumber\\
&&6n-2\Big)\times\psi_{0}(a+n+1)+\frac{1}{2}(a-1)a\psi_{0}^{2}(a+1)+\frac{1}{4}\Big(\!-6a^2+2a+\nonumber\\
&&2\Big)\psi_{0}(a+1)+\frac{1}{4}n(-6a+n+5)\end{aligned}$$
$$\begin{aligned}
\label{eq:A8}
\fl\sum_{k=1}^{n}k^2\psi_{0}^{2}(k+a)&=&\frac{1}{6}\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{0}^{2}(a+n+1)-\frac{1}{18}\Big(22a^3+\nonumber\\
&&12a^{2}n-21a^2-6an^2-24an-a+4n^3+15n^2+17n+3\Big)\times\nonumber\\
&&\psi_{0}(a+n+1)-\frac{1}{6}a\left(2a^2-3a+1\right)\psi_{0}^{2}(a+1)+\frac{1}{18}\Big(22a^3-\nonumber\\
&&21a^2-a+3\Big)\psi_{0}(a+1)+\frac{1}{108}n\Big(132a^2-30an-192a+8n^2+\nonumber\\
&&+39n+79\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A9}
\fl\sum_{k=1}^{n}k^3\psi_{0}^{2}(k+a)&=&-\frac{1}{4}\left(a^4-2a^3+a^2-n^4-2n^3-n^2\right)\psi_{0}^{2}(a+n+1)+\frac{1}{24}\Big(25a^4+\nonumber\\
&&12a^{3}n-38a^3-6a^{2}n^2-30a^{2}n+11a^2+4an^3+18an^2+26an+\nonumber\\
&&2a-3n^4-14n^3-21n^2-10n\Big)\psi_{0}(a+n+1)+\frac{1}{4}(a-1)^{2}a^2\times\nonumber\\
&&\psi_{0}^{2}(a+1)-\frac{1}{24}a\left(25a^3-38a^2+11a+2\right)\psi_{0}(a+1)+\frac{1}{288}n\times\nonumber\\
&&\Big(\!-300a^3+78a^{2}n+606a^2-28an^2-162an-410a+9n^3+\nonumber\\
&&50n^2+111n+118\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A10}
\fl\sum_{k=1}^{n}\psi_{0}^{3}(k+a)&=&-\frac{1}{2}\psi_{1}(a+n+1)+\frac{1}{2}\psi_{1}(a+1)+(a+n)\psi_{0}^{3}(a+n+1)-\frac{3}{2}(2a+\nonumber\\
&&2n+1)\psi_{0}^{2}(a+n+1)+3(2a+2n+1)\psi_{0}(a+n+1)-a\times\nonumber\\
&&\psi_{0}^{3}(a+1)+\frac{3}{2}(2a+1)\psi_{0}^{2}(a+1)-3(2a+1)\psi_{0}(a+1)-6n\end{aligned}$$
$$\begin{aligned}
\label{eq:A11}
\fl\sum_{k=1}^{n}k\psi_{0}^{3}(k+a)&=&\frac{1}{4}\left(2a-1\right)\psi_{1}(a+n+1)+\frac{1}{4}\left(-2a+1\right)\psi_{1}(a+1)+\frac{1}{2}\Big(\!-a^2+\nonumber\\
&&a+n^2+n\Big)\psi_{0}^{3}(a+n+1)+\frac{3}{4}\left(3a^2+2an-a-n^2-3n-1\right)\times\nonumber\\
&&\psi_{0}^{2}(a+n+1)+\frac{1}{8}\left(-42a^2-36an+6a+6n^2+30n+14\right)\times\nonumber\\
&&\psi_{0}(a+n+1)+\frac{1}{2}(a-1)a\psi_{0}^{3}(a+1)+\frac{3}{4}\left(-3a^2+a+1\right)\times\nonumber\\
&&\psi_{0}^{2}(a+1)+\frac{1}{4}\left(21a^2-3a-7\right)\psi_{0}(a+1)+\frac{1}{8}\Big(42an-3n^2-\nonumber\\
&&27n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A12}
\fl\sum_{k=1}^{n}k^{2}\psi_{0}^{3}(k+a)&=&\frac{1}{12}\left(-6a^2+6a-1\right)\psi_{1}(a+n+1)+\frac{1}{12}\left(6a^2-6a+1\right)\psi_{1}(a+1)-\nonumber\\
&&\frac{1}{6}\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{0}^{3}(a+n+1)-\frac{1}{12}\Big(22a^3+\nonumber\\
&&12a^{2}n-21a^2-6an^2-24an-a+4n^3+15n^2+17n+3\Big)\times\nonumber\\
&&\psi_{0}^{2}(a+n+1)+\frac{1}{36}\Big(170a^3+132a^{2}n-123a^2-30an^2-192an-\nonumber\\
&&47a+8n^3+39n^2+79n+33\Big)\psi_{0}(a+n+1)-\frac{1}{6}a\left(2a^2-3a+1\right)\times\nonumber\\
&&\psi_{0}^{3}(a+1)+\frac{1}{12}\left(22a^3-21a^2-a+3\right)\psi_{0}^{2}(a+1)-\frac{1}{36}\Big(170a^3-\nonumber\\
&&123a^2-47a+33\Big)\psi_{0}(a+1)+\frac{1}{216}\Big(\!-1020a^{2}n+114an^2+\nonumber\\
&&1248an-16n^3-105n^2-365n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A13}
\fl\sum_{k=1}^{n}k^{3}\psi_{0}^{3}(k+a)&=&\frac{1}{4}a\left(2a^2-3a+1\right)\psi_{1}(a+n+1)+\frac{1}{4}a\left(-2a^2+3a-1\right)\psi_{1}(a+1)-\nonumber\\
&&\frac{1}{4}\left(a^4-2a^3+a^2-n^4-2n^3-n^2\right)\psi_{0}^{3}(a+n+1)+\frac{1}{16}\Big(25a^4+\nonumber\\
&&12a^{3}n-38a^3-6a^{2}n^2-30a^{2}n+11a^{2}+4an^{3}+18an^2+26an+\nonumber\\
&&2a-3n^4-14n^3-21n^2-10n\Big)\psi_{0}^{2}(a+n+1)-\frac{1}{96}\Big(415a^4+\nonumber\\
&&300a^{3}n-530a^3-78a^{2}n^{2}-606a^{2}n+17a^2+28an^3+162an^2+\nonumber\\
&&410an+146a-9n^4-50n^3-111n^2-118n-36\Big)\psi_{0}(a+n+1)+\nonumber\\
&&\frac{1}{4}(a-1)^{2}a^{2}\psi_{0}^{3}(a+1)-\frac{1}{16}a\left(25a^3-38a^2+11a+2\right)\psi_{0}^{2}(a+1)+\nonumber\\
&&\frac{1}{96}\left(415a^4-530a^3+17a^2+146a-36\right)\psi_{0}(a+1)+\frac{1}{1152}\times\nonumber\\
&&\Big(4980a^{3}n-690a^{2}n^{2}-8850a^{2}n+148an^3+1134an^2+4790an-\nonumber\\
&&27n^4-182n^3-525n^2-850n\Big)\end{aligned}$$
$$\label{eq:A14}
\fl\sum_{k=1}^{n}\psi_{1}(k+a)=(a+n)\psi_{1}(a+n+1)-a\psi_{1}(a+1)+\psi_{0}(a+n+1)-\psi_{0}(a+1)$$
$$\begin{aligned}
\label{eq:A15}
\fl\sum_{k=1}^{n}k\psi_{1}(k+a)&=&\frac{1}{2}\Big(\!\left(-a^2+a+n^2+n\right)\psi_{1}(a+n+1)+(a-1)a\psi_{1}(a+1)+\nonumber\\
&&(-2a+1)\psi_{0}(a+n+1)+(2a-1)\psi_{0}(a+1)+n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A16}
\fl\sum_{k=1}^{n}k^{2}\psi_{1}(k+a)&=&\frac{1}{6}\Big(\!\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{1}(a+n+1)+a(a-1)\times\nonumber\\
&&(-2a+1)\psi_{1}(a+1)+\left(6a^2-6a+1\right)\psi_{0}(a+n+1)+\Big(\!-6a^2+\nonumber\\
&&6a-1\Big)\psi_{0}(a+1)-4an+n^2+4n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A17}
\fl\sum_{k=1}^{n}k^{3}\psi_{1}(k+a)&=&\frac{1}{24}\Big(6\left(-a^4+2a^3-a^2+n^4+2n^3+n^2\right)\psi_{1}(a+n+1)+6\Big(a^4-\nonumber\\
&&2a^3+a^2\Big)\psi_{1}(a+1)-12a\left(2a^2-3a+1\right)\psi_{0}(a+n+1)+12a\times\nonumber\\
&&\left(2a^2-3a+1\right)\psi_{0}(a+1)+18a^{2}n-6an^2-30an+2n^3+9n^2+\nonumber\\
&&13n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A18}
\fl\sum_{k=1}^{n}\psi_{2}(k+a)&=&(a+n)\psi_{2}(a+n+1)-a\psi_{2}(a+1)+2\psi_{1}(a+n+1)-\nonumber\\
&&2\psi_{1}(a+1)\end{aligned}$$
$$\begin{aligned}
\label{eq:A19}
\fl\sum_{k=1}^{n}k\psi_{2}(k+a)&=&\frac{1}{2}\left(-a^2+a+n^2+n\right)\psi_{2}(a+n+1)+\frac{1}{2}a(a-1)\psi_{2}(a+1)+\nonumber\\
&&(-2a+1)\psi_{1}(a+n+1)+(2a-1)\psi_{1}(a+1)-\psi_{0}(a+n+1)+\nonumber\\
&&\psi_{0}(a+1)\end{aligned}$$
$$\begin{aligned}
\label{eq:A20}
\fl\sum_{k=1}^{n}k^{2}\psi_{2}(k+a)&=&\frac{1}{6}\Big(\!\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{2}(a+n+1)+a\Big(\!-2a^2+\nonumber\\
&&3a-1\Big)\psi_{2}(a+1)+2\left(6a^2-6a+1\right)\psi_{1}(a+n+1)+2\Big(\!-6a^2+\nonumber\\
&&6a-1\Big)\psi_{1}(a+1)+6(2a-1)\psi_{0}(a+n+1)+6(-2a+1)\times\nonumber\\
&&\psi_{0}(a+1)-4n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A21}
\fl\sum_{k=1}^{n}k^{3}\psi_{2}(k+a)&=&\frac{1}{4}\Big(\!\left(-a^4+2a^3-a^2+n^4+2n^3+n^2\right)\psi_{2}(a+n+1)+(a-1)^{2}a^{2}\times\nonumber\\
&&\psi_{2}(a+1)+4a\left(-2a^2+3a-1\right)\psi_{1}(a+n+1)+4a\left(2a^2-3a+1\right)\times\nonumber\\
&&\psi_{1}(a+1)-2\left(6a^2-6a+1\right)\psi_{0}(a+n+1)+2\left(6a^2-6a+1\right)\times\nonumber\\
&&\psi_{0}(a+1)+6an-n^2-5n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A22}
\fl\sum_{k=1}^{n}\psi_{0}(k+a)\psi_{1}(k+a)&=&(a+n)\psi_{0}(a+n+1)\psi_{1}(a+n+1)-a\psi_{0}(a+1)\psi_{1}(a+1)-\nonumber\\
&&\frac{1}{2}(2a+2n+1)\psi_{1}(a+n+1)+\frac{1}{2}(2a+1)\psi_{1}(a+1)+\nonumber\\
&&\frac{1}{2}\psi_{0}^{2}(a+n+1)-\psi_{0}(a+n+1)-\frac{1}{2}\psi_{0}^{2}(a+1)+\nonumber\\
&&\psi_{0}(a+1)\end{aligned}$$
$$\begin{aligned}
\label{eq:A23}
\fl\sum_{k=1}^{n}k\psi_{0}(k+a)\psi_{1}(k+a)&=&\frac{1}{4}\Big(2\left(-a^2+a+n^2+n\right)\psi_{0}(a+n+1)\psi_{1}(a+n+1)+\nonumber\\
&&2(a-1)a\psi_{0}(a+1)\psi_{1}(a+1)-\Big(\!-3a^2-2an+a+n^2+\nonumber\\
&&3n+1\Big)\psi_{1}(a+n+1)+\left(-3a^2+a+1\right)\psi_{1}(a+1)+\nonumber\\
&&(1-2a)\psi_{0}^{2}(a+n+1)+(6a+2n-1)\psi_{0}(a+n+1)+\nonumber\\
&&(2a-1)\psi_{0}^{2}(a+1)+(1-6a)\psi_{0}(a+1)-3n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A24}
\fl\sum_{k=1}^{n}k^{2}\psi_{0}(k+a)\psi_{1}(k+a)&=&\frac{1}{36}\Big(6\left(2a^3-3a^2+a+2n^3+3n^2+n\right)\psi_{0}(a+n+1)\times\nonumber\\
&&\psi_1(a+n+1)-6a\left(2a^2-3a+1\right)\psi_{0}(a+1)\psi_{1}(a+1)+\nonumber\\
&&\Big(\!-22a^3-12a^{2}n+21a^2+6an^2+24an+a-4n^3-\nonumber\\
&&15n^2-17n-3\Big)\psi_{1}(a+n+1)+\left(22a^3-21a^2-a+3\right)\times\nonumber\\
&&\psi_{1}(a+1)+3\left(6a^2-6a+1\right)\psi_{0}^{2}(a+n+1)+\Big(\!-66a^2-\nonumber\\
&&24an+42a+6n^2+24n+1\Big)\psi_{0}(a+n+1)-3\Big(6a^2-6a+\nonumber\\
&&1\Big)\psi_{0}^{2}(a+1)-\left(-66a^2+42a+1\right)\psi_{0}(a+1)+44an-\nonumber\\
&&5n^2-32n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A25}
\fl\sum_{k=1}^{n}k^{3}\psi_{0}(k+a)\psi_{1}(k+a)&=&\frac{1}{288}\Big(\!-72\left(a^4-2a^3+a^2-n^4-2n^3-n^2\right)\psi_{0}(a+n+1)\times\nonumber\\
&&\psi_{1}(a+n+1)+72(a-1)^{2}a^2\psi_{0}(a+1)\psi_{1}(a+1)+\nonumber\\
&&\Big(150a^4+72a^{3}n-228a^3-36a^{2}n^{2}-180a^{2}n+66a^2+\nonumber\\
&&24an^3+108an^2+156an+12a-18n^4-84n^3-126n^2-\nonumber\\
&&60n\Big)\psi_{1}(a+n+1)+\left(-150a^4+228a^3-66a^2-12a\right)\times\nonumber\\
&&\psi_{1}(a+1)-72a\left(2a^2-3a+1\right)\psi_{0}^{2}(a+n+1)-12\Big(\!-\nonumber\\
&&50a^3-18a^{2}n+57a^2+6an^2+30an-11a-2n^3-9n^2-\nonumber\\
&&13n-1\Big)\psi_{0}(a+n+1)+72a\left(2a^2-3a+1\right)\psi_{0}^{2}(a+1)+\nonumber\\
&&12\left(-50a^3+57a^2-11a-1\right)\psi_{0}(a+1)-450a^{2}n+\nonumber\\
&&78an^2+606an-14n^3-81n^2-205n\Big)\end{aligned}$$
Semi closed-form expressions {#sec:ap1-sc}
----------------------------
$$\begin{aligned}
\label{eq:A26}
\fl\sum_{k=1}^{n}\psi_{0}(k+a)\psi_{0}(k)&=&a\sum_{k=1}^{n}\frac{\psi_{0}(k)}{k+a}+n\psi_{0}(a+n+1)\psi_{0}(n+1)-(a+n+1)\psi_{0}(a+n\nonumber\\
&&+1)-n\psi_{0}(n+1)+(a+1)\psi_{0}(a+1)+2n\end{aligned}$$
$$\begin{aligned}
\label{eq:A27}
\fl\sum_{k=1}^{n}k\psi_{0}(k+a)\psi_{0}(k)&=&-\frac{a(a-1)}{2}\sum_{k=1}^{n}\frac{\psi_{0}(k)}{k+a}+\frac{1}{4}\Big(2n(n+1)\psi_{0}(a+n+1)\psi_{0}(n+1)+\nonumber\\
&&\left(a^2-a-n^2-3n-2\right)\psi_{0}(a+n+1)+(2a-n-3)n\nonumber\\
&&\psi_{0}(n+1)-(a-2)(a+1)\psi_{0}(a+1)-3an+n^2+5n\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A28}
\fl\sum_{k=1}^{n}k^{2}\psi_{0}(k+a)\psi_{0}(k)&=&\frac{1}{6}a(a-1)(2a-1)\sum_{k=1}^{n}\frac{\psi_{0}(k)}{k+a}+\frac{1}{108}\Big(18n(n+1)(2n+1)\times\nonumber\\
&&\psi_{0}(a+n+1)\psi_{0}(n+1)-3\Big(4a^3-3a^2-a+4n^3+15n^2+\nonumber\\
&&17n+6\Big)\psi_{0}(a+n+1)+3n\Big(\!-12a^2+6an+24a-4n^2-\nonumber\\
&&15n-17\Big)\psi_{0}(n+1)+3\left(4a^3-3a^2-a+6\right)\psi_{0}(a+1)+\nonumber\\
&&n\left(48a^2-15an-96a+8n^2+39n+79\right)\!\Big)\end{aligned}$$
$$\begin{aligned}
\label{eq:A29}
\fl\sum_{k=1}^{n}k^{3}\psi_{0}(k+a)\psi_{0}(k)&=&-\frac{1}{4}a^{2}(a-1)^{2}\sum_{k=1}^{n}\frac{\psi_{0}(k)}{k+a}+\frac{1}{288}\Big(72n^{2}(n+1)^{2}\psi_{0}(a+n+1)\times\nonumber\\
&&\psi_{0}(n+1)+6\Big(3a^4-2a^3-3a^2+2a-3n^4-14n^3-21n^2-\nonumber\\
&&10n\Big)\psi_{0}(a+n+1)+6n\Big(12a^3-6a^{2}n-30a^2+4an^2+18an+\nonumber\\
&&26a-3n^3-14n^2-21n-10\Big)\psi_{0}(n+1)-6a(a-1)(a+1)\times\nonumber\\
&&(3a-2)\psi_{0}(a+1)+n\Big(\!-90a^3+27a^{2}n+219a^2-14an^2-\nonumber\\
&&81an-205a+9n^3+50n^2+111n+118\Big)\Big)\end{aligned}$$
Remarks on the first type summation {#sec:ap1-re}
-----------------------------------
The principles of evaluating finite sums of the first type (\[eq:A1\]), that led to the above listed formulas, are simple. The idea is to change the order of sums by first replacing one polygamma function at a time by the definition (\[eq:pg\]) and make use of the obtained lower order summation formulas in a recursive manner. We demonstrate the principles by considering the sum below as an example $$\label{A1:eg}
\sum_{k=1}^{n}k^{c}\psi_{0}^{b}(k+a),$$ which is a special case of (\[eq:A1\]). We first show the recursion in parameter $b$ by the example $b=2$, $$\begin{aligned}
\fl\sum_{k=1}^{n}\psi_{0}^{2}(k+a)&=&\sum_{k=1}^{n}\sum_{j=1}^{k+a-1}\frac{\psi_{0}(k+a)}{j}-\gamma\sum_{k=1}^{n}\psi_{0}(k+a) \\
&=&\sum_{k=1}^{n}\sum_{j=a+1}^{k+a-1}\frac{\psi_{0}(k+a)}{j}+\left(\sum_{j=1}^{a}\frac{1}{j}-\gamma\right)\sum_{k=1}^{n}\psi_{0}(k+a) \\ &=&\sum_{k=1}^{n}\sum_{j=1}^{k-1}\frac{\psi_{0}(k+a)}{j+a}+\psi_{0}(a+1)\sum_{k=1}^{n}\psi_{0}(k+a) \\
&=&\sum_{j=1}^{n-1}\frac{1}{j+a}\sum_{k=j+1}^{n}\psi_{0}(k+a)+\psi_{0}(a+1)\sum_{k=1}^{n}\psi_{0}(k+a) \\
&=&\sum_{j=1}^{n-1}\frac{1}{j+a}\left(\sum_{k=1}^{n}\psi_{0}(k+a)-\sum_{k=1}^{j}\psi_{0}(k+a)\right)+\psi_{0}(a+1)\sum_{k=1}^{n}\psi_{0}(k+a), \nonumber\end{aligned}$$ which reduces to evaluating the sum (\[A1:eg\]) for $b=1$. To illustrate the recursion in parameter $c$, we consider the example $c=1$ in (\[A1:eg\]) $$\begin{aligned}
\sum_{k=1}^{n}k\psi_{0}^{2}(k+a)&=&\sum_{k=1}^{n}\sum_{j=1}^{k}\psi_{0}^{2}(k+a) \\
&=&\sum_{j=1}^{n}\sum_{k=j}^{n}\psi_{0}^{2}(k+a) \\
&=&\sum_{j=1}^{n}\left(\sum_{k=1}^{n}\psi_{0}^{2}(k+a)-\sum_{k=1}^{j-1}\psi_{0}^{2}(k+a)\right),\end{aligned}$$ which reduces to evaluating the sum (\[A1:eg\]) for $c=0$.
Using the principles as shown in the above examples, the listed formulas in this appendix can be derived, which are in fact valid for any non-negative real number $a$. Some of these formulas can be found in the literature. In particular, the formulas (\[eq:A2\])–(\[eq:A6\]), (\[eq:A10\]), (\[eq:A22\]) are available in [@Brychkov08 Chap. 5.1]. By keeping in mind the relation between harmonic numbers and polygamma functions (\[eq:pg\]), the formulas (\[eq:A2\])–(\[eq:A13\]) and (\[eq:A14\])–(\[eq:A21\]) may be also derived from the result [@Spiess90 Th. 2.2] and the result [@Spiess90 Th. 2.1], respectively.
Note that the results (\[eq:A22\])–(\[eq:A25\]) can be also obtained by the relation $$\frac{\partial}{\partial a}\sum_{k=1}^{n}k^{c}\psi_{0}^{2}(k+a)=2\sum_{k=1}^{n}k^{c}\psi_{0}(k+a)\psi_{1}(k+a),$$ and that the semi closed-form expressions (\[eq:A26\])–(\[eq:A29\]) reduce to the corresponding closed-form ones (\[eq:A6\])–(\[eq:A9\]) when $a=0$.
We also point out that currently the computer algebra system Mathematica is only able to evaluate into closed-form expressions the sum of the first type (\[eq:A1\]) when $b_{1}=1$, $b_{2}=\cdots=b_{m}=0$, i.e., the sum [@KG] $$\sum_{k=1}^{n}k^{c}\psi_{j}(k+a).$$ We are working with Wolfram Research to implement the polygamma summation (\[eq:A1\]) in an algorithmic manner into future versions of Mathematica.
Polygamma summation identities of the second type {#App2}
=================================================
In this appendix, we list finite summation identities of the type $$\label{eq:B1}
S_{f}(m,n)=\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}f(k),~~~~~~m\leq n,$$ (hereinafter referred to as the second type) that are utilized in the simplification process in Section \[sec:simp\]. Here, $f(k)$ is referred to as the test function that may involve a polygamma function and $(n-k)!/(m-k)!$ is referred to as the summation kernel.
We list the corresponding closed-form and semi closed-form identities in \[sec:ap2-c\] and \[sec:ap2-sc\], respectively. The strategy in deriving the listed formulas is discussed in \[sec:ap2-re\].
Close-form expressions {#sec:ap2-c}
----------------------
$$\label{eq:B2}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}=\frac{n!}{(m-1)!}\frac{1}{n-m+1}$$
$$\label{eq:B3}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{1}{k}=\frac{n!}{m!}\left(\psi_{0}\left(n+1\right)-\psi_{0}\left(n-m+1\right)\right)$$
$$\begin{aligned}
\label{eq:B4}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\psi_{0}(k)&=&\frac{n!}{(m-1)!(n-m+1)}\Bigg(\psi_{0}(n+1)-\psi_{0}(n-m+1)+\psi_{0}(1)-\nonumber\\
&&\frac{1}{n-m+1}\Bigg)\end{aligned}$$
$$\begin{aligned}
\label{eq:B5}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{\psi_{0}(k)}{k}&=&\frac{n!}{m!}\Bigg(\frac{1}{2}\Big(\psi_{1}(n+1)-\psi_{1}(n-m+1)+\psi_{0}^{2}(n+1)+\nonumber\\
&&\psi_{0}^{2}(n-m+1)\Big)+\psi_{0}(1)\left(\psi_{0}(n+1)-\psi_{0}(n-m+1)\right)-\nonumber\\
&&\psi_{0}(n+1)\psi_{0}(n-m+1)\Bigg)\end{aligned}$$
$$\begin{aligned}
\label{eq:B6}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{\psi_{0}(n+1-k)}{k}&=&\frac{n!}{m!}(\psi_{1}(n+1)-\psi_{1}(n-m+1)+\psi_{0}(n+1)(\psi_{0}(n+1)-\nonumber\\
&&\psi_{0}(n-m+1)))\end{aligned}$$
Semi closed-form expressions {#sec:ap2-sc}
----------------------------
$$\begin{aligned}
\label{eq:B7}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{1}{k+a}&=&\frac{(a+n)!}{(a+m)!}\sum_{k=1}^{m}\frac{(k+n-m-1)!(k+a-1)!}{(k-1)!(k+a+n-m)!}\end{aligned}$$
$$\begin{aligned}
\label{eq:B8}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{1}{k^{2}}&=&\frac{n!}{m!}\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k}+\frac{n!}{m!}\Bigg(\frac{1}{2}\Big(\psi_{1}(n-m+1)-\psi_{1}(n+1)+\nonumber\\
&&\psi_{0}^{2}(n-m+1)-\psi_{0}^{2}(n+1)\Big)+\psi_{0}(n-m)(\psi_{0}(n+1)-\psi_{0}(m+1)-\nonumber\\
&&\psi_{0}(n-m+1)+\psi_{0}(1))\Bigg)\end{aligned}$$
$$\begin{aligned}
\label{eq:B9}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{\psi_{0}(k)}{k^{2}}&=&\frac{n!}{2m!}\sum_{k=1}^{m}\frac{\psi_{1}(k+n-m)+\psi_{0}^{2}(k+n-m)}{k}-\frac{n!}{m!}\Big(\psi_{0}(n-m)-\nonumber\\
&&\psi_{0}(1)\Big)\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k}+\frac{n!}{2m!}\Bigg(\!-\frac{1}{3}\bigg(\psi_{2}(n+1)-\nonumber\\
&&\psi_{2}(n-m+1)+\psi_{0}^{3}(n+1)-\psi_{0}^{3}(n-m+1)+3\psi_{0}(n+1)\times\nonumber\\
&&\psi_{1}(n+1)-3\psi_{0}(n-m+1)\psi_{1}(n-m+1)\bigg)+(\psi_{0}(n-m)-\nonumber\\
&&\psi_{0}(1))\Big(\psi_{1}(n+1)-\psi_{1}(n-m+1)+\psi_{0}^{2}(n+1)-\nonumber\\
&&\psi_{0}^{2}(n-m+1)\Big)-\Big(\psi_{1}(n-m)-\psi_{0}^{2}(n-m)+2\psi_{0}(1)\times\nonumber\\
&&\psi_{0}(n-m)\Big)\Big(\psi_{0}(m+1)-\psi_{0}(n+1)+\psi_{0}(n-m+1)-\nonumber\\
&&\psi_{0}(1)\Big)\Bigg)\end{aligned}$$
$$\begin{aligned}
\label{eq:B10}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{\psi_{0}(n+1-k)}{k^{2}}&=&\frac{n!}{m!}\sum_{k=1}^{m}\frac{\psi_{1}(k+n-m)+\psi_{0}(n+1)\psi_{0}(k+n-m)}{k}+\frac{n!}{m!}\times\nonumber\\
&&\Bigg(\frac{1}{2}\psi_{2}(n-m+1)-\frac{1}{2}\psi_{2}(n+1)+\psi_{0}(n-m+1)\times\nonumber\\
&&\psi_{1}(n-m+1)+\psi_{0}(n+1)\Bigg(\frac{1}{2}\Big(\psi_{1}(n-m+1)-\nonumber\\
&&\psi_{1}(n+1)+\psi_{0}^{2}(n-m+1)-\psi_{0}^{2}(n+1)\Big)+\psi_{0}(n-m)\times\nonumber\\
&&(\psi_{0}(n+1)-\psi_{0}(n-m+1)-\psi_{0}(m+1)+\psi_{0}(1))+\nonumber\\
&&\psi_{1}(n-m)-\psi_{1}(n+1)\Bigg)-\psi_{1}(n-m)(\psi_{0}(n-m+1)+\nonumber\\
&&\psi_{0}(m+1)-\psi_{0}(1))+\psi_{0}(n-m)(\psi_{1}(n+1)-\nonumber\\
&&\psi_{1}(n-m+1))\Bigg)\end{aligned}$$
$$\begin{aligned}
\label{eq:B11}
\fl\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\psi_{1}(k)&=&-\frac{n!}{(m-1)!(n-m+1)}\sum_{k=1}^{m}\frac{\psi_{0}(k+n-m)}{k}-\nonumber\\
&&\frac{n!}{(m-1)!(n-m+1)}\Bigg(\frac{1}{2}\Big(\psi_{1}(n-m+1)-\psi_{1}(n)-\psi_{0}^{2}(n)+\nonumber\\
&&\psi_{0}^{2}(n-m+1)\Big)+\psi_{0}(n-m)(\psi_{0}(n)-\psi_{0}(m)-\psi_{0}(n-m+1)+\nonumber\\
&&\psi_{0}(1))-\psi_{1}(1)-\frac{\psi_{0}(n)-\psi_{0}(n-m+1)}{n}-\frac{\psi_{0}(n)}{m}\Bigg)\end{aligned}$$
Remarks on the second type summation {#sec:ap2-re}
------------------------------------
The generic approach in deriving the summation identities of the second type (\[eq:B1\]) relies on finding the recurrence relation between $S_{f}(m,n)$ and $S_{f}(m-1,n-1)$, where the summation terminates after $m$ recursions since $S_{f}(0,n-m)=0$. For a given test function $f(k)$, the recurrence relation can often be found by first rewriting (\[eq:B1\]) as $$S_{f}(m,n)=\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-1-k)!}\frac{n-k}{m-k}f(k),$$ where the term $$\frac{(n-1-k)!}{(m-1-k)!}$$ is understood as the new kernel of the sum $S_{f}(m-1,n-1)$ associated with the new test function $$\label{eq:mtf}
\frac{n-k}{m-k}f(k).$$ The relation between $S_{f}(m,n)$ and $S_{f}(m-1,n-1)$ can then be obtained by partial fraction decomposition in the variable $k$ of this modified test function (\[eq:mtf\]).
To illustrate the above approach, we show in details the derivation of some of the listed formulas as examples. The first example is when $f(k)=1/k$, where the modified test function is decomposed as $$\label{eq:Beg1}
\frac{n-k}{m-k}\frac{1}{k}=\frac{n}{m}\frac{1}{k}+\frac{n-m}{m}\frac{1}{m-k}.$$ The corresponding recurrence relation is deduced as $$\begin{aligned}
S_{f}(m,n)&=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{n-m}{m}\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-k)!} \\
&=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{(n-1)!}{m!},\end{aligned}$$ where we have used the formula (\[eq:B2\]). Iterating $m$ times the above relation leads to the desired expression (\[eq:B3\]). The next example is the case $f(k)=\psi_{0}(k)/k$, where similarly as in (\[eq:Beg1\]) the test function is decomposed as $$\frac{n-k}{m-k}\frac{\psi_{0}(k)}{k}=\frac{n}{m}\frac{\psi_{0}(k)}{k}+\frac{n-m}{m}\frac{\psi_{0}(k)}{m-k}.$$ The recurrence relation is then calculated as $$\begin{aligned}
\fl~\!S_{f}(m,n)&\!\!\!\!\!\!\!\!\!\!\!\!\!=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{n-m}{m}\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-k)!}\psi_{0}(k) \\
&\!\!\!\!\!\!\!\!\!\!\!\!\!=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{(n-1)!}{m!}\left(\psi_{0}(n)-\psi_{0}(n-m+1)+\psi_{0}(1)\right),\end{aligned}$$ where we have made use of the formula (\[eq:B4\]). After $m$ recursions of the above relation, one obtains the claimed identity (\[eq:B5\]). We also consider the case $f(k)=1/(k+a)$ as an example, where the modified test function is decomposed as $$\frac{n-k}{m-k}\frac{1}{k+a}=\frac{n+a}{m+a}\frac{1}{k+a}+\frac{n-m}{m+a}\frac{1}{m-k}.$$ The resulting recurrence relation is $$\begin{aligned}
S_{f}(m,n)&=&\frac{n+a}{m+a}S_{f}(m-1,n-1)+\frac{n-m}{m+a}\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-k)!}\\
&=&\frac{n+a}{m+a}S_{f}(m-1,n-1)+\frac{(n-1)!}{(m-1)!(m+a)},\label{eq:ka}\end{aligned}$$ where we have utilized the result (\[eq:B2\]). The claimed identity (\[eq:B7\]) is established after $m$ iterations of the recurrence relation (\[eq:ka\]). Note that the formula (\[eq:B3\]) is a special case of the formula (\[eq:B7\]), which is a useful identity that could transform a summation into one of the listed sums in the appendices. The last example is when $f(k)=1/k^{2}$, where partial fraction decomposition of the modified test function gives $$\frac{n-k}{m-k}\frac{1}{k^{2}}=\frac{n}{m}\frac{1}{k^{2}}+\frac{n-m}{m^{2}}\frac{1}{k}+\frac{n-m}{m^{2}}\frac{1}{m-k}.$$ The recurrence relation is then obtained as $$\begin{aligned}
\fl~\!S_{f}(m,n)&\!\!\!\!\!\!\!\!\!\!\!\!\!=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{n-m}{m^{2}}\left(\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-1-k)!}\frac{1}{k}+\sum_{k=1}^{m}\frac{(n-1-k)!}{(m-k)!}\right)\nonumber\\
&\!\!\!\!\!\!\!\!\!\!\!\!\!=&\frac{n}{m}S_{f}(m-1,n-1)+\frac{(n-1)!(n-m)}{m!m}(\psi_{0}(n)-\psi_{0}(n-m)),\end{aligned}$$ where we have used the identities (\[eq:B2\]) and (\[eq:B3\]). Iterating $m$ times the above relation gives the desired result (\[eq:B8\]).
Some of formulas in \[App2\] exist in the literature: The formula (\[eq:B2\]) is the well-known Chu-Vandermonde identity [@Luke p. 99] and the formula (\[eq:B3\]) can be also obtained via the connection to a hypergeometric function of unit argument as [@Luke p. 111] $$\begin{aligned}
\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{1}{k}&=&\frac{(n-1)!}{(m-1)!}~\!_{3}F_{2}\left(1,1,1-m;2,1-n;1\right)\\
&=&\frac{n!}{m!}\left(\psi_{0}(n+1)-\psi_{0}(n-m+1)\right).\end{aligned}$$ Moreover, the identity (\[eq:B8\]) recently appears in [@Wei17 Eq. (A12)].
Finally, we note that the listed formulas in \[App2\] are in fact valid for any positive real number $n$ greater than the integer $m$. This fact provides an alternative derivation of the formulas (\[eq:B6\]) and (\[eq:B10\]) via the derivative with respect to $n$ on (\[eq:B3\]) and (\[eq:B8\]), respectively, i.e., $$\frac{\partial}{\partial n}\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{1}{k^{c}}=\sum_{k=1}^{m}\frac{(n-k)!}{(m-k)!}\frac{\psi_{0}(n+1-k)}{k^{c}},~~~~~~c=1,2.$$
References {#references .unnumbered}
==========
[99]{}
Page D N 1993 Average entropy of a subsystem [**]{}[**71**]{} 1291-4
Bianchi E and Donà P 2019 Typical entropy of a subsystem: Page curve and its variance arXiv:1904.08370v2
Foong S K and Kanno S 1994 Proof of Page’s conjecture on the average entropy of a subsystem [**]{}[**72**]{} 1148-51
Sánchez-Ruiz J 1995 Simple proof of Page’s conjecture on the average entropy of a subsystem [*E*]{} [**52**]{} 5653-5
Vivo P, Pato M P and Oshanin G 2016 Random pure states: Quantifying bipartite entanglement beyond the linear statistics [*E*]{} [**93**]{} 052106
Wei L 2017 Proof of Vivo-Pato-Oshanin’s conjecture on the fluctuation of von Neumann entropy [*E*]{} [**96**]{} 022106
Majumdar S N 2011 [*Extreme eigenvalues of Wishart matrices: application to entangled bipartite system (The Oxford Handbook of Random Matrix Theory)*]{} ed G Akemann, J Baik and P Di Francesco (Oxford, Oxford University Press)
Wei L 2019 A note on Bianchi-Donà’s proof to the variance formula of von Neumann entropy arXiv:1906.10303
Kapadia D and Germundsson R 2019 private communication (Wolfram Research)
Cramér H 1946 [*Mathematical Methods of Statistics*]{} (Princeton: Princeton University Press)
Forrester P 2010 [*Log-gases and Random Matrices*]{} (Princeton: Princeton University Press)
Schr[ö]{}dinger E 1926 Quantisierung als eigenwertproblem [**]{}[**80**]{} 437-90
Milgram M 2017 On some sums of digamma and polygamma functions arXiv:0406338v3
Brychkov Y A 2008 [*Handbook of Special Functions: Derivatives, Integrals, Series and Other Formulas*]{} (Boca Raton: CRC Press)
Spie[ß]{} J 1990 Some identities involving harmonic numbers [*Math. Comp.*]{} [**55**]{} 839-63
Luke Y L 1969 [*The Special Functions and Their Approximations Vol. 1*]{} (Academic Press, New York)
[^1]: Note that since the composite system is in a random pure state, the von Neumann entropy of the full system is zero [@Page93].
[^2]: A linear spectral statistics can be defined as $\sum_{k=1}^{m}f(x_{i})$, where $x_{i}, i=1,\dots,m$, are the eigenvalues of an $m\times m$ Hermitian random matrix.
[^3]: For convenience of the discussion, we refer to the random variable $T$ as an induced entropy, which may not have the physical meaning of an entropy.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a Bell-like inequality performed using various instances of multi-photon entangled states to demonstrate that losses occurring [*after*]{} the unitary transformations used in the nonlocality test can be counteracted by enhancing the “size" of such entangled states. In turn, this feature can be used to overcome detection inefficiencies affecting the test itself: a slight increase in the size of such states, pushing them towards a more [*macroscopic*]{} form of entanglement, significantly improves the state robustness against detection inefficiency, thus easing the closing of the detection loophole. Differently, losses [*before*]{} the unitary transformations cause decoherence effects that cannot be compensated using macroscopic entanglement.'
author:
- 'Youngrong Lim,$^1$ Mauro Paternostro,$^2$ Minsu Kang,$^1$ Jinhyoung Lee,$^3$ and Hyunseok Jeong$^1$'
title: Using macroscopic entanglement to close the detection loophole in Bell inequality
---
Quantum mechanics reveals a world quite different from the classical one. Probably the most surprising consequence of the basic assumptions of quantum mechanics is that local realism is no longer tenable. This is proved by violation of Bell’s inequality [@Bell]. However, the undoubtedly impressive successes in the experimental violation of local realism have not yet reached the stage where the violation of Bell’s inequality free from well-known experimental loopholes is possible: while the locality loophole can be handled by using ultra-fast analyzers and photonic information carriers, which guarantee the space-like distribution of particles outside the light cone [@weihs], considerable efforts have been directed towards the closure of the detection loophole through the use of various sorts of quantum states and measurement schemes [@qudit; @atom].
While this sets the underlying motivations for our investigation, the following identifies the somehow counterintuitive path that we propose to take in order to solve the issue of the detection loophole: we consider an optical setting for the violation of Bell’s inequality and argue that only a slight increase in the “macroscopic nature” of the two-mode optical resource to be used in the test is sufficient to close the detection loophole. Here, “macroscopic entanglement” should be intended as the entanglement between macroscopically distinguishable states (we use the term “macroscopicity” accordingly) [@note]. The achievement of macroscopic entanglement is one of the most stimulating topics, at all levels, in modern quantum mechanics in light of the intrinsic interest in the observation of quantum phenomena on a macroscopic scale. Endeavors in this sense are made difficult due to the fact that entanglement between macroscopically distinguishable states appears to be possible only under very selective conditions [@atomic; @mauro] and it would easily be destroyed by the interaction with the surrounding world [@zurek]. However, we show that by considering optical states that are very close to experimental realization and are nevertheless endowed with a non-negligible macroscopic character, the experimental requirements for the closure of the detection loophole can be significantly lowered.
We first focus on entanglement between photon number states $$|\psi_n\rangle=\frac{1}{\sqrt{2}}(|n_H\rangle|n_V\rangle+|n_V\rangle|n_H\rangle),
\label{ens}$$ where $|n_H\rangle$ ($|n_V\rangle$) denotes a state of $n$ photons, all with horizontal (vertical) polarization in a single spatial mode. State is equivalent to a GHZ-type entangled state $|H\rangle^{\otimes 2n}+|V\rangle^{\otimes 2n}$ [@Gao2010]: it is obvious that local unitary operations and spatial mode rearrangement can convert one to the other. These types of states have been experimentally demonstrated for values of $n$ up to $5$ with fidelity 0.56 ($n=4$ with 0.78) [@Walther2004; @Gao2010]. The components $|n_H\rangle$ and $|n_V\rangle$ of Eq. (\[ens\]) can be considered macroscopically distinguishable when $n$ is large.
Suppose now that, experimentally, we only have extremely inefficient detectors, such as human eyes, so that only “macroscopic difference” can be noticed. Notice that such a possibility has been recently considered in the context of experimental revelation of entanglement and non-locality in entangled states involving macroscopic components [@sticatzi; @main; @Stobinska2011] as well as quantum-to-classical transition [@main; @Buzek1995; @brukner08; @Rae2011]. If $n$ is large enough, the two states can still be discriminated using two inefficient detectors, $A$ and $B$, and a polarization beam splitter (PBS) as shown in Fig. 1. If only detector $A$ ($B$) clicks, one can tell that the input state was $|n_H\rangle$ ($|n_V\rangle$). The measurement scheme fails only when the relevant detector misses all the incoming photons. Assuming that efficiency of each detector is $\eta$, the success probability $P_s$ for the scheme is $$P_s=1-(1-\eta)^n,$$ which can be made arbitrarily close to unity by increasing $n$, for any non-zero value of $\eta$. Distinguishability between the two states can be made arbitrarily good by increasing the photon number.
Let us introduce the measurement operator $$\hat{O}=\sum_{k=1}^n\big(|k_H\rangle\langle k_H|-|k_V\rangle\langle k_V|\big)+|0\rangle\langle 0|,
\label{op}$$ whose last term is necessary to include a “no-click” event at both the $A$ and $B$ photodetectors in Fig. \[schemi\][**(a)**]{}, a case that should be included in a loophole-free Bell test. The correlation function is constructed as $$E^p(\theta_a,\theta_b,\eta){=}
{\rm Tr}_{aba^\prime b^\prime}[\hat{O}_a \otimes \hat{O}_b |\Psi_n(\theta_a,\theta_b,\eta)\rangle
\langle\Psi_n(\theta_a,\theta_b,\eta)|]
\label{correl}$$ where the superscript $p$ is used to indicate that polarization-entangled states are used and $$\label{stato}
|\Psi_n(\theta_a,\theta_b,\eta)\rangle{=}\hat{\cal B}^p_{a a^\prime}(\eta,\theta_a)\otimes
\hat{\cal B}^p_{b b^\prime}(\eta,\theta_b)|\psi_n\rangle_{ab}|00\rangle_{a' b^\prime}.$$ We have introduced $\hat{\cal B}^p_{j j^\prime}(\eta,\theta_j)=\hat{B}_{j j^\prime}(\eta)\hat{U}^p(\theta_j)$ $(j=a,b)$ with ${\hat B}_{ab}(\eta)=e^{\frac{\zeta}{2}({\hat a}^\dagger {\hat b}
-{\hat a} {\hat b}^\dagger)}$ the operator of a beam splitter of transmittivity $\eta{=}(\cos\zeta)^2$ and $\hat{U}^p(\theta_j)=\exp[i\theta_j(|n_H\rangle_j\langle n_V|+h.c.)]$ a rotation about the $x$-axis of the Bloch sphere of a polarization qubit $\{|n_H\rangle,|n_V\rangle\}_j$ encoded in mode $j$ ($\forall{n}{\in}{\mathbb N}$). As such unitary operation depends on the photon number $n$, it needs the non-linear Hamiltonian $\hat H_n =g( {{\hat a^n_H}}{\hat a_V^{\dag n}}e^{i\phi} + h.c.)$ to be realized: by choosing the interaction time, any value of $\theta_j$ can be obtained. One can in principle implement this type of highly non-linear Hamiltonian by decomposing it into series of Gaussian unitaries and cubic operations [@Seckin; @Petr].
Using the correlation function (\[correl\]), it is straightforward to construct the Bell function [@CHSH] $B^p(\theta_a,\theta_b,\theta'_a,\theta'_b,\eta)=E^p(\theta_a,\theta_b,\eta)+E^p(\theta_a,\theta_b^\prime,\eta)
+E^p(\theta_a^\prime,\theta_b,\eta)-E^p(\theta_a^\prime,\theta_b^\prime,\eta)$, which should satisfy $|B^p|\leq 2$ under the assumptions of local realism. In our case, the correlation function is $$E^p({\theta _a},{\theta _b},\eta ){=}(1-\eta )^{2 n}-[1-(1-\eta )^n]^2 \cos\left[2 \left(\theta _a+\theta _b\right)\right].$$ In Fig. \[polar\][**(a)**]{} we have plotted the optimized Bell function $|B^p|_{\rm max}$ against detection efficiency $\eta$ for several values of $n$. For $n=1$, the efficiency of minimum violation (detection-inefficiency threshold) is 82.8$\%$, which is a well known value for the maximally entangled qubit states [@82]. As it can be clearly seen, when $n$ grows (hence increasing the ‘macroscopic’ nature of the entangled state at hand), the detection-inefficiency threshold decreases, thus showing that low efficiencies can be compensated by an enhanced macroscopic character.
We now address an important point. As we are interested in the effects of low-efficiency detectors, we have so far considered losses after the unitary rotations and prior to the detection, [*i.e.*]{}, only BS2 was considered in Fig. \[schemi\][**(a)**]{}. What happens if losses occur before the unitary operations as modeled by BS1 in Fig. \[schemi\][**(a)**]{}? In order to assess this case, we only need to replace $\hat{\cal B}_{jj'}(\eta,\theta_j)$ in Eq. with $\hat{\cal B}^{p}_{j j^\prime}(\eta,\theta_j)=\hat{U}^{p}(\theta_j)\hat{B}_{j j^\prime}(\eta)$ $(j{=}a,b)$. In this case \[see Fig. \[polar\][**(b)**]{}\], the configuration with increasing photon number becomes very fragile against photon losses. This is due to the fact that losses before the unitary operations must be treated as decoherence rather than detection inefficiency. As the effective decoherence rate is faster for larger $n$, simply increasing it and using local unitary operations cannot cure the corresponding spoiling effects. For more plausible result we need to consider both effect at once. This is the reason why we do not indicate the case of higher $n$ despite the fact that $n=4$ is not sufficiently [*macroscopic*]{}. More discussions will be presented at the end of the manuscript.
As a second significant example, we consider the entangled coherent state (ECS) [@sanders] $${\left\vert\text{ECS}\right\rangle}=N(|\alpha,\alpha\rangle+|-\alpha,-\alpha\rangle),$$ where $|{\pm}\alpha\rangle$ is a coherent state of amplitude $\pm\alpha\in{\mathbb C}$ and $N{=}[2(1 + {e^{ - 4{{\left| \alpha \right|}^2}}})]^{-1/2}$ is a normalization factor. Such states can be generated using a 50:50 beam splitter and coherent-state superpositions, $|\alpha\rangle+|-\alpha\rangle$, which have been experimentally demonstrated [@cat-gen]. Recently, a nonlocal generation of the entangled coherent state was successfully demonstrated [@ecs-gen].
The set-up that we consider in this case is illustrated in Fig. \[schemi\][**(b)**]{} but BS1 is ignored for now. Similarly to our previous example, we take unitary operations $\hat{U}^{ecs}_{jj'}(\theta_j)~(j{=}a,b)$ embodied by effective rotation performed in the space spanned by the basis $\{\left| \alpha \right\rangle, \left| -\alpha \right\rangle\}$. For large values of $|\alpha|$, such transformations can be performed approximately using a properly arranged cascade of single-mode Kerr-like nonlinearities and displacement operations [@paternostro10; @SJR; @main]. A Bell test can then be constructed by using homodyne measurements, whose outcomes are dichotomized in a way that a logical outcome $+1$ ($-1$) is associated to positive (negative) expectation values of the in-phase quadrature operator of each mode. In Refs. [@main; @SJR; @paternostro10; @mckeown10; @mckeown11], some of us have proven that this approach can be successfully applied to the exploration of non-classicality tests, including multipartite non-locality and quantum contextuality.
Here, we model inefficient homodyne detectors in a way fully analogous to what has been done previously for photodetectors \[cf. Fig. \[schemi\][**(b)**]{}\]. For simplicity and without limitations to the validity of our analysis, in what follows we assume $\alpha\in{\mathbb R}$. We study the behavior of the optimized Bell parameter $|B^{ecs}(\theta_a,\theta_b,\theta'_a,\theta'_b,\eta)|$, which has been constructed using the correlation functions ${E}^{ecs}(\theta_a,\theta_b,\eta)=P_{++}+P_{--}-P_{+-}-P_{-+}$, where we have defined the probabilities $${P}_{kl}{=}\!\!\int^{k_s}_{k_i}\!\!\!dx\!\!\int^{l_s}_{l_i}\!\!\!dy\langle{x_ax_b}|{\rm Tr}_{a'b'}[|\Phi(\theta_a,\theta_b,\eta)\rangle\langle\Phi(\theta_a,\theta_b,\eta)|]|x_ax_b\rangle$$ with $|\Phi(\theta_a,\theta_b,\eta_{a^\prime b^\prime})\rangle{=}\hat{\cal B}^{ecs}_{aa^\prime}(\eta,\theta_a)\otimes\hat{\cal B}^{ecs}_{bb^\prime}(\eta,\theta_b)|\text{ECS}\rangle_{ab}|00\rangle_{a'b'}$, $\hat{\cal B}^{ecs}_{j j^\prime}(\eta,\theta_j){=}\hat{B}_{j j^\prime}(\eta)\hat{U}^{ecs}(\theta_j)$ $(j{=}a,b)$, the subscripts $k,l{=}\pm$ that correspond to the assigned measurement outcomes $\pm{1}$ and the integration limits $+_s=\infty,~+_i=-_s=0$ and $-_i=-\infty$. Moreover, $|x_{a}\rangle$ ($|x_{b}\rangle$) is an eigenstate of the quadrature operator $\hat{x}_a=\hat{a}+\hat{a}^\dag$ ($\hat{x}_b=\hat{b}+\hat{b}^\dag$) with eigenvalue $x_a$ ($x_b$).
In Fig. \[ECS\] we plot the Bell parameter against the homodyne detector efficiency $\eta$ for various choices of $\alpha$. It is clear in this case too that by increasing the macroscopic character of the state (i.e. by increasing $\alpha$), we gain in robustness and the violation of the Bell inequality is possible even for large losses. For instance, at $\alpha=1$, violation is possible for $50\%$ detection efficiencies, which sets an important benchmark. Comparing with the previous case, ECS is more robust to detection inefficiency than polarization entangled state under same average photon number. For example, the detection-inefficiency threshold is about 0.83 for the polarization entangled state of two photons, but it is about 0.5 for the ECS. Moreover, by addressing the case of photon losses occurring before the local unitary operations, modeled by BS1 in Fig. \[schemi\][**(b)**]{}, one can also see that an opposite effect, very much similar to the one observed for polarization entangled state, is in order: larger values $\alpha$ imply a larger detection-inefficiency threshold for the violation of the Bell-CHSH inequality. This can be easily understood and verified by studying the Markovian master equation $\partial_t\varrho=\gamma\hat{a}\varrho\hat{a}^\dag-\frac{\gamma}{2}
(\hat{a}^\dag\hat{a}\varrho+\varrho\hat{a}^\dag\hat{a})$ with $\varrho$ the density matrix of a boson with associated annihilation (creation) operator $\hat{a}$ ($\hat{a}^\dag$), $\gamma$ the loss rate and $t$ the interaction time. Decoherence described by this maser equation is equivalent to the beam splitter loss: the relation between the beam splitter loss $\eta$ before the unitary operation and the loss rate $\gamma$ in the master equation is given by $\eta=\exp[-\gamma t]$.
We now extend our analysis to an entangled thermal state (ETS) [@jr06] $$\label{ruotato}
\begin{split}
&\rho_{ab}^{ets}(\theta_a,\theta_b){=}N_+\!\!\int\!\!\int{d}^2\alpha{d}^2
\beta{P}^{th}_{\alpha}(V,d){P}^{th}_{\beta}(V,d){\left\vert\text{ecs}\right\rangle}_{ab}\!{\left\langle\text{ecs}\right\vert}
\end{split}$$ with ${\left\vert\rm ecs\right\rangle}=|{\alpha,\beta}\rangle+|{-\alpha,-\beta}\rangle$, $N_\pm=[2(1\pm{e^{-{4d^2}/{V}}}/{V^2})]^{-1}$ and $P^{th}_{\alpha}(V,d)=\frac{2}{\pi(V-1)}e^{-\frac{2|\alpha-d|^2}{(V-1)}}$ the Gaussian thermal distribution with variance $V=2({\overline{n}}-d^2)+1$ (${\overline{n}}$ is the mean photon number) and center $d$ (with respect to the origin of the phase space). This state can be created by entangling two single-mode thermal states mutually displaced by $d$ and has been used to prove the possibility to violate local realism with coarse grained homodyne measurements and thermal local states [@main].
We subject Eq. to the local rotations $\hat{U}^{ecs}_{jj'}(\theta_j)$ used in order to assess an ECS and eventually arrive at the following expression for the correlation function $C^{ets}(\theta_a,\theta_b)$ that enters in the Bell parameter $B^{ets}(\theta_a,\theta_b,\theta_a^\prime,\theta_b^\prime)=
C^{ets}(\theta_a,\theta_b)+C^{ets}(\theta_a^\prime,\theta_b)+C^{ets}(\theta_a,\theta_b^\prime)-C^{ets}(\theta_a^\prime,\theta_b^\prime)$
$$\begin{split}
C^{ets}(\theta_a,\theta_b)&=
V_1V_2\left\{e^{4i\theta_a}g_\eta(\theta_a)
\left[
Qg_\eta(\theta_b)s(\theta_b)
+ie^{\frac{2d^2}{V}}V_1h(\theta_b)
\left(f_{-,\eta}(\theta_b)-e^{8i\theta_b}f_{+,\eta}(\theta_b)\right)\right]\right.\\
&\left.+V_1h(\theta_a)\left[ie^{2\theta_b(2i+\frac{V\theta_b}{d^2})}g_{\eta}(\theta_b)
s(\theta_b)\left(f_-(\theta_a)-e^{8i\theta_b}f_+(\theta_a)\right)
+4V_1h(\theta_b)\left(e^{8i\theta_a}f_{-,\eta}(\theta_b)f_{+,\eta}(\theta_a)
+e^{8i\theta_a}f_{-,\eta}(\theta_a)f_{+,\eta}(\theta_b)\right)
\right]\right\}
\end{split}$$
where $s(\theta_i)={\rm sign(\theta_i)}$ and $$\begin{split}
&h(\theta_i)=e^{\frac{2(d^4+\theta_i^2)}{d^2V}},~g_\eta(\theta_j){=}{\rm Erfi}\left[\frac{\sqrt{2}\eta\theta_j}{d\sqrt{V^2-\eta^2V(V-1)}}\right],\\
&V_1=(1/8)(1+V^2e^{\frac{4d^2}{V}})^{-1},~V_2=e^{{-4i(\theta_a+\theta_b)-\frac{2(1+V^2)(\theta_a^2+\theta_b^2)}{d^2V}}},\\
&Q=8e^{{4i\theta_b+\frac{2V(\theta_a^2+\theta_b^2)}{d^2}}},~f_{\pm,\eta}(\theta_j){=}{\rm Erf}\left[\frac{\sqrt{2}\eta(d^2\pm i V\theta_j)}{d\sqrt{1+\eta^2(V-1)}}\right].
\end{split}$$
[**(a)**]{}
While a detailed study of the behavior of $B^{ets}$ against $V$ and $d$ is provided in Ref. [@main], here we mention that the effects that detection inefficiencies have on the violation of the Bell-CHSH inequality can indeed be compensated for a large violation of Bell’s inequality by increasing the displacement amplitude $d$. This means that photon losses at detectors can be overcome by increasing the amplitude $d$ of the initial thermal state. However, photon losses [*before*]{} the measurements at the detectors (for example, during the generation process of the ETS), which will cause decoherence of the state itself, may reduce entanglement and destroy the violation of Bell-CHSH inequality. This is what we now ascertain.
We can exploit the formulation of the lossy evolution of a bosonic system given by Phoenix [@phoenix] to find out that the correlation function to use for the evaluation of the Bell parameter of an ETS affected by losses at a rate $\gamma$ can be explicitly calculated, although the analytic expression turns out to be quite lengthy. In Fig. \[fig:dec\][**(a)**]{}, we show the behavior of the Bell parameter against the separation $d$ between the thermal components and the dimensionless dissipation parameter $\gamma{t}$ for $V=10$ (arbitrary choice). Evidently, already for modest values of $d$ an ETS violates the Bell-CHSH inequality for any value of the decoherence parameter. Larger values of $V$ simply require a larger threshold in $d$ to show violation of the Bell-CHSH inequality, which is therefore still a property of our genuinely macroscopic states affected by decoherence. In panel [**(b)**]{} we have compared the Bell functions obtained assigning the value of $V$ and $d$. We see that the increase of $V$ does not boost the decrease of $B^{ets}$ with $\gamma{t}$. Indeed, the slopes of the curves in Fig. \[fig:dec\][**(b)**]{} of an ETS with $V=10$ and $d=5$ (dashed curve) and of a (nearly) pure ECS with $V=1.001$ and $d=5$ (solid curve) are very close to each other. Clearly, a [larger separation in phase space]{} (i.e. larger $d$) causes a quicker destruction of the violation of the Bell-CHSH inequality.
We finally consider a more realistic situation where losses before the unitary operation is present together with detection inefficiency. We include both the effects in our calculations by using beam splitter operations before [*and*]{} after the unitary operation as $\hat{B}_{j j^\prime}(\eta_2)\hat{U}(\theta_j)\hat{B}_{j j^{\prime\prime}}(\eta_1)$, where $\eta_1$ ($\eta_2$) is the parameter determining losses before the unitary operation (the effects of detection efficiency). The results in Fig. \[fig:both\][**(a)**]{} show that the entangled polarization state of $n=3$ causes the required detection efficiency to be larger than 61% for 5% losses before the unitary operation. As shown in Fig. \[fig:both\][**(b)**]{}, the ECS is found to be significantly more efficient: at $\alpha=2$, an ECS shows Bell violations when detection efficiency is larger than $\simeq17\%$ for 15% losses. Considering that about 250m length for traveling photons would be sufficient to be free from the locality loophole [@weihs], this range of values is not far from experimental reality when using telecom fibers [@Gisin02].
We discussed several examples of entangled multi-photon states to show that “macroscopic quantum correlations" may indeed be used in order to overcome limitations in fundamental tests of physics performed with inefficient detectors: the detection loophole can be made less relevant in Bell tests conducted by using states involving a large number of excitations and, as such, verging towards classicality. In contrast to this, losses occurring before the local operations needed to run a Bell-like test affect the quality of the test itself in a way that cannot be corrected by simply considering a more “macroscopic" resource. Our results reveal a previously unknown relation between macroscopic entanglement and measurement efficiency in Bell inequality tests. They reinforce in an unexpected way the idea that quantum features are indeed observable at larger, quasi-macroscopic scales. As we have shown, state macroscopicity can in fact be used to magnify such features and ascertain them in a more reliable way.
The authors thank M. S. Kim for stimulating discussions. This work was supported by the NRF grant funded by the Korea government (MEST) (No. 3348-20100018), the World Class University (WCU) program and the UK EPSRC.
[99]{}
J. S. Bell, Physics [**1**]{}, 195 (1964).
G. Weihs, T. Jennewein, C. Simon, H. Weinfurter, and A. Zeilinger, Phys. Rev. Lett. [**81**]{}, 5039 (1998).
D. N. Matsukevich, P. Maunz, D. L. Moehring, S. Olmschenk, and C. Monroe, Phys. Rev. Lett. [**100**]{}, 150404 (2008).
T. Vértesi, S. Pironio, and N. Brunner, Phys. Rev. Lett. [**104**]{}, 060401 (2010).
J. Hald, J. L. Sørensen, C. Schori, and E. S. Polzik, Phys. Rev. Lett. [**83**]{}, 1319 (1999).
The word “macroscopic” in this paper is used in a relative sense rather than defining a boundary between macroscopic and microscopic states.
M. Paternostro, D. Vitali, S. Gigan, M. S. Kim, C. Brukner, J. Eisert, and M. Aspelmeyer, Phys. Rev. Lett. [**99**]{}, 250401 (2007).
W. H. Zurek, Phys. Today [**44**]{}, 36 (1991).
P. Walther, J.-W. Pan, M. Aspelmeyer, R. Ursin, S. Gasparoni, and A. Zeilinger, Nature [**429**]{}, 158 (2004).
W.-B. Gao, C.-Y. Lu, X.-C. Yao, P. Xu, O. Gühne, A. Goebel, Y.-A. Chen, C.-Z. Peng, Z.-B. Chen, and J.-W. Pan, Nature Phys. [**6**]{}, 331 (2010).
F. De Martini, F. Sciarrino, and C. Vitelli, Phys. Rev. Lett. [**100**]{}, 253601 (2008); P. Sekatski, N. Brunner, C. Branciard, N. Gisin, and C. Simon, Phys. Rev. Lett. [**103**]{}, 113601 (2009); E. Pomarico [*et al.*]{}, arXiv:1104.2212; N. Spagnolo, C. Vitelli, M. Paternostro, F. De Martini, and F. Sciarrino, Phys. Rev. A [**84**]{}, 032102 (2011).
H. Jeong, M. Paternostro, T. C. Ralph, Phys. Rev. Lett. [**102**]{}, 060403 (2009).
M. Stobińska, P. Sekatski, A. Buraczewski, N. Gisin, G. Leuchs, Phys. Rev. A [**84**]{}, 034104 (2011).
V. Bužek, M. S. Kim, and M. G. Kim, J. Kor. Phys. Soc. [**28**]{}, 123 (1995).
J. Kofler and Č. Brukner, Phys. Rev. Lett. [**99**]{}, 180403 (2007); Phys. Rev. Lett. [**101**]{}, 090403 (2008).
S. Raeisi, P. Sekatski, and C. Simon, Phys. Rev. Lett. [**107**]{}, 250401 (2011).
S. Sefi and P. van Loock, Phys. Rev. Lett. [**107**]{}, 170501 (2011).
P. Marek, R. Filip, and A. Furusawa, Phys. Rev. A [**84**]{}, 053802 (2011).
J. F. Clauser, M. A. Horne, A. Shimony, and R. A. Holt, Phys. Rev. Lett. [**23**]{}, 880 (1969).
A. Garg and N. D. Mermin, Phys. Rev. D [**35**]{}, 3831 (1987).
B. C. Sanders, Phys. Rev. A [**45**]{}, 6811 (1992).
A. Ourjoumtsev, H. Jeong, R. Tualle-Brouri, and Ph. Grangier, Nature (London) **448**, 784 (2007); H. Takahashi [*et al.*]{}, Phys. Rev. Lett. **101**, 233605 (2008).
A. Ourjoumtsev, F. Ferreyrol, R. Tualle-Brouri, and P. Grangier, Nature Physics [**5**]{}, 189 (2009).
M. Stobińska, H. Jeong, and T. C. Ralph, Phys. Rev. A [**75**]{}, 052105 (2007).
M. Paternostro and H. Jeong, Phys. Rev. A [**81**]{}, 032115 (2010).
G. McKeown, F.L. Semiao, H. Jeong, and M. Paternostro, Phys. Rev. A [**82**]{}, 022315 (2010); C.-W. Lee, M. Paternostro, and H. Jeong, Phys. Rev. A [**83**]{}, 022102 (2011).
G. McKeown, M.G.A. Paris, and M. Paternostro, Phys. Rev. A [**83**]{}, 062105 (2011).
H. Jeong and T.C. Ralph, Phys.Rev.Lett. [**97**]{}, 100401 (2006); Phys.Rev.A [**76**]{}, 042103 (2007).
S. J. D. Phoenix, Phys. Rev. A [**41**]{}, 5132 (1990).
N. Gisin, G. Ribordy, W. Tittel, and H. Zbinden, Rev. Mod. Phys. [**74**]{}, 145 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Ioan Ghisoiu,'
- 'Tyler Gorda,'
- 'Aleksi Kurkela,'
- 'Paul Romatschke,'
- 'Matias Säppi,'
- and Aleksi Vuorinen
title: 'On high-order perturbative calculations at finite density'
---
Introduction \[sec:intro\]
==========================
Understanding the properties of cold and dense strongly interacting matter is known to be a very challenging task. With the Sign Problem preventing a lattice approach [@deForcrand:2010ys], the first-principles methods available for describing, e.g., the internal composition of neutron stars, are limited to Chiral Effective Theory at low baryon densities [@Machleidt:2011zz] and perturbative QCD at ultrahigh density [@Kraemmer:2003gd]. In order to decrease the currently sizable error bars of the Equation of State (EoS) of neutron star matter [@Hebeler:2013nza], it is thus clear that these two approaches should be actively pushed to higher orders. Indeed, it has been shown recently that a systematic interpolation between the low- and high-density limits can be efficiently used to restrict the behavior of the neutron star EoS at all densities, provided that the asymptotic limits are accurate enough [@Kurkela:2014vha; @Fraga:2015xha].
The current state-of-the-art result for the perturbative EoS of zero-temperature quark matter is from a three-loop, or ${\mathcal O}(\alpha_s^2)$, calculation that was first performed at vanishing quark masses [@Freedman:1976ub; @Vuorinen:2003fs], but later generalized to nonzero quark masses [@Kurkela:2009gj] (see also [@Fraga:2004gz]) and small but nonvanishing temperatures [@Kurkela:2016was]. Extending these zero-temperature results to higher orders, however, presents a considerable technical challenge. Similarly to the case of high temperatures [@Kajantie:2002wa; @DiRenzo:2006nh], part of the problem in extending these results lies in understanding how to handle the contributions of the soft momentum scales to the quantity. These difficulties will be addressed in a forthcoming publication, containing the logarithmic $\alpha_s^3\ln\,\alpha_s$ and $\alpha_s^3\ln^2\,\alpha_s$ contributions to the perturbative EoS [@nextpaper]. A more challenging part of the full ${\mathcal O}(\alpha_s^3)$ result is, however, the contribution of the hard energy scale $\mu_B$, i.e. the baryon chemical potential, which is obtained from the sum of all four-loop bubble diagrams of QCD. The high-temperature counterpart of this computation has turned out to be extremely challenging, and has only been worked out in $\phi^4$ theory [@Gynther:2007bw] as well as for the large-$N_f$ limit of QCD [@Gynther:2009qf].
In the paper at hand, we present a new technical tool for perturbative calculations at zero temperature but finite chemical potentials that we argue enables a high-order determination of many important thermodynamic quantities. This tool is referred to as a set of “cutting rules”, which were proposed but not explicitly derived in ref. [@Kurkela:2009gj]. They concern Feynman integrals at zero temperature and finite fermionic chemical potentials, and reduce the evaluation of the original One-Particle-Irreducible (1PI) Feynman graph to the computation of three-dimensional phase space integrals over on-shell vacuum ($T=\mu=0$) amplitudes. This represents a remarkable simplification for practical calculations, as there is a vast amount of literature on vacuum amplitudes that can be directly taken over. This significantly streamlines the evaluation of multi-loop Feynman graphs.
Although our derivation of the cutting rules utilizes the imaginary-time formalism of thermal field theory, it is interesting to note that the result appears to have an intimate connection to the real-time formalism as well. In particular, it can be shown that (modulo some simple additional assumptions) our cutting rules would emerge from a naive replacement of Euclidean propagators by the time-ordered propagators of the real-time formalism, closely reminiscent of eq. (4) of ref. [@Andersen:2000zn]. This is sometimes referred to as the “naive real-time formalism”. This result dates back to the much earlier work of Dashen et al. [@Dashen:1969ep], where a connection between certain statistical-physics quantities and scattering amplitudes was proposed, and it has since then been developed, e.g., in [@Frenkel:1992az; @Bugrii:1995vn]. It is, however, important to note that the formalism has been proposed only for vacuum diagrams, and even there no proof to all orders exists; rather, the validity of the replacement has only been checked on a case-by-case basis up to partial three-loop order. In contrast, our proof of the zero-temperature cutting rules covers all Euclidean $n$-point functions up to an arbitrary order in perturbation theory, thereby validating the use of the naive real-time formalism for these quantities.
Our paper is organized as follows. In section \[sec:cut\], we introduce our notation and state the cutting rules. In addition, as an illustration, we present a simple two-loop computation in two ways: both without and with the help of the cutting rules. Section \[sec:proof\] then contains a detailed proof of the rules, as well as two intermediate lemmas that are each interesting in their own right. After this, we discuss the regularization of the most common divergences occurring in the cut graphs in section \[sec:reg\], while section \[sec:conc\] presents an outlook on the applications of the cutting rules, in particular in dense QCD. Lastly, many details of the more subtle parts of our proof have been relegated to the appendixes.
Cutting rules \[sec:cut\]
=========================
Notation and statement of the rules \[sec:state\]
-------------------------------------------------
We work with Euclidean signature Feynman graphs at zero temperature and finite chemical potentials. This means that we can think of our diagrams as consisting of two types of propagators, “fermionic” $1/((q_0+i\mu)^2+(E_q^i)^2)$ and “bosonic” $1/(q_0^2+(E_q^i)^2)$, where $E_q^i\equiv \sqrt{q^2+m_i^2}$ and $\mathbf{q}$ represents a spatial momentum vector. Consistently with the fermionic nature of the chemical potential, we assume $\mu$ to be larger than the mass of the corresponding field. Divergences are finally regulated via dimensional regularization by working in $d=3-2\epsilon$ spatial dimensions, defining $$\begin{aligned}
\int_Q\,\equiv\, \int_{-\infty}^\infty \frac{dq_0}{2\pi}\int_q \,\equiv\,\int_{-\infty}^\infty \frac{dq_0}{2\pi} \int\frac{d^{d}q}{(2\pi)^{d}}\, ,\end{aligned}$$ where $Q$ denotes a Euclidean four-vector, such that $Q^2\equiv q_0^2+q^2$. Before stating the cutting rules, we make a few simplifying assumptions, the purpose of which is to keep our presentation as concise and readable as possible:
- There is no structure in the numerator of the original Feynman integral, i.e. we consider scalar propagators and trivial vertex functions.
- No individual propagator is raised to a power higher than one.
- There is only one chemical potential appearing in the graph.
- In the external momenta $P_k=(p_0^k,\mathbf{p}_k)$, the $p_0^k$ are always real-valued, corresponding to imaginary frequencies $\omega_k$.
As will be discussed in section \[sec:conc\], the first three of these assumptions can be easily relaxed, but at the cost of making the notation somewhat more convoluted. Note, however, that we have made absolutely no assumptions about the masses of the propagators, so that they can and will be considered independent.
Under the above assumptions, let us consider an arbitrary 1PI $N$-loop $n$-point Feynman graph $F(\{P_k\},\mu)$, where the $P_k$, $k=1,2,...,n$ stand for the external momenta. According to the cutting rules, we may write this function in the form $$\begin{aligned}
F(\{P_k\},\mu)&=&F_\text{0-cut}(\{P_k\})+F_\text{1-cut}(\{P_k\},\mu)+\cdots+F_\text{$N$-cut}(\{P_k\},\mu), \end{aligned}$$ where $F_\text{0-cut}(\{P_i\})$ is simply the original graph evaluated at vanishing chemical potential, $\mu=0$, while the remaining pieces result from the cutting procedure. In particular, $F_\text{$j$-cut}(\{P_k\})$ denotes the sum of all so-called “$j$-cut” diagrams, in which exactly $j$ of the internal fermionic propagators have been cut off. This cutting procedure involves the following steps:
1. Removing the cut propagators from the original graph.
2. Evaluating the resulting $N-j$ -loop $n+2j$ -point amplitude at $T=\mu=0$, assuming all external momenta to be real-valued.
3. Setting the cut momenta $Q_i$ on shell, i.e. writing $q_0^i=iE_i$ for each of them.
4. Integrating the resulting expression over the cut three-dimensional momenta with the weights $-\theta(\mu-E_i)/(2E_i)$.
An important additional rule is that those cuts that divide the original graph into two or more disconnected pieces are to be thrown out.
The usefulness of the cutting rules originates from the fact that they isolate the chemical-potential dependence of the original graph in the $\theta$-function factors in the three-dimensional “phase space” integrations. Owing to the abundance of analytic results for vacuum amplitudes in the literature, one typically only needs to perform (some of) these phase space integrations numerically, which is an enormous simplification.
Example calculation: standard technique
---------------------------------------
![\[fig:mot1\] A two-loop diagram contributing to the partition functions of QED and QCD, with the wiggly line corresponding to a gauge boson (photon or gluon) and the solid lines to fermions (electrons or quarks).](twoloop_gluons.pdf){width=".8\linewidth"}
To illustrate the use of the cutting rules as well as motivate their introduction, we consider next an elementary practical example in the form of a two-loop integral encountered when considering the graph of fig. \[fig:mot1\], appearing in the ${\mathcal O}(\alpha_s)$ contribution to the EoS of QED or QCD matter. At zero temperature, the most nontrivial master integral resulting from this diagram is the two-loop “sunset”
[2loop]{} $$\begin{aligned}
\label{eq:abc2}
\!\!I_2(\mu)&\equiv&\!\!\!\!\!\!
\begin{gathered}
\begin{fmfgraph*}(65,60)
\fmfleft{in1}
\fmfright{out1}
\fmf{phantom,tension=5}{in1,vin1}
\fmf{phantom,tension=5}{vout1,out1}
\fmf{plain,left,tension=0.4}{vin1,vout1,vin1}
\fmf{dots,width=2}{vin1,vout1}
\end{fmfgraph*}
\end{gathered}
\!\!\!\!\!=\int_P \int_Q \frac{1}{(p_0+i\mu)^2+E_p^2}\frac{1}{(q_0+i\mu)^2+E_q^2}\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2} \\
&=&\int_P \int_Q \int_K \frac{(2\pi)^4\delta^{(4)}(P-Q-K)}{\left((p_0+i\mu)^2+E_p^2\right)\left((q_0+i\mu)^2+E_k^2\right)\left(k_0^2+k^2\right)}\,,\quad E_k\equiv\sqrt{k^2+m^2}\,, \nonumber\end{aligned}$$ where the solid lines in the graph denote a massive fermion propagator and the dotted line a massless boson one. In order to reduce the integral to a more manageable form, we first perform the integrations over the 0-components of the different momenta, which is a rather straightforward task at such a low loop order.
The standard way of evaluating the $p_0$- and $q_0$-integrals proceeds by writing the $\delta$-function in eq. (\[eq:abc2\]) in the form [@Kapusta:2006pm] $$\label{eq:delta}
2\pi \delta(p_0-q_0-k_0)=\int_{-\infty}^\infty d\alpha\, e^{i\alpha(p_0-q_0-k_0)}\,,$$ which allows us to perform the $p_0$, $q_0$, and $k_0$ integrations independently using the Residue theorem. Choosing the integration contours to lie on the upper or lower half of the complex plane depending on the sign of the exponent in eq. (\[eq:delta\]), we obtain after quite some algebra $$\begin{aligned}
I_2(\mu)&=&\int_{p,q}\frac{1}{2|\mathbf{p}-\mathbf{q}|\,2E_p \,2E_q}\frac{2}{|\mathbf{p}-\mathbf{q}|+E_p+E_q}\nonumber\\
&&-\int_{p,q}\frac{\theta(\mu-E_p)}{2|\mathbf{p}-\mathbf{q}|\, 2E_p\, 2E_q}\frac{2 (|\mathbf{p}-\mathbf{q}|+E_q)}{(|\mathbf{p}-\mathbf{q}|+E_q)^2-E_p^2}\nonumber\\
&&-\int_{p,q}\frac{\theta(\mu-E_q)}{2|\mathbf{p}-\mathbf{q}|\, 2E_p\, 2E_q}\frac{2 (|\mathbf{p}-\mathbf{q}|+E_p)}{(|\mathbf{p}-\mathbf{q}|+E_p)^2-E_q^2}\nonumber\\
&&+\int_{p,q}\frac{\theta(\mu-E_p) \theta(\mu-E_q)}{2E_p\,2E_q} \frac{1}{(\mathbf{p}-\mathbf{q})^2-(E_p-E_q)^2}\, . \label{I2int} \end{aligned}$$ While perfectly correct, this result is unfortunately rather unpractical, as the first three lines contain complicated UV divergent integrals of highly non-standard objects that we would need to determine in $3-2\epsilon$ dimensions. Only the last of the four terms is of a form that may be directly evaluated as a numerical integral.
Some insights into how the above integrals might become tractable can be gained by observing that the first term of eq. (\[I2int\]), containing no step functions, may be recast in the form of the original diagram evaluated at $\mu=0$, $$\begin{aligned}
&&\int_{p,q}\frac{1}{2|\mathbf{p}-\mathbf{q}|\,2E_p \,2E_q}\frac{2}{|\mathbf{p}-\mathbf{q}|+E_p+E_q} \nonumber \\
&=& \int_P \int_Q \frac{1}{\left(p_0^2+E_p^2\right)\left(q_0^2+E_k^2\right)\left((p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2\right)} = I_2(0).
\label{eq:mot1}\end{aligned}$$ Apart from being of limited physical interest, this term can be easily evaluated using the standard tools of perturbative zero-temperature field theory.
Perhaps even more interestingly, we note that the second term of eq. (\[I2int\]) is expressible in terms of a two-point function, $$\begin{aligned}
\int_Q \frac{1}{q_0^2+E_q^2}\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Big{|}_{p_0\to iE_p}=\int_q\frac{1}{2|\mathbf{p}-\mathbf{q}|\,2E_q}\frac{2 (|\mathbf{p}-\mathbf{q}|+E_q)}{(|\mathbf{p}-\mathbf{q}|+E_q)^2-E_p^2},\nonumber \\ \label{eq:I21}\end{aligned}$$ where the substitution $p_0\to iE_p$ is made only after performing the integral on the left-hand side (prior to this substitution, $p_0$ is considered real). Again, we note that the integral on the left-hand side is of a form often encountered in standard $T=0$ quantum field theory calculations. A similar result clearly exists for the third term of eq. (\[I2int\]) as well, via the substitution $\mathbf{p}\leftrightarrow\mathbf{q}$.
Example calculation: cutting rules
----------------------------------
The observations made in the previous subsection clearly suggest that the two-loop integral of eq. (\[eq:abc2\]) may be written in a form reminiscent of the cutting rules, i.e. as a sum of phase-space integrals of higher-point amplitudes. To make this statement more precise, we shall now explicitly show that the cutting rules indeed exactly reproduce the above results.
According to the cutting rules, the integral $I_2(\mu)$ of eq. (\[eq:abc2\]) can be directly written in the form $$\begin{aligned}
I_2(\mu)&\equiv& I_2^\text{0-cut}+ I_2^\text{1-cut}(\mu)+I_2^\text{2-cut}(\mu)\,.
\end{aligned}$$ Here, the first term reads $$\begin{aligned}
I_2^\text{0-cut}&=& I_2(0)\:=\!\!\!\!
\begin{gathered}
\begin{fmfgraph*}(60,60)
\fmfleft{in2}
\fmfright{out2}
\fmf{phantom,tension=5}{in2,vin2}
\fmf{phantom,tension=5}{vout2,out2}
\fmf{dashes,left,tension=0.4}{vin2,vout2,vin2}
\fmf{dots,width=2}{vin2,vout2}
\end{fmfgraph*}
\end{gathered}\!\!\!\!,\end{aligned}$$ where we have denoted a $\mu=0$ massive propagator by a dashed line. This clearly agrees with eq. (\[eq:mot1\]). For the one-cut piece, on the other hand, we obtain $$\begin{aligned}
I_2^\text{1-cut}(\mu)&=& -\int_p \frac{\theta(\mu-E_p)}{2E_p} \Bigg[\int_Q \frac{1}{q_0^2+E_q^2}\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Bigg]_{p_0\to iE_p} \nonumber \\
&-&\int_q \frac{\theta(\mu-E_q)}{2E_q} \Bigg[\int_P \frac{1}{p_0^2+E_p^2}\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Bigg]_{q_0\to iE_q} \nonumber
\\
&\equiv&
-\:2\intop_{p}\frac{\theta(\mu - E_{p})}{2 E_{p}}\;
\begin{gathered}
\begin{fmfgraph*}(60,60)
\fmfleft{in3}
\fmfright{out3}
\fmf{dashes,tension=5}{in3,vin3}
\fmf{dashes,tension=5}{vout3,out3}
\fmf{dashes,left,tension=0.4}{vin3,vout3}
\fmf{dots,width=2}{vin3,vout3}
\end{fmfgraph*}
\end{gathered}\,
\Bigg{|}_{p_{0}\rightarrow iE_{p}}\,, \label{sunset1}\end{aligned}$$ where we have used the symmetry of the two terms in the first form of the result. This can be identified with the second and third terms of eq. (\[I2int\]), using eq. (\[eq:I21\]). Finally, the two-cut part of the graph takes the form $$\begin{aligned}
I_2^\text{2-cut}(\mu)&=& \int_p \frac{\theta(\mu-E_p)}{2E_p}\int_q \frac{\theta(\mu-E_q)}{2E_q} \Bigg[\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Bigg]_{p_0\to iE_p,\,q_0\to iE_q}\;\;\nonumber
\\
&\equiv&
\intop_{p}\frac{\theta(\mu - E_{p}) }{2 E_{p}}
\intop_{q}\frac{\theta(\mu - E_{q}) }{2 E_{q}}
\begin{gathered}
\begin{fmfgraph*}(80,40)
\fmfleft{ind4,inu4}
\fmfright{outd4,outu4}
\fmf{dashes,tension=0.3}{ind4,vin4,inu4}
\fmf{dashes,tension=0.3}{outd4,vout4,outu4}
\fmf{dots,width=2,tension=0.2}{vin4,vout4}
\end{fmfgraph*}
\end{gathered}
\Bigg{|}_{p_{0}\rightarrow iE_{p},q_{0}\rightarrow iE_{q}}, \label{sunset2}\end{aligned}$$ which is nothing but the last term of eq. (\[I2int\]). Thus, the cutting rules do indeed reproduce the full result for the graph considered. It is worth pointing out here that this result is by no means new and only serves as a pedagogical introduction to our discussion; in a finite-$T$ context, a similar result has been obtained, e.g., in Appendix A of [@Laine:2006cp].
Proof of the rules \[sec:proof\]
================================
Organization of the proof
-------------------------
In this section, we provide a proof of the cutting rules in the case of a generic 1PI Feynman graph at zero temperature but finite chemical potential. For reasons of clarity, we do this in three sequential steps, of which the first two can be considered useful lemmas while the connection to the cutting rules is made in the third part of the proof:
1. *Vanishing-chemical-potential case*: Consider a generic Euclidean Feynman integral at zero temperature and vanishing chemical potential, corresponding to a 1PI $N$-loop $n$-point diagram. Such a graph consists of some (potentially large) number $M$ of internal lines, which we enumerate by the index $\alpha=1,2,...,M$. The corresponding propagators can be written in the form $$\begin{aligned}
\frac{1}{(r_0^\alpha)^2+E_\alpha^2}, \quad E_\alpha\equiv\sqrt{r_\alpha^2+m_\alpha^2}\, ,
\end{aligned}$$ of which exactly $N$ can be chosen to correspond to the loop momenta $Q_i$, $i=1,...,N$, that are integrated over. The remaining $M-N$ $R_\alpha$ are then linear combinations of the loop momenta and the external momenta $P_k$, $k=1,...,n$, as dictated by momentum conservation at the vertices (see appendix \[sec:mom\] for a discussion of this issue).
Our claim is now that the integral $$\begin{aligned}
I&\equiv& \int_{-\infty}^\infty \prod_{i=1}^{N} \frac{dq_0^i}{2\pi}\, \prod_{\alpha=1}^M \frac{1}{(r_0^\alpha)^2+E_{\alpha}^2}, \label{Ires}
\end{aligned}$$ where we have made some arbitrary choice for the loop momenta, can be written in a simple form that is explicitly independent of this choice. To write down this result, we introduce the following notation: denoting the set of all propagators by $P\equiv\{1,2,...,M\}$, we define $S$ to be the set of all possible choices of loop momenta, such that each element $S_r\in S$ corresponds to some subset of $N$ indices from $P$. The sets $S_r$ are limited only by momentum conservation, and several examples of them are given in appendix \[sec:mom\].
With the above notation, our proposed result for $I$ reads $$\begin{aligned}
I&=&\sum_{S_r\in S}\;\prod_{i\in S_r} \frac{1}{2E_i} \prod_{\alpha\in P\setminus S_r} \frac{1}{(r_0^\alpha(S_r))^2+E_\alpha^2(S_r)}\Big{|}_{\{q_0^i=iE_i\}}, \label{Iresmu0}\end{aligned}$$ where $P\setminus S_r$ denotes the propagators that do not belong to the set $S_r$ and the explicit forms of the $R_\alpha$ in terms of the $Q_i$ and $P_k$ are dictated by $S_r$. Note that each set $S_r$ is to be counted only once here, and the labeling of the momenta within $S_r$ (including changing the direction of some loop momenta, $Q_i\to-Q_i$) plays no role: it is only the choice of propagators that counts.
2. *Generalization to finite density*: Assume next that some of the internal propagators of the graph are fermionic in the sense that they carry a chemical potential $\mu$ in the way stated in the previous section. The above result for $I$ is then modified only by the factors $1/(2E_i)$ corresponding to the internal fermion lines changing according to $$\begin{aligned}
\frac{1}{2E_i}&\to& \frac{\theta(E_i-\mu)}{2E_i} \, .\end{aligned}$$ Note that for different $S_r\in S$, the numbers of fermionic momenta and thus $\theta$-function factors are typically different.
3. *Connection to the original cutting rules*: Writing the $\theta$-functions in the form $\theta(E_i-\mu) = 1-\theta(\mu-E_i)$ and rearranging terms, the above results reduce to the cutting rules stated in section \[sec:cut\].
We now proceed to prove these three claims, thereby deriving the cutting rules.
Vanishing-chemical-potential case
---------------------------------
Given a random choice of loop momenta $S_r\in S$, we may clearly express the integral we wish to evaluate \[cf. eq. (\[Ires\])\] in the form $$\begin{aligned}
I(S_r)&\equiv& \int_{-\infty}^\infty \prod_{i \in S_r}\frac{dq_0^i}{2\pi}\, \prod_{\alpha=1}^M \frac{1}{(r_0^\alpha(S_r))^2+E_{\alpha}^2(S_r)}\,, \label{Iresx}\end{aligned}$$ where our notation highlights the fact that the $R_\alpha$ and corresponding $E_\alpha$ depend on $S_r$. Picking some $i \in S_r$ as the first integration to be performed gives a sum of residues of the form $\frac{1}{2E_{\alpha}}\left(\ldots\right)|_{q_0^i = iE_\alpha+...}$, where each $\mathbf{r}_\alpha$ depends linearly on the $\mathbf{q}_i$. Placing the momentum in question on shell shifts the poles of some of the remaining propagators but does not affect the corresponding residues. Upon subsequent integrations, complicated combinations of $\theta$-functions typically appear in the numerator of the result due to these shifts. What remains unchanged, however, is that each integration produces an additional factor of $1/(2E_{\alpha'})$, originating from the residue of one of the remaining propagators.
To illustrate the above reasoning, consider the following simple example, where we perform three $q_0$ integrations, always picking up the pole from the highlighted propagator: $$\begin{aligned}
&&\int_{q_0^1}\int_{q_0^2}\int_{q_0^3}\mathbf{\frac{1}{(q_0^1)^2+E_1^2}}\frac{1}{(q_0^2)^2+E_2^2}\frac{1}{(q_0^3)^2+E_3^2} \frac{1}{(q_0^1+q_0^2)^2+E_4^2}\frac{1}{(q_0^1+q_0^2-q_0^3)^2+E_5^2} \nonumber \\
&=& \int_{q_0^2}\int_{q_0^3}\frac{1}{2E_1}\frac{1}{(q_0^2)^2+E_2^2}\frac{1}{(q_0^3)^2+E_3^2}\mathbf{\frac{1}{(iE_1+q_0^2)^2+E_4^2}}\frac{1}{(iE_1+q_0^2-q_0^3)^2+E_5^2}+\cdots \nonumber \\
&=& \int_{q_0^3}\frac{1}{2E_1}\frac{\theta(E_4-E_1)}{2E_4}\frac{1}{(iE_4-iE_1)^2+E_2^2}\frac{1}{(q_0^3)^2+E_3^2}\mathbf{\frac{1}{(iE_4-q_0^3)^2+E_5^2}} +\cdots\nonumber \\
&=& \frac{1}{2E_1}\frac{\theta(E_4-E_1)}{2E_4}\frac{\theta(E_5-E_4)}{2E_5}\frac{1}{(iE_4-iE_1)^2+E_2^2}\frac{1}{(iE_4-iE_5)^2+E_3^2} +\cdots\, .\end{aligned}$$ Note that additional terms of the exact same form but with different $\theta$-function factors originate from taking the same poles in a different order.
From the above exercise, we see that $I(S_r)$ obtains the form of a lengthy sum of terms of a similar kind: a product of residues $1/(2E_\alpha)$ from some set ${\mathcal S}_s$ of $N$ propagators, multiplied by a complicated sum of products of $\theta$ functions along with the product of the remaining propagators, with the ${\mathcal S}_s$ momenta placed on shell. Defining a function $A_{S_r}({\mathcal S}_s)$ to stand for the latter part of the result, we may write it in the form $$\begin{aligned}
I(S_r)&=&\sum_{{\mathcal S}_s} \prod_{k\in {\mathcal S}_s}\frac{1}{2E_k}\,A_{S_r}({\mathcal S}_s)\Big{|}_{\{q_0^k=iE_k\}}.\end{aligned}$$ A crucial realization is now the following: the sets of $N$ propagators obtained above, i.e. the ${\mathcal S}_s$, cannot contain any sets that are not part of the “superset” $S$, defined as all the possible choices of loop momenta in the original graph. This is a simple consequence of momentum conservation: we cannot take residues of a set of propagators whose momenta are linearly dependent. This implies that we may directly write the result in a form reminiscent of eq. (\[Iresmu0\]), $$\begin{aligned}
I(S_r)&=&\sum_{S_{r'}\in S}\;\prod_{i\in S_{r'}} \frac{1}{2E_i}\, \tilde{A}_{S_r}(S_{r'})\Big{|}_{\{q_0^i=iE_i\}}, \label{Iresmu1}\end{aligned}$$ where the tilde on $A$ highlights the fact that the summation now goes over the sets $S_{r'}$.
The remaining step in relating the above result to eq. (\[Iresmu0\]) is to use the known independence of $I(S_r)$ on the random initial set $S_r$, i.e. the fact that $I(S_{r})=I$. For any given $S_r\in S$, there is one term in the sum of eq. (\[Iresmu1\]) that is particularly simple, namely that where $S_{r'}=S_{r}$. For this term, each of the $q_0^i$ integrations picks up a pole from a propagator of the simple form $1/((q_0^i)^2+E_i^2)$, so that no $\theta$-functions arise, producing $$\begin{aligned}
\tilde{A}_{S_r}(S_{r})&=& \prod_{\alpha\in P\setminus S_r} \frac{1}{(r_0^\alpha(S_r))^2+E_\alpha^2(S_r)}\, . \label{ASr}\end{aligned}$$ Owing to the independence of the $E_\alpha$, we on the other hand know that the different terms in the sum of eq. (\[Iresmu1\]) must be unique (see appendix \[sec:ind\] for a detailed discussion of this point), so that $\tilde{A}_{S_r}(S_{r'})=\tilde{A}_{S_{r'}}(S_{r'})$ for all $S_r, S_{r'}$. This implies that also the coefficients $\tilde{A}_{S_r}(S_{r'})$, $r\neq r'$, must reduce to the simple form $$\begin{aligned}
\tilde{A}_{S_r}(S_{r'})&=& \prod_{\alpha\in P\setminus S_{r'}} \frac{1}{(r_0^\alpha(S_{r'}))^2+E_\alpha^2(S_{r'})}\,, \label{ASr2}\end{aligned}$$ which — together with eq. (\[Iresmu1\]) — leads us directly to the result (\[Iresmu0\]).
Generalization to finite density
--------------------------------
Somewhat surprisingly, the generalization of the above result to the presence of nonzero $\mu$ in some of the propagators is by far the simplest part of our proof. Namely, the exact same reasoning goes through as in the $\mu=0$ case, with the only modification being a shift in the poles and residues of the fermion propagators originating from the $\mu$-dependence of the integral $$\begin{aligned}
\int_{-\infty}^\infty dq_0^i\, \frac{1}{(q_0^i+i\mu)^2+E_i^2}&=&\frac{\theta(E_i-\mu)}{2E_i}\, .\end{aligned}$$ In other words: whenever the pole of a fermionic propagator is taken, we need to multiply the corresponding residue in the product $\prod_{i\in S_{r'}} \frac{1}{2E_i}$ by the factor $\theta(E_i-\mu)$.
Connection to the original cutting rules
----------------------------------------
The previous step of the proof brought us to a somewhat cumbersome result, featuring a sum over all possible labelings of loop momenta in the original Feynman graph, with each term in the sum further containing a product of some number of $\theta(E_i-\mu)$ factors. To move forward, we write each of the $\theta$-functions in the form $1-\theta(\mu-E_i)$, and then reassemble the result for $I$ in the form $$\begin{aligned}
I&=& \Big(\text{terms with 0 $\theta(\mu-E_i)$'s}\Big)+\Big(\text{terms with 1 $\theta(\mu-E_i)$}\Big)+\cdots \nonumber \\
&&+ \Big(\text{terms with $N$ $\theta(\mu-E_i)$'s}\Big)\, .\end{aligned}$$ It is self-evident that the first term in the above sum corresponds to the $\mu=0$ version of the same graph, but a little more effort is required to see what happens to the terms with one or more $\theta$-functions.
In the single-$\theta$ part of the result, we first group together terms according to the argument of the $\theta(\mu-E_i)$ function they contain, which clearly correspond to all the fermionic propagators in the original graph. Singling out one of them (and the associated $-\frac{1}{2E_i}$ factor), we note that it is multiplied by a sum of terms, each of which contains a product of $N-1$ factors of $1/(2E_j)$ as well as the product of the rest of the propagators with the $E_i$ and $E_j$ lines placed on shell. Recalling the result of the first part of our proof, we recognize this as the result for an $N-1$ -loop $n+2$ -point function that is obtained by removing the line $i$ from the original graph, i.e. $$\begin{aligned}
I&=&\cdots -\frac{\theta(\mu-E_i)}{2E_i} \times \Big(\text{original graph with $Q_i$-propagator removed}\Big)\Big{|}_{q_0^i=iE_i} + \cdots \, , \nonumber \\\end{aligned}$$ so that the sum of all such terms exactly corresponds to the sum of all 1-cut graphs in the cutting rules. In evaluating this expression, the $Q_i$ line clearly must be placed on-shell only after computing the associated $n+2$ -point function, as one of the assumptions of the $\mu=0$ cutting rules above was that all external momenta in the original graph be real-valued (modulo a possible $\mu$ in the external legs of the original graph).
Moving on to the terms with two or more $\theta$-function factors, the above reasoning goes through in each case, leaving us with the 2-, 3-, ..., and $N$-cut contributions to the original graph. In each case, the generated amplitudes are to be evaluated assuming the external momenta to be real: only afterwards are the cut momenta placed on-shell.
One final comment is in order. Each time some number of fermion lines are cut in a given Feynman graph, it follows from the construction presented above that these propagators must form a subset of some possible choice of integration momenta $S_r\in S$. This means that the cuts can never split the original 1PI graph into two (or more) disconnected pieces: for this to happen, we would need to cut propagators whose momenta are not linearly independent, which is not possible for any subset of $S_r$.
Regularization of the integrals \[sec:reg\]
===========================================
Before we can successfully apply the cutting procedure to an arbitrary Feynman diagram, there is one further issue that needs to be discussed. This is related to the regularization of unphysical divergences appearing in the calculations, of which there are two distinct variations. They differ in that the first kind of divergence appears in the very definition of the finite-$\mu$ integrals, while the latter is a byproduct of the cutting procedure and therefore more artificial.
The first type of singularity is related to the divergence of the fermionic propagator $1/((p_0+i\mu)^2+E_p^2)$ when $p_0=0$ and $E_p=\mu$, i.e. it appears along the original integration contour. It gets realized only in the limit where the $\theta(\mu-E_p)$ function in the integration measure gets saturated, but one might nevertheless worry that it makes the $p_0$ integrations ill-defined. The most natural resolution to this problem turns out to involve the use of an infinitesimal but nonzero temperature $T$ as a regulator. As we shall show in detail in appendix \[sec:matsu\], the effect of this prescription amounts to interpreting all $p_0$ integrations in the principal value sense, i.e. writing $$\begin{aligned}
\int_{-\infty}^\infty \frac{dp_0}{2\pi}\int_p &\to&\mathcal{P}\int_{-\infty}^\infty \frac{dp_0}{2\pi}\int_p\;=\; \lim_{\epsilon\to 0^+}\Bigg\{\int_{-\infty}^{-\epsilon} \frac{dp_0}{2\pi}+\int_{\epsilon}^{\infty} \frac{dp_0}{2\pi}\Bigg\}\int_p \nonumber \\
&=&\frac{1}{2}\,\Bigg\{\int_{-\infty+i0^+}^{\infty+i0^+} \frac{dp_0}{2\pi}+\int_{-\infty-i0^+}^{\infty-i0^+} \frac{dp_0}{2\pi}\Bigg\}\int_p\,, \label{principal}\end{aligned}$$ where we have assumed $p_0=0$ to be the only singular point on the real axis. While this does not affect the practical application of the cutting rules, it demonstrates that the integrand is well-defined on the entire integration contour, so that no imaginary parts can be generated in bubble graphs due to the divergence.
Another frequently occurring problem is the emergence of spurious poles in some of the cut graphs that would automatically cancel, should all of the $p_0$ integrations in the diagram be computed at the same time and the results added together. A simple example of this is seen in the integral $$\begin{aligned}
&&\int_{-\infty}^\infty\frac{dp_0}{2\pi}\int_{-\infty}^\infty\frac{dq_0}{2\pi}\frac{1}{p_0^2+E_1^2}\frac{1}{q_0^2+E_2^2}\frac{1}{(p_0-q_0)^2+E_3^2}\nonumber \\
&=& \frac{1}{2E_1}\frac{1}{2E_2}\frac{1}{(iE_1-iE_2)^2+E_3^2}+\frac{1}{2E_1}\frac{1}{2E_3}\frac{1}{(iE_1+iE_3)^2+E_2^2}+\frac{1}{2E_2}\frac{1}{2E_3}\frac{1}{(iE_2+iE_3)^2+E_1^2}\nonumber \\
&=&\frac{1}{4E_1 E_2 E_3(E_1+E_2+E_3)}\,,\end{aligned}$$ where the intermediate stage corresponds to the outcome of the cutting rules. Even though the initial integral as well as its final form are both perfectly well-defined for all real-valued $E_i$, we see that the intermediate result contains a sum of three terms that each diverge when the three energies satisfy the linear relation $E_1-E_2=\pm E_3$. This is clearly a deeply unphysical problem.
The simplest manifestation of the second type of divergence is seen in the two-loop sunset graph, introduced already in sec. \[sec:cut\], where we now set the mass of one of the two fermion lines to zero. Considering the two-cut contribution corresponding to eq. (\[sunset2\]), we are left to evaluate the integral $$\begin{aligned}
\!\!\!\!I_2^\text{2-cut}(\mu)= \int_p \frac{\theta(\mu-E_p)}{2E_p}\int_q \frac{\theta(\mu-q)}{2q} \Bigg[\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Bigg]_{p_0\to iE_p,\,q_0\to iq}\,, \label{pvex}\end{aligned}$$ where the integrand $$\begin{aligned}
\Bigg[\frac{1}{(p_0-q_0)^2+(\mathbf{p}-\mathbf{q})^2}\Bigg]_{p_0\to iE_p,\,q_0\to iq}&=&\frac{1}{2E_p\, q-2\,\mathbf{p}\cdot\mathbf{q}-m^2}\end{aligned}$$ contains a singularity that cannot be regulated using dimensional regularization. For bubble diagrams and those $N$-point functions that are known to be real-valued, the choice of regulator is in principle free, but the most straightforward prescription is to interpret the diverging integrations in a principal value sense. For correlators that might develop physical imaginary parts upon a specific $i\epsilon$ prescription, the procedure is, however, more tricky and one must be careful not to discard any physically meaningful contributions.
Discussion and outlook \[sec:conc\]
===================================
The cutting rules we stated and proved in the previous three sections apply as such only under the assumptions listed in sec. \[sec:state\]. It is, however, straightforward to see that the first three of the conditions can be easily relaxed:
- The appearance of external or internal momenta in the numerator of the Feynman graph does not prohibit the application of the cutting rules. The only potential problem might originate from 0-components of internal momenta, but even these will simply be replaced by the corresponding $iE_i$ factors according to the Residue Theorem.
- If a scalar propagator is raised to a higher power, care must be taken when evaluating the Feynman integral. The most straightforward way to proceed is by first evaluating the corresponding graph with the propagator raised to power 1, and then (repeatedly) differentiating the result with respect to the mass squared of the propagator in question, relying on the formula $$\begin{aligned}
\frac{1}{(Q^2+m_i^2)^n}&=&\frac{(-1)^{n-1}}{(n-1)!}\frac{d}{dm_i^2} \frac{1}{Q^2+m_i^2}\,.
\end{aligned}$$ A possible caveat here has to do with massless propagators raised to higher powers and the associated physical IR divergences. If we introduce a mass term for such a line and then differentiate with respect to it, this will in general produce a $1/m_i^k$ term in the $m_i\to 0$ limit. Some extra effort will then be required to convert this divergence into a $1/\epsilon$ pole, as expected in dimensional regularization.
- Having several closed fermion loops in the graph, each with an independent chemical potential, clearly produces a mere notational complication, and the form of the result stays exactly the same as above. We only need to keep track of the correspondence of the chemical potentials with the cut fermion lines.
Together, these three generalizations allow us to tackle all Feynman integrals encountered in gauge field theories coupled to Dirac fermions, such as QED or QCD.
As discussed already in section \[sec:intro\], the cutting rules become increasingly important when one tries to extend perturbative studies of the thermodynamics of cold and dense systems to higher loop orders. The rules were an integral part of the determination of the three-loop EoS of cold quark matter in ref. [@Kurkela:2009gj], and it is because of them that an extension of this result to the full four-loop order is feasible. In this context, there are in fact two separate challenges: in addition to the evaluation of all four-loop bubble diagrams one needs to determine (a specific component of) the gluon polarization tensor to two-loop order. The latter of these two computations is alone sufficient for determining the logarithmic contributions $\alpha_s^3\ln^2\,\alpha_s$ and $\alpha_s^3\ln\,\alpha_s$ to the EoS. This work is near completion, and the results will be presented in a separate publication later [@nextpaper].
To conclude, let us briefly return to the connection between our work and the naive real-time formalism discussed in sec. \[sec:intro\]. In proving the validity of the zero-temperature cutting rules, we have, in effect, also shown that the naive real-time formalism is applicable not only for vacuum diagrams, contributing to the free energy, but also for Euclidean $n$-point functions in the $T=0$ limit. At the same time, we know that the multitude of various Minkowskian correlators (retarded, advanced, time-ordered, etc.) at nonzero temperature can only be reproduced using the Feynman rules of the full real-time formalism, featuring, in particular, a doubling of field variables. Trying to gain a detailed understanding of the conditions, under which the full real-time formalism reduces to its naive version, is clearly an intriguing avenue for future research.
Acknowledgments {#acknowledgments .unnumbered}
===============
IG and AV were supported in part by the Academy of Finland, grant no. 1273545 and 1303622. PR was supported in part by the Department of Energy, DOE, award no. DE-SC0008132.
On the choice of loop momenta \[sec:mom\]
=========================================
When expressing an $N$-loop Feynman diagram in momentum space, there are a number of possible choices for the integration (or loop) momenta. We choose each of them to correspond to the momentum flowing along one of the propagators, in which case their assignment is limited by two rules, both related to momentum conservation:
1. All internal lines meeting at a given vertex or subdiagram cannot be chosen to correspond to independent loop momenta, as they are linearly dependent.
2. For each closed loop in the graph, at least one of the propagators forming the loop must be chosen to correspond to a loop momentum.
Besides these rules, the choice is arbitrary, and each choice merely corresponds to a slightly different way of writing the original graph. However, they must all lead to the same result.
![\[fig:topos\] Topologies of three vacuum graphs appearing in the determination of the 3-loop EoS of cold quark matter [@Kurkela:2009gj] and discussed in the main text.](topologies_labelled.pdf){width="10.5cm"}
To illustrate this point as well as our notation for the sets $S_r\in S$ introduced in section \[sec:proof\], let us first consider the topology $a$ of fig. \[fig:topos\]. According to our notation, we then have $P_a=\{1,2,3,4,5,6\}$ as well as $$\begin{aligned}
S_a&=&\{\{1,2,3\},\{1,2,4\},\{1,2,5\},\{1,3,4\},...\},\end{aligned}$$ where the only sets of three indices missing from $S$ are $\{1,2,6\},\, \{2,3,4\},\,\{1,3,5\}$, and $\{4,5,6\}$, corresponding to the four three-vertices of the graph (cf. fig. \[fig:topos\]). Writing down the first few terms of eq. (\[Iresmu0\]), we similarly obtain: $$\begin{aligned}
I_a&=&\frac{1}{(2E_1)(2E_2)(2E_3)}\frac{1}{(iE_1-iE_2)^2+E_6^2}\frac{1}{(iE_2-iE_3)^2+E_4^2}\frac{1}{(iE_3-iE_1)^2+E_5^2}\nonumber \\
&+&\frac{1}{(2E_1)(2E_2)(2E_4)}\frac{1}{(iE_1-iE_2)^2+E_6^2}\frac{1}{(iE_2-iE_4)^2+E_3^2}\frac{1}{(iE_2-iE_1-iE_4)^2+E_5^2}\nonumber \\
&+&\cdots\, .\end{aligned}$$ For the diagrams $b$ and $c$, the corresponding supersets $S$ read $$\begin{aligned}
S_b&=&\{\{1,2,3\},\{1,2,4\},\{1,2,5\},\{1,3,4\},\{1,4,5\},\{2,3,5\},\{2,4,5\},\{3,4,5\}\}\,, \nonumber \\
S_c&=&\{\{1,2\},\{1,3\},\{2,3\}\}\,.\end{aligned}$$
On the uniqueness of the cutting rules \[sec:ind\]
==================================================
In this appendix, we provide a detailed argument for the final form of the cutting rules, eq. (\[ASr2\]), starting from the earlier result of eq. (\[ASr\]). To this end, we define the function $$\begin{aligned}
J(S_r) &\equiv& 2^M \prod_{i=1}^M E_i \times I(S_r) \nonumber \\
&=& 2^M \prod_{i=1}^M E_i \times\sum_{S_{{r'}}\in S}\;\prod_{i\in S_{{r'}}} \frac{1}{2E_i}\, \tilde{A}_{S_r}(S_{{r'}})\Big{|}_{\{q_0^i=iE_i\}} \nonumber \\
&\equiv& \sum_{S_{r'} \in \mathcal S}\prod_{\alpha \in P \backslash S_{r'}} \left. \frac{E_{\alpha}(S_{r'})\theta_{r{r'}}}{(r_0^{\alpha}(S_{r'}))^2+E_{\alpha}^2(S_{r'})} \right|_{\{q_0^i=iE_i,\,i\in S_{r'}\}}\,.\end{aligned}$$ where we denote by $\theta_{rr'}$ dimensionless coefficients composed of $\theta$-functions that may in principle depend both on $S_r$ and $S_{r'}$. From eq. (\[ASr\]) we know that $\theta_{rr}=1$ for all $r$, and we shall now show that the independence of the $E_\alpha$ implies that $\theta_{rr'}=1$ even when $r\neq r'$.
To achieve the above goal, we choose another $S_{r''}\in S$ and multiply the function $J(S_r)$ by the product $\prod_{\beta \in P \backslash S_{r''}} E_\beta(S_{r''})$, after which we take the limit where the $E_\beta$ approach infinity: $$\begin{aligned}
&&\lim_{E_{\beta}\to\infty} \prod_{\beta \in P \backslash S_{r''}} E_\beta(S_{r''})\times J(S_r) \nonumber \\
&=& \sum_{S_{r'} \in \mathcal S} \lim_{E_{\beta} \to \infty} \prod_{\alpha \in P \backslash S_{r'}}
\prod_{\beta \in P \backslash S_{r''}}
\left. \frac{E_{\alpha}(S_{r'})E_{\beta}(S_{r''})\theta_{rr'}}{(r_0^{\alpha}(S_{r'}))^2+E_{\alpha}^2(S_{r'})} \right|_{\{q_0^i=iE_i,\,i\in S_{r'}\}}\,.\end{aligned}$$ At this point, we notice that in those terms of the sum where $r'\neq r''$ we have at least one index $\beta$ that belongs to the set $S_{r'}$. The corresponding $E_\beta$ thus appears only linearly in the numerator, but quadratically in the denominator. This implies that the corresponding limit must tend to 0, leaving us with $$\begin{aligned}
&&\lim_{E_{\beta}\to\infty} \prod_{\beta \in P \backslash S_{r''}} E_\beta(S_{r''})\times J(S_r) \nonumber \\
&=& \lim_{E_{\beta} \to \infty}
\prod_{\beta \in P \backslash S_{r''}}
\left. \frac{E_{\beta}^2(S_{r''})\theta_{rr''}}{(r_0^{\beta}(S_{r''}))^2+E_{\beta}^2(S_{r''})} \right|_{\{q_0^i=iE_i,\,i\in S_{r''}\}}\; = \; \theta_{rr''}\,.\end{aligned}$$ Knowing that $J(S_r)$ must be independent of $r$ — just as $I(S_r)$ is — we conclude from here that $\theta_{rr'} = \theta_{r'r'}=1$ and hence $\tilde{A}_{S_r}(S_{r'})=\tilde{A}_{S_{r'}}(S_{r'})$, which is what we wanted to show.
Zero-temperature limit of a fermionic Matsubara contour \[sec:matsu\]
=====================================================================
In this appendix, we demonstrate that the use of an infinitesimal temperature as a regulator of finite-$\mu$ Feynman graphs leads to the handling of divergences along the $p_0$ integration contour in terms of a principal value prescription.
To begin, we consider a generic fermionic Matsubara sum, denoted by $T \sum_n h\left(i \omega_n\right)$, where $\omega_n=(2n+1)\pi T$ and the chemical potential resides in the function $h(z)$ that is taken to vanish sufficiently rapidly at large $|z|$. As usual, we assume that this function may be analytically continued to a meromorphic function $h:\mathbb{C}\rightarrow\mathbb{C}$. Letting $\varepsilon>0$, we then denote by $\Omega$ the $\varepsilon$-strip $\Omega\approx\left(-\varepsilon,\varepsilon\right)\times\mathbb{R}$, noting that if $h$ is holomorphic on $\Omega$, we may evaluate the sum by multiplying $h$ by an appropriately normalized Fermi distribution function that has poles at $\omega=i\omega_n$. This leads to the usual integral representation $$\begin{aligned}
T \sum_{\left\{ \omega_n\right\} }h\left(i \omega_n\right)&=&-\sum_{\left\{ \omega_n\right\} }\mathrm{Res}\left[h\left(z\right)n_F\left(z\right)|z=i\omega_n\right] = \frac{1}{2\pi i}\lim_{\varepsilon\rightarrow0^{+}}\ointop_{\Gamma_{\varepsilon}}dz\,h\left(z\right)n_F\left(z\right), \label{matsu1}\end{aligned}$$ where $n_F(z)\equiv 1/(e^{z/T}+1)$ and $\Gamma_{\varepsilon}$ denotes a *clockwise* rectangular contour whose long sides lie on $\left\{ -\varepsilon\right\} \times\mathbb{R}$ and $\left\{ \varepsilon\right\} \times\mathbb{R}$, respectively. As the horizontal sides of the rectangle produce vanishing contributions, we may equivalently close the vertical contours by two infinite semicircles on the left and right halves of the complex plane.
Proceeding to the zero-temperature limit, we may easily take $\varepsilon\rightarrow0^{+}$, which makes the two vertical contours pinch together. Taking advantage of the relation $\lim_{T\rightarrow0^{+}}n_F(z)=\theta(-{\rm Re}\,z)$ then leads to the simple result $\intop_{-i\infty}^{i\infty}\frac{dz}{2\pi i}\,h\left(z\right)$, where it is customary to redefine the integration variable as $z=i\tilde{z}$ so that we obtain a Euclidean signature integral along the real axis. This is a well-known result that we used as the starting point in our derivation of the cutting rules. Unfortunately however, not all physically interesting functions $h$ are holomorphic on the strip $\Omega$, as they may develop poles along the imaginary axis. This means that special care must be applied when proceeding to the $T\rightarrow0^{+}$ limit in the Matsubara sum, as we shall presently demonstrate.
Let us now choose $\delta\in\left(0,\frac{\pi T}{2}\right)$, and make the simplifying assumption that the only problematic pole of the function $h$ resides at the origin, $z=0$.[^1] In this case, the integral over $\Gamma_{\varepsilon}$ has an unphysical contribution not present in the original Matsubara sum that can be removed by integrating the function $h(z) n_F(z)$ clockwise over the boundary of the rectangle $\Psi_{\varepsilon,\delta}=\left(-\varepsilon,\varepsilon\right)\times\left(-i\delta,i\delta\right)$ (note that this function is holomorphic on $\bar{\Psi}_{\varepsilon,\delta}\backslash\left\{ 0\right\}$, since $\underset{z\in\bar{\Psi}_{\varepsilon,\delta}}{\sup}\left|\mathrm{Im}z\right| \leq \frac{\pi T}{2}<\pi T$). This yields as the generalization of eq. (\[matsu1\]) $$\begin{aligned}
T \sum_{\left\{ \omega_n\right\} }h\left(i\omega_n\right)&=&\frac{1}{2\pi i}\lim_{\delta\rightarrow0^{+}}\lim_{\varepsilon\rightarrow0^{+}}\bigg\{ \ointop_{\Gamma_{\varepsilon}}dz\,h\left(z\right)n_F\left(z\right)-\ointop_{\partial\bar{\Psi}_{\varepsilon,\delta}}dz\,h\left(z\right)n_F\left(z\right)\bigg\}\,, \label{matsu2}\end{aligned}$$ which we depict in fig. \[fig:matsu\] and where we have at the end taken the limit that $\delta$, too, tends to zero.
![\[fig:matsu\] An illustration of the procedure by which we convert a Matsubara sum into a contour integral when the summand has a pole at $z=0$. The red color represents the original integration contour $\Gamma_\epsilon$, while the blue rectangle stands for $\partial\bar{\Psi}_{\varepsilon,\delta}$. ](matsubara.pdf){width="1\linewidth"}
Let us now inspect the form of eq. (\[matsu2\]) in detail. Considering first the horizontal sides of the rectangular integral, we note that the integrand is regular along them, so that they independently vanish in the $\epsilon\to 0^+$ limit, $$\begin{aligned}
\lim_{\varepsilon\rightarrow0^{+}}\intop_{\left(-\varepsilon,\pm i\delta\right)}^{\left(\varepsilon,\pm i\delta\right)}dz\,h\left(z\right)n_F\left(z\right)=0\;\;\;\;\forall\ T>0,\ \delta\in\Big(0,\frac{\pi T}{2}\Big)\,.\end{aligned}$$ At the same time, the arc integrals in $\Gamma_{\varepsilon}$ are unchanged (i.e. they still vanish at infinity), so for the first term in eq. (\[matsu2\]) we are left with the usual result $\intop_{-i\infty-\varepsilon}^{i\infty-\varepsilon}dz\,h\left(z\right)n_F\left(z\right)-\intop_{-i\infty+\varepsilon}^{i\infty+\varepsilon}dz\,h\left(z\right)n_F\left(z\right)$. Subtracting from here the vertical sides of the second term of eq. (\[matsu2\]) yields then $\intop_{-i\infty\pm\varepsilon}^{-i\delta\pm\varepsilon}dz\,h\left(z\right)n_F\left(z\right)+\intop_{i\delta\pm\varepsilon}^{i\infty\pm\varepsilon}dz\,h\left(z\right)n_F\left(z\right)$, so that in total, we obtain $$\begin{aligned}
T\sum_{\left\{ \omega_n\right\} }h\left(i\omega_n\right)
=\frac{1}{2\pi i} \lim_{\delta\rightarrow0^{+}}\lim_{\varepsilon\rightarrow0^{+}}&\,& \left[ \intop_{-i\infty-\varepsilon}^{-i\delta-\varepsilon}dz\,h\left(z\right)n_F\left(z\right)+\intop_{i\delta-\varepsilon}^{i\infty-\varepsilon}dz\,h\left(z\right)n_F\left(z\right) \right. \nonumber\\
&-& \left. \intop_{-i\infty+\varepsilon}^{-i\delta+\varepsilon}dz\,h\left(z\right)n_F\left(z\right)-\intop_{i\delta+\varepsilon}^{i\infty+\varepsilon}dz\,h\left(z\right)n_F\left(z\right) \right].\;\;\;\;\end{aligned}$$ Taking now advantage of the fact that the integrand is regular along the integration contour, we may proceed to the $T\rightarrow0^{+}$ limit in the usual manner. This gives as the zero-temperature limit of the Matsubara sum $$\begin{aligned}
T\sum_{\left\{ \omega_n\right\} }h\left(i\omega_n\right)&\overset{T\rightarrow0^{+}}{\rightarrow}&\frac{1}{2\pi i}\lim_{\delta\rightarrow0^{+}}\bigg\{ \intop_{-i\infty}^{-i\delta}dz\,h\left(z\right)+\intop_{i\delta}^{i\infty}dz\,h\left(z\right)\bigg\} \equiv\mathcal{P}\int_{-\infty}^{\infty}\frac{d\tilde{z}}{2\pi}\, h\left(i\tilde{z}\right)\,,\;\;\;\; \label{prinval}\end{aligned}$$ where we have arrived at a principal value type integral. This result implies that the correct starting point for the derivation of the cutting rules is to define the integration measure as in eq. (\[principal\]).
[99]{}
P. de Forcrand, PoS LAT [**2009**]{} (2009) 010 \[arXiv:1005.0539 \[hep-lat\]\]. R. Machleidt and D. R. Entem, Phys. Rept. [**503**]{} (2011) 1 doi:10.1016/j.physrep.2011.02.001 \[arXiv:1105.2919 \[nucl-th\]\]. U. Kraemmer and A. Rebhan, Rept. Prog. Phys. [**67**]{} (2004) 351 doi:10.1088/0034-4885/67/3/R05 \[hep-ph/0310337\]. K. Hebeler, J. M. Lattimer, C. J. Pethick and A. Schwenk, Astrophys. J. [**773**]{} (2013) 11 doi:10.1088/0004-637X/773/1/11 \[arXiv:1303.4662 \[astro-ph.SR\]\].
A. Kurkela, E. S. Fraga, J. Schaffner-Bielich and A. Vuorinen, Astrophys. J. [**789**]{} (2014) 127 doi:10.1088/0004-637X/789/2/127 \[arXiv:1402.6618 \[astro-ph.HE\]\]. E. S. Fraga, A. Kurkela and A. Vuorinen, Eur. Phys. J. A [**52**]{} (2016) no.3, 49 doi:10.1140/epja/i2016-16049-6 \[arXiv:1508.05019 \[nucl-th\]\]. B. A. Freedman and L. D. McLerran, Phys. Rev. D [**16**]{} (1977) 1169. doi:10.1103/PhysRevD.16.1169 A. Vuorinen, Phys. Rev. D [**68**]{} (2003) 054017 doi:10.1103/PhysRevD.68.054017 \[hep-ph/0305183\].
A. Kurkela, P. Romatschke and A. Vuorinen, Phys. Rev. D [**81**]{} (2010) 105021 doi:10.1103/PhysRevD.81.105021 \[arXiv:0912.1856 \[hep-ph\]\]. E. S. Fraga and P. Romatschke, Phys. Rev. D [**71**]{} (2005) 105014 doi:10.1103/PhysRevD.71.105014 \[hep-ph/0412298\]. A. Kurkela and A. Vuorinen, Phys. Rev. Lett. [**117**]{} (2016) no.4, 042501 doi:10.1103/PhysRevLett.117.042501 \[arXiv:1603.00750 \[hep-ph\]\]. K. Kajantie, M. Laine, K. Rummukainen and Y. Schroder, Phys. Rev. D [**67**]{} (2003) 105008 doi:10.1103/PhysRevD.67.105008 \[hep-ph/0211321\]. F. Di Renzo, M. Laine, V. Miccio, Y. Schroder and C. Torrero, JHEP [**0607**]{} (2006) 026 doi:10.1088/1126-6708/2006/07/026 \[hep-ph/0605042\]. I. Ghisoiu, T. Gorda, A. Kurkela, P. Romatschke, and A. Vuorinen, In preparation.
A. Gynther, M. Laine, Y. Schroder, C. Torrero and A. Vuorinen, JHEP [**0704**]{} (2007) 094 doi:10.1088/1126-6708/2007/04/094 \[hep-ph/0703307 \[HEP-PH\]\]. A. Gynther, A. Kurkela and A. Vuorinen, Phys. Rev. D [**80**]{} (2009) 096002 doi:10.1103/PhysRevD.80.096002 \[arXiv:0909.3521 \[hep-ph\]\]. J. O. Andersen, E. Braaten and M. Strickland, Phys. Rev. D [**62**]{} (2000) 045004 doi:10.1103/PhysRevD.62.045004 \[hep-ph/0002048\].
R. Dashen, S. K. Ma and H. J. Bernstein, Phys. Rev. [**187**]{} (1969) 345. doi:10.1103/PhysRev.187.345
A. I. Bugrii and V. N. Shadura, hep-th/9510232. J. Frenkel, A. V. Saa and J. C. Taylor, Phys. Rev. D [**46**]{} (1992) 3670. doi:10.1103/PhysRevD.46.3670 M. Laine and Y. Schroder, Phys. Rev. D [**73**]{} (2006) 085009 doi:10.1103/PhysRevD.73.085009 \[hep-ph/0603048\].
J. I. Kapusta and C. Gale, “Finite-temperature field theory: Principles and applications.”
[^1]: Other isolated poles not coinciding with the imaginary Matsubara frequencies can be easily removed in the same way, so this is not a restriction.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The retarded van-der-Waals potential, as first obtained by Casimir and Polder, is usually computed on the basis of nonrelativistic QED. The hamiltonian describes two infinitely heavy nuclei, charge $e$, separated by a distance $R$ and two spinless electrons, charge $-e$, nonrelativistically coupled to the quantized radiation field. Casimir and Polder use the dipole approximation and small coupling to the Maxwell field. We employ here the full hamiltonian and determine the asymptotic strength of the leading $-R^{-7}$ potential, which is valid for all $e$. Our computation is based on a path integral representation and expands in $1/R$, rather than in $e$.'
author:
- |
Tadahiro Miyao and Herbert Spohn\
[*Zentrum Mathematik,*]{} [*Technische Universität München,*]{}\
[*D-85747 Garching, Germany*]{}\
e-mail: [email protected], [email protected]
---
\[section\] \[define\][Theorem]{} \[define\][Proposition]{} \[define\][Lemma]{} \[define\][Remark]{} \[define\][Corollary]{}
Introduction
============
Neutral atoms and molecules interact through the long range, attractive van der Waals potential which has a decay as $-R^{-6}$ for large separation $R$. The quantum origin of this force was first recognized by London [@London]. Lieb and Thirring [@LT] supply a non-perturbative proof valid for very general charge configurations. If one goes beyond the static Coulomb interaction in quantizing the Maxwell field, then the action is no longer instantaneous but travels with the speed of light between atoms. In a now very famous paper [@CP] Casimir and Polder establish that thereby the effective interaction potential decays somewhat faster, namely as $-R^{-7}$, which is known as the retarded van der Waals potential. For two hydrogen atoms, the cross over between $R^{-6}$ and $R^{-7}$ sets in at roughly 100 Bohr radii. There are both direct and indirect measurements [@Ex] which confirm the theoretical prediction.
The starting point of Casimir and Polder is nonrelativistic QED for two atoms separated by a distance $R$. Within dipole approximation they expand to fourth order in the coupling to the Maxwell field and obtain a prefactor of $-R^{-7}$ which is proportional to the square of the electric dipole moment of a single atom. Later on alternative routes and simplified derivations were proposed. For an extensive discussion we refer to the book by Milonni [@Milonni], see also the monograph by Margenau and Kestner [@MK] and the lecture notes of Martin and Bünzli [@MB]. Feinberg and Sucher [@FS0; @FS] reconsider the issue by employing a dispersion-theoretic approach. Their prefactor turns out to be quadratic in the electric and magnetic dipole moment of a single atom. Somewhat later Boyer [@Boyer] rederived the same prefactor using quantum zero-point energy and semiclassical expressions for the level shifts due to the presence of the atoms. In our note we stick to nonrelativistic QED, no dipole approximation and no assumption on small coupling, and expand in $1/R$. We use the path integral formulation, in which the subtraction of the ground state energy at $R=\infty$ is particularly transparent. As in previous studies the strength of the retarded van der Waals potential is quadratic in the electric and magnetic dipole moments, but with modified coefficients as compared to [@FS; @Boyer].
In mathematical physics there has been a revived interest in nonrelativistic QED [@Spohn]. It is conceivable that some parts of the argument can be elevated to a rigorous proof. In our paper we mostly ignore this line of research, but will provide a more detailed discussion in the conclusions.
Hamiltonian and van der Waals potential
=======================================
We consider a single hydrogen atom with an infinitely heavy nucleus located at the origin. The nucleus has charge $e$, $e>0$, the electron has charge $-e$. We will use units in which $\hbar=1, c=1$, and the bare mass of the electron $m=1$. In Section \[DL\] we will restore the proper physical units. Let $x,p$ be position and momentum of the electron. Then the nonrelativistic QED hamiltonian for this system reads $$\begin{aligned}
H=\frac{1}{2}\big(p-eA(x)\big)^2-e^2V(x)+{H_{\mathrm{f}}}.\label{1PL}\end{aligned}$$ For $H$ to make sense the electron is assumed to have a prescribed charge distribution ${\varphi}$ with the following properties: ${\varphi}$ is normalized, $\int \mathrm{d}x\, {\varphi}(x)=1$, rotation invariant, $ {\varphi}(x)={\varphi}_{\mathrm{rad}}(|x|)$, of rapid decrease, and its Fourier transform, $\hat{{\varphi}}$, is real. Then $V$ is the smeared Coulomb potential $$\begin{aligned}
V(x)=\int\mathrm{d}k\, |\hat{{\varphi}}(k)|^2|k|^{-2} {\mathrm{e}}^{-{\mathrm{i}}k\cdot x}.\end{aligned}$$ $A(x)$ is the quantized vector potential and ${H_{\mathrm{f}}}$ is the field energy. These are defined through a two-component Bose field $a(k, \lambda), k\in {\mathbb{R}}^3, \lambda=1,2,$ with commutation relation $$\begin{aligned}
[a(k, \lambda), a(k', \lambda')^*]= \delta_{\lambda\lambda'}\delta(k-k').\end{aligned}$$ Explicitly $$\begin{aligned}
{H_{\mathrm{f}}}=\sum_{\lambda=1,2}\int_{{\mathbb{R}}^3}\mathrm{d}k\, \omega(k)a(k, \lambda)^*a(k, \lambda)\end{aligned}$$ with dispersion relation $$\begin{aligned}
\omega(k)=|k|\end{aligned}$$ and $$\begin{aligned}
A(x)=\sum_{\lambda=1,2}\int_{{\mathbb{R}}^3}\mathrm{d}k\, \frac{\hat{{\varphi}}(k)}{\sqrt{2\omega(k)}}{\varepsilon}(k, \lambda)\big({\mathrm{e}}^{{\mathrm{i}}k\cdot x}a(k, \lambda)+{\mathrm{e}}^{-{\mathrm{i}}k\cdot x}a(k, \lambda)^*\big)\end{aligned}$$ with the standard dreibein ${\varepsilon}(k,1), {\varepsilon}(k, 2), \hat{k}=k/|k|$. Thus the Hilbert space for $H$ is $$\begin{aligned}
\mathcal{H}=L^2({\mathbb{R}}^3_x)\otimes \mathfrak{F},\end{aligned}$$ where $\mathfrak{F}$ is the bosonic Fock space over $
L^2({\mathbb{R}}^3)\otimes {\mathbb{C}}^2$. From the quantization of the classical system of charges coupled to the Maxwell field it follows that for the smearing of $A(x)$ and of $V$ the same charge distribution has to be used. We refer to [@Spohn] for details. As proved by Griesemer, Lieb, and Loss [@BFS; @LL; @GLL], $H$ has a unique ground state, denoted here by ${\psi}$, with ground state energy $E$, $H{\psi}=E{\psi}$.
The asymptotic strength $\kappa $ of the van der Waals potential depends on the properties of a single hydrogen atom only through its electric and magnetic dipole moment, $\alpha_{\mathrm{E}}$ and $\alpha_{\mathrm{M}}$. They are defined through the energy, $W$, of our system for weak external uniform electric and magnetic fields according to $$\begin{aligned}
W=-\frac{1}{2} \alpha_{\mathrm{E}} E_{\mathrm{ex}}^2-\frac{1}{2} \alpha_{\mathrm{M}} B_{\mathrm{ex}}^2.\end{aligned}$$ To say, $H$ is perturbed by $e E_{\mathrm{ex}}\cdot x$ and the vector potential is perturbed by $\frac{1}{2} B_{\mathrm{ex}}\wedge x$. Then by second order perturbation theory it follows that $$\begin{aligned}
\alpha_{\mathrm{E}}&=2 \big( \tfrac{1}{3} {\langle}{\psi}, x\cdot (H-E)^{-1} x{\psi}{\rangle}\big)e^2,
{\nonumber \\}\alpha_{\mathrm{M}}&= -\frac{1}{4} \big( \tfrac{1}{3} {\langle}{\psi}, x^2{\psi}{\rangle}\big)e^2.\end{aligned}$$ As a convention, ${\langle}\cdot, \cdot {\rangle}$ denotes always the inner product on the respective Hilbert space.
To investigate the van der Waals potential we consider two hydrogen atoms, one located at $0$ and the other at $r=(0,0,R), R\ge 0$. It will be convenient to define the position of the second electron relative to $r$. Then $x_1,
x_2+r$ are positions and $p_1, p_2$ the momenta of the two electrons. The two-electron hamiltonian reads $$\begin{aligned}
H_R=&\frac{1}{2}\big(p_1-eA(x_1)\big)^2-e^2V(x_1)+\frac{1}{2}\big(p_2-eA(x_2+r)\big)^2-e^2V(x_2)\nonumber \\
&+{H_{\mathrm{f}}}+e^2 V_R(x_1, x_2)\label{FullHami}\end{aligned}$$ with the interaction potential $$\begin{aligned}
V_R(x_1, x_2)&= -V(x_1-r)-V(x_2+r)+V(r)+V(r+x_2-x_1)\nonumber \\
&=\int_{{\mathbb{R}}^3}\mathrm{d}k\, |\hat{{\varphi}}(k)|^2{\mathrm{e}}^{{\mathrm{i}}k\cdot r}
|k|^{-2}
(1-{\mathrm{e}}^{-{\mathrm{i}}k\cdot x_1})(1-{\mathrm{e}}^{{\mathrm{i}}k\cdot x_2}).\end{aligned}$$ $H_R$ acts on the Hilbert space $L^2({\mathbb{R}}^3_{x_1})\otimes
L^2({\mathbb{R}}^3_{x_2})\otimes \mathfrak{F}$. $H_R$ has a unique ground state with energy $E(r)$.
In the Born-Oppenheimer approximation $E(r)$ is the effective potential between the two neutral hydrogen atoms in their ground state. Thus the issue at hand is to investigate $E(r)$ for large $R$. By rotation invariance $E(r)$ depends only on $|r|=R$ and we also write $E(r)=E(R)$. For $R\to \infty$ the two atoms become independent and one can show that $E(R)$ converges to $2E$. The Casimir-Polder result is that, for small $e$, $$\begin{aligned}
\lim_{R\to \infty}R^7(E(R)-2E)=-\frac{23}{4\pi}\big(\frac{1}{2\pi}\big)^2
\big(\frac{1}{2}\alpha_{\mathrm{E, at}}\big)^2 .\label{CP}\end{aligned}$$ The factor $(1/2\pi)^2$ results from our use of the Lorentz-Heaviside units. We remark that Casimir and Polder omit in their definition of $\alpha_{\mathrm{E}}$ the factor $2$, which accounts for the extra $1/2$. $\alpha_{\mathrm{E, at}}$ is the dipole moment of a decoupled hydrogen atom. It is defined through $$\begin{aligned}
\alpha_{\mathrm{E}, \mathrm{at}}= 2 \big(\tfrac{1}{3} {\langle}\psi_{\mathrm{at}},
x\cdot (H_{\mathrm{at}}-E_{\mathrm{at}})^{-1} x\psi_{\mathrm{at}}{\rangle}\big)e^2
\label{Prefactor}\end{aligned}$$ with $$\begin{aligned}
H_{\mathrm{at}}=\frac{1}{2}p^2-e^2 V(x)\end{aligned}$$ and $\psi_{\mathrm{at}}$ the ground state of the hydrogen atom, $H_{\mathrm{at}}\psi_{\mathrm{at}}=E_{\mathrm{at}}\psi_{\mathrm{at}}$.
In our set-up, the natural dimensionless coupling constant is the Sommerfeld fine-structure constant $$\begin{aligned}
\alpha=\frac{e^2}{4\pi \hbar c}.
\end{aligned}$$ The energy unit is set by the ionization energy of the hydrogen atom, which is $\alpha^2 m c^2$ and the length unit is the Bohr radius $r_{\mathrm{B}}=\hbar/\alpha m c$. Anticipating a decay as $R^{-7}$, the dimensionless coupling strength, $\kappa$, is defined through $$\begin{aligned}
E(R)-2E =-\kappa(\alpha, \lambda_{\mathrm{c}}\Lambda) \alpha m c^2 (R/r_{\mathrm{B}})^{-7}
\end{aligned}$$ valid for large $R$. $\kappa$ depends on $\alpha$ and on the ultraviolet cutoff $\Lambda$ in units of the inverse Compton wave length $\lambda_{\mathrm{c}}=\hbar / mc$. For this interpretation the form factor $\hat{{\varphi}}$ is chosen as $\hat{{\varphi}}(k)=\hat{{\varphi}}_1(\Lambda^{-1}k)$, where $\hat{{\varphi}}_1$ rapidly interpolates between $\hat{{\varphi}}_1=(2\pi)^{-3/2}$ for $|k|\le 1-\delta$ and $\hat{{\varphi}}_1=0$ for $|k|\ge 1$.
The goal of our note is to obtain an exact expression for the strength $\kappa$. Readers not so much interested in the details of the computation may skip ahead to Section \[DL\] where the result is discussed.
Path integration
================
As noted by Feynman [@Feynman], in the functional integral representation of ${\mathrm{e}}^{-tH_R}, t\ge 0$, the interaction with the radiation field is linear in $A$. Therefore one can carry out the Gaussian integration over the fluctuating photon field. This form will be particularly convenient for the Born-Oppenheimer energy $E(R)-2E$. After such a detour we will return to operators. Our notation is formal, but rigorous versions are available [@Hiroshima1; @Hiroshima2].
We denote by $q_j(t)\in {\mathbb{R}}^3$ the path of electron $j$. In case of a single electron we omit the index $j$. For the ground state energy of the two-electron system one obtains $$\begin{aligned}
E(r)=&-\lim_{T\to \infty}\frac{1}{2T}\log
\int\Big[\Pi\mathrm{d}q_1(\cdot)\Big]\int\Big[\Pi\mathrm{d}q_2(\cdot)\Big]
\nonumber \\
&\exp\Big[-\int_{-T}^T\mathrm{d}t\Big(\sum_{j=1,2}\big(\tfrac{1}{2}\dot{q}_j(t)^2-e^2V(q_j(t))
\big)
+e^2 V_R(q_1(t), q_2(t))
\Big)\nonumber \\
&-\sum_{j=1,2}\tfrac{1}{2}e^2\int_{-T}^T\mathrm{d}s\int_{-T}^T\mathrm{d}t\,
\dot{q}_j(s)\cdot W_0\big(q_j(s)-q_j(t), s-t\big)\dot{q}_j(t)\nonumber
\\
&- e^2\int_{-T}^T\mathrm{d}s\int_{-T}^T\mathrm{d}t\,
\dot{q}_1 (s)\cdot W_R\big(q_1(s)-q_2(t), s-t\big)\dot{q}_2(t)
\Big].\end{aligned}$$ Here $\int\big[\Pi\mathrm{d}q_j(\cdot)\big]$ is the “sum over all paths" and $W_R$ is the photon propagator, $$\begin{aligned}
W_R(q,t)=\int_{{\mathbb{R}}^3}\mathrm{d}k\,
|\hat{{\varphi}}(k)|^2\frac{1}{2\omega(k)}
({{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}-| \hat{k}{\rangle}{\langle}\hat{k}|)\, {\mathrm{e}}^{-\omega(k)|t|}{\mathrm{e}}^{{\mathrm{i}}k\cdot
r}{\mathrm{e}}^{-{\mathrm{i}}k\cdot q}\label{Propagator}\end{aligned}$$ as a $3\times 3$ matrix. Here ${{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}$ is the unit matrix and $|\hat{k}{\rangle}{\langle}\hat{k}|$ the orthogonal projection onto $\hat{k}$, $\hat{k}=k/|k|$.
Correspondingly for a single electron $$\begin{aligned}
E=&-\lim_{T\to \infty}\frac{1}{2T}\log \int
\Big[\Pi \mathrm{d}q(\cdot)\Big]
\exp\Big[
-\int_{-T}^T \mathrm{d}t\, \big(\tfrac{1}{2}\dot{q}(t)^2-e^2 V(q(t))\big)
{\nonumber \\}&-\tfrac{1}{2}e^2\int_{-T}^T\mathrm{d}s\int_{-T}^T\mathrm{d}t\,
\dot{q}(s)\cdot W_0\big(q(s)-q(t), s-t\big)\dot{q}(t)
\Big].\label{SingleEnergy}\end{aligned}$$ Therefore $$\begin{aligned}
E(R)-2E=&-\lim_{T\to \infty}\frac{1}{2T}\log {\mathbb{E}}\times {\mathbb{E}}\Big[
\exp\Big[
-e^2\int_{-T}^T\mathrm{d}t\, V_R(q_1(t), q_2(t))
{\nonumber \\}&-e^2\int_{-T}^T\mathrm{d}s\int_{-T}^T\mathrm{d}t\,
\dot{q}_1(s)\cdot W_R\big(q_1(s)-q_2(t), s-t\big)\dot{q}_2(t)
\Big]
\Big].\label{PathBindingEr}\end{aligned}$$ Here $q_1(t)$ and $q_2(t)$ are two independent copies of the ground state process $q(t)$ with a path measure as written in (\[SingleEnergy\]). $q(t)$ is stationary and the distribution of $q(t)$ at a fixed time is the electronic density computed from the ground state ${\psi}$ of $H$. The average with respect to the ground state process is denoted by ${\mathbb{E}}[\cdot]$ and the average over the two independent copies by ${\mathbb{E}}\times {\mathbb{E}}[\cdot]$
We note that in the expression (\[PathBindingEr\]) the only $R$ dependence sits in $V_R$ and $W_R$, which in a certain sense are small. Thus it is natural to use the cumulant expansion. Denoting the exponent by $X_T(R)$, one arrives at $$\begin{aligned}
-\frac{1}{2T}\log {\mathbb{E}}\times {\mathbb{E}}[ {\mathrm{e}}^{- e^2 X_T(R)}]
&=-\frac{1}{2T}\big(
e^2 C_1(R,T)+\frac{1}{2}e^4 C_2(R, T)+\cdots\label{2ndCum}
\big),\nonumber\\
C_1(R, T)&= -{\mathbb{E}}\times {\mathbb{E}}[X_T(R)],\nonumber \\[1ex]
C_2(R,T)&={\mathbb{E}}\times {\mathbb{E}}[X_T(R)^2] -{\mathbb{E}}\times {\mathbb{E}}[X_T(R)]^2.\end{aligned}$$ As we will see below ${\mathbb{E}}\times {\mathbb{E}}[ X_T(R)^2]/2T=\mathcal{O}(R^{-6})$, not using the cancellation between $V_R$ and $W_R$ terms. Since we are heading for a decay as $R^{-7}$ for large $R$, it should suffice to stop the expansion at the second cumulant $C_2(R, T)$. [*This will be our main assumption.*]{} ${\mathbb{E}}\times {\mathbb{E}}[ X_T(R)]/2T$ is exponentially small. The large $R$ behavior of the second cumulant will be investigated in detail in the following section.
We note that the expectation under the $k$-integral always factorizes with respect to $q_1, q_2$. Using their independence we will need to compute only a few expectations for the ground state process of a single electron. They are listed now for later convenience and proved in Appendix \[(i)to(v)\]. We set ${\varepsilon}={\varepsilon}(k, \lambda)$, ${\varepsilon}_j={\varepsilon}(k_j, \lambda_j)$ for either $\lambda_j=1$ or $\lambda_j=2$ and $\theta(t)$ the step function, $\theta(t)=-1$ for $t\le 0$, $\theta(t)=1$ for $t> 0$.
- $\displaystyle {\mathbb{E}}[ {\mathrm{e}}^{{\mathrm{i}}k\cdot q(t)}]={\langle}{\psi},
{\mathrm{e}}^{{\mathrm{i}}k\cdot x}{\psi}{\rangle}$,
- $\displaystyle {\mathbb{E}}[ {\mathrm{e}}^{{\mathrm{i}}k\cdot q(t)}({\varepsilon}(k,
\lambda)\cdot \dot{q}(t))]=-{\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot
x}(H-E)({\varepsilon}(k, \lambda)\cdot x){\psi}{\rangle}=0 ,$
- $\displaystyle {\mathbb{E}}[ {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q(s)}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot q(t)}]={\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot
x}{\mathrm{e}}^{-|s-t|(H-E)}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x}{\psi}{\rangle}$,
- $ \displaystyle {\mathbb{E}}[({\varepsilon}_1\cdot
\dot{q}(s)){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q(s)}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q(t)})]\vspace{3mm}
\\=
\theta(t-s){\langle}{\psi}, ({\varepsilon}_1\cdot x)(H-E){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} {\mathrm{e}}^{-|s-t|(H-E)}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle},
$
- $\displaystyle
{\mathbb{E}}[({\varepsilon}_1\cdot \dot{q}(s)){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q(s)}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q(t)}({\varepsilon}_2\cdot \dot{q}(t))] \vspace{3mm}
\\
=-{\langle}{\psi}, ({\varepsilon}_1\cdot x )(H-E){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot
x}{\mathrm{e}}^{-|s-t|(H-E)}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x} (H-E)({\varepsilon}_2\cdot
x){\psi}{\rangle}\vspace{3mm}\\
+\delta(t-s)({\varepsilon}_1\cdot {\varepsilon}_2){\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}
{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}{\psi}{\rangle}. \label{(v)}
$
In (v) the second term arises because locally $q(t)$ is like a standard Brownian motion for which $\mathrm{d}q(s)\otimes
\mathrm{d}q(t)={{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}\delta(s-t)\mathrm{d}s\mathrm{d}t.
$
The first cumulant can be dealt with immediately. Using (ii) in the above list one obtains $$\begin{aligned}
{\mathbb{E}}\times {\mathbb{E}}\big[ \dot{q}_1(s)\cdot W_R\big(q_1(s)-q_2(t), s-t\big)\dot{q}_2(t)\big]=0.\end{aligned}$$ For the potential term it holds $$\begin{aligned}
&{\mathbb{E}}\times {\mathbb{E}}[ V_R(q_1(t), q_2(t))]
{\nonumber \\}&=\int_{{\mathbb{R}}^3}\mathrm{d}k\,
|\hat{{\varphi}}(k)|^2\omega(k)^{-2}{\mathrm{e}}^{{\mathrm{i}}k\cdot r}
{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k\cdot x}){\psi}{\rangle}{\langle}{\psi}, (1-{\mathrm{e}}^{{\mathrm{i}}k\cdot
x}){\psi}{\rangle}{\nonumber \\}&=\int_{{\mathbb{R}}^3}\mathrm{d}x\int_{{\mathbb{R}}^3}\mathrm{d}y\big((\varrho-\delta)*{\varphi}\big)(x)
(4\pi |x+r-y|)^{-1}\big((\varrho-\delta)*{\varphi}\big)(y).\end{aligned}$$ Here $\varrho(x)$ is the electron density for ${\psi}$, which is known to have an exponential decay [@Gr]. $\delta$ is the Dirac delta and $*$ denotes convolution. Since $\int_{{\mathbb{R}}^3}\mathrm{d}x\,
(\varrho(x)-\delta(x))=0$ and since $\varrho$ decays rapidly, by Newton’s theorem it holds that there are suitable constants $c_1,
c_2$ such that $$\begin{aligned}
\big|{\mathbb{E}}\times {\mathbb{E}}[ V_R(q_1(t), q_2(t))]\big| \le c_1 {\mathrm{e}}^{-c_2R}.\end{aligned}$$
The second cumulant
===================
\[Computation\]
The second cumulant consists of three terms. They are all proportional to $2T$ by the stationarity of $q_1(t), q_2(t)$. We could take the limit $T\to \infty$ first, but the symmetric version is more convenient. All expectations are written in terms of $H$ and its ground state ${\psi}$. For notational symplicity we replace $H-E$ by $H$, hence $H{\psi}=0$. For inverses as ${\langle}\phi_1,
H^{-n}\phi_2{\rangle}$ we make sure that either ${\langle}\phi_1, {\psi}{\rangle}=0$ or ${\langle}\phi_2, {\psi}{\rangle}=0$. Note that the ground state is nondegenerate [@Hiroshima1]. But ${\langle}\phi_1, H^{-n}\phi_2{\rangle}$ could still be infinite. If $H$ is replaced by $H_{\mathrm{at}}$, then $H_{\mathrm{at}}$ has a spectral gap and therefore an inverse on the orthogonal complement of $\psi_{\mathrm{at}}$.
We set $$\begin{aligned}
C_2(R, T)=I_{VV}+2I_{VW}+I_{WW}\label{Dec2ndCum}\end{aligned}$$ and compute each term separately. The second step is a partial time-integration through which one can understand how the $R^{-6}$ decay from the interaction potential is canceled. In a final step we collect terms according to their number of time-integrations and discuss their $R$-dependence.
Expectations
------------
a\) $I_{VV}$. Setting $\hat{{\varphi}}(k_j)=\hat{{\varphi}}_j,
\omega(k_j)=\omega_j$, we have $$\begin{aligned}
I_{VV}=&\int\mathrm{d}t_1\mathrm{d}t_2\int\mathrm{d}k_1\mathrm{d}k_2\,
|\hat{{\varphi}}_1|^2|\hat{{\varphi}}_2|^2(|k_1|^2|k_2|^2)^{-1}{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
\nonumber\\
&\times \Big\{
\big|
{\mathbb{E}}[(1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q_1(t_1)})(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q_1(t_2)})]
\big|^2
\nonumber\\
&-\big|
{\mathbb{E}}[ (1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q_1(t_1)})]{\mathbb{E}}[(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q_1(t_2)})]
\big|^2
\Big\}
{\nonumber \\}=&\int\mathrm{d}t_1\mathrm{d}t_2\int\mathrm{d}k_1\mathrm{d}k_2\,
|\hat{{\varphi}}_1|^2|\hat{{\varphi}}_2|^2(|k_1|^2 |k_2|^2)^{-1}{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
\nonumber\\
&\times \Big\{
\big|
{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}){\mathrm{e}}^{-|t_1-t_2|H}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}\big|^2
\nonumber\\
&-\big|
{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}){\psi}{\rangle}{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}\big|^2
\Big\}. \label{IVV}\end{aligned}$$
Note that the integrand in (\[IVV\]) decays to zero, since $$\begin{aligned}
{\mathrm{e}}^{-|t_1-t_2|H}\to |{\psi}{\rangle}{\langle}{\psi}|\end{aligned}$$ as $|t_1-t_2|\to \infty$.\
\
b) $I_{VW}$. Setting ${\varepsilon}_j={\varepsilon}(k_j, \lambda_j)$ and noting $$\begin{aligned}
\sum_{\lambda=1,2}| {\varepsilon}(k, \lambda){\rangle}{\langle}{\varepsilon}(k, \lambda)|={{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}-| \hat{k}{\rangle}{\langle}\hat{k}|, \label{Polarization1}\end{aligned}$$ one has $$\begin{aligned}
I_{VW}=&\int{\mathrm{d}}t_1{\mathrm{d}}t_2 {\mathrm{d}}t_3\sum_{\lambda_1=1,2} \int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2|{\hat{{\varphi}}}_2|^2 (2\omega_1 |k_2|^2)^{-1}
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r} {\mathrm{e}}^{-\omega_1|t_1-t_2|}
\nonumber\\
&\times {\mathbb{E}}[ ({\varepsilon}_1\cdot \dot{q}_1(t_1)){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q_1(t_1)}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q_1(t_3)})]
\nonumber\\
&\times {\mathbb{E}}[ ({\varepsilon}_1\cdot \dot{q}_2(t_2)){\mathrm{e}}^{{\mathrm{i}}k_1\cdot q_2(t_2)}(1-{\mathrm{e}}^{{\mathrm{i}}k_2\cdot q_2(t_3)})]
{\nonumber \\}=&\int{\mathrm{d}}t_1{\mathrm{d}}t_2 {\mathrm{d}}t_3\sum_{\lambda_1=1,2} \int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2|{\hat{{\varphi}}}_2|^2 (2\omega_1 |k_2|^2)^{-1}
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r} {\mathrm{e}}^{-\omega_1|t_1-t_2|}
{\nonumber \\}&\times \theta(t_1-t_3){\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}{\mathrm{e}}^{-|t_1-t_3|H}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}{\nonumber \\}&\times \theta(t_2-t_3){\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x}{\mathrm{e}}^{-|t_2-t_3|H}(1-{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}.
\end{aligned}$$ As proved in Appendix \[GrEx\], it holds $$\begin{aligned}
{\langle}{\psi}, ({\varepsilon}_1\cdot x)H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}{\psi}{\rangle}={\mathrm{i}}{\langle}{\psi}, ({\varepsilon}_1\cdot (p-eA(x))){\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}{\psi}{\rangle}=0.\end{aligned}$$ Thus no truncation of the expectation is needed.\
\
c) $I_{WW}$. One has $$\begin{aligned}
I_{WW}
=&\int {\mathrm{d}}t_1{\mathrm{d}}t_2{\mathrm{d}}t_3{\mathrm{d}}t_4 \sum_{\lambda_1, \lambda_2}\int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2 |{\hat{{\varphi}}}_2|^2(2\omega_12\omega_2)^{-1}
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
{\nonumber \\}&\times {\mathrm{e}}^{-\omega_1|t_1-t_2|}{\mathrm{e}}^{-\omega_2|t_3-t_4|}
{\mathbb{E}}\big[ ({\varepsilon}_1 \cdot \dot{q}_1(t_1)) {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q_1(t_1)}({\varepsilon}_2 \cdot \dot{q}_1(t_3)){\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q_1(t_3)}\big]
{\nonumber \\}&\times {\mathbb{E}}\big[ ({\varepsilon}_1 \cdot \dot{q}_2(t_2)) {\mathrm{e}}^{{\mathrm{i}}k_1\cdot q_2(t_2)}({\varepsilon}_2 \cdot \dot{q}_2(t_4)){\mathrm{e}}^{{\mathrm{i}}k_2\cdot q_2(t_4)}\big]
{\nonumber \\}=&\int {\mathrm{d}}t_1{\mathrm{d}}t_2{\mathrm{d}}t_3{\mathrm{d}}t_4 \sum_{\lambda_1, \lambda_2}\int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2 |{\hat{{\varphi}}}_2|^2(2\omega_12\omega_2)^{-1}
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
{\nonumber \\}&\times {\mathrm{e}}^{-\omega_1|t_1-t_2|}{\mathrm{e}}^{-\omega_2|t_3-t_4|}
{\nonumber \\}&\times\big(
-{\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} {\mathrm{e}}^{-|t_1-t_3|H}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}&+ \delta(t_1-t_3)({\varepsilon}_1\cdot {\varepsilon}_2) {\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}(k_1+k_2)\cdot x}{\psi}{\rangle}\big)
{\nonumber \\}&\times \big(
- {\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x} {\mathrm{e}}^{-|t_2-t_4|H}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}&+ \delta(t_2-t_4)({\varepsilon}_1\cdot {\varepsilon}_2) {\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot x}{\psi}{\rangle}\big).
\end{aligned}$$
Partial time integration
------------------------
The next step is a partial time integration for $I_{VW} $ and $I_{WW}$. For the integrand of $I_{VW}$ we use the identity $$\begin{aligned}
& \theta(s-t) {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} {\mathrm{e}}^{-|s-t|H} (1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}{\nonumber \\}&=-\frac{\partial }{\partial s} {\langle}{\psi}, ( {\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} H^{-1} {\mathrm{e}}^{-|s-t|H} (1-{\mathrm{e}}^{-{\mathrm{i}}k_2 \cdot x}){\psi}{\rangle}\end{aligned}$$ and for the integrand of $I_{WW}$ the identity $$\begin{aligned}
&-{\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x }{\mathrm{e}}^{-|s-t|H} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x} H ({\varepsilon}_2\cdot x){\psi}{\rangle}{\nonumber \\}&=\frac{\partial}{\partial s}\frac{\partial}{\partial t} {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x }H^{-2} {\mathrm{e}}^{-|s-t|H} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x} H ({\varepsilon}_2\cdot x){\psi}{\rangle}{\nonumber \\}& -2 \delta(s-t){\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}H^{-1} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}H ({\varepsilon}_2 \cdot x){\psi}{\rangle}.\end{aligned}$$ We insert these expectations in $I_{VW}, I_{WW}$, integrate partially in time, and use $$\begin{aligned}
\frac{\partial}{\partial s}\frac{\partial}{\partial t}{\mathrm{e}}^{-\omega|s-t|}=-\omega^2{\mathrm{e}}^{-\omega|s-t|}+2\omega\delta(s-t).\end{aligned}$$ The boundary terms vanish. For $I_{VW} $ one obtains $$\begin{aligned}
I_{VW}=&\int{\mathrm{d}}t_1{\mathrm{d}}t_2{\mathrm{d}}t_3 \sum_{\lambda_1}\int{\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2|{\hat{{\varphi}}}_2|^2 (2\omega_1|k_2|^2)^{-1} {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
{\nonumber \\}&\times \big(-\omega_1^2{\mathrm{e}}^{-\omega_1|t_1-t_2|}+2\omega_1 \delta(t_1-t_2)\big){\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}H^{-1}{\mathrm{e}}^{-|t_1-t_3|H}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x} H^{-1} {\mathrm{e}}^{-|t_2-t_3|H} (1-{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x}){\psi}{\rangle}.\end{aligned}$$ Finally $I_{WW}$ is given by $$\begin{aligned}
I_{WW}=&\int {\mathrm{d}}t_1{\mathrm{d}}t_2{\mathrm{d}}t_3{\mathrm{d}}t_4 \sum_{\lambda_1, \lambda_2}\int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2 |{\hat{{\varphi}}}_2|^2(2\omega_12\omega_2)^{-1}
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
{\nonumber \\}&\times\Big\{ \big(-\omega_1^2 {\mathrm{e}}^{-\omega_1|t_1-t_2|}+2\omega_1 \delta(t_1-t_2)\big)\big(-\omega_2^2 {\mathrm{e}}^{-\omega_2|t_3-t_4|}+2\omega_2 \delta(t_3-t_4)\big)
{\nonumber \\}&
\times {\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} H^{-2} {\mathrm{e}}^{-|t_1-t_3|H}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}& \times {\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x} H^{-2}{\mathrm{e}}^{-|t_2-t_4|H}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}\
{\nonumber \\}&+\theta(t_1-t_2)\omega_1 \theta(t_3-t_4)\omega_2 {\mathrm{e}}^{-\omega_1|t_1-t_2|} {\mathrm{e}}^{-\omega_2|t_3-t_4|}
{\nonumber \\}&\times \Big(
{\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}H^{-2} {\mathrm{e}}^{-|t_1-t_3|H}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}&\times\delta(t_2-t_4)B_+(k_1, \lambda_1; k_2, \lambda_2)
{\nonumber \\}& + {\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x} H^{-2}{\mathrm{e}}^{-|t_2-t_4|H}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}\
{\nonumber \\}& \times \delta(t_1-t_3)B_-(k_1, \lambda_1; k_2, \lambda_2)
\Big)
{\nonumber \\}&+{\mathrm{e}}^{-\omega_1|t_1-t_2|}{\mathrm{e}}^{-\omega_2|t_3-t_4|}\delta(t_1-t_3)\delta(t_2-t_4)
{\nonumber \\}&\times B_-(k_1, \lambda_1; k_2, \lambda_2)B_+(k_1, \lambda_1; k_2, \lambda_2) \Big\},\end{aligned}$$ where $$\begin{aligned}
&B_-(k_1, \lambda_1; k_2, \lambda_2)
{\nonumber \\}&=({\varepsilon}_1\cdot {\varepsilon}_2){\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}(k_1+k_2)\cdot x}{\psi}{\rangle}-2{\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} H^{-1}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}H ({\varepsilon}_2\cdot x){\psi}{\rangle},
\\
&B_+(k_1, \lambda_1; k_2, \lambda_2)
{\nonumber \\}&=({\varepsilon}_1\cdot {\varepsilon}_2){\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot x}{\psi}{\rangle}-2{\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x} H^{-1}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x}H ({\varepsilon}_2\cdot x){\psi}{\rangle}\end{aligned}$$
We use (\[Dec2ndCum\]) and collect the terms of $C_2(R, T)$ according to the number of their time-integrations, divide by $1/2T$, and take the limit as $T\to \infty$. To prepare for the limit $R\to \infty$, we rescale the momentum integration as $k_j\leadsto k_j/R, j=1,2$, and introduce the unit vector $\hat{n}=r/R=(0,0,1)$. Note that ${\varepsilon}(k/R, \lambda)={\varepsilon}(k, \lambda)$. The two-time, three-time, and four-time integrations are treated separately.
Two-time integrations {#TwoTime}
---------------------
The sum of all terms involving two-time integrations is denoted by $I_2(R)$. One has $$\begin{aligned}
I_2(R)=&\int{\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2 {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
{\nonumber \\}& \times \Big[ \int {\mathrm{d}}t\Big\{
R^{-2} (|k_1|^2 |k_2|^2)^{-1} \Big(
\big|
{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}){\mathrm{e}}^{-|t|H}(1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R}){\psi}{\rangle}\big|^2
{\nonumber \\}& -\big|
{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}){\psi}{\rangle}{\langle}{\psi}, (1-{\mathrm{e}}^{-{\mathrm{i}}k\cdot x_2/R}){\psi}{\rangle}\big|^2
\Big)
{\nonumber \\}&+2\sum_{\lambda_1} R^{-4} |k_2|^{-2} {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}H^{-1}{\mathrm{e}}^{-|t|H} (1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R}){\psi}{\rangle}{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x/R} H^{-1}{\mathrm{e}}^{-|t|H} (1-{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x/R}){\psi}{\rangle}{\nonumber \\}&+\sum_{\lambda_1, \lambda_2}R^{-6} {\langle}{\psi}, ({\varepsilon}_1\cdot x) H{\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t|H} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R} H({\varepsilon}_2\cdot x)
{\psi}{\rangle}{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H{\mathrm{e}}^{{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t|H} {\mathrm{e}}^{{\mathrm{i}}k_2\cdot x/R} H({\varepsilon}_2\cdot x)
{\psi}{\rangle}\Big\}
{\nonumber \\}&+\sum_{\lambda_1, \lambda_2}R^{-3} (2\omega_1\omega_2 (\omega_1+\omega_2))^{-1}
{\nonumber \\}&\times B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)B_+(k_1/R, \lambda_1; k_2/R, \lambda_2)
\Big]\label{TwoTime1}
{\nonumber \\}=& I_{2,1}(R)+I_{2,2}(R).\end{aligned}$$
We consider the sum, $I_{2,1}(R)$, of the first three terms and expand in $1/R$, which yields expectations of the form ${\langle}{\psi}, (a\cdot x) {\mathrm{e}}^{-|t|H} (b\cdot x){\psi}{\rangle}$ with $a,b\in {\mathbb{R}}^3$. By rotation invariance $$\begin{aligned}
{\langle}{\psi}, (a\cdot x) {\mathrm{e}}^{-|t|H}(b\cdot x) {\psi}{\rangle}=(a\cdot b)\tfrac{1}{3}{\langle}{\psi}, x\cdot {\mathrm{e}}^{-|t|H}x{\psi}{\rangle}.\end{aligned}$$ Using $$\begin{aligned}
\sum_{\lambda_1, \lambda_2}({\varepsilon}_1\cdot {\varepsilon}_2)^2=1+(\hat{k}_1\cdot \hat{k}_2)^2,\ \ \sum_{\lambda_1} ({\varepsilon}_1\cdot \hat{k}_2)^2=1-(\hat{k}_1\cdot \hat{k}_2)^2,\end{aligned}$$ one arrives at the lowest order $$\begin{aligned}
R^{-6}\big(\tfrac{1}{3}{\langle}{\psi}, x\cdot {\mathrm{e}}^{-|t|H}x{\psi}{\rangle}\big)^2\big((\hat{k}_1\cdot \hat{k}_2)^2+2-2(\hat{k}_1\cdot \hat{k}_2)^2+1+(\hat{k}_1\cdot \hat{k}_2)^2\big). \label{Cal1}\end{aligned}$$ The nonsmooth contributions, containing $(\hat{k}_1\cdot
\hat{k}_2)^2$, are canceled exactly, while the smooth part, at this order, is given by $$\begin{aligned}
&I_{2,1}(R)
{\nonumber \\}&=3\int{\mathrm{d}}t\int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} R^{-6} \big(\tfrac{1}{3}{\langle}{\psi}, x\cdot {\mathrm{e}}^{-|t|H}x{\psi}{\rangle}\big)^2,\end{aligned}$$ which inherits the rapid decay in $R$ from ${\varphi}$. At the next order one picks up the quadratic contributions $k_1^2, k_1\cdot k_2, k_2^2$ with coefficients integrable in $t$. By power counting one arrives at $I_{2, 1}(R)\simeq R^{-8}$ as $R\to \infty$.
For the second summand, $I_{2,2}(R)$, we use that to leading order in $1/R$, $$\begin{aligned}
B_{\pm}(k_1/R, \lambda_1; k_2/R, \lambda_2)= R^{-2} |k_1| |k_2| ({\varepsilon}_1\cdot\hat{ k}_2) ({\varepsilon}_2\cdot \hat{k}_1)(\tfrac{1}{3}{\langle}{\psi}, x^2{\psi}{\rangle})+\mathcal{O}(R^{-4}),\end{aligned}$$ as in shown in Appendix \[GrEx\]. Using that $|{\hat{{\varphi}}}(0)|^2=(2\pi)^{-3}$ one arrives at $$\begin{aligned}
&I_{2,2}(R)
{\nonumber \\}&= R^{-7}\big(\tfrac{1}{3} {\langle}{\psi}, x^2 {\psi}{\rangle}\big)^2(2\pi)^{-6}\frac{1}{2} \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 (\omega_1\omega_2(\omega_1+\omega_2))^{-1}|k_1|^2 |k_2|^2
{\nonumber \\}&\times {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
(1-(\hat{k}_1\cdot
\hat{k}_2)^2)^2
+\mathcal{O}(R^{-9})
{\nonumber \\}&= \frac{128}{\pi}\big(\frac{1}{2\pi}\big)^2 e^{-4 } \alpha_{\mathrm{M}}^2 R^{-7}
+\mathcal{O}(R^{-9}).\end{aligned}$$ The evaluation of the numerical coefficient is discussed in Appendix \[NumC\]. Altogether $$\begin{aligned}
\lim_{R\to \infty} R^7I_2(R)= \frac{128}{\pi}\big(\frac{1}{2\pi}\big)^2 e^{-4} \alpha_{\mathrm{M}}^2. \label{I2}
\end{aligned}$$
Three-time integrations
-----------------------
The sum of all terms involving three-time integrations is denoted by $I_3(R)$. One obtains $$\begin{aligned}
I_3(R)=&\int{\mathrm{d}}t_1{\mathrm{d}}t_2 \int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2 {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
{\nonumber \\}&\times \Big\{
-R^{-5} \omega_1 |k_2|^{-2} {\mathrm{e}}^{-\omega_1|t_1-t_2|/R}
{\nonumber \\}&\times\Big( \sum_{\lambda_1} {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}H^{-1}
{\mathrm{e}}^{-|t_1|H} (1-{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R}){\psi}{\rangle}{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x/R}H^{-1}
{\mathrm{e}}^{-|t_2|H} (1-{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x/R}){\psi}{\rangle}\Big)
{\nonumber \\}&-R^{-7}\big(
\tfrac{1}{2}\omega_1 {\mathrm{e}}^{-\omega_1|t_1-t_2|/R}+\tfrac{1}{2}\omega_2 {\mathrm{e}}^{-\omega_2|t_1-t_2|/R}\big) {\nonumber \\}&\times \Big(\sum_{\lambda_1, \lambda_2}
{\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t_1|H}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R} H({\varepsilon}_2\cdot x){\psi}{\rangle}{\nonumber \\}&\times
{\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t_2|H}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x/R} H({\varepsilon}_2\cdot x){\psi}{\rangle}\Big)\nonumber\end{aligned}$$
$$\begin{aligned}
&+R^{-6}\tfrac{1}{4} \theta(t_1) {\mathrm{e}}^{-\omega_1 |t_1|/R} \theta(t_2){\mathrm{e}}^{-\omega_2 |t_2|/R}
{\nonumber \\}&\times\Big(\sum_{\lambda_1, \lambda_2} \big(
{\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t_1-t_2|H}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x/R }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}&\times B_+(k_1, \lambda_1; k_2, \lambda_2)
{\nonumber \\}&+{\langle}{\psi}, ({\varepsilon}_1 \cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x/R}H^{-2} {\mathrm{e}}^{-|t_1-t_2|H}{\mathrm{e}}^{{\mathrm{i}}k_2\cdot x/R }H ({\varepsilon}_2 \cdot x){\psi}{\rangle}{\nonumber \\}&\times B_-(k_1, \lambda_1; k_2, \lambda_2)
\big)\Big)
\Big\}{\nonumber \\}=& I_{3,1}(R)+I_{3,2}(R).\end{aligned}$$
We expand the integrand in $1/R$ which yields $$\begin{aligned}
&\int{\mathrm{d}}t_1{\mathrm{d}}t_2 \int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2 {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
{\nonumber \\}&\times \Big\{
-R^{-7} \omega_1 {\mathrm{e}}^{-\omega_1|t_1-t_2|/R} (1-(\hat{k}_1\cdot\hat{k_2})^2)
{\nonumber \\}&\times \tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_1|H} x{\psi}{\rangle}\tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_2|H} x{\psi}{\rangle}{\nonumber \\}&-R^{-7}\big(
\tfrac{1}{2}\omega_1 {\mathrm{e}}^{-\omega_1|t_1-t_2|/R}+\tfrac{1}{2}\omega_2 {\mathrm{e}}^{-\omega_2|t_1-t_2|/R}\big) (1+(\hat{k}_1\cdot \hat{k}_2)^2)
{\nonumber \\}&\times \tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_1|H} x{\psi}{\rangle}\tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_2|H} x{\psi}{\rangle}{\nonumber \\}&+R^{-6} \tfrac{1}{4}\theta(t_1) {\mathrm{e}}^{-\omega_1 |t_1|/R} \theta(t_2){\mathrm{e}}^{-\omega_2 |t_2|/R}
{\nonumber \\}&\times \Big(\sum_{\lambda_1, \lambda_2}
\tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_1-t_2|H} x{\psi}{\rangle}{\nonumber \\}&\times ({\varepsilon}_1\cdot {\varepsilon}_2)\big(
B_+(k_1/R, \lambda_1; k_2/R, \lambda_2)
+ B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)
\big)\Big)
\Big\}\label{ThreeT1}
{\nonumber \\}=&-R^{-7}\int{\mathrm{d}}t_1{\mathrm{d}}t_2 \int{\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} \omega_1 {\mathrm{e}}^{-\omega_1|t_1-t_2|/R}
{\nonumber \\}&\times \tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_1|H} x{\psi}{\rangle}\tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_2|H} x{\psi}{\rangle}{\nonumber \\}&+ R^{-6}\int{\mathrm{d}}t_1{\mathrm{d}}t_2 \int{\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2\Big(
\sum_{\lambda_1, \lambda_2}
{\nonumber \\}&\times {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}\tfrac{1}{4}\theta(t_1){\mathrm{e}}^{-\omega_1|t_1|/R} \theta(t_2) {\mathrm{e}}^{-\omega_2 |t_2|/R}
\tfrac{1}{3}{\langle}{\psi}, x\cdot
{\mathrm{e}}^{-|t_1-t_2|H} x{\psi}{\rangle}{\nonumber \\}&\times ({\varepsilon}_1\cdot {\varepsilon}_2)\big(
B_+(k_1/R, \lambda_1; k_2/R, \lambda_2)
+ B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)
\big)\Big).\end{aligned}$$ For the first summand the nonsmooth contributions are canceled exactly, as in Section \[TwoTime\], while the smooth contribtions decay rapidly since ${\varphi}$ does so. Its next order picks up an extra factor $R^{-2}$. Thus $I_{3,1}(R)\simeq R^{-9} $ for large $R$. For the second summand we perform explicitly the time integration with the result $$\begin{aligned}
I_{3,2}(R)=&R^{-6} \int {\mathrm{d}}k_1{\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2\Big(
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
{\nonumber \\}&\times R (\omega_1+\omega_2)^{-1}\tfrac{1}{3}
{\langle}{\psi}, x\cdot \big(H (H+R^{-1}\omega_1)^{-1} (H+R^{-1}\omega_2)^{-1}\big)
x{\psi}{\rangle}{\nonumber \\}&\times\sum_{\lambda_1, \lambda_2}
({\varepsilon}_1\cdot {\varepsilon}_2)\big(
B_+(k_1/R, \lambda_1; k_2/R, \lambda_2)
+ B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)\big)\Big).\end{aligned}$$ Since $B_{\pm}=R^{-2} |k_1| |k_2|({\varepsilon}_1\cdot \hat{k}_2) ({\varepsilon}_2\cdot \hat{k}_1)(\tfrac{1}{3}{\langle}{\psi}, x^2{\psi}{\rangle})+\mathcal{O}(R^{-4})$, as shown in Appendix \[GrEx\], we conclude that $$\begin{aligned}
I_{3,2}(R)=&-R^{-7}\big(\tfrac{1}{3}{\langle}{\psi}, x^2{\psi}\big{\rangle}\big)\big(\tfrac{1}{3}{\langle}{\psi}, x\cdot H^{-1}x{\psi}{\rangle}\big)
{\nonumber \\}&\times 2 (2\pi)^{-6} \int {\mathrm{d}}k_1 {\mathrm{d}}k_2\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot \hat{n}} (\omega_1\omega_2(\omega_1+\omega_2))^{-1}
{\nonumber \\}&\times |k_1|^2 |k_2|^2 (\hat{k}_1\cdot \hat{k}_2)
(1-(\hat{k}_1\cdot \hat{k}_2)^2)
+\mathcal{O}(R^{-9})
{\nonumber \\}=&\frac{52}{\pi}\big(\frac{1}{2\pi}\big)^2 e^{-4}\alpha_{\mathrm{E}}\alpha_{\mathrm{M}} R^{-7}+\mathcal{O}(R^{-9}).\end{aligned}$$ The evaluation of the numerical coefficient is discussed in Appendix \[NumC\]. Altogether one has $$\begin{aligned}
\lim_{R\to \infty} R^7 I_3(R)=\frac{52}{\pi}\big(\frac{1}{2\pi}\big)^2 e^{-4 } \alpha_{\mathrm{E}}\alpha_{\mathrm{M}}. \label{I3Lim}
\end{aligned}$$
Four-time integrations
----------------------
There is only a single term with four-time integrations, namely $$\begin{aligned}
I_4(R)=& \int{\mathrm{d}}t_1{\mathrm{d}}t_2{\mathrm{d}}t_3 \sum_{\lambda_1, \lambda_2}\int{\mathrm{d}}k_1 {\mathrm{d}}k_2 |{\hat{{\varphi}}}_1|^2|{\hat{{\varphi}}}_2|^2
{\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot r}
{\nonumber \\}&\times \frac{1}{4}\omega_1\omega_2 {\mathrm{e}}^{-\omega_1|t_1-t_2+t_3|}{\mathrm{e}}^{-\omega_2 |t_3|}
{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}H^{-2}{\mathrm{e}}^{-|t_1|H} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x} H ({\varepsilon}_2\cdot x){\psi}{\rangle}{\nonumber \\}&\times {\langle}{\psi}, ({\varepsilon}_1\cdot x) H {\mathrm{e}}^{{\mathrm{i}}k_1\cdot x}H^{-2}{\mathrm{e}}^{-|t_2|H} {\mathrm{e}}^{{\mathrm{i}}k_2\cdot x} H ({\varepsilon}_2\cdot x){\psi}{\rangle}.
\end{aligned}$$ We rescale $k_j \leadsto k_j/R$ as before and in addition $t_3\leadsto t_3R$. Then the exponentially decaying terms are $$\begin{aligned}
{\mathrm{e}}^{-\omega_1|t_3+(t_1-t_2)/R|}{\mathrm{e}}^{-\omega_2 |t_3|}\simeq {\mathrm{e}}^{-\omega_1 |t_3|} {\mathrm{e}}^{-\omega_2 |t_3|}\end{aligned}$$ for large $R$. Expanding in $1/R$ yields $$\begin{aligned}
I_4(R)= &R^{-7} \big(\tfrac{1}{3} {\langle}{\psi}, x\cdot H^{-1}x{\psi}{\rangle}\big)^2 2 \int_0^{\infty}{\mathrm{d}}t \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 |{\hat{{\varphi}}}(k_1/R)|^2|{\hat{{\varphi}}}(k_2/R)|^2
{\nonumber \\}&\times {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}}
\omega_1\omega_2 {\mathrm{e}}^{-t (\omega_1+\omega_2)} (1+(\hat{k}_1\cdot \hat{k}_2)^2) +\mathcal{O}(R^{-9})
{\nonumber \\}= &R^{-7} (2\pi)^{-6}\big(\tfrac{1}{3} {\langle}{\psi}, x\cdot H^{-1}x{\psi}{\rangle}\big)^2 2\int {\mathrm{d}}k_1 {\mathrm{d}}k_2
{\nonumber \\}&\times {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} \frac{\omega_1\omega_2}{\omega_1+\omega_2} (1+(\hat{k}_1\cdot \hat{k}_2)^2)+\mathcal{O}(R^{-9})
{\nonumber \\}=& R^{-7} \big(\frac{1}{2\pi}\big)^2 e^{-4}\frac{23}{8\pi} \alpha_{\mathrm{E}}^2+\mathcal{O}(R^{-9}). \label{I4}
\end{aligned}$$ The prefactor of $R^{-7}$ agrees with the one computed already by Casimir and Polder.
Sum of all terms
----------------
We add the limits listed in (\[I2\]), (\[I3Lim\]), and (\[I4\]), which yields $$\begin{aligned}
\kappa(e)= \lim_{R\to \infty} R^7 \Big(\lim_{T\to \infty} \frac{1}{2T}e^4 \frac{1}{2}C_2(R,T)\Big)=\big(\frac{1}{2\pi}\big)^2 \Big(\frac{23}{16\pi}\alpha_{\mathrm{E}}^2
+\frac{26}{\pi}\alpha_{\mathrm{E}}\alpha_{\mathrm{M}}
+\frac{64}{\pi}\alpha_{\mathrm{M}}^2\Big).\label{SumAll}
\end{aligned}$$
Comparison with previous results
--------------------------------
Magnetic contributions to the $-R^{-7}$ decay were first considered by Feinberg and Sucher [@FS0; @FS] and by Boyer [@Boyer] . They find that $\alpha^2_{\mathrm{E,at }}$ and $\alpha_{\mathrm{M, at}}^2$ have the same coefficient, namely $23/16\pi$, while the one of $\alpha_{\mathrm{E, at}}
\alpha_{\mathrm{M, at}}$ is $7/8\pi$. This raises the issue on the origin for the discrepancy.
The $\alpha_{\mathrm{E, at}}^2$ term can be most easily obtained through the dipole approximated hamiltonian $$\begin{aligned}
H_{\mathrm{dip}}=&\frac{1}{2} (p_1-eA(0))^2-e^2 V(x_1)+\frac{1}{2} (p_2-e A(r))^2-e^2 V(x_2)+{H_{\mathrm{f}}}{\nonumber \\}&+ e^2 V_R(x_1, x_2). \label{Dipole}\end{aligned}$$ One shifts $p_1$ by $eA(0)$ and $p_2$ by $eA(r)$ through the unitary $U=\exp[-{\mathrm{i}}(x_1 \cdot A(0) +x_2 \cdot A(r))]$. Then $$\begin{aligned}
U H_{\mathrm{dip}} U^{-1} =& \frac{1}{2}p_1^2-e^2 V(x_1)+\frac{1}{2}p_2^2-e^2 V(x_2)+{H_{\mathrm{f}}}{\nonumber \\}&-e x_1\cdot E_{\perp}(0)-e x_2\cdot E_{\perp}(r)+e^2 V_R(x_1, x_2)
{\nonumber \\}&+e^2 \int \mathrm{d}k\, |\hat{{\varphi}}(k)|^2 \Big(\frac{1}{3} x_1^2+\frac{1}{3}x_2^2+{\mathrm{e}}^{{\mathrm{i}}k\cdot r}\big(
x_1\cdot x_2-|k|^{-2} (x_1\cdot k)(x_2\cdot k)
\big)\Big),\label{UniTra}
\end{aligned}$$ where $E_{\perp}(r)$ is the quantized transverse electric field. Note that the long range part of $V_R$ is cancelled. From the $4$-th order perturbation in $-e x_1\cdot E_{\perp}(0)-ex_2 \cdot E_{\perp}(r)$ one obtains the Casimir-Polder result.
To include magnetic effects one proceeds to the next order of the multipole expansion and defines $$\begin{aligned}
H_{\mathrm{mul}}=&\frac{1}{2} \big(p_1-eA(0)-e x_1\cdot \nabla_r A(0)\big)^2-e^2 V(x_1)
{\nonumber \\}&+\frac{1}{2} \big(p_2-e A(r)-e x_2\cdot \nabla_r A(r) \big)^2-e^2 V(x_2)
{\nonumber \\}&+{H_{\mathrm{f}}}+ e^2 V_R(x_1, x_2),
\end{aligned}$$ which, as before, is unitarily transformed to $$\begin{aligned}
U H_{\mathrm{mul}}U^{-1}=&\frac{1}{2} \big(p_1-e x_1\cdot \nabla_r A(0)\big)^2-e^2 V(x_1)
{\nonumber \\}&+\frac{1}{2} \big(p_2-e x_2\cdot \nabla_r A(r) \big)^2-e^2 V(x_2)
{\nonumber \\}&+{H_{\mathrm{f}}}-e x_1\cdot E_{\perp}(0)-e x_2\cdot E_{\perp}(r)+e^2 V_R^{\mathrm{cor}}(x_1, x_2), \label{UniTra2}
\end{aligned}$$ where $V_R^{\mathrm{cor}}$ is the interaction potential from (\[UniTra\]). Note that $\nabla_x\wedge (x\cdot \nabla_r A(r))=B(r)=\nabla_r\wedge A(r)$. Thus the hamiltonian (\[UniTra2\]) clearly displays the fluctuating electric and magnetic fields. Expanding in the terms proportional to $\nabla_r A(0), \\
\nabla_r A(r), E_{\perp}(0), E_{\perp}(r)$ to $4$-th order leads to an energy with a large $R$ asymptotics in agreement with (\[SumAll\]). In spirit the authors of [@FS0; @FS; @Boyer] make a further gauge transformation through the unitary $
\exp[ -{\mathrm{i}}(x_1\cdot \nabla_r (x_1\cdot A(0))+x_2 \cdot \nabla_r(x_2\cdot A(r)))].
$ Thereby $x_1\cdot \nabla_rA(0)$ is transformed to $\frac{1}{2} x_1\wedge B(0)$ and correspondingly for $x_2\cdot \nabla_r A(0)$. In addition there are terms coming from the shifting of $E_{\perp}$. In the $4$-th order perturbation only the former terms are taken into account. More precisely the term ${\langle}{\psi}_{\mathrm{at}}\otimes \Omega, (x_1\wedge B(0))^2 P H^{-1} P (x_2\wedge B(r))^2{\psi}_{\mathrm{at}}\otimes \Omega{\rangle}$, which yields indeed the integrand (\[SS1\]) and thus $23/16\pi$ for the prefactor. Here $P={{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}-|{\psi}{\rangle}{\langle}{\psi}|$ and $\Omega$ is the Fock vacuum. For the cross-term they use $
{\langle}{\psi}_{\mathrm{at}}\otimes \Omega, x_1\cdot E_{\perp}(0) P H^{-1} P x_1\cdot E_{\perp}(0) P H^{-1} P (x_2\wedge B(r))^2{\psi}_{\mathrm{at}}\otimes \Omega{\rangle}$, which yields the integral (\[SS2\]) with the term $(1-(\hat{k}_1\cdot \hat{k}_2)^2)$ omitted and thus $7/8\pi$ for the prefactor of $\alpha_{\mathrm{E, at}} \alpha_{\mathrm{M, at}}$.
The dimensionless strength
==========================
\[DL\]
We restore the physical units in $H$ of (\[1PL\]). Then $$\begin{aligned}
H=\frac{1}{2m} \big(p-\tfrac{e}{c} A(x)\big)^2-e^2 V(x)+{H_{\mathrm{f}}}\end{aligned}$$ with $p=-{\mathrm{i}}\hbar \nabla_x$, $$\begin{aligned}
{H_{\mathrm{f}}}=\sum_{\lambda=1,2} \int_{{\mathbb{R}}^3} {\mathrm{d}}k\, \hbar c |k| a(k, \lambda)^* a(k, \lambda),
\end{aligned}$$ and $$\begin{aligned}
A(x)=\sum_{\lambda=1,2} \int_{{\mathbb{R}}^3} {\mathrm{d}}k \sqrt{c/2|k|} \hat{{\varphi}}(k){\varepsilon}(k, \lambda) \big({\mathrm{e}}^{{\mathrm{i}}k\cdot x}a(k, \lambda)+{\mathrm{e}}^{-{\mathrm{i}}k\cdot x}a(k, \lambda)^*\big) .
\end{aligned}$$ $H$ is transformed to atomic units through the canonical transformation $U$ defined as $$\begin{aligned}
U^* a(k, \lambda)U&=(\alpha^{-2} \lambda_{\mathrm{c}})^{3/2} a(\alpha^{-2} \lambda_{\mathrm{c}} k, \lambda),
{\nonumber \\}U^* x U& =\alpha^{-1}r_{\mathrm{B}}x, \ \ \ U^* pU =\alpha r _{\mathrm{B}}^{-1} p.
\end{aligned}$$ Then $$\begin{aligned}
U^* H U&= \alpha^2 m c^2 \Big(
\frac{1}{2} \big(
-{\mathrm{i}}\nabla_x - \sqrt{4\pi} \alpha^{3/2} \tilde{A}(\alpha x)
\big)^2 - \tilde{V}(x) +\tilde{H}_{\mathrm{f}}\Big)
{\nonumber \\}&= \alpha^2 m c^2 \tilde{H},
\end{aligned}$$ where $$\begin{aligned}
\tilde{H}_{\mathrm{f}}= \sum_{\lambda=1,2} \int_{{\mathbb{R}}^3} {\mathrm{d}}k\, |k| a(k, \lambda)^* a(k, \lambda) \label{THf}
\end{aligned}$$ and $$\begin{aligned}
\tilde{V}(x)&=4\pi \int_{{\mathbb{R}}^3}{\mathrm{d}}k\, \hat{{\varphi}}(\alpha^2 \lambda_{\mathrm{c}}^{-1} k)^2 |k|^{-2} {\mathrm{e}}^{-{\mathrm{i}}k\cdot x},\label{TV}
\\
\tilde{A}(x)&=\sum_{\lambda=1,2}\int_{{\mathbb{R}}^3} {\mathrm{d}}k\, \frac{\hat{{\varphi}}(\alpha^2 \lambda_{\mathrm{c}}^{-1} k) }{\sqrt{2|k|}}{\varepsilon}(k, \lambda) \big({\mathrm{e}}^{{\mathrm{i}}k\cdot x}a(k, \lambda)+{\mathrm{e}}^{-{\mathrm{i}}k\cdot x}a(k, \lambda)^*\big) . \label{TA}
\end{aligned}$$ Energy is now in units of $\alpha^2 m c^2$ and distances are in units of the Bohr radius $r_{\mathrm{B}}$.
From (\[THf\]), (\[TA\]) the photon propagator in atomic units is obtained as
$$\begin{aligned}
\tilde{W}_R(q,t)=4\pi
\int_{{\mathbb{R}}^3}\mathrm{d}k\,
|\hat{{\varphi}}(\alpha^2 \lambda_{\mathrm{c}}^{-1}k)|^2\frac{1}{2|k|}
({{\mathchoice {\rm 1\mskip-4mu l} {\rm 1\mskip-4mu l}
{\rm 1\mskip-4.5mu l} {\rm 1\mskip-5mu l}}}-| \hat{k}{\rangle}{\langle}\hat{k}|)\, {\mathrm{e}}^{-\alpha^{-1} |k||t|}{\mathrm{e}}^{{\mathrm{i}}k\cdot
r}{\mathrm{e}}^{-{\mathrm{i}}k\cdot q}.
\end{aligned}$$
Comparing with (\[Propagator\]) this amounts to replacing $e^2$ by $4\pi$ and $t$ by $t/\alpha$. One merely has to follow through this change in the computation of Section \[Computation\]. The final result is $$\begin{aligned}
E(R)-2E=-\kappa (\alpha, \lambda_{\mathrm{c}} \Lambda) \alpha m c^2 (R/r_{\mathrm{B}})^{-7}
\end{aligned}$$ for large $R$ with the strength $\kappa$ obtained as $$\begin{aligned}
\kappa (\alpha, \lambda_{\mathrm{c}}\Lambda)= \frac{4}{\pi} \Big(\frac{23}{16}\tilde{\alpha}_{\mathrm{E}}^2
+26\tilde{\alpha}_{\mathrm{E}}\tilde{\alpha}_{\mathrm{M}}
+64\tilde{\alpha}_{\mathrm{M}}^2\Big) .
\end{aligned}$$ Here, the dimensionless electric and magnetic dipole moments are $$\begin{aligned}
\tilde{\alpha}_{\mathrm{E}}=2 \big(\tfrac{1}{3} {\langle}{\psi}, x\cdot (\tilde{H}-E)^{-1} x{\psi}{\rangle}\big)
\end{aligned}$$ and $$\begin{aligned}
\tilde{\alpha}_{\mathrm{M}} =-\alpha^2 \frac{1}{4} \big(\tfrac{1}{3} {\langle}{\psi}, x^2{\psi}{\rangle}\big)
\end{aligned}$$ with ${\psi}$ the ground state of $\tilde{H}$, $\tilde{H}{\psi}=E{\psi}$. The ultraviolet cut-off is implemented by replacing $\hat{{\varphi}}(k)$ by the form factor $\hat{{\varphi}}_1(k/ \Lambda )$: $\hat{{\varphi}}_1$ decreases rapidly at $|k|=1$ from $(2\pi)^{-3/2}$ to $0$.
Following Bethe [@Bethe] a physically reasonable choice is $\lambda_{\mathrm{c}} \Lambda=1$. One would like to remove the ultraviolet cut-off through $\Lambda\to \infty$. But this limit is not well understood. In any case, the bare mass $m$ would have to substituted by the renormalized mass. For fixed $\Lambda,\
\lambda_{\mathrm{c}}\Lambda =1$, in the limit $\alpha\to 0$, $\tilde{H}$ decouples from the radiation field and $\kappa(\alpha,
\lambda_{\mathrm{c}} \Lambda)$ tends to the strength obtained by Casimir and Polder. Then $\tilde{\alpha}_{\mathrm{E}}\simeq 9/4$ and $\tilde{\alpha}_{\mathrm{M}}\simeq -\alpha^2$ with quality for the strict Coulomb potential. For a systematic expansion in $\alpha$ one would need the correction of order $\alpha$ to $\tilde{\alpha}_{\mathrm{E}}$ and to the bare mass $m$.
Conclusions and Outlook
=======================
The reader may wonder how much is still missing to a complete proof. For the full model, a central difficulty is that the ground state process $q(t)$ is not so well under control. Since $H$ has no spectral gap, the two-point correlation has a slow decay, presumably as ${\mathbb{E}}[ q(t)\cdot q(0)] \simeq |t|^{-4}$ for large $t$. This means that higher order cumulants are difficult to control. In fact, we cannot even prove that $\alpha_{\mathrm{E}}<\infty$.
From the statistical mechanics point of view an interesting case would be to replace the ground state process $q(t)$ by $q_{\mathrm{at}}(t)$, namely the one governed by $H_{\mathrm{at}}$. This process is exponentially mixing which should help in controlling the error term. On the other hand, the double stochastic integral in the action causes extra difficulties. Unfortunately there is no obvious hamiltonian corresponding to this approximation.
A further variant would be the dipole approximation of equation (\[Dipole\]). In the path integral (\[PathBindingEr\]) this corresponds to replacing $W_R(x, t)$ by $W(0, t)$. The effective action is quadratic in $\dot{q}_j(t)$ and the partial time integration can be easily implemented. It results in a diagonal term which cancels the slowly decaying part of $V_R$ and the remainder action is given by $\int {\mathrm{d}}s {\mathrm{d}}t q(s)\cdot \ddot{W}_R(0, s-t) q(t)$. Since $\ddot{W}_R(0, t)\simeq t^{-4}$ for large $t$, the difficulties mentioned above remain. To have an exactly solvable Gaussian model, on top one would have to use the quadratic approximation for $V$ and $V_R$.
For ground state properties a powerful method is the Feshbach projection together with a successive integration over high $k$-modes of the radiation field [@BCFS; @GH]. It would be interesting to understand whether this technique could be used for a rigorous control on the van der Waals potential.
[**Acknowledgement:**]{} We are grateful to P. Milonni for very useful hints on the literature.
Proof of (i) to (v)
===================
\[(i)to(v)\] The proof of (i) to (v) is based on the identity $$\begin{aligned}
{\mathbb{E}}\Big[\prod_{j=1}^m f_j(q(t_j))\Big]={\langle}{\psi}, f_1(x) {\mathrm{e}}^{-|t_1-t_2|(H-E)}\cdots f_{m-1}(x) {\mathrm{e}}^{-|t_{m-1}-t_m|(H-E)}f_m(x){\psi}{\rangle}\end{aligned}$$ for the time order $t_1<t_2<\cdots< t_m$. $\dot{q}(t){\mathrm{d}}t$ is the Ito stochastic integral as defined through the forward discretization. For this purpose we introduce the lattice spacing $\delta$ and let $[t]_{\delta}$ be the integer part of $t$ modulo $\delta$. Then $\dot{q}(t)= \lim_{\delta\to 0}\delta^{-1}(q([t]_{\delta}+\delta)-q([t]_{\delta}))$. We only establish (ii) and (v). The other items are proved by the same procedure.\
ad (ii): Using stationarity,\
$$\begin{aligned}
{\mathbb{E}}[
{\mathrm{e}}^{{\mathrm{i}}k\cdot q(t)}{\varepsilon}\cdot \dot{q}(t)
]&=\lim_{\delta\to 0} \frac{1}{\delta}{\mathbb{E}}[{\mathrm{e}}^{{\mathrm{i}}k\cdot q(0)} {\varepsilon}\cdot (q(\delta)-q(0))]
{\nonumber \\}&=\lim_{\delta\to 0}\frac{1}{\delta} {\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x}({\mathrm{e}}^{-\delta (H-E)}-1) ({\varepsilon}\cdot x){\psi}{\rangle}{\nonumber \\}&=-{\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x} (H-E)({\varepsilon}\cdot x){\psi}{\rangle},
\end{aligned}$$ which vanishes as proved in Proposition \[VGREXP\].\
ad (v): For $t\neq s$ we proceed as in ad (ii), which yields the first summand of (v). For the “diagonal part" one has, for small $\delta$, $$\begin{aligned}
&{\mathbb{E}}\big[
({\varepsilon}_1\cdot (q(\delta)-q(0))) {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot q(0)} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot q(0)} ({\varepsilon}_2\cdot (q(\delta)-q(0)))
\big]
{\nonumber \\}&= {\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}\big({\mathrm{e}}^{-\delta (H-E)} ({\varepsilon}_1\cdot x)({\varepsilon}_2\cdot x)-({\varepsilon}_1\cdot x){\mathrm{e}}^{-\delta (H-E)}
({\varepsilon}_2\cdot x)
{\nonumber \\}&- ({\varepsilon}_1\cdot x) {\mathrm{e}}^{-\delta (H-E)}({\varepsilon}_2\cdot x)+({\varepsilon}_1\cdot x)({\varepsilon}_2\cdot x){\mathrm{e}}^{-\delta (H-E)}
\big){\psi}{\rangle}{\nonumber \\}&= -\delta {\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x} {\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x} [[\delta^{-1}(1-{\mathrm{e}}^{-\delta(H-E)}), ({\varepsilon}_1\cdot x)], ({\varepsilon}_2\cdot x)]{\psi}{\rangle}{\nonumber \\}&\simeq \delta ({\varepsilon}_1\cdot {\varepsilon}_2){\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k_1\cdot x}{\mathrm{e}}^{-{\mathrm{i}}k_2\cdot x}{\psi}{\rangle}.\end{aligned}$$
Ground state expectations
=========================
\[GrEx\] In this appendix, we mainly work in the Schrödinger representation $L^2({\mathbb{R}}^3_x)\otimes \mathfrak{F}=L^2({\mathbb{R}}^3_x)\otimes L^2(Q)$, see [@Spohn; @Hiroshima1]. In this representation it holds
- $A(x)$ is a real-valued multiplication operator.
- Let $\vartheta=\Gamma({\mathrm{e}}^{{\mathrm{i}}\pi/2})$. Then $\vartheta {\mathrm{e}}^{-tH}\vartheta^{-1}$ is positivity improving. Hence ${\tilde{\psi}}=\vartheta {\psi}$ is strictly positive.
Here for a unitary operator $U$, $\Gamma(U)$ is defined by $$\Gamma(U)a(f_1)^*\cdots a(f_n)^*\Omega=a(Uf_1)^*\cdots a(Uf_n)^*\Omega$$ for all $n\in {\mathbb{N}}$ and $f_1, \dots, f_n\in {\mathbb{C}}^2\otimes L^2({\mathbb{R}}^3)$, where $\Omega$ is the Fock vacuum. A linear operator $A$ “improves the positivity" if, for all $0\le f\in (L^2({\mathbb{R}}^3_x)\otimes L^2(Q))\backslash \{0\}$, ${\mathrm{e}}^{-t A}f>0$ for all $t>0$. Let $J$ be the natural conjugation in $L^2({\mathbb{R}}^3_x)\otimes L^2(Q)$, namely, $J\Psi=\overline{\Psi}$ for all $\Psi\in L^2({\mathbb{R}}^3_x)\otimes L^2(Q)$. A linear operator $A$ is called to be $J$-real if $A$ commutes with $J$, i.e., $JA=AJ$. Then, from (b), it follows that\
(c) $\tilde{H}=\vartheta H\vartheta^{-1}$ is a $J$-real operator, i.e., $J\tilde{H}=\tilde{H}J$.\
\[VanisngGrlemm\](Vanishing ground state expectation I). Let $F(x)$ a measurable function on ${\mathbb{R}}^3_x$ such that ${\psi}\in {\mathrm{dom}}(F)$. Then one has $$\begin{aligned}
{\langle}{\psi}, F(x) A(x) {\psi}{\rangle}=0. \label{VGr1}\end{aligned}$$
[From the proof one infers the stronger property $${\langle}{\psi}, F(x) A(x)_{i_1}{\mathrm{e}}^{-s_1 H}\cdots {\mathrm{e}}^{-s_{2n}H} A(x)_{i_{2n+1}} {\psi}{\rangle}=0$$ for all $n\in \{0\}\cup {\mathbb{N}}, s_1, \dots, s_{2n}>0$ and $F\in L^{\infty}({\mathbb{R}}^3)$. (Note that we have to use (c).) ]{}
[*Proof.*]{} Note that $$\begin{aligned}
J \vartheta =\Gamma({\mathrm{e}}^{{\mathrm{i}}\pi}) \vartheta J.\label{JCommu}\end{aligned}$$ Since $A(x)$ is a real-valued multiplication operator, one has $A(x)J=JA(x)$. Hence using (\[JCommu\]), one sees that $E(x)=\vartheta A(x)\vartheta^{-1}$ is purely $J$-imaginary, that is, $JE(x)=-E(x)J$. Remark that $${\langle}{\psi}, A(x) {\psi}{\rangle}={\langle}{\tilde{\psi}}, E(x) {\tilde{\psi}}{\rangle}.$$ Since ${\tilde{\psi}}$ is real-valued by (b), the right hand side is purely imaginary. On the other hand, by the self-adjointness of $A(x)$, the left hand side is real. Thus the only possibility is ${\langle}{\psi}, A(x) {\psi}{\rangle}=0$. Similarly one can show that $${\langle}{\psi}, \Re F (x)A(x) {\psi}{\rangle}=0={\langle}{\psi}, \Im F(x)A(x) {\psi}{\rangle}$$ which implies the assertion. $\Box$
\[VanisngGrlemm2\](Vanishing ground state expectation II). One has the following:
- For all $a\in {\mathbb{R}}^3$, $
{\langle}{\psi}, (a\cdot x) ^{2n+1}{\psi}{\rangle}=0.
$
- If $m+n$ is odd, then $
{\langle}{\psi}, (a\cdot x) H (b\cdot x)^m H^{-1} (c\cdot x)^n H(d\cdot x){\psi}{\rangle}=0
$ for all $a,b,c, d\in {\mathbb{R}}^3$.
[*Proof.*]{} For the proof, we return to the Fock representation $L^2({\mathbb{R}}^3_x)\otimes {\mathfrak{F}}$. Let $J_2$ be the involution defined by $$\begin{aligned}
J_2 \Psi=\Sumoplus \overline{\Psi}_n(-x; k_1,\lambda_1,\dots, k_n, \lambda_n )\end{aligned}$$ for all $\Psi=\sum_{n\ge 0}^{\oplus}\Psi_n(x; k_1, \lambda_1,\dots, k_n, \lambda_n)\in L^2({\mathbb{R}}^3_x)\otimes {\mathfrak{F}}$. Then as proved in [@LMS], we can check that $$\begin{aligned}
J_2 x&=-xJ_2,
\\
J_2 H&=HJ_2, \label{J2H}
\end{aligned}$$ namely, $x$ is purely $J_2$-imaginary and $H$ is $J_2$-real. Note that $J_2{\psi}={\mathrm{e}}^{{\mathrm{i}}\theta}{\psi}$ for some $\theta\in [0, 2\pi)$ by (\[J2H\]) and the uniqueness of the ground state. Hence $$\begin{aligned}
{\langle}{\psi}, (a\cdot x) ^{2n+1}{\psi}{\rangle}&=-{\langle}J_2 {\psi}, (a\cdot x) ^{2n+1}J_2 {\psi}{\rangle}{\nonumber \\}&=-{\langle}{\psi}, (a\cdot x) ^{2n+1}{\psi}{\rangle},\end{aligned}$$ which implies ${\langle}{\psi}, (a\cdot x) ^{2n+1}{\psi}{\rangle}=0$. Similarly if $m+n$ is odd, then $$\begin{aligned}
&{\langle}{\psi}, (a\cdot x) H (b\cdot x)^m H^{-1} (c\cdot x)^n H(d\cdot x){\psi}{\rangle}{\nonumber \\}&=-{\langle}J_2{\psi}, (d\cdot x) H (c\cdot x)^m H^{-1} (b\cdot x)^n H(a\cdot x)J_2 {\psi}{\rangle}{\nonumber \\}&=-{\langle}{\psi}, (d\cdot x) H (c\cdot x)^n H^{-1} (b\cdot x)^m H(a\cdot x) {\psi}{\rangle}.\label{VGr2Eq}\end{aligned}$$ On the other hand, in the Schrödinger representation, we can see that $$\begin{aligned}
&{\langle}\tilde{{\psi}}, (d\cdot x) \tilde{H} (c\cdot x)^n \tilde{H}^{-1} (b\cdot x)^m \tilde{H}(a\cdot x) \tilde{{\psi}}{\rangle}{\nonumber \\}&={\langle}\tilde{{\psi}}, (a\cdot x) \tilde{H} (b\cdot x)^m \tilde{H}^{-1} (c\cdot x)^n \tilde{H}(d\cdot x)\tilde{{\psi}}{\rangle}\end{aligned}$$ because every operator appearing in the expectation is $J$-real. Therefore $$\begin{aligned}
&{\langle}{\psi}, (d\cdot x) H (c\cdot x)^n H^{-1} (b\cdot x)^m H(a\cdot x) {\psi}{\rangle}{\nonumber \\}&={\langle}{\psi}, (a\cdot x) H (b\cdot x)^m H^{-1} (c\cdot x)^n H(d\cdot x){\psi}{\rangle}.\end{aligned}$$ Combining this with (\[VGr2Eq\]), we conclude ii). $\Box$
\[VGREXP\] One has the following
- $\displaystyle {\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x} H({\varepsilon}(k, \lambda)\cdot x){\psi}{\rangle}=0$,
- $\displaystyle B_{\pm}(k_1/R, \lambda_1; k_2/R, \lambda_2)=R^{-2} |k_1| |k_2|({\varepsilon}_1\cdot \hat{k}_2)({\varepsilon}_2\cdot \hat{k}_1)(\tfrac{1}{3}{\langle}{\psi}, x^2{\psi}{\rangle})+\mathcal{O}(R^{-4})$ as $R\to \infty$.
[*Proof.*]{} ad i) By (\[Identity3\]) below and by ${\langle}{\psi}, {\mathrm{e}}^{-{\mathrm{i}}k\cdot x} A(x){\psi}{\rangle}=0$ according to Lemma \[VanisngGrlemm\], one has $$\begin{aligned}
{\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x} H({\varepsilon}(k, \lambda)\cdot x){\psi}{\rangle}=&-{\mathrm{i}}{\langle}{\psi}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x}({\varepsilon}(k, \lambda)\cdot (p-e A(x))){\psi}{\rangle}\\
=&-{\mathrm{i}}{\langle}{\tilde{\psi}}, {\mathrm{e}}^{{\mathrm{i}}k\cdot x}({\varepsilon}(k, \lambda)\cdot p){\tilde{\psi}}{\rangle}\\
=&- \frac{{\mathrm{i}}}{2}\int {\mathrm{d}}x{\mathrm{d}}\mu (A) {\mathrm{e}}^{{\mathrm{i}}k\cdot x} {\varepsilon}(k, \lambda)\cdot (-{\mathrm{i}}\nabla_x{\tilde{\psi}}^2)(x, A)\\
=&- \frac{{\mathrm{i}}}{2} \int {\mathrm{d}}x{\mathrm{d}}\mu (A) {\mathrm{e}}^{{\mathrm{i}}k\cdot x} ({\varepsilon}(k, \lambda)\cdot k ) {\tilde{\psi}}^2(x, A)\\
=&0,\end{aligned}$$ since $k\cdot {\varepsilon}(k, \lambda)=0$.
ad ii) By Lemma \[VanisngGrlemm2\], the order $R^{-1}$ vanishes and it suffices to check the order $R^0$ and $R^{-2}$.\
ii-a) order $R^0$. One has $$\begin{aligned}
\lim_{R\to \infty}B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)=({\varepsilon}_1\cdot {\varepsilon}_2)-2 {\langle}{\psi}, ({\varepsilon}_1\cdot x) H ({\varepsilon}_2\cdot x){\psi}{\rangle}.\end{aligned}$$ Furthermore, using the identities $$\begin{aligned}
[a\cdot \tilde{p}, b\cdot x]&=-{\mathrm{i}}a\cdot b,\\
[H, a\cdot x]&=-{\mathrm{i}}a\cdot \tilde{p},\\
H (a\cdot x){\psi}&=-{\mathrm{i}}a\cdot \tilde{p}{\psi}\label{Identity3}\end{aligned}$$ for all $a,b\in {\mathbb{C}}^3$, where $\tilde{p}=p-eA(x)$, one obtains $$\begin{aligned}
2 {\langle}{\psi}, ({\varepsilon}_1\cdot x) H ({\varepsilon}_2\cdot x){\psi}{\rangle}={\mathrm{i}}{\langle}{\psi}, \big(
({\varepsilon}_1\cdot \tilde{p}) ({\varepsilon}_2\cdot x)- ({\varepsilon}_1\cdot x) ({\varepsilon}_2\cdot \tilde{p})
\big){\psi}{\rangle}=({\varepsilon}_1\cdot {\varepsilon}_2).\end{aligned}$$ Remark that we have used (\[VGr1\]) to conclude that ${\langle}{\psi}, ({\varepsilon}_1\cdot A(x)) ({\varepsilon}_2\cdot x){\psi}{\rangle}=0\\={\langle}{\psi}, ({\varepsilon}_1\cdot x)({\varepsilon}_2\cdot A(x)) {\psi}{\rangle}$.\
ii-b) order $R^{-2}$. We expand as $$\begin{aligned}
R^2 B_-(k_1/R, \lambda_1; k_2/R, \lambda_2)=c_1 k_1^2+c_2k_2^2 +d (k_1, \lambda_1 ; k_2, \lambda_2)+\mathcal{O}(R^{-2}).
\end{aligned}$$ For $c_1k_1^2$, the expansion coefficient is $$\begin{aligned}
({\varepsilon}_1\cdot {\varepsilon}_2){\langle}{\psi}, (k_1\cdot x)^2 {\psi}{\rangle}-2 {\langle}{\psi}, ({\varepsilon}_1\cdot x)H (k_1\cdot x)^2 ({\varepsilon}_2\cdot x){\psi}{\rangle}.
\end{aligned}$$ Using (\[VGr1\]) and that ${\tilde{\psi}}$ and $\tilde{H}$ can be chosen as real, for the second term one obtains $$\begin{aligned}
&{\langle}{\psi}, ({\varepsilon}_1\cdot x) H (k_1\cdot x)^2 ({\varepsilon}_2\cdot x){\psi}{\rangle}+{\langle}{\psi}, ({\varepsilon}_2\cdot x) (k_1\cdot x)^2 H ({\varepsilon}_1\cdot x){\psi}{\rangle}{\nonumber \\}&= {\mathrm{i}}{\langle}{\psi}, ({\varepsilon}_1\cdot \tilde{p}) (k_1\cdot x)^2 ({\varepsilon}_2\cdot x){\psi}{\rangle}-{\mathrm{i}}{\langle}{\psi}, ({\varepsilon}_2\cdot x) (k_1\cdot x)^2 ({\varepsilon}_1\cdot \tilde{p}){\psi}{\rangle}{\nonumber \\}&={\mathrm{i}}{\langle}{\psi}, (k_1\cdot x)^2 [({\varepsilon}_1\cdot\tilde{p}), ({\varepsilon}_2\cdot x)]{\psi}{\rangle}{\nonumber \\}&= ({\varepsilon}_1\cdot {\varepsilon}_2) {\langle}{\psi}, (k_1\cdot x)^2 {\psi}{\rangle}.
\end{aligned}$$ Thus $c_1=0=c_2$.
The expansion coefficient of the mixed term $d(k_1, \lambda_1;k_2, \lambda_2)$ is given by $$\begin{aligned}
-({\varepsilon}_1\cdot {\varepsilon}_2) {\langle}{\psi}, (k_1\cdot x) (k_2\cdot x) {\psi}{\rangle}+ 2 {\langle}{\psi}, ({\varepsilon}_1\cdot x) H (k_1\cdot x) H^{-1} (k_2\cdot x) H({\varepsilon}_2\cdot x){\psi}{\rangle}.
\end{aligned}$$ Using (\[VGr1\]) and that $\tilde{H}$ and ${\tilde{\psi}}$ can be chosen to be real, we have $$\begin{aligned}
&{\langle}{\psi}, ({\varepsilon}_1\cdot x) H (k_1\cdot x) H^{-1} (k_2\cdot x)H ({\varepsilon}_2\cdot x){\psi}{\rangle}+(1\leftrightarrow 2)
{\nonumber \\}&=\frac{1}{2}\Big\{
-{\mathrm{i}}{\langle}{\psi}, ({\varepsilon}_1\cdot x)(k_1\cdot x) (k_2\cdot x) ({\varepsilon}_2\cdot \tilde{p}){\psi}{\rangle}{\nonumber \\}&-{\langle}{\psi}, ({\varepsilon}_1\cdot x)(k_1\cdot \tilde{p})H^{-1} (k_2\cdot x) ({\varepsilon}_2\cdot \tilde{p}){\psi}{\rangle}{\nonumber \\}& +{\mathrm{i}}{\langle}{\psi}, ({\varepsilon}_1\cdot \tilde{p})(k_1\cdot x) (k_2\cdot x) ({\varepsilon}_2\cdot x){\psi}{\rangle}{\nonumber \\}&+{\langle}{\psi}, ({\varepsilon}_1\cdot \tilde{p})(k_1\cdot x)H^{-1} (k_2\cdot \tilde{p}) ({\varepsilon}_2\cdot x) {\psi}{\rangle}+(1 \leftrightarrow 2)
\Big\}
{\nonumber \\}&= \frac{1}{2} \Big\{
{\mathrm{i}}{\langle}{\psi}, (k_1\cdot x) \big(
({\varepsilon}_1\cdot \tilde{p})({\varepsilon}_2\cdot x)-({\varepsilon}_1\cdot x)({\varepsilon}_2\cdot \tilde{p})
\big)(k_2\cdot x){\psi}{\rangle}{\nonumber \\}&-{\langle}{\psi}, ({\varepsilon}_1\cdot x) (k_1\cdot \tilde{p}) H^{-1} (k_2\cdot x)({\varepsilon}_2\cdot \tilde{p}){\psi}{\rangle}{\nonumber \\}&+{\langle}{\psi}, ({\varepsilon}_1\cdot \tilde{p}) (k_1\cdot x)
H^{-1} ({\varepsilon}_2\cdot x) (k_2\cdot \tilde{p}){\psi}{\rangle}+(1 \leftrightarrow 2)
\Big\}
{\nonumber \\}&= ({\varepsilon}_1\cdot {\varepsilon}_2) {\langle}{\psi}, (k_1\cdot x)(k_2\cdot x){\psi}{\rangle}+({\varepsilon}_1\cdot k_2)({\varepsilon}_2\cdot k_1) (\tfrac{1}{3}{\langle}{\psi}, x^2 {\psi}{\rangle})
{\nonumber \\}&= \tfrac{1}{3} {\langle}{\psi}, x^2{\psi}{\rangle}\big(
({\varepsilon}_1\cdot {\varepsilon}_2)(k_1\cdot k_2)+({\varepsilon}_1\cdot k_2) ({\varepsilon}_2\cdot k_1)
\big).
\end{aligned}$$ This proves the assertion. $\Box$
Numerical coefficients
======================
\[NumC\] In this appendix, we will explain how to compute the following integrals appearing in the main text: $$\begin{aligned}
S_1=& \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 \frac{|k_1| |k_2|}{|k_1|+|k_2|}\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} (1+(\hat{k}_1\cdot \hat{k}_2)^2),\label{SS1}\\
S_2=& \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 \frac{|k_1||k_2|}{|k_1|+|k_2|}\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} (\hat{k}_1\cdot \hat{k}_2)(1-(\hat{k}_1\cdot \hat{k}_2)^2),\label{SS2}\\
S_3=& \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 \frac{|k_1||k_2|}{|k_1|+|k_2|}\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} (1-(\hat{k}_1\cdot \hat{k}_2)^2)^2.\label{SS3}
\end{aligned}$$ These integrals are of the form $$\begin{aligned}
S_j=& \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 \frac{|k_1||k_2|}{|k_1|+|k_2|}\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} F_j(\hat{k}_1\cdot \hat{k}_2),
\end{aligned}$$ which we rewrite as $$\begin{aligned}
S_j=&\int_0^{\infty} {\mathrm{d}}t \int {\mathrm{d}}k_1 {\mathrm{d}}k_2 |k_1| |k_2| {\mathrm{e}}^{-t(|k_1|+|k_2|)}\, {\mathrm{e}}^{{\mathrm{i}}(k_1+k_2)\cdot {\hat{n}}} F_j(\hat{k}_1\cdot \hat{k}_2).
\end{aligned}$$
Let us switch to polar coordinates $(r, {\varphi}, \vartheta)$ by $$\begin{aligned}
\hat{k}&=(Y \cos {\varphi}, Y \sin {\varphi}, X),
{\nonumber \\}X&=\cos \vartheta,\ \ Y=\sin \vartheta.
\end{aligned}$$ Clearly $X^2+Y^2=1$. Then we have $$\begin{aligned}
\hat{k}_1\cdot \hat{k}_2= \cos({\varphi}_1-{\varphi}_2) Y_1 Y_2+X_1 X_2,
\end{aligned}$$ and hence $$\begin{aligned}
S_j
=&\int_0^{\infty}{\mathrm{d}}t \int_0^{\infty} {\mathrm{d}}r_1 r_1^3 {\mathrm{e}}^{-t r_1} \int_{-1}^1{\mathrm{d}}X_1\, {\mathrm{e}}^{{\mathrm{i}}r_1 X_1}
{\nonumber \\}&\times \int_0^{\infty} {\mathrm{d}}r_2 r_2^3 {\mathrm{e}}^{-t r_2} \int_{-1}^1{\mathrm{d}}X_2\, {\mathrm{e}}^{{\mathrm{i}}r_2 X_2}
\mathfrak{S}_j (X_1, X_2),
\end{aligned}$$ where $$\begin{aligned}
\mathfrak{S}_j(X_1, X_2)&=\int_0^{2\pi}{\mathrm{d}}{\varphi}_1 \int_0^{2\pi}{\mathrm{d}}{\varphi}_2 F_j\big( \cos({\varphi}_1-{\varphi}_2)Y_1Y_2+X_1X_2\big)
{\nonumber \\}&=2\pi \int_0^{2\pi} {\mathrm{d}}{\varphi}F_j(\cos {\varphi}Y_1 Y_2+X_1X_2). \label{VphiInt}
\end{aligned}$$ Set $$\begin{aligned}
{\langle}A{\rangle}=2\pi \int_0^{\infty} {\mathrm{d}}r\, r^3\, {\mathrm{e}}^{-t r} \int_{-1}^1 {\mathrm{d}}X\, {\mathrm{e}}^{{\mathrm{i}}r X}A(X).
\end{aligned}$$ After performing ${\varphi}$-integration in (\[VphiInt\]), $S_j$ can be expressed as $$\begin{aligned}
S_1=&\int_0^{\infty} {\mathrm{d}}t\Big\{
\tfrac{3}{2} {\langle}1{\rangle}{\langle}1{\rangle}-{\langle}1{\rangle}{\langle}X^2{\rangle}+\tfrac{3}{2}{\langle}X^2{\rangle}{\langle}X^2{\rangle}\Big\},\label{S1}
\\
S_2=&\int_0^{\infty} {\mathrm{d}}t\Big\{
-\tfrac{1}{2} {\langle}X{\rangle}{\langle}X{\rangle}+ 3 {\langle}X{\rangle}{\langle}X^3{\rangle}-\tfrac{5}{2}{\langle}X^3{\rangle}{\langle}X^3{\rangle}\Big\}, \label{S2}
\\
S_3=&\int_0^{\infty} {\mathrm{d}}t\Big\{
\tfrac{3}{8} {\langle}1{\rangle}{\langle}1 {\rangle}+ \tfrac{1}{2} {\langle}1{\rangle}{\langle}X^2{\rangle}+\tfrac{3}{2}{\langle}X^2{\rangle}{\langle}X^2{\rangle}+\tfrac{3}{4}{\langle}1{\rangle}{\langle}X^4{\rangle}-\tfrac{15}{2} {\langle}X^2{\rangle}{\langle}X^2{\rangle}{\nonumber \\}&+\tfrac{35}{8} {\langle}X^4{\rangle}{\langle}X^4{\rangle}\Big\}.\label{S3}
\end{aligned}$$ Using Mathematica, one obtains $$\begin{aligned}
{\langle}1{\rangle}&= \frac{-4+12t}{(1+t^2)^3},\\
{\langle}X{\rangle}&= {\mathrm{i}}\frac{16 t }{(1+t^2)^3},\\
{\langle}X^2{\rangle}&= \frac{4(-3 +t^2)}{(1+t^2)^3},\\
{\langle}X^3{\rangle}&=4{\mathrm{i}}\Big\{
\frac{t(9+8t^2+3t^4)}{(1+t^2)^3}-3\, \mathrm{arccot}( t)
\Big\}
,
\\
{\langle}X^4{\rangle}&=
\frac{4(3+27 t^2+32 t^4+12 t^6)}{(1+t^2)^3}-48 t\, \mathrm{arccot}( t)
.
\end{aligned}$$ Inserting these formulas to (\[S1\])-(\[S3\]) and using Mathematica again, one arrives at $$\begin{aligned}
S_1 =92 \pi^3,\ \ S_2=208 \pi^3,\ \ S_3=256 \pi^3.
\end{aligned}$$
[100]{} F. London, Zur Theorie und Systematik der Molekularkräfte, Z. Physik [**63**]{}, 245–79 (1930). E. H. Lieb, W. Thirring, Universal nature of van der Waals forces for Coulomb systems, Phys. Rev. A [**34**]{}, 40–46 (1986). H. B. G. Casimir, D. Polder, The influence of retardation on the London-van der Waals forces, Phys. Rev. [**73**]{}, 360–372 (1948). F. S. Levin, D. A. Micha, Long-Range Casimir Forces, Theory and Experiments on Atomic Systems. Plenum Press 1993. P. W. Milonni, The Quantum Vacuum. An introduction to quantum electrodynamics. Academic Press 1994. H. Margenau, N. R. Kestner, Theory of Intermolecular Forces, Pergamon Press 1969. P. A. Martin, P. R. Bünzli, The Casimir effect, Acta. Phys. Polonica B [**37**]{}, 2503-2559 (2006), arXiv:cond-mat/0602559v1. G. Feinberg, J. Sucher, General form of the retarded van der Waals potential, J. Chem. Phys. [**48**]{}, 3333 (1968). G. Feinberg, J. Sucher, General theory of the van der Waals interaction: a model-independent approach, Phys. Rev. A [**2**]{}, 2395–2415 (1970). T. H. Boyer, Van der Waals forces and zero-point energy for dielectric and permeable materials, Phys. Rev. A [**9**]{}, 2078–2084 (1974). H. Spohn, Dynamics of Charged Particles and Their Radiation Field. Cambridge University Press, Cambridge, 2004. V. Bach, J. Fröhlich, I. M. Sigal, Spectral analysis for systems of atoms and molecules coupled to the quantized radiation field, Comm. Math. Phys. [**207**]{}, 249-290 (1999). E. H. Lieb, M. Loss, Existence of atoms and molecules in non-relativistic quantum electrodynamics, Adv. Theor. Math. Phys. [**7**]{}, 667–710 (2003). M. Griesemer, E. H. Lieb, M. Loss, Ground states in non-relativistic quantum electrodynamics, Invent. Math. [**145**]{}, 557–595 (2001). R. P. Feynman, Space-time approach to non-relativistic quantum mechanics, Rev. Mod. Phys. [**20**]{}, 367–387 (1948). F. Hiroshima, Ground states of a model in nonrelativistic quantum electrodynamics II, J. Math. Phys. [ **41**]{}, 661–674 (2000). F. Hiroshima, Functional integral representation of a model in quantum electrodynamics, Rev. Math. Phys. [**9**]{}, 489–530 (1997). M. Griesemer, Exponential decay and ionization thresholds in non-relativistic quantum electrodynamics, J. Funct. Anal. [**210**]{}, 321–340 (2004). H. A. Bethe, The electromagnetic shift of energy level, Phys. Rev. [**72**]{}, 339–341 (1947). V. Bach, T. Chen, J. Fröhlich, I. M. Sigal, The renormalized electron mass in non-relativistic quantum electrodynamics, J. Funct. Anal. [**243**]{} , 426–535 (2007). M. Griesemer, D. Hasler, On the smooth Feshbach-Schur map, J. Funct. Anal. [**254**]{}, 2329–2335 (2008). M. Loss, T. Miyao, H. Spohn, Kramers degeneracy theorem in nonrelativistic QED, arXiv:0809.4471.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
[^1]\
Ludwig-Maximillians Universität, Munich, Germany\
E-mail:
title: Electroweak physics at the Tevatron
---
Introduction
============
The analysis of single and pair production of the weak vector bosons $Z$ and $W$ in hadron-hadron collisions provides not only insight into the electroweak interaction (e.g. through the measurements of the couplings of quarks and leptons to the $Z$ boson) but also allows to study the strong interaction (e.g. with measurements of differential boson production cross sections) or even to possibly detect new phenomena beyond the Standard Model (SM, e.g. via an observation of anomalous triple-gauge couplings). Furthermore, a precise measurement of the $W$ boson mass $M_W$ provides an indirect constraint on the mass of the long-sought Higgs boson given the radiative corrections to $M_W$.
In $p\bar{p}$ collisions at the Fermilab Tevatron collider, the weak boson decays in leptonic final states exhibit very clear experimental signatures. In fall 2008, both the CDF and D0 experiments have collected data sets with integrated luminosities beyond 4[$\mathrm{fb}^{-1}$]{} at the collision energy of $\sqrt{s} = 1.96\,{\ensuremath{\mathrm{TeV}}}$. In the following, we summarize recent results of both collaborations on differential single boson production cross sections, on the mass of the $W$ boson and on diboson production.
Single boson production
=======================
Inclusive cross sections for the production of $Z$ and $W$ bosons decaying into final states with $e$, $\mu$ or $\tau$ were previously published or presented as preliminary results [@CDFweb; @D0web]. The D0 collaboration has recently updated their measurement of the inclusive $Z$ boson production cross section in the $Z\rightarrow \tau \tau$ decay channel using a $1\,{\ensuremath{\mathrm{fb}^{-1}}}$ data set collected with an inclusive muon trigger [@Abazov:2008ff]. In this analysis, one of the two $\tau$ candidates is reconstructed as a decay muon whereas the second is identified as a hadronically or electronically decaying $\tau$ using neural networks. The measurement yields $\sigma_{Z} \cdot Br(Z{\ensuremath{\rightarrow}}{{\ensuremath{\tau^+\tau^-}}}) =
237 \pm 15 \mathrm{(stat)}
\pm 18 \mathrm{(syst)} \pm 15 \mathrm{(lum)}\,{\ensuremath{\mathrm{pb}}}$, which is in good agreement with the SM prediction. Furthermore, it verifies the capability to identify isolated $\tau$ leptons, which is critical in the search for the Higgs boson in supersymmetric extensions of the SM.
The large data sets being accumulated by both experiments enable precise measurements of differential weak boson production cross sections, which allow to set additional constraints on parton distribution functions (PDFs) and to test corrections to the lowest-order predictions of Quantum Chromodynamics (QCD).
$Z$ boson transverse momentum distribution
------------------------------------------
QCD corrections to vector boson production manifest themselves in the radiation of additional quarks or gluons in the final state which gives rise not only to the production of associated jets, but also to a substantial transverse momentum $p_T$ of the produced boson. At large $p_T$, fixed-order calculations in perturbative QCD (pQCD) are applicable and have been derived up to next-to-next-to-leading order (NNLO) [@Melnikov:2006kv]. For low $p_T$, where the emission of multiple soft gluons becomes important, the leading logarithms in the perturbative expansion can be resummed. The Monte Carlo generator [resbos]{} accounts for additional non-perturbative corrections using a form factor [@Landry:2002ix].
The D0 collaboration has recently published a measurement of the $p_T$ distribution in $Z/\gamma^*{\ensuremath{\rightarrow}}{{\ensuremath{e^+e^-}}}$ events based on a data set with 1[$\mathrm{fb}^{-1}$]{} [@Abazov:2007nt]. The NNLO pQCD calculation is found to describe the shape of the distribution at $p_T>30\,{\ensuremath{\mathrm{GeV}}}$, but underestimates the measured rate by $\sim 25\%$. The data in the low $p_T$ region is well modeled by [resbos]{} and disfavors a recently suggested modification of the form factor (small-$x$ broadening) [@Berge:2004nt].
The D0 collaboration has also presented a preliminary measurement of $g_2$, a parameter in the non-perturbative form factor, using a data set of 2[$\mathrm{fb}^{-1}$]{} and both ${{\ensuremath{e^+e^-}}}$ and ${{\ensuremath{\mu^+\mu^-}}}$ decay signatures [@D05755]. A new method with a reduced sensitivity to the lepton $p_T$ resolution is applied, which yields a measurement of $g_2$ as precise as the previous world average. Future measurements of $M_W$ would benefit from an increased precision on the form factor as the systematic uncertainty due to the modeling of weak boson production would decrease.
$Z$ boson rapidity distribution
-------------------------------
At leading order the boson rapidity $y$ is directly related to the momentum fraction of the scattering partons $x_{1,2} = M_Z/\sqrt{s} \cdot e^{\pm y}$. Thus, its distribution at large $|y|$ probes the PDFs at high momentum transfer $Q^2 \approx M_Z^2$ and at both very large and low $x$. Similar to a previous D0 publication [@Abazov:2007jy] a preliminary measurement by the CDF collaboration [@CDFweb] finds a good agreement with higher-order pQCD. The data is compared to various predictions using different PDF sets, but its current precision does not provide significant additional constraints on PDFs.
$W$ boson charge asymmetry
--------------------------
As $u$ quarks carry on average a higher momentum fraction $x$ than $d$ quarks, $W^+$ ($W^-$) bosons are preferentially boosted along the $p$ ($\bar{p}$) direction, thus resulting into a charge asymmetry $$A(y_W) =\frac{{\mathrm{d}}\sigma^+/{\mathrm{d}}y_W - {\mathrm{d}}\sigma^-/{\mathrm{d}}y_W}{{\mathrm{d}}\sigma^+/{\mathrm{d}}y_W + {\mathrm{d}}\sigma^-{\mathrm{d}}y_W}$$ in the $W$ boson rapidity distribution. In leptonic $W$ decays $y_W$ cannot be directly determined since the longitudinal momentum of the decay neutrino is unmeasured. However, the pseudorapidity $\eta_\ell$ of the charged lepton from the $W$ decay is correlated with $y_W$ given the $V-A$ coupling of the weak interaction. In addition, for the lepton asymmetry $$A(\eta_\ell) = \frac{{\mathrm{d}}\sigma(\ell^+)/{\mathrm{d}}\eta_\ell - {\mathrm{d}}\sigma(\ell^-)/{\mathrm{d}}\eta_\ell}{{\mathrm{d}}\sigma(\ell^+)/{\mathrm{d}}\eta_\ell + {\mathrm{d}}\sigma(\ell^-)/{\mathrm{d}}\eta_\ell}$$ the systematic uncertainty related to lepton reconstruction largely cancels.
![The electron charge asymmetry distribution $A(\eta_e)$ in two $E_T^e$ bins: $25\,{\ensuremath{\mathrm{GeV}}}< E_T^e < 35\,{\ensuremath{\mathrm{GeV}}}$ for (a) and $E_T^e > 35\,{\ensuremath{\mathrm{GeV}}}$ for (b) compared to the predictions based on the CTEQ6.6 (including uncertainty band) and MRST04NLO PDF sets [@Abazov:2008qv].[]{data-label="fig1"}](figures/w_asym_d0_2.eps "fig:"){width=".46\textwidth"} ![The electron charge asymmetry distribution $A(\eta_e)$ in two $E_T^e$ bins: $25\,{\ensuremath{\mathrm{GeV}}}< E_T^e < 35\,{\ensuremath{\mathrm{GeV}}}$ for (a) and $E_T^e > 35\,{\ensuremath{\mathrm{GeV}}}$ for (b) compared to the predictions based on the CTEQ6.6 (including uncertainty band) and MRST04NLO PDF sets [@Abazov:2008qv].[]{data-label="fig1"}](figures/w_asym_d0_3.eps "fig:"){width=".46\textwidth"}
Fig. \[fig1\] shows $A(\eta_e)$ as measured by the D0 collaboration in two bins of electron transverse energy $E_T^e$, which probe different regions of $y_W$ (and thus $x$) for fixed $\eta_e$ [@Abazov:2008qv]. For nearly all data points and in particular for $E_T^e>35\,{\ensuremath{\mathrm{GeV}}}$ the experimental uncertainties are significantly smaller that the PDF uncertainty band on the theoretical prediction, which is calculated using the CTEQ6.6 error PDF sets [@Nadolsky:2008zw].
The CDF collaboration has developed a method to directly measure the $W$ boson asymmetry $A(y_W)$ by reconstructing the $y_W$ distribution using a constraint on $M_W$ [@CDF8942]. The two possible solutions for the neutrino momentum are weighted with a probability which is iteratively determined from simulation. This new method has an improved statistical sensitivity compared to the measurement of $A(\eta_\ell)$, but since for fixed values of $y_{W^\pm}$ the corresponding lepton pseudorapidities $\eta_\ell^+$ and $\eta_\ell^-$ are distinct, the geometric acceptances to reconstruct $W^+$ or $W^-$ bosons, respectively, can differ by large amounts and the systematic uncertainty due to the lepton reconstruction is increased. Similar to the D0 result on $A(\eta_e)$, CDF’s preliminary measurement of $A(y_W)$ based on $W{\ensuremath{\rightarrow}}e\nu$ decays and an integrated luminosity of $1\,{\ensuremath{\mathrm{fb}^{-1}}}$ has a higher precision than the PDF uncertainty on the prediction [@CDF8942].
Forward-backward charge asymmetry in $Z/\gamma^*\rightarrow e^+ e^-$ events
---------------------------------------------------------------------------
A measurement of the forward-backward charge asymmetry $A_{FB}$ as function of the dilepton mass $M_{{\ensuremath{\ell^+\ell^-}}}$ in $Z/\gamma^*\rightarrow \ell^+ \ell^-$ events probes the interference between the $\gamma^*$ and $Z$ propagators and is sensitive to the vector and axial-vector couplings $g_V^{u,d,\ell}$ and $g_A^{u,d,\ell}$ of the $u$ and $d$ quarks and the final-state lepton to the $Z$ boson. Therefore, $A_{FB}$ also constrains the effective weak mixing angle $\sin^2\theta_W^{\mathrm{eff}}$ and a hypothetical heavy $Z^\prime$ boson would alter $A_{FB}$ for $M_{{\ensuremath{\ell^+\ell^-}}}\approx M_{Z^\prime}$.
A previous measurement of $A_{FB}$ based on only $72\,{\ensuremath{\mathrm{pb}^{-1}}}$ of integrated luminosity performed by the CDF collaboration put constraints on $g_V^{u,d}$ and $g_A^{u,d}$ complementary to the results from LEP and HERA [@Acosta:2004wq]. The D0 collaboration recently published a measurement of $A_{FB}$ up to dielectron masses $M_{ee}>300\,{\ensuremath{\mathrm{GeV}}}$ using their $1\,{\ensuremath{\mathrm{fb}^{-1}}}$ data set and extracted $\sin^2\theta_W^{\mathrm{eff}}$ with a precision comparable to that of the LEP measurements of the inclusive hadronic charge asymmetry [@Abazov:2008xq].
Measurement of the $W$ boson mass and width
===========================================
The new combination of the measurements of the $W$ mass $M_W$ and width $\Gamma_W$ from the Tevatron Runs and LEP [@CDFD0:2008ut] is shown in Fig. \[fig2\]. The RunI results have been corrected to account for their outdated assumptions on PDFs and $\Gamma_W$ or $M_W$, respectively.
![Comparison of the measurements of the $W$ boson mass (left) and width (right) together with their averages [@CDFD0:2008ut].[]{data-label="fig2"}](figures/mw_plot.eps "fig:"){width=".46\textwidth"} ![Comparison of the measurements of the $W$ boson mass (left) and width (right) together with their averages [@CDFD0:2008ut].[]{data-label="fig2"}](figures/gw_plot.eps "fig:"){width=".46\textwidth"}
The most precise single measurement of $M_W$ has been achieved by the CDF collaboration using $W{\ensuremath{\rightarrow}}e\nu$ and $W{\ensuremath{\rightarrow}}\mu\nu$ events collected in a data set with 200[$\mathrm{pb}^{-1}$]{} integrated luminosity [@Aaltonen:2007ypa]. For this result, the $W$ mass is derived from distributions of the transverse mass $m_T$, the lepton $p_T$ and the missing transverse energy ${\mbox{\ensuremath{E \kern-0.6em\slash_{\rm T}}}}$. The statistical and systematic uncertainty contribute equally to the precision on the measured $W$ boson mass $\Delta M_W = 48\,{\ensuremath{\mathrm{MeV}}}$.
A preliminary CDF analysis based on a more than tenfold integrated luminosity $L$ shows that the statistical uncertainty on $M_W$ scales with approximately $\sqrt{L}$, which demonstrates that the measurement of $M_W$ does not degrade with the larger energy pile-up in the calorimeter due to the increasing instantaneous luminosity [@CDFweb]. Whereas the lepton energy scale of the published measurement is mostly constrained by $J/\psi$ and $\Upsilon$ decays due to the limited number of $Z$ boson decays, the improved statistical precision on $M_Z$ with increasing data sets is expected to significantly reduce the scale uncertainty, which is one of the dominating systematic errors in the measurement of $M_W$.
Diboson production
==================
The pair production of the electroweak bosons, $W$, $Z$, and $\gamma$, probes the trilinear gauge boson couplings predicted by the non-Abelian structure of the SM. Any deviation from the expected couplings, commonly referred to as anomalous couplings, would indicate the presence of new physics beyond the SM. The Tevatron measurements are complementary to those performed at LEP, since the former probe higher energies and are sensitive to different combinations of couplings. As a compilation of all diboson measurements of both CDF and D0 collaborations can be found elsewhere [@CDFweb; @D0web], only a few recent results will be presented below.
Observation of $ZZ$ production
------------------------------
Following CDF’s measurement of $ZZ$ production with a significance of $4.4\sigma$ [@Aaltonen:2008mv], the D0 collaboration has reported the first observation of this process with a significance of $5.7\sigma$ [@Abazov:2008gya]. Two selections are applied. For $ZZ\rightarrow \ell\ell\ell^\prime\ell^\prime$ ($\ell,\ell^\prime = e$ or $\mu$) production, three candidate events, with an expected background of 0.14 events are found in 1.7[$\mathrm{fb}^{-1}$]{} of data corresponding to a significance of $5.3\sigma$. Fig. \[fig3\] shows the four lepton invariant mass for these events compared to the expected signal and background distributions. For the other channel, $ZZ\rightarrow \ell\ell\ell\nu\nu$, a new estimator for the missing transverse energy with a reduced sensitivity to instrumental mismeasurements resulting in an improved background rejection has been developed (Fig. \[fig3\]) [@Abazov:2008yf]. After the signal is discriminated from the dominating $WW$ background using a likelihood a significance of $2.7\sigma$ is observed. The combined $ZZ$ production cross section is $\sigma = 1.60\pm 0.63\mathrm{(stat)}^{+0.16}_{-0.17}\mathrm{(syst)}\,{\ensuremath{\mathrm{pb}}}$, consistent with the SM prediction of $1.4\pm 0.1$pb derived at NLO.
![Left: Distribution of four lepton invariant mass in the $ZZ\rightarrow \ell\ell\ell^\prime\ell^\prime$ selection [@Abazov:2008gya]. Right: Distribution of the missing transverse energy estimator ${\mbox{\ensuremath{E \kern-0.6em\slash_{\rm T}}}}^\prime$ in the $\rightarrow \ell\ell\nu\nu$ selection before applying the ${\mbox{\ensuremath{E \kern-0.6em\slash_{\rm T}}}}^\prime$ requirement [@Abazov:2008yf].[]{data-label="fig3"}](figures/ZZ_4l_d0.eps "fig:"){width=".4775\textwidth"} ![Left: Distribution of four lepton invariant mass in the $ZZ\rightarrow \ell\ell\ell^\prime\ell^\prime$ selection [@Abazov:2008gya]. Right: Distribution of the missing transverse energy estimator ${\mbox{\ensuremath{E \kern-0.6em\slash_{\rm T}}}}^\prime$ in the $\rightarrow \ell\ell\nu\nu$ selection before applying the ${\mbox{\ensuremath{E \kern-0.6em\slash_{\rm T}}}}^\prime$ requirement [@Abazov:2008yf].[]{data-label="fig3"}](figures/ZZ_eemet_d0.eps "fig:"){width=".4425\textwidth"}
Limits on anomalous triple gauge couplings
------------------------------------------
The analysis of $WZ$ production allows to study the $WWZ$ triple gauge coupling independently of the $WW\gamma$ vertex contribution. For the $WWZ$ vertex three CP conserving coupling parameters are defined, which take the following values in the SM: $g_1^Z=1$, $\kappa_Z=1$, $\lambda_Z=0$. In a recent preliminary measurement the CDF collaboration utilizes the $p_T$ distribution of $Z$ bosons in $WZ$ events (Fig. \[fig4\]) to constrain anomalous contributions to the $WWZ$ triple gauge coupling [@CDFweb]. As an example, the one and two-dimensional limits on $g_1^Z$ and $\kappa_Z$ are shown in Fig. \[fig4\].
![Left: The $p_T$ distribution of the $Z$ boson in $WZ$ candidate events. Right: The one and two-dimensional limits on $\Delta g_1^Z$ and $\Delta\kappa_Z$ (deviations from SM values) derived from the distribution shown on the left side [@CDFweb].[]{data-label="fig4"}](figures/trilepton_anaWZ_Signal_Z_Pt.eps "fig:"){width=".47\textwidth"} ![Left: The $p_T$ distribution of the $Z$ boson in $WZ$ candidate events. Right: The one and two-dimensional limits on $\Delta g_1^Z$ and $\Delta\kappa_Z$ (deviations from SM values) derived from the distribution shown on the left side [@CDFweb].[]{data-label="fig4"}](figures/g_kappa-2TeV.eps "fig:"){width=".45\textwidth"}
Conclusions
===========
The increasing luminosities at Tevatron enable the precise study of single and diboson production as well as the measurement of fundamental parameters like $M_W$. Furthermore precise measurements are expected from the continuously increasing data sets which are collected by both the CDF and D0 experiments.
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to thank my colleagues from the CDF and D0 collaborations for providing their excellent results and the organizers for the stimulating conference.
[99]{}
CDF Collaboration, [http://www-cdf.fnal.gov/physics/ewk/]{}.
D0 Collaboration, [http://www-d0.fnal.gov/Run2Physics/WWW/results/ew.htm]{}.
V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Lett. B. [**670**]{}, 292 (2009). K. Melnikov and F. Petriello, Phys. Rev. D [**74**]{}, 114017 (2006). F. Landry, R. Brock, P. M. Nadolsky and C. P. Yuan, Phys. Rev. D [**67**]{}, 073016 (2003). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett. [**100**]{}, 102002 (2008). S. Berge, P. M. Nadolsky, F. Olness and C. P. Yuan, Phys. Rev. D [**72**]{}, 033015 (2005). D0 Collaboration, Note 5755-CONF (2008).
V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. D [**76**]{}, 012003 (2007). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett. [**101**]{}, 211801 (2008). P. M. Nadolsky [*et al.*]{}, Phys. Rev. D [**78**]{}, 013004 (2008). CDF Collaboration, Note 8942 (2007).
D. E. Acosta [*et al.*]{} \[CDF Collaboration\], Phys. Rev. D [**71**]{}, 052002 (2005). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett. [**101**]{}, 191801 (2008). CDF and D0 Collaborations, arXiv:0808.0147 \[hep-ex\].
T. Aaltonen [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett. [**99**]{}, 151801 (2007) and Phys. Rev. D [**77**]{}, 112001 (2008). T. Aaltonen [*et al.*]{} \[CDF Collaboration\], Phys. Rev. Lett. [**100**]{}, 201801 (2008). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. Lett. [**101**]{}, 171803 (2008). V. M. Abazov [*et al.*]{} \[D0 Collaboration\], Phys. Rev. D [**78**]{}, 072002 (2008).
[^1]: for the DØ and CDF Collaborations
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Institute of Theoretical and Experimental Physics, 117259 Moscow, Russia'
author:
- 'A. BOGATYREV'
title: 'OPEN AND HIDDEN CHARM PRODUCTION WITH THE HERA-B EXPERIMENT'
---
Introduction
============
The mechanisms by which charmonium and open charm states are produced in hadronic collisions are not well understood and are the focus of much interest. Although most of the results from previous experiments are qualitatively described by different theoretical models, there are some quantitative discrepancies that may be resolved with a better understanding of non-pertubative effects in the production of heavy quarks. Taking into account that the production mechanisms can be identified by their strong kinematic dependences, it is crucial to have new measurements with high statistics and broad kinematic coverage to accomplish comprehensive descriptions of charm production in nuclei. In this context the HERA-B experiment provides new data on open and hidden charm production.
The HERA-B Experiment
=====================
HERA-B is a fixed target experiment at the HERA storage ring at DESY. Diverse hidden and open charm states are produced in proton-nucleus inelastic collisions. The $pN$ energy in the center of mass corresponds to 41.6 GeV.
The spectrometer emphasizes vertexing, tracking, particle identification and features a dedicated multilevel $J / \psi$ trigger.
The HERA-B target is a fixed target system consisting of two stations separated by 4 cm along the beam direction. Each station houses 4 target wires which can be moved independently into the beam halo.
The side view of the detector is shown in Fig. \[fig:det\]. The protons come from the left and interact in the target region. The target station is followed by the Vertex Detector System(VDS), which provides a precise reconstruction of primary and secondary vertices. The main tracking system(ITR+OTR) is placed behind the VDS. It uses different technologies, depending on the distance to the beam center.
Particle identification is performed by a Ring Imaging Cherenkov Counter(RICH), an Electromagnetic Calorimeter(ECAL) and a MUON detector(MUON).
The $J / \psi$ trigger [@herabtrigger] is initiated by di-lepton signatures from either MUON or ECAL systems. The information on regions of interest defined by pretriggers is used as seeds for the following search for tracks in the tracking system. Then the trigger performs the extrapolation of found tracks through the magnet and follows them through the VDS. In addition a vertex constraint is applied. Events with at least two fully reconstructed di-lepton candidates which come from a common vertex pass the trigger requirements.
Data Set 2002/3
===============
The HERA-B collaboration collected data during 3 month data-taking period in 2002/2003. The data samples were obtained using carbon, titanium and tungsten wires. During that time 150 million di-lepton triggered events were acquired. About 170,000 of $J / \psi$ were reconstructed in muon channel and about 150,000 of $J / \psi$ were found in electron channel. The trigger performance has been significantly improved since the year 2000. For comparison, in 2000 the $J / \psi$ rate was 25-35 per hour and in 2002/2003 we already had 1000-1500 $J / \psi$ per hour. Significant improvements in the data acquisition system had been made as well, which resulted in achieving high recording rates, which in turn allowed accumulation of 210 million minimum bias events in just one week of running. The Minimum Bias trigger selected events satisfying a requirement on either a minimum number of hits in RICH or a minimum energy deposited in ECAL. It allowed to suppress events with no interaction without introducing any significant bias.
Preliminary results. The Minimum Bias Data
==========================================
The following open charm decays are fully reconstructed in minimum bias data:
$D^0 \rightarrow K^- \pi^+$
$D^+ \rightarrow K^- \pi^+ \pi^+$
$D^{*+} \rightarrow D^0 \pi^+$ 0.4cm $D^0 \rightarrow K^- \pi^+$\
Throughout this paper charge conjugate decays are also implied.
The invariant mass plots for $D^+$ and $D^0$ are shown in Fig. \[fig:dplus\] and Fig. \[fig:d0\] respectively. The bump seen to the left of the $D^0$ in Fig. \[fig:d0\] is produced by reflection of not fully reconstructed charm decays, for instance $D^0 \rightarrow K^- \pi^+ \pi^0$. The shape of this reflection was obtained from MC generated $c\overline{c}$ events and was used in the fit as a component of the background in $K \pi$ invariant mass plot.
The mass difference between the mass of $D^{*+}$ and the mass of $D^0$ is shown in Fig. \[fig:dstar\]. This quantity peaks close to the pion mass and has a very small width. The peak is narrow due to the fact that the errors in calculation of the invariant mass of three particles (in case of $D^{*+}$) and two particles (in case of $D^0$) are highly correlated, and therefore cancel out to a large extent.
The signal yields for the fully reconstructed $D^+$, $D^0$ and $D^{*+}$ decays are shown in Table \[tab:dyields\].
$D^+ \rightarrow K^- \pi^+ \pi^+$ $D^0 \rightarrow K^- \pi^+$ $D^{*+} \rightarrow D^0 \pi^+$
----------------------------------- ----------------------------- --------------------------------
$98\pm12$ $189\pm20$ $43\pm8$
: Signal yields for $D^0$, $D^+$ and $D^{*+}$.\[tab:dyields\]
The sample of open charm collected by HERA-B constitutes one of the highest statistics available in proton beam experiments. Therefore the HERA-B will provide the precise measurement of $D^0$, $D^+$ and $D^{*+}$ cross sections as well as $\frac{D^+}{D^0}$ and $\frac{D^{*+}}{D^0}$ ratios.
$-0.1<x_F<0.05$ full $x_F$
-------------------------- ------------------------------------- -------------------------------------
$ \sigma(D^+)\mu b/nucl$ $11.5 \pm 1.7_{stat} \pm 2.2_{sys}$ $30.2 \pm 4.5_{stat} \pm 5.8_{sys}$
$ \sigma(D^0)\mu b/nucl$ $21.4 \pm 3.2_{stat} \pm 3.6_{sys}$ $56.3 \pm 8.5_{stat} \pm 9.5_{sys}$
: Results on $D^+$ and $D^0$ production cross sections.\[tab:dcross\]
The preliminary results on $D^+$ and $D^0$ cross sections are summarized in Table \[tab:dcross\]. The $x_F$ range of this analysis extends from -0.1 to 0.05. Also shown are the total cross sections which were calculated by extrapolating over the full $x_F$ range. In the calculation of the ratio between $D^+$ and $D^0$ cross sections the systematic errors related to detector efficiencies are minimized and the dependence on the absolute luminosity determination is removed. This allows the accurate measurement of the cross section ratio. Result for the $\frac{D^+}{D^0}$ ratio is $0.54 \pm 0.11_{stat} \pm 0.14_{sys}$.
The comparison of our measurement for $\frac{D^+}{D^0}$ ratio with those of other experiments is shown in Fig. \[fig:wratio\]. For the ratio the experimental situation is unclear, and the uncertainties are large particularly for proton beam experiments. The line on the plot shows the expected theoretical value for the ratio of charged to neutral $D$ mesons which is approximately equal to $\frac{1}{3}$ and is the same for the proton and pion induced interactions [@frix]. The big difference between the $D^+$ and $D^0$ production results from the fact that $D$ mesons can be produced either directly or through feed-down from $D^*$ decays. The prediction for the ratio is obtained by taking into account the well known branching ratios for the decays of $D^*$ into $D$ mesons.
(-130,135)[$\frac{D^+}{D^0}$]{}
(-133,137)[$D\overline{D}$($\mu b/nucl$)]{}
Our result on $D\overline{D}$ cross section is shown in Fig. \[fig:cross\]. It is compatible with previous data and has smaller experimental uncertainty in comparison with the experiments at nearby proton beam energies.
Preliminary results. The Di-lepton Data
=======================================
The Di-lepton invariant mass spectra {#subsec:prod}
------------------------------------
The $\mu^+\mu^-$ and $e^+e^-$ invariant mass spectra are shown in Fig. \[fig:jpsimumu\] and Fig. \[fig:jpsiee\] respectively. Clear $J / \psi$ and $\psi(2S)$ resonance peaks are visible. Lower mass particle decays of $\rho$, $\omega$, $\phi$ into $l^+l^-$ pair are also seen.
$J / \psi$ Differential Distributions {#subsec:prod}
-------------------------------------
The huge sample of $J / \psi$ decays accumulated in 2002/2003 and the large acceptance covered by the HERA-B spectrometer provide a broad coverage in $p_T$, up to 4.8 GeV. The differential distribution for $J / \psi$ as a function of transverse momentum is shown in Fig. \[fig:pt\]. Superimposed on the plot is a fit to the parameterization function:
$$\frac{d \sigma}{d p_\mathrm{T} ^2 } = A \cdot
\left[ 1 + \left( \frac{35 \cdot \pi \cdot p_{T} }{256 \; \cdot <p_\mathrm{T}> } \right)^2
\right] ^{-6}$$
Our results on the average transverse momentum as well as the results [@e771; @e789au] from E771 and E789 experiments are shown in Table \[tab:jpsipt\]. The results obtained in electron and muon channel are compatible with each other for each target material. The HERA-B measurement confirms the trend of increasing average $p_T$ with increasing atomic number and energy in the center of mass.
The differential distribution of $J / \psi$ versus Feynman $x_F$ is plotted in Fig. \[fig:xf\]. The points and errors are shown along with a fit to the parameterization function:
$$\frac{d \sigma}{d x_F } = A \cdot ( 1 - | x_F | )^C$$
The HERA-B differs from all other proton-nucleus collision experiments by its large acceptance in the negative $x_F$ region. Our measurement for exponential parameter and the results [@e771; @e789au; @e789cu] from previous experiments are shown in Table \[tab:jpsixf\]. Due to the ongoing acceptance estimations for different running periods, at the moment only a range for the slope is quoted.
Experiment Beam Energy Target $p_T$ Range $<p_{T}> (e^+e^-)$ $<p_{T}> (\mu^+\mu^-)$
------------ ------------- -------- --------------- ---------------------- ------------------------
HERA-B 920 GeV C $p_T<4.8$ GeV $1.22\pm0.01_{stat}$ $1.22\pm0.01_{stat}$
HERA-B 920 GeV W $p_T<4.8$ GeV $1.29\pm0.01_{stat}$ $1.30\pm0.01_{stat}$
E771 800 GeV Si $p_T<3.5$ GeV $1.20\pm0.01$
E789 800 GeV Au $p_T<2.6$ GeV $1.29\pm0.009$
: The HERA-B results on the average transverse momentum for $J / \psi$. Comparison with data from E771 and E789 experiments.\[tab:jpsipt\]
Experiment Beam Energy Target $x_F$ Range c
------------ ------------- -------- ------------------ --------------------------
HERA-B 920 GeV C,W $-0.35<x_F<0.15$ $(5.0-6.5)\pm0.3_{stat}$
E771 800 GeV Si $-0.05<x_F<0.25$ $6.54\pm0.23$
E789 800 GeV Au $-0.03<x_F<0.13$ $4.91\pm0.18$
E789 800 GeV Cu $0.30<x_F<0.95$ $5.21\pm0.04$
: The HERA-B results on the exponential parameter for $J / \psi$ distribution in $x_F$. Comparison with data from E771 and E789 experiments.\[tab:jpsixf\]
A-Dependence for $J / \psi$ {#subsec:prod}
---------------------------
The suppression of charmonium production in heavy-ion collisions is predicted to be a signature for the formation of the quark-gluon plasma. And it is therefore of particular interest to study these mechanisms in the absence of quark-gluon plasma. In this context the HERA-B presents the new measurement of A-Dependence for $J / \psi$.
Nuclear effects are usually considered in a power-law parameterization:
$$\sigma_{pA} = \sigma_{pN} \cdot A^{\alpha}$$
The parameter $\alpha$ can be determined from measurements with two materials:
$$\alpha = \frac{1}{log(A_W/A_C)} \cdot log \left( \frac{N_W}{N_C}
\frac{\mathcal{L}_C}{\mathcal{L}_W}
\frac{\epsilon_C}{\epsilon_W} \right)$$
where $N$ is the number of reconstructed $J / \psi$’s, $L$ is the integrated luminosity, $\epsilon$ - is the detector and trigger efficiency and the acceptance. To reduce the systematic effects only runs with C and W wires operating simultaneously were used.
A-dependence exhibits strong kinematic dependences with Feynman $x_F$ and transverse momentum $p_T$ of the produced $J / \psi$. Therefore we plot $\alpha$ as a function of $x_F$, as shown in Fig. \[fig:alpha\]. Since the ratio of luminosities is under investigation at the moment, the results were normalized to the E866 data [@e866].
Production ratio of $\psi(2S)$ to $J / \psi$ {#subsec:prod}
--------------------------------------------
The production ratio of $\psi(2S)$ to $J / \psi$ is determined according to: $$R_{\psi(2S)} = \frac{\sigma(\psi(2S))}{\sigma(J / \psi)} =
\frac{N(\psi(2S))}{N(J / \psi )} \cdot
\frac{Br(J / \psi \rightarrow l^+l^- )}{Br(\psi(2S) \rightarrow l^+l^- )} \cdot
\frac{\epsilon(J / \psi)}{\epsilon(\psi(2S))}$$ The electron channel (see Fig. \[fig:psiee\]) gives $0.13\pm0.02_{stat}$ for the $R_{\psi(2S)}$ ratio, which is in good agreement with the result from the muon channel. We also measure the $\psi(2S)$ cross section relative to the $J / \psi$. The advantage of this method is that detector and trigger efficiencies largely cancel. The value for the $J / \psi$ cross section is based on E771 and E789 measurements [@e771jpsicross; @e789jpsicross] and corresponds to $\sigma(J / \psi) = 357\pm8_{stat}\pm27_{sys}nb/nucl$ [@bb]. The resulting value for the $\psi(2S)$ cross section is $\sigma(\psi(2S)) = 46\pm12_{stat}nb/nucl$ and is shown in Fig. \[fig:psicross\] along with the results from previous experiments.
Fraction of $J / \psi$ produced via $\chi_c$ decays {#subsec:prod}
---------------------------------------------------
The production ratio of $\chi_c$ to $J / \psi$ can help to discriminate between different theoretical models. The $\chi_c$ is reconstructed in the decay $\chi_c \rightarrow J / \psi \gamma \rightarrow l^+ l^- \gamma$. Due to the small branching ratio of $ \chi_{c0} \rightarrow J / \psi \gamma$, only $\chi_{c1}$ and $\chi_{c2}$ contribute to the observed $\chi_c$ signal. The energy resolution in the HERA-B is insufficient to distinguish $\chi_{c1}$ and $\chi_{c2}$ states. The fraction of $J / \psi$ produced via radiative $\chi_c$ decays is determined from:
$$R_{ \chi_c } = \frac{ \sum_{i=1}^{2} \sigma(\chi_{ci}) \cdot Br( \chi_{ci}
\rightarrow J / \psi \gamma) }{\sigma(J / \psi)} =
\frac{N(\chi_c)}{N(J / \psi)} \cdot
\frac{\epsilon (J / \psi)}{\epsilon (\chi_c) \cdot \epsilon (\gamma ) }$$
The published result based on data from an earlier HERA-B run with limited statistics for $\chi_c$ is $ R_{\chi_c} = 0.32 \pm 0.06_{stat} \pm 0.04_{sys} $ [@chic], which is compatible with most of the previous data and the NRQCD prediction.
In the 2002/2003 HERA run, HERA-B collected a $\chi_c$ sample which significantly exceeds the statistics obtained by previous experiments. The mass difference distribution $\Delta M=M(l^+l^-\gamma)-M(l^+l^-)$ is shown in Fig. \[fig:chic\]. An excess of events with respect to combinatorial background determines the number of $\chi_c$ candidates. The same plot, but after background subtraction is shown in Fig. \[fig:chicdifference\]. The shape of the background is taken by combining $J / \psi$ and photon candidates from different events. The plot corresponds to $15 \%$ of statistics available in muon channel. We expect about 15,000 $\chi_c$ for the full data set combined from electron and muon samples. The preliminary result for $ R_{\chi_c}$ from the 2002/2003 data is $0.21\pm0.05_{stat}$.
Summary
=======
HERA-B has finished data-taking and now focuses on analysis of large accumulated data samples: 150 million of events with di-lepton trigger and 210 million of events with minimum bias trigger. These data were obtained on a variety of nuclear targets. About 300,000 $J / \psi$’s decaying into electron and muon pairs were reconstructed in the di-lepton sample. HERA-B features broad coverage in $p_T$ and explores the negative $x_F$ region for the first time. Differential distributions for $J / \psi$ as functions of $p_T$ and $x_F$ are shown. New data on $J / \psi$ suppression are presented. Results on the production ratios of $\psi(2S)$ to $J / \psi$ and $\chi_c$ to $J / \psi$ are also shown. Clear signals of $D^+$, $D^0$ and $D^{*+}$ mesons were obtained in the minimum bias data. Results on $D^+$ and $D^0$ production cross sections and the $D^+$ to $D^0$ ratio are discussed.
Lots of physics analyses are ongoing and we are looking forward to new results.
References {#references .unnumbered}
==========
[99]{}
M. T. Nunez Pardo de Vera, [*Nucl. Phys. Proc. Suppl.*]{} [120:166-172]{}[ (2003)]{}. S. Frixione, M. L. Mangano, P. Nason, G. Ridolfi, [*Adv. Ser. Direct. High Energy Phys.*]{}[ 15]{}[ 609-706]{}[ (1998)]{}. T. Alexopoulos [*et al*]{}, . M.H. Schub [*et al*]{}, . M.S. Kowitt [*et al*]{}, . M.J. Leich [*et al*]{}, . T. Alexopoulos [*et al*]{}, . D. M. Jansen [*et al*]{}, . I. Abt [*et al*]{}, . I. Abt [*et al*]{}, .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a first-principles study of the electronic structures and properties of ideal (atomically sharp) LaAlO$_3$/ SrTiO$_3$ (001) heterointerfaces and their variants such as a new class of quantum well systems. We demonstrate the insulating-to-metallic transition as a function of the LaAlO$_3$ film thickness in these systems. After the phase transition, we find that conduction electrons are bound to the $n$-type interface while holes diffuse away from the $p$-type interface, and we explain this asymmetry in terms of a large hopping matrix element that is unique to the $n$-type interface. We build a tight-binding model based on these hopping matrix elements to illustrate how the conduction electron gas is bound to the $n$-type interface. Based on the ‘polar catastrophe’ mechanism, we propose a new class of quantum wells at which we can manually control the spatial extent of the conduction electron gas. In addition, we develop a continuous model to unify the LaAlO$_3$/SrTiO$_3$ interfaces and quantum wells and predict the thickness dependence of sheet carrier densities of these systems. Finally, we study the external field effect on both LaAlO$_3$/SrTiO$_3$ interfaces and quantum well systems. Our systematic study of the electronic reconstruction of LaAlO$_3$/SrTiO$_3$ interfaces may serve as a guide to engineering transition metal oxide heterointerfaces.'
author:
- 'Hanghui Chen$^{1,3}$ Alexie Kolpak$^{2,3}$ and Sohrab Ismail-Beigi$^{1,2,3}$'
bibliography:
- 'sto-lao-long-prb11.bib'
title: 'A First-Principles Study of the Electronic Reconstructions of $\textrm{LaAlO}_3/\textrm{SrTiO}_3$ Heterointerfaces and Their Variants'
---
Introduction
============
With the advance of techniques to control thin film growth on the atomic scale, the study of epitaxial oxide heterostructures is a rapidly developing area of materials science [@Dagotto-Science-2007]. Due to the ability to produce a well-defined single terminated substrate surface [@Kawasaki-Science-1994], interfaces that are nearly atomically sharp can be formed. In many cases, the properties of these interfaces turn out to be much richer than those of the bulk constituents. One of the most interesting examples is the (001) interface between LaAlO$_3$ (LAO) and SrTiO$_3$ (STO), both constituents of which are conventional band insulators in the bulk. Among the intriguing phenomena observed at this interface is the presence of a high-mobility quasi two-dimentional electron gas [@Hwang-Nature-2004], which is sensitive to the LaAlO$_3$ film thickness on the nanoscale [@Mannhart-Science-2006; @Huijben-NatMat-2006; @Bell-APL-2009], and can be tuned via an external field [@Mannhart-Science-2006; @Triscone-Nature-2008]. In addition, both magnetism [@Brinkman-NatMat-2007] and superconductivity [@Mannhart-Science-2007] have been observed at LaAlO$_3$/SrTiO$_3$ interfaces.
Due to these interface properties, the LaAlO$_3$/SrTiO$_3$ system shows great promise for the development of novel applications in nano-scale oxide electronics [@Cen-NatMat-2008; @Cen-Science-2009]. However, though significant efforts have been made in both theory [@Pickett-PRB-2006; @Pickett-PRB-2008; @Freeman-PRB-2006; @Albina-PRB-2007; @Tsymbal-PRL-2009; @Kelly-EPL-2008; @Demkov-PRB-2008] and experiment [@Willmott-PRL-2007; @Kala-PRB-2007; @Herranz-PRL-2007; @Siemons-PRL-2007; @Sing-PRL-2009; @Salluzzo-PRL-2009; @Thiel-PRL-2009], the origin of the new phases emerging at the LaAlO$_3$/SrTiO$_3$ interface has not been completely resolved yet. In part, this is due to the fact that more than one mechanism may play a role in determining the behavior, and furthermore, the dominant mechanism can depend on external conditions [@Eckstein-NatMat-2007].
One mechanism that is generally believed to play a role in the formation of the two-dimensional electron gas is known as the polar catastophe. This mechanism is a direct result of the charge discontinuity that occurs at an abrupt interface between a nonpolar (SrTiO$_3$) and a polar (LaAlO$_3$) material. Both materials are perovskite oxides with an $ABO_3$ structure that forms alternating planes of $AO$ and $BO_2$ stacked along the (001) direction. Consequently, there are two types of LaAlO$_3$/SrTiO$_3$ interfaces along this direction: TiO$_2$/LaO (known as the $n$-type interface) and SrO/AlO$_2$ (the $p$-type interface) [@Hwang-Nature-2004; @Hwang-NatMat-2006]. In the ionic limit, SrTiO$_3$ is composed of charge neutral atomic layers (SrO)$^0$ and (TiO$_2$)$^0$, while LaAlO$_3$ consists of positively charged (LaO)$^{+}$ and negatively charged (AlO)$^{-}$ atomic layers. When LaAlO$_3$ is grown epitaxially on the SrTiO$_3$ substrate along the (001) direction, an internal electric field directed from (LaO)$^+$ to (AlO$_2$)$^-$ is formed through LaAlO$_3$, resulting in a potential difference that, mathematically, diverges with increasing LaAlO$_3$ thickness. In reality, to offset the diverging potential, an electronic reconstruction is expected to occur, with electrons transferring from the film surface (or a $p$-type interface) to the $n$-type interface. The polar catastrophe mechanism [@Hwang-Nature-2004; @Hwang-NatMat-2006] has been confirmed in density functional theory (DFT) calculations [@Demkov-PRB-2008; @Chen-PRB-2009; @Pickett-PRL-2009; @Son-PRB-2009], and could be responsible for the presence of the two-dimensional electron gas at the atomically sharp $n$-type interface (i.e., with no defects).
However, the polar catastrophe mechanism alone can not explain all of the electronic properties of the LaAlO$_3$/SrTiO$_3$ interface, such as the observed confinement of the transferred electrons within several nanometers of the $n$-type interface [@Basletic-NatMat-2008]. In previous work [@Chen-PRB-2009], we have shown that the electrons are bound to the $n$-type interface as a result of the chemical environment of the interface, which produces a large hopping matrix element between La and Ti.
In this paper, we use both DFT and phenomenological modeling to provide a more detailed and complete picture of the electronic reconstruction at ideal LaAlO$_3$/SrTiO$_3$ interfaces. We predict the behavior of the sheet carrier density as a function of the LaAlO$_3$ thickness by constructing a continuous model which approximates the LaAlO$_3$ as a homogeneous media. We also use both DFT calculations and model calculations to give a simple yet quantitative picture of the external field effect that has been experimentally realized at the $n$-type interface [@Mannhart-Science-2006; @Triscone-Nature-2008]. In addition, we propose a new class of quantum well systems, based on the polar catastrophe mechanism, at which the spatial extent of the two-dimensional electron gas can be manually controlled.
The remainder of the paper is organized as follows. In Section \[method\], we discuss the technical details of our DFT calculations. We present our first-principles results in Section \[DFT\]. Section \[DFT-symmetric\] gives a brief discussion on the symmetric superlattices. In \[DFT-polar\], we briefly describe the polar catastrophe at various types of interfaces, and we introduce the new quantum well systems. Section \[DFT-bound\] discusses the computation of the on-site and hopping matrix elements and a tight-binding model. An exceptionally large Ti-La hopping is found at the $n$-type interface and its origin is explained. Section \[oxygen-repulsion\] studies the behavior of oxygen vacancies at the interfaces. Section \[DFT-thickness\] examines the thickness dependence of the sheet carrier density. Simulations of the external field effect are presented in \[DFT-external\]. In Section \[model\], we present a continuous model which gives a simple yet quantitative description of both the sheet carrier density and the external field effect. We conclude in Section \[conclusion\].
Method Details {#method}
==============
Our calculations are performed using density functional theory within the *ab initio* supercell plane-wave approach [@Payne-RMP-1992], with the code PWscf in the Quantum-ESPRESSO package [^1]. We employ the local density approximation (LDA) [@Kohn-PR-1965] and ultrasoft pseudopotentials [@Vanderbilt-PRB-1990], which include semi-core states in Sr, Ti and La. The reference configurations of the pseudopotentials are: Sr $4s^24p^65s^2$ ($r^s_c=2.0 a_0,
r^p_c=1.8 a_0$), Ti $3s^23p^63d^14s^2$ ($r^s_c=r^p_c=r^d_c=1.8
a_0$), La $5s^25p^65d^16s^{1.5}6p^{0.5}$ ($r^s_c=r^d_c=2.2 a_0,
r^p_c=2.0 a_0)$, Al $3s^23p^1$ ($r^s_c=1.8 a_0, r^p_c=1.82 a_0$) and O $2s^22p^4$ ($r^s_c=r^p_c=1.3 a_0$), where $a_0$ is the Bohr radius. The plane wave basis energy cutoff and charge cutoff are 30 Ry and 180 Ry, respectively. We use a Gaussian smearing width of 5 mRy when sampling the Brillouin zone. The $k$-grid sampling of the Brillouin zone is $10\times10\times 1$ where the $z$-axis is orthogonal to the interface. The convergence of the total energy and total charge density has been checked with $k$-grids of up to $20\times20\times1$. Periodic copies are separated by $\sim$15 Å of vacuum. We also extend the vacuum to $30$ Å to check the convergence. The effect of the artifical electric fields in the vacuum due to the periodic boundary condition, which turns out to be quite small, is discussed in Appendix \[appendix-PBC-intrinsic\]. The force convergence threshold is 26 meV/Å. In some key results we reduce the threshold to 13 meV/Å to check the convergence.
In all our calculations, the interfaces are along the (001) direction so that the $z$ axis is perpendicular to the interface. The $x$ and $y$ directions of the simulation cell are subject to periodic boundary conditions and their lengths are fixed to the theoretical lattice constant of SrTiO$_3$ $a=3.85$ Å (1.5% smaller than the experimental value). The atomic coordinates are relaxed [^2] until every force component is smaller than the convergence threshold. The detailed configurations for different calculations are specified in each section below.
DFT results {#DFT}
===========
Symmetric double $n$-type and $p$-type superlattices {#DFT-symmetric}
----------------------------------------------------
The simplest approach to studying the $n$-type or $p$-type [SrTiO$_3$/LaAlO$_3$ ]{}interface is to use a symmetric superlattice approach [@Pickett-PRB-2006; @Pickett-PRB-2008; @Freeman-PRB-2006; @Albina-PRB-2007; @Gemming-Mater-2006; @Tsymbal-JAP-2008; @Tsymbal-PRL-2009; @Satpathy-PRL-2008; @Kelly-EPL-2008], as exemplified by Fig. \[fig:symif\], which shows a simulation cell without vacuum in which a number of [SrTiO$_3$ ]{}unit cells are adjacent to a number of [LaAlO$_3$ ]{}unit cells and periodic boundary conditions are imposed. Therefore, a non-stoichiometric supercell contains two identical [SrTiO$_3$/LaAlO$_3$ ]{}interfaces: [*e.g.*]{} in Fig. \[fig:symif\]a, we have an additional TiO$_2$ layer in [SrTiO$_3$ ]{}and an additional LaO layer in LaAlO$_3$, and in Fig. \[fig:symif\]b, there is an additional SrO layer in [SrTiO$_3$ ]{}and an additional AlO$_2$ in LaAlO$_3$. In this way, the superlattice contains two $n$-type interfaces (TiO$_2$/LaO) or two $p$-type interfaces (SrO/AlO$_2$).
The advantage of the superlattice approach is that no vacuum is needed in the simulation cell, making computation easier. However, due to the imposed symmetry and non-stoichiometry of the LaAlO$_3$ film, this geometry does not result in a polar field, so the evolution of the polar catastrophe can not be modeled. In addition, the non-stoichiometry of the LaAlO$_3$ also imposes a fixed carrier doping: in the ionic limit, an LaO (AlO$_2$) layer has a charge of +1 (-1), and there is one extra electron (hole) present in the conduction (valence) band of LaAlO$_3$, which is shared evenly by the two interfaces. Therefore each interface is doped by 0.5 electrons ($n$-type) or 0.5 holes ($p$-type) per two-dimensional unit cell. These values are precisely those needed to fully compensate the polar field of LaAlO$_3$ [@Hwang-NatMat-2006], so the symmetric supercell approach is equivalent to studying the properties of the interfaces when the LaAlO$_3$ film is very thick (infinite thickness limit).
We briefly present results on the symmetric interfaces and highlight some key observations and questions that we will answer in later sections. We show in Fig. \[fig:chargespatial\]a the $xy$-plane integrated *conduction* electron density for the double $n$-type superlattice, and in Fig \[fig:chargespatial\]b the hole density for the double $p$-type superlattice. The method of how to calculate the transferred charge density is detailed in Appendix \[appendix-density\]. The integrated densities are 0.49 electrons per $n$-type interface and 0.49 holes per $p$-type interface, showing that the ionic picture for LaAlO$_3$ is highly accurate, and that one should expect 0.5 electrons (holes) at the fully compensated $n$-type ($p$-type) interface for very thick LaAlO$_3$ films.
The density profiles in Fig. \[fig:chargespatial\] display an interesting asymmetry. One can characterize the electrons in the double $n$-type system as being bound to the interface. As the figure shows, they are confined in the [SrTiO$_3$ ]{}, with the density decreasing away from the interface. On the other hand, in the double $p$-type system, the holes reside in both the LaAlO$_3$ and SrTiO$_3$, do not show a strong preference for the interface itself, and have a slightly higher amplitude in the LaAlO$_3$.
Our calculated band offsets place the LaAlO$_3$ valence band maximum $\simeq$ 0.1 eV above the SrTiO$_3$ valence band maximum ($p$-type) and the LaAlO$_3$ conduction band minimum $\sim$ 2 eV above the SrTiO$_3$ conduction band minimum ($n$-type) [@Albina-PRB-2007; @Pickett-PRB-2008]. [*A priori*]{}, one would expect that any introduced carriers would occupy the most energetically favorable band edge available, [*i.e.*]{} holes would migrate to the LaAlO$_3$ valence band edge and electrons to the SrTiO$_3$ conduction band edge. Furthermore, the minimization of kinetic energy would lead to a relatively uniform distribution of carriers. In Fig. \[fig:chargespatial\], we see that, overall, the holes prefer to be in the LaAlO$_3$ at the $p$-type interface, while the electrons reside in SrTiO$_3$ at the $n$-type interface. For the $p$-type interface, the small value of the band offset and the large hole density result in some “leakage” into the SrTiO$_3$ valence band states.
However, the binding of the electrons at the $n$-type interface cannot be explained by the above arguments. As no polar fields are present in this system to create a potential that will bind the electrons, what mechanism overcomes the kinetic cost inherent in localizing them to the interface? Whatever mechanism is present is absent for the holes, whose spatial profile is relatively uniform in both materials. We provide the answer to this fundamental question in Section \[DFT-bound\], when we discuss the tight-binding hopping matrix elements across both interfaces.
Before we move onto the next section, we would like to give a short discussion on the valence band offset (VBO) of LaAlO$_3$/SrTiO$_3$ interfaces. The VBO of the $n$-type interface has been calculated by several groups [@Albina-PRB-2007; @Lee-MRS-2007; @Pickett-PRB-2008; @Cen-NatMat-2008; @Satpathy-PRL-2008]. Although all of the computed values place the conduction band edge of LaAlO$_3$ significantly higher than that of SrTiO$_3$ at the $n$-type interface, the large range of computed VBO values (shown in Table I), is puzzling. To investigate this issue, we use two different approaches to calculate the VBO of the $n$-type interface and the $p$-type interface. The valence band offset is defined by:
$$\label{vbo}
E_{\textrm{VBO}}=E_{\textrm{V}}^{\textrm{LAO}}-E_{\textrm{V}}^{\textrm{STO}}$$
where $E_{\textrm{V}}^{\textrm{LAO}}$ and $E_{\textrm{V}}^{\textrm{STO}}$ are the valence band edges of LaAlO$_3$ and SrTiO$_3$, respectively. The simulation cell is a symmetric nonstoichiometric superlattice with 12.5 layers of SrTiO$_3$ and 4.5 layers of LaAlO$_3$. The first method to determine the VBO is to analyze the local density of states, referred to as the LDOS approach [@Balereschi-JPD-1998; @Albina-PRB-2007]. The other method is to use the macroscopic average potential, denoted further as the MA approach [@Baldereschi-PRL-1988; @Baroni-PRB-1991; @Albina-PRB-2007]. We compare our results with other theoretical calculations and the recent experiment in Table \[tab:vbo\].
$n$-type (TiO$_2$/LaO) $p$-type (SrO/AlO$_2$)
--------------------------- ------------------------- ------------------------
present study -0.47 (MA) -0.39 (LDOS) -0.02 (MA) 0.11 (LDOS)
Ref. [@Albina-PRB-2007] 0.51 (MA) 0.39 (LDOS) 0.19 (MA) 0.10 (LDOS)
Ref. [@Cen-NatMat-2008] 1.1 NA
Ref. [@Pickett-PRB-2008] -0.15 (LDOS) NA
Ref. [@Satpathy-PRL-2008] -0.9 (LDOS) NA
Ref. [@Lee-MRS-2007] -0.1 (MA) NA
Expt. [@Segal-PRB-2009] -0.35$\pm$0.18 NA
: \[tab:vbo\] The valence band offset (VBO) of LaAlO$_3$/SrTiO$_3$ $n$-type and $p$-type interfaces in eV. Each value is followed by the method used to determine it in parentheses.
We can see from Table \[tab:vbo\] that for the $n$-type interface not only the magnitude of VBO differs, but the sign is not unanimous. One possible origin for the range of theoretical values is likely that the lattice constants have minute differences (due to different pseudopotentials), which causes a variation in strain and probably affects the alignment of valence band edges. Another possibility is that we, as well as a few other groups [@Albina-PRB-2007; @Lee-MRS-2007; @Pickett-PRB-2008; @Satpathy-PRL-2008], use symmetric nonstoichiometric superlattices, in which the macroscopic average potentials are flat in both LaAlO$_3$ and SrTiO$_3$, to perform the calculations. However, if a stoichiometric LaAlO$_3$ slab is employed [@Cen-NatMat-2008] so that a net internal electric field results, the determined VBO will likely be different from the symmetric nonstoichiometric case. Although our results for the $n$-type interface agree well with the available experiments, the value as well as variations of the VBO for the $n$-type interface is not a closed subject and needs further work. In contrast, the computed values of the VBO for the $p$-type interface are in general agreement and $\simeq$ 0.1 eV.
The polar catastrophe {#DFT-polar}
---------------------
### $n$-type, $p$-type and $np$-type interfaces
As discussed in the Introduction, there are two types of LaAlO$_3$/SrTiO$_3$ interfaces: the $n$-type (TiO$_2$/LaO) interface and the $p$-type (SrO/AlO$_2$) interface. If a stoichiometric LaAlO$_3$ layer on an SrTiO$_3$ substrate also has a SrTiO$_3$ capping layer, then we have both $n$-type and $p$-type interfaces in the same system. Below, we refer to this type of geometry as the $np$-type interface. Experimentally, the $n$-type, $p$-type and $np$-type interfaces have all been fabricated. In this study, we simulate all three types of interfaces, the configurations of which are illustrated in Fig. \[fig:interface\]. For the $np$-type interface, we use 5 unit cells of SrTiO$_3$ to simulate the substrate and another 5 unit cells of SrTiO$_3$ as a capping layer. For both the $n$-type and $p$-type interfaces, we use 11 unit cells of SrTiO$_3$ to simulate the substrate. In all cases, the thickness of LaAlO$_3$ is varied from 1 to 7 unit cells. In addition, the first two SrTiO$_3$ unit cells in the substrate are fixed at the ideal perovskite positions to simulate the bulk-like interior of the substrate [@Chen-PRB-2009]. The termination of the SrTiO$_3$ substrate and capping layers is always SrO. As illustrated in our previous work [^3], the TiO$_2$ termination has a surface O-$p$ state which is $\simeq$ 0.5 eV higher than the valence band edge of bulk SrTiO$_3$, while the SrO termination does not have this surface state. Since we are interested in the evolution of polar fields of LaAlO$_3$ on very thick SrTiO$_3$ substrates, this pure surface state should be avoided when simulating the SrTiO$_3$ substrate and thick SrTiO$_3$ capping layers. However, for thin SrTiO$_3$ capping layers, the presence of this O-$p$ state on the TiO$_2$-terminated surface will lower the critical separation, which has recently been discussed [@Pentcheva-2009].
![\[fig:interface\] Schematics of [the supercells]{} and energy diagrams for different types of interfaces. **a)** The $np$-type interface. **b)** The $n$-type interface. **c)** The $p$-type interface. **d)** The new quantum well systems.](fullconfig.ps){width="13cm"}
We observe that an insulating-to-metallic transition occurs in our simulations when the [LaAlO$_3$ ]{}film reaches a critical thickness, the value of which depends on the system geometry. The phase transition can be explained in terms of an energy diagram shown in Fig. \[fig:interface\]. For the $np$-type and $n$-type interfaces, due to the polar structure of LaAlO$_3$, the electric field through the LaAlO$_3$ film lifts up the valence band edge of the LaAlO$_3$ and reduces the energy gap. The energy gap of the $np$-type interface is the energy difference between the Ti-$d$ states of the SrTiO$_3$ substrate and the O-$p$ states of the SrTiO$_3$ capping layer, while the energy gap of the $n$-type interface is given by the energy difference between the Ti-$d$ states of the SrTiO$_3$ substrate and the O-$p$ states on the surface. For the $p$-type interface, since the polarity of LaAlO$_3$ is reversed, the electric field through the LaAlO$_3$ film decreases the valence band edge of the LaAlO$_3$. However, the energy gap of the $p$-type interface is not the most relevant quantity to monitor for the insulating-to-metallic transition. Rather, the LaAlO$_3$ film reduces the energy difference between the La states on the surface [^4] and the O $p$-states of the SrTiO$_3$ substrate, which from now on we call the ‘La-O energy difference’. In all three cases, the energy gap ($np$-type and $n$-type) or La-O energy difference ($p$-type) diminishes with the LaAlO$_3$ film thickness, finally disappearing when the insulating-to-metallic transition occurs. The minimum number of LaAlO$_3$ unit cells necessary to induce this phase transition is referred to as the ‘critical separation’.
Our DFT calculations, as well as previous studies [@Demkov-PRB-2008; @Chen-PRB-2009; @Pickett-PRL-2009; @Son-PRB-2009], support the above schematics. The values of the calculated energy gap versus the LaAlO$_3$ thickness are listed in Table \[tab:gap\] for the $np$-type, $n$-type and $p$-type interfaces. The corresponding ‘critical separation’ is the smallest thickness of LaAlO$_3$ that makes the interface conducting. The increasing thickness of the LaAlO$_3$ film reduces the energy gap until it disappears and the interface becomes metallic. However, the critical separation depends on the band gap of SrTiO$_3$, which is underestimated in DFT calculations (the calculated band gap is 1.85 eV and the experimental band gap is 3.2 eV). Taking into account the difference between the calculated and experimental band gaps, the realistic critical separation is two more unit cells of LaAlO$_3$ in addition to the calculated one, resulting in 5, 6 and 8 u.c. for the $np$-type, $n$-type and $p$-type interfaces, respectively.
![\[fig:potdos\]The average potential (left) and density of states (right) of the $np$-type interface with different thicknesses. The upper two panels correspond to the $np$-type interface with 1 unit cell of LaAlO$_3$, and the lower two panels correspond to the $np$-type interface with 3 unit cells of LaAlO$_3$. In the right panels, the red, blue, and orange lines are the total density of states (DOS), the atomic projected DOS of Ti-$d$ states, and the atomic projected DOS of O $p$-states, respectively. The vertical green line in the right panels is the Fermi level.](pot-dos.ps){width="10cm"}
In order to give a clearer illustration of how the polar field reduces the energy gap, we show in the left column of Fig. \[fig:potdos\], the $xy$ planar average electrostatic potential [@Baldereschi-PRL-1988] and the associated macro-averaged smoothed potential along the direction perpendicular to the interface for the $np$-type geometry. The $xy$ planar average potential is obtained by averaging the raw three-dimensional total potential over the interface plane. The associated macro-averaged smoothed potential is to convolute the $xy$ planar average potential with a Gaussian function using the filting width $\simeq 0.6$Å. As the figure shows, the macro-averaged smoothed potential in SrTiO$_3$ is flat, indicating that there is no internal electric field through the SrTiO$_3$ ([*i.e.*]{}, it is nonpolar). In the LaAlO$_3$ layers, the macro-averaged smoothed potential is increasing and lifts up the valence band edge of the capping SrTiO$_3$. The energy difference between the O-$p$ states of the SrTiO$_3$ capping layer and those of the SrTiO$_3$ substrate increases with increasing LaAlO$_3$ film thickness. The right column of Fig. \[fig:potdos\] shows the corresponding density of states (DOS). As the upper panel shows, since the potential increase by 1 unit cell of LaAlO$_3$ is not large enough to overcome the band gap of SrTiO$_3$, the system is still insulating. In the lower panel, the potential increase by 3 unit cells of LaAlO$_3$ is larger than the band gap of SrTiO$_3$, so that the Ti-$d$ states in the SrTiO$_3$ substrate and O-$p$ states in the capping SrTiO$_3$ layer overlap. Then the system becomes metallic.
Further evidence of the insulating-to-metallic transition is shown by the local density of states (LDOS) at the Fermi level of the $n$-type interface in Fig. \[fig:ldosf\]a. From the character of these states, we can see that we have Ti-$d$ states in the SrTiO$_3$ substrate and O-$p$ states on the surface. Fig. \[fig:chargespatial\] shows the spatial distributions of the conducting electrons and the holes integrated over $xy$-plane at the $n$-type and $p$-type interfaces, respectively. The details of how to extract out these transferred charge densities are provided in Appendix \[appendix-density\]. From Fig. \[fig:chargespatial\], it is clear that the conducting electrons and holes behave very differently in the SrTiO$_3$ substrate. The conducting electrons, which occupy Ti-$d$ states, decay away from the $n$-type interface over a length scale of $\simeq$ 10 unit cells, indicating that the conducting electron gas is several nm thick. This result is in qualitative agreement with experimental measurements of the superconducting electron gas [@Mannhart-Science-2007] and other theoretical calculations [@Tsymbal-PRL-2009; @Son-PRB-2009]. Unlike the conducting electrons, the spatial distribution of the holes resembles that of a particle in a box, implying that the holes are very weakly bound to the interface. As elucidated in our previous work [@Chen-PRB-2009], the reason for this asymmetry is that the conducting electron gas is trapped at the $n$-type interface due to a large interfacial hopping matrix element. We will discuss this phenomenon further in Section \[DFT-bound\].
### The quantum well system {#sec:quantumwell}
Though the conduction electrons are bound to the $n$-type interface, they still spread over several nanometers. The observed thickness of the electron gas at the $n$-type interface varies widely in experiments, from nanometers to microns, depending on the growth conditions [@Chen-Adv-2010]. Thus it would be desirable to be able to manually control the thickness of the electron gas. Based on the polar catastrophe mechanism, we propose a new class of quantum well systems that will allow this functionality. The quantum well system, which is illustrated in Fig. \[fig:interface\]d, has the following composition: [SrTiO$_3$ ]{}substrate/2 u.c. [LaAlO$_3$ ]{}/$m$ u.c. [SrTiO$_3$ ]{}/[LaAlO$_3$ ]{}capping layer, where $m$ can be varied to control the thickness of the conducting electron gas, as we discuss below. Along the (001) direction, this configuration results in three interfaces: one $p$-type interface between the SrTiO$_3$ substrate and the 2 u.c. [LaAlO$_3$ ]{}buffer layer, and two $n$-type interfaces on both sides of the inserted thin [SrTiO$_3$ ]{}film. Both of the LaAlO$_3$ thin films are stoichiometric in this sytem, while the embedded SrTiO$_3$ thin film is not. As we have two $n$-type interfaces, the quantum well systems have one extra TiO$_2$ layer in the inserted SrTiO$_3$ film. Like the $n$-type and $np$-type interfaces, the two SrTiO$_3$ unit cells of the substrate facing the vacuum are fixed to simulate a bulk-like substrate and all the other atoms are fully relaxed.
The reason that this configuration forms a quantum well can also be explained in terms of an energy diagram. Fig. \[fig:interface\]d shows that the polar fields in the two LaAlO$_3$ thin films point in opposite directions. Therefore, the energy of the Ti-$d$ states in the embedded SrTiO$_3$ thin film is lowered relative to the conduction band edge of the SrTiO$_3$ substrate. (We choose two unit cells of LaAlO$_3$ as the buffer layer in order to bring down the Ti-$d$ states of the embedded SrTiO$_3$ film into the band gap of the SrTiO$_3$ substrate). As the capping LaAlO$_3$ layer thickens, the energy of the O-$p$ states eventually becomes higher than that of the Ti-$d$ states of the embedded SrTiO$_3$ film, and hence an insulating-to-metallic transition occurs. The difference between the quantum well systems and the $n$-type interface is that once the insulating-to-metallic transition happens, the conduction electrons get trapped in the Ti-$d$ states of the embedded SrTiO$_3$ film rather than those of the SrTiO$_3$ substrate. Therefore, by changing the thickness of the embedded SrTiO$_3$ film, we can manually control the spatial extent of the electron gas while holding the number of transferred electrons constant. Consequently, we have two parameters to vary: 1) the thickness of embedded SrTiO$_3$ film to control the spatial extent of electron gas, and 2) the thickness of LaAlO$_3$ capping layer to control the insulating-to-metallic transition. For simplicity, in the following, we focus on one particular subclass in which the embeded SrTiO$_3$ is narrowed down to a single TiO$_2$ layer. The thickness of the LaAlO$_3$ capping layer will be varied. The quantum wells (QW) in the following refer to this particular subclass.
Like the $p$-type interface, at this new class of quantum well systems, the energy gap is not the most direct quantity to monitor for the insulating-to-metallic transition. As the LaAlO$_3$ capping layer is thin, the energy gap of the QW is the energy difference between the Ti-$d$ states of the embedded TiO$_2$ layer and the O-$p$ states of the SrTiO$_3$ substrate, which remains at a constant energy. The more relevant quantity is the energy difference between the Ti-$d$ states of the embedded TiO$_2$ layer and the O-$p$ states on the surface. We refer to this quantity as the ‘Ti-O energy difference’ in the following. Increasing the thickness of the [LaAlO$_3$ ]{}capping layer reduces the Ti-O energy difference until it disappears and electron transfer occurs. The minimum thickness of the LaAlO$_3$ capping layer that enables this electron transfer is defined as the critical separation of the QW systems.
Table \[tab:gap\] shows the Ti-O energy difference and the critical separation of the QW systems. The critical separation of the QW coincides with that of the $n$-type interface, reflecting the fact that the band gap of SrTiO$_3$ is mainly determined by the Ti-$d$ states and the O-$p$ states. Therefore, the absence of Sr atoms does not change the critical separation. The local density of states at the Fermi level of the QW are shown in Fig. \[fig:ldosf\]b. Unlike the $n$-type interface, at which the filled Ti $d$-states extend over $\sim$10 unit cells in the [SrTiO$_3$ ]{}substrate, the only metallic states in the QW systems are localized in the single embedded TiO$_2$ layer, as anticipated.
This new class of quantum wells displays a number of appealing properties that are absent at the LaAlO$_3$/SrTiO$_3$ $n$-type interface. First, the thickness of conduction electrons is controlled by the inserted SrTiO$_3$ film and can be in principle reduced to a single atomic TiO$_2$ layer. Second, the electronic properties of the new quantum wells largely depend on the polar structure of LaAlO$_3$. Therefore, the SrTiO$_3$ substrate can be replaced by other materials on which LaAlO$_3$ can be epitaxially grown, *e.g.* LSAT (La$_{0.29}$Sr$_{0.71}$Al$_{0.65}$Ta$_{0.35}$O$_3$). This provides more choices of substrates on which to grow this new quantum well structure. Finally, the new quantum well serves as a practical way to test the recently proposed hypothesis that some electrons do not contribute to transport due to Anderson localization. Popović *et al.* [@Satpathy-PRL-2008] argued that since all the two dimensional states are Anderson localized by disorder, the electrons that occupy the lowest Ti-$d$ bands do not conduct and therefore the observed sheet carrier density should be much smaller than the 0.5$e$ per two-dimensional unit cell predicted by the polar catastrophe mechanism [@Hwang-NatMat-2006]. Nevertheless, it is not trivial to distinguish from which band conduction electrons originate. The thinnest of this new type of quantum wells has only one Ti-$d$ band at the Fermi level, corresponding to the single embedded TiO$_2$ layer, which has a strong two-dimensional Ti-$d_{xy}$ character. Therefore it should show strong localization and not contribute to conductivity. By thickening the embedded SrTiO$_3$, one should be changing only the extent of the wave functions and see the localization properties. Fig. \[fig:bands\_compare\] shows a comparison of the band structures of the $n$-type interface and the QW. These two systems both have approximately 0.1$e$ per two-dimensional unit cell (from Fig. \[fig:charge\]). However, the Fermi level of the $n$-type interface crosses three Ti-$d$ bands, while the QW system has only one Ti-$d$ band at the Fermi level, as expected.
![\[fig:bands\_compare\] A comparison of the band structures of [**a)**]{} the $n$-type interface and [**b)**]{} the new QW. The $n$-type interface has 5 unit cells of LaAlO$_3$ and the QW has a capping layer of 6 unit cells of LaAlO$_3$. The red solid line is the Fermi level.](bands_compare.ps){width="16cm"}
$np$-type $n$-type $p$-type QW
----- -------------- ----------- ----- -------------- ----------- ----- -------------- ----------- ----- -------------- -----------
$i$ $\Delta$(eV) $V_i$(eV) $i$ $\Delta$(eV) $V_i$(eV) $i$ $\delta$(eV) $V_i$(eV) $i$ $\delta$(eV) $V_i$(eV)
0 1.85 0 1.85 0 N/A 0 1.85
1 1.25 0.60 1 1.71 0.14 1 2.59 1 1.83 0.02
2 0.55 0.70 2 1.28 0.43 2 1.89 0.70 2 1.30 0.53
3 metallic 3 0.57 0.71 3 1.18 0.71 3 0.59 0.71
4 metallic 4 0.49 0.69 4 metallic
5 metallic
: \[tab:gap\] DFT-LDA computed energy gaps and potential changes versus the number of LaAlO$_3$ layers $i$ for the $np$-type, $n$-type and $p$-type interfaces ($i$=0 refers to a pure SrTiO$_3$ substrate). $\Delta$ is the energy gap of the interface systems ($np$-type and $n$-type) with the corresponding number of LaAlO$_3$ layers. $\delta$ is the ‘La-O energy difference’ ($p$-type interface) or ‘Ti-O energy difference’ (quantum well) with the corresponding number of LaAlO$_3$ layers. $V_i$ is the macroscopic potential change due to adding the $i$-th LaAlO$_3$ layer. Note that ‘La-O energy difference’ is not well defined when $i=0$ (no LaAlO$_3$ layer).
Bound versus unbound carriers at the interfaces {#DFT-bound}
-----------------------------------------------
In this section, we provide a microscopic picture of the quantum states inhabited by the carriers at the $n$-type and $p$-type interfaces, based on a first-principles extraction of tight-binding parameters. As reported in our previous work [@Chen-PRB-2009], the main result is that a significant Ti-La interfacial hopping element unique to the $n$-type interface causes electron to bind there. No such mechanism is operative at the $p$-type interface, explaining the delocalization of the holes into the SrTiO$_3$ substrate. As shown below, the Ti-La hopping is significantly enhanced at the $n$-type interface due to the relatively large size of La $d$-orbitals as well as the spatial proximity of the La and Ti atoms in the neighboring atomic planes. Here we provide an expanded and more detailed explanation and analysis of the hopping elements.
To calculate the on-site and hopping matrix elements, we employ Löwdin atomic orbitals [@Lowdin-JChem-1950] where $\langle \textbf{r}|\alpha \textbf{R}\rangle=
\phi_{\alpha}(\textbf{r}-\textbf{R})$ is a Löwdin orbital of type $\alpha$ localized around lattice position $\textbf{R}$. However, as the Kohn-Sham wave functions are Bloch states indexed by wave vectors **k**, it is more fruitful to employ Bloch-like superpositions $\varphi^{\textbf{k}}_{\alpha}(\textbf{r})$ defined as $$\label{hopping-1} \varphi^{\textbf{k}}_{\alpha}(\textbf{r})
=\frac{1}{\sqrt{N_{\textbf{k}}}}
\sum_{\textbf{R}}e^{i\textbf{k}\cdot\textbf{R}}
\phi_{\alpha}(\textbf{r}-\textbf{R})\,.$$ $N_{\textbf{k}}$ is the total number of $k$-points in the Brillouin zone sampling. To extract a tight-binding model for the Bloch states at **k**, we need to calculate the matrix elements $\langle
\varphi^{\textbf{k}}_{\alpha}|H| \varphi^{\textbf{k}}_{\beta}\rangle$. Making use of the completeness of the Hamiltionian, we obtain: $$\label{hopping-4} \langle \varphi^{\textbf{k}}_{\alpha}|H| \varphi^{\textbf{k}}_{\beta}\rangle
=\sum_{n}\langle \varphi^{\textbf{k}}_{\alpha}|n\textbf{k}\rangle E_{n\textbf{k}}
\langle n\textbf{k}| \varphi^{\textbf{k}}_{\beta}\rangle$$ where $\langle \textbf{r}|n\textbf{k}\rangle=\psi_{n\textbf{k}}(\textbf{r})$ are the actual Bloch eigenstates of the Hamiltonian $H$. The overlap $\langle\varphi^{\textbf{k}}_{\alpha}|n\textbf{k}\rangle$ can be recast easily as follows: $$\label{hopping-5}\langle\varphi^{\textbf{k}}_{\alpha}|n\textbf{k}\rangle
=\sqrt{N_{\textbf{k}}}\int
d\textbf{r}\phi^{*}_{\alpha}(\textbf{r})\psi_{n\textbf{k}}(\textbf{r})
=\langle \alpha \textbf{0}|n\textbf{k}\rangle$$ Our final operational formula is $$\label{hopping-6} \langle \varphi^{\textbf{k}}_{\alpha}|H|
\varphi^{\textbf{k}}_{\beta}\rangle =\sum_{n}\langle
\alpha\textbf{0}|n\textbf{k}\rangle E_{n\textbf{k}} \langle
n\textbf{k}| \beta \textbf{0}\rangle$$ where the overlaps $\langle \alpha \textbf{0}| n \textbf{k} \rangle$ are automatically computed and reported by the PWscf code. Choosing $\alpha=\beta$ gives the on-site energies while $\alpha\ne\beta$ are the hopping elements. The orthonormality relation can also be obtained by replacing $H$ with the identity $I$ operator:
$$\label{hopping-7} \langle \varphi^{\textbf{k}}_{\alpha}|
\varphi^{\textbf{k}}_{\beta}\rangle =\sum_{n}\langle
\alpha\textbf{0}|n\textbf{k}\rangle \langle
n\textbf{k}| \beta \textbf{0}\rangle=\delta_{\alpha\beta}$$
Eq.(\[hopping-7\]) is a good criterion to check the truncation in the infinite summation over the band index $n$. In our calculations, we include bands with energies up to 29 eV above the Fermi level, so that $|\langle \varphi^{\textbf{k}}_{\alpha}| \varphi^{\textbf{k}}_{\alpha}\rangle|
>0.99$ for all Löwdin orbitals considered and $|\langle \varphi^{\textbf{k}}_{\alpha}|\varphi^{\textbf{k}}_{\beta}\rangle|
<5\times 10^{-4}$ for all the pairs of two different atomic orbitals.
At the $n$-type interface, direct projection of the Bloch states $\psi_{n\textbf{k}}$ onto the atomic orbitals shows that the character of the lowest band accommodating the conduction electrons is mainly Ti-$d_{xy}$ with a small component of La-$d_{xy}$. The minimum of these occupied bands is at $\Gamma$ $({\bf k}={\bf 0})$. Therefore we build our tight-binding model on the subspace composed of Ti-$d_{xy}$ and La-$d_{xy}$ orbitals and calculate the following on-site and hopping matrix elements:
$$\label{hopping-8} H_{00}=\langle
\varphi^{\Gamma}_{\textrm{La-}d_{xy}}|H|
\varphi^{\Gamma}_{\textrm{La-}d_{xy}}\rangle$$
$$\label{hopping-9} H_{jj}=\langle
\varphi^{\Gamma}_{\textrm{Ti}^j\textrm{-}d_{xy}}|H|
\varphi^{\Gamma}_{\textrm{Ti}^j\textrm{-}d_{xy}}\rangle$$
$$\label{hopping-10} H_{0j}=H^{*}_{j0}=\langle
\varphi^{\Gamma}_{\textrm{La}\textrm{-}d_{xy}}|H|
\varphi^{\Gamma}_{\textrm{Ti}^j\textrm{-}d_{xy}}\rangle$$
$$\label{hopping-11} H_{ij}=H^{*}_{ji}=\langle
\varphi^{\Gamma}_{\textrm{Ti}^i\textrm{-}d_{xy}}|H|
\varphi^{\Gamma}_{\textrm{Ti}^j\textrm{-}d_{xy}}\rangle\\$$
where ‘La’ in Eq. (\[hopping-8\]) and Eq. (\[hopping-10\]) is the La atom in the LaAlO$_3$ layer at the $n$-type interface. $i$ (or $j$) labels the Ti atoms in the SrTiO$_3$ where $i=1$ (or $j=1$) is in the very TiO$_2$ layer at the $n$-type interface, and increasing $i$ (or $j$) refers to the TiO$_2$ layers that are further away from the interface.
The $n$-type interface with 2 u.c. of LaAlO$_3$ is insulating (before the polar catastrophe) and that with 4 u.c. of LaAlO$_3$ is conducting (after the polar catastrophe). Table \[tab:onsite\] shows the on-site and hopping matrix elements of the $n$-type interface before and after the polar catastrophe, respectively. The on-site matrix elements are also shown in Fig. \[fig:onsite\]. We can see that the polar catastrophe does not result in significant difference in either the on-site or the hopping matrix elements. Therefore, in the following we mainly focus on the tight-binding Hamiltonian after the polar catastrophe, which is more relevant to describe the conduction electron gas.
![\[fig:onsite\] The on-site matrix elements of the $n$-type interface. The squares are the on-site matrix elements. Site 0 is the La atom. Sites 1-6 are the Ti atoms. The triangles are the ‘new’ onsite matrix elements after taking into account the hopping effects. The data can be read from Table \[tab:onsite\]. $a$ is the lattice constant of SrTiO$_3$. [**a)**]{} The $n$-type interface with 2 unit cells of LaAlO$_3$. [**b)**]{} The $n$-type interface with 4 unit cells of LaAlO$_3$. The purple arrows illustrate the effect of Ti-La hopping as per Eq. (\[hopping-14\]) and Eq. (\[hopping-15\]).](onsite.ps){width="15cm"}
Fig. \[fig:onsite\] and Table \[tab:onsite\] show that the on-site matrix element of the first Ti atom (closest to the interface) is $not$ the lowest one before or after the polar catastrophe; instead, the second Ti atom has the lowest potential. Therefore, one would expect to find the highest electron density on the second Ti atom, in direct contrast to what is actually found in Fig. \[fig:chargespatial\]c. However, there is a very large hopping matrix element between the La and Ti atoms at the interface, at least 100 times larger than all the other hopping matrix elements between Ti atoms. This is due to two factors: one is that La-$5d$ orbitals are more extended than Ti-$3d$ orbitals, which leads to a larger overlap, and the other is that the distance between La and Ti atom at the $n$-type interface is $\sqrt{3}/2$ times the distance between neighboring Ti atoms [@Chen-PRB-2009]. Consequently, to the leading order [^5], we neglect the hopping between adjacent Ti atoms and other higher order hopping matrix elements. The Hamiltonian becomes block-diagonal and the only block we need to diagonalize is the leading $2\times2$ Ti-La submatrix right at the interface: $$\label{hopping-13} h=\left(\begin{matrix}H_{00} & H_{01} \\
H_{10} & H_{11}\end{matrix}\right)$$ Diagonalization of Eq. (\[hopping-13\]) with the tabulated values gives the following eigenvalues and eigenvectors:
$$\label{hopping-14} \varepsilon_1=12.99 \textrm{eV}, \ \
|\xi_1\rangle=0.98|\textrm{La}\textrm{-}d_{xy}\rangle-0.18
|\textrm{Ti}^1\textrm{-}d_{xy}\rangle$$
$$\label{hopping-15} \varepsilon_2=9.04\textrm{eV}, \ \
|\xi_2\rangle=0.18|\textrm{La}\textrm{-}d_{xy}\rangle+0.98
|\textrm{Ti}^1\textrm{-}d_{xy}\rangle$$
After diagonalization, we find a pair of bonding and anti-bonding states at the $n$-type interface. The energy of the binding state $\varepsilon_2$ is even lower than $H_{22}$, showing that the electrons prefer the Ti site right at the interface. Therefore, after the insulating-to-metallic transition occurs, most electrons reside in this bonding state, which leads to an electron gas bound to the interface. In other words, the large Ti-La hopping matrix element is critical in creating a strongly bound state for the electrons at the interface. Moreover, this hopping is intrinsic to the $n$-type interface geometry as it does not significantly change its magnitude before and after the polar catastrophe. This also explains the overall similarity of the electron distributions at the symmetric double $n$-type superlattice (Fig. \[fig:chargespatial\]a) and the stoichiometric $n$-type interface (Fig. \[fig:chargespatial\]c). Despite differing in the lack or presence of oppositely charged carriers and polar fields, the two systems share the same Ti-La hopping and thus the same binding force to the interface.
On the other hand, at the $p$-type interface, holes are found to occupy bands with essentially pure O-$p$ characters and the maximum of those bands is located at M ($\textbf{k}=(\frac{\pi}{a},\frac{\pi}{a},0)$). The tight-binding model is based on the following matrix elements: $$\label{hopping-16} H^{\mu\nu}_{ij}=
\langle \varphi^{\textrm{M}}_{\textrm{O}^i\textrm{-}p_{\mu}}|H|\varphi^{\textrm{M}}_{\textrm{O}^j\textrm{-}p_{\nu}}\rangle$$ where $i,j$ refer to different O atoms and $\mu,\nu=x,y,z$ refer to different O-$p$ orbitals. We calculate all these matrix elements at the $p$-type interface before and after the polar catastrophe. We find that the hopping elements between nearest neighbor O atoms are all close to 0.6 eV, that there is no order of magnitude difference in the hopping elements, and that there is no discontinuity in going across the $p$-type interface. This is no surprise because the O atoms form a continuous network across the $p$-type interface and the hopping matrix elements do not sensitively depend on the nature of the surrounding cations. Therefore, the only difference we expect from the symmetric double $p$-type results (compare Fig. \[fig:chargespatial\]b and Fig. \[fig:chargespatial\]d) is that the presence of the polar field at the $p$-type interface will drive the holes from the LaAlO$_3$ into the SrTiO$_3$. However, once in the SrTiO$_3$, the holes do not feel any strong preference for the interface and therefore diffuse further into the SrTiO$_3$ substrate. As our calculations have finite film thicknesses, we expect that the profile of holes resembles the lowest state of a free particle in a box and thus have a maximum density in the middle of the film, as borne out by Fig. \[fig:chargespatial\]d.
----- ------------ ------------ ------ ------------ ------------
$i$ $l=2$ u.c. $l=4$ u.c. $ij$ $l=2$ u.c. $l=4$ u.c.
0 12.99 12.87 0 1 -697 -696
1 9.13 9.18 1 2 -6.39 -1.3
2 9.09 9.14 2 3 -6.80 -0.2
3 9.10 9.15 3 4 -7.03 -0.3
4 9.10 9.19 4 5 -7.07 -0.3
5 9.10 9.16 5 6 -7.07 -0.4
6 9.10 9.17 6 7 -7.05 -0.4
7 9.10 9.18 7 8 -7.03 -0.5
8 9.10 9.19 8 9 -6.98 -0.5
9 9.10 9.19 0 2 -0.87 -7.2
----- ------------ ------------ ------ ------------ ------------
: \[tab:onsite\] On-site and hopping matrix elements of the $n$-type interface before and after the polar catastrophe, respectively. $l$ is the thickness of LaAlO$_3$. $l=2$ and 4 unit cells (u.c.) correspond to before and after the polar catastrophe, respectively. We list all the nearest neighbor hopping matrix elements and the largest next nearest neighbor hopping matrix element. The first and third columns are the on-site matrix elements of Ti $d$-states and hopping matrix elements before the polar catastrophe. The second and fourth columns are the on-site matrix elements of Ti $d$-states and hopping matrix elements after the polar catastrophe.
Oxygen vacancies repulsion from the interfaces {#oxygen-repulsion}
----------------------------------------------
The theoretical results we have discussed to this point have concerned ideal interfaces with sharp boundaries and $1\times1$ in-plane periodicity. An important question is whether these assumptions are reasonable for describing the experimentally realized interfaces. This question is obviously very difficult to answer in general as the number of possible ways of perturbing the ideal interfaces is enormous and includes intermixtures, impurities, vacancies, interstitials, off-stoichiometry, etc. Below, we focus on answering one simple but important question for the most prevalent type of imperfections: do oxygen vacancies present in the SrTiO$_3$ substrate have a strong preference for the SrTiO$_3$ region close to the interface? Based on both our first principles results and general physical considerations, the answer appears to be no. As we show below, the interface SrTiO$_3$ region should be relatively free of oxygen vacancies compared to the bulk SrTiO$_3$ substrate.
Our approach is motivated by experiments that infer a concentration of one oxygen vacancy per four two-dimensional unit cells at the $p$-type interface [@Hwang-NatMat-2006]. If one assumes each oxygen vacancy donates fully two electrons, this is precisely the amount required to compensate the 0.5 holes per two-dimensional unit cell from the polar catastrophe. The same experiments are also used to infer a non-zero (but smaller) concentration of oxygen vacancies at the $n$-type interface [@Hwang-NatMat-2006]. To determine the most energetically favorable position for oxygen vacancies, we compute the formation energy of one monolayer of oxygen vacancies at the same areal density (1/4 per two-dimensional unit cell) at various positions inside the SrTiO$_3$ close to the interface with LaAlO$_3$. We perform calculations on both $p$-type and $n$-type interfaces using the unit cells shown in Fig. \[fig:ovp2x2\].
Our simulation cells include a $2\times2\times5$ SrTiO$_3$ film and a $2\times2\times1$ LaAlO$_3$ film and $\simeq$ 25 Å of vacuum. The substrate termination is SrO in order to minimize any surface effects [@Chen-PRB-2009]. The SrTiO$_3$ unit cell next to the vacuum is fixed to simulate the bulk-like substrate and all the other atoms are relaxed. For a reference, we compute the formation energy of an isolated oxygen vacancy in bulk SrTiO$_3$ in a $2\times2\times5$ supercell with one oxygen vacancy. Starting from the interface, we place the oxygen vacancy in the SrO or TiO$_2$ layer of the first three SrTiO$_3$ unit cells and compute the formation energy (see Fig. \[fig:ovp2x2\] for the numbering nomenclature of the layers). We show the results in Table \[table:ovenergies\]. As the oxygen vacancies move away from the $p$-type interface, the formation energy decreases, approaching a constant value of 5.4-5.5 eV in the bulk-like regions of the SrTiO$_3$ films. The computed bulk formation energy is slightly lower (5.18 eV); presumably the small difference is due to the fact that the SrTiO$_3$ films in the interface geometry are rather thin (only 5 unit cells) and not yet fully in the substrate limit. Nevertheless the energetic *trend* clearly shows the interface region repels the vacancies.
Our results show that there is no energetically favorable binding of oxygen vacancies to either the $p$-type or $n$-type interface. A simple physical picture can explain the repulsion of the oxygen vacancies from both types of interface. An isolated, neutral oxygen vacancy in SrTiO$_3$ is a donor that binds two electrons (O has formal charge O$^{2-}$). The electrons reside on localized states composed of the $d$-orbitals of the vacancy’s surrounding Ti atoms; the energy of these states is close to the SrTiO$_3$ conduction band edge. Thus, oxygen vacancies can be described as a type of hydrogenic system with bound electrons. As the vacancy approaches the interface, the bound electrons experience what are essentially hard wall boundary conditions since the conduction band edge of LaAlO$_3$ is $\sim$ 2 eV higher than that of SrTiO$_3$. As the vacancy approaches the hard wall, its energy increases due to the electron confinement effect [@Satpathy-PRB-1983]. Given the basic physical principles behind this argument, it is clear that it applies to both interfaces. Futhermore, it suggests that even though our first principles results are for relatively high densities of oxygen vacancies in a particular configuration, lower densities and more positionally disordered oxygen vacancies in the SrTiO$_3$ will still be repelled from the interface region.
The fact that oxygen vacancies are repelled from the interfaces provides a self-consistent picture for our computations of ideal interfaces. Namely, the dominant defects (oxygen vacancies) should not be present in the immediate vicinity of the interfaces. For example, for the $p$-type interface, our theory predicts extended band states for the holes that diffuse substantially into the SrTiO$_3$ substrate. Since the vacancies are themselves in the bulk of the substrate, they can trap the holes, rendering them immobile. For a relatively thick SrTiO$_3$ substrate, this situation is likely since only one oxygen vacancy per four two-dimensional unit cells, distributed over the three-dimensional volume of the substrate, is required to provide the compensating number of electrons.
We would like to comment that experimentally Nakagawa *et al.* [@Hwang-NatMat-2006] found $32 \pm 6\%$ of oxygen vacancies per two-dimensional unit cell at the $p$-type interface. Their conclusion is based on a least-square fit of EELS spectra to the reference spectra of bulk SrTiO$_3$, bulk LaAlO$_3$ and oxygen-deficient SrTiO$_{3-\delta}$ with $\delta = 1/4$. The discrepancy between our results and those experiments may be due to the following two non-mutually exclusive reasons. One is that in our simulations, we include only one unit cell of LaAlO$_3$ owing to the computational limitations. More realistic simulations would include thicker LaAlO$_3$ films to simulate the system after the polar catastrophe when carriers appear in the vicinity of the interfaces. The presence of carriers might in principle affect the formation energy of oxygen vacancies, which needs to be checked in much larger calculations in future work. The other reason is the interpretation of EELS spectra in experiment. The experimental fitting is based on the assumption that any deviation of O-K edge EELS from the bulk SrTiO$_3$ is caused by the presence of oxygen vacancies. However, even without oxygen vacancies at the $p$-type interface, we observe in our calculations that the atoms close to the interface move away from their bulk positions due to the broken symmetry at the interface. The EELS spectra of the distorted SrTiO$_3$ could be different from that of the reference bulk structure. Therefore, a more accurate (and more difficult) fit would take into account the deviation of the EELS spectra of distorted SrTiO$_3$ from the bulk and other possible factors which could also affect the EELS spectra, such as intermixture and nonstoichiometry.
![\[fig:ovp2x2\]Schematics of the simulation cell for $\frac{1}{4}$ monolayer oxygen vacancy calculations. One oxygen vacancy is placed in either SrO and TiO$_2$ of the first, second and third layer, respectively. **a)** The $p$-type interface. **b)** The $n$-type interface. ](ov-2x2.ps){width="12cm"}
---------------------------- ---------- ---------------------------- ----------
position of oxygen vacancy $E$ (eV) position of oxygen vacancy $E$ (eV)
1st SrO layer 6.25 1st TiO$_2$ layer 5.65
1st TiO$_2$ layer 5.85 1st SrO layer 5.65
2nd SrO layer 5.69 2nd TiO$_2$ layer 5.56
2nd TiO$_2$ layer 5.56 2nd SrO layer 5.46
3rd SrO layer 5.60 3rd TiO$_2$ layer 5.57
3rd TiO$_2$ layer 5.50 3rd SrO layer 5.41
---------------------------- ---------- ---------------------------- ----------
: \[table:ovenergies\] Formation energies $E$ of $\frac{1}{4}$ monolayer of oxygen vacancy at the $p$-type and $n$-type interface. $\frac{1}{4}$ monolayer of oxygen vacancy is in either $i$th SrO atomic layer or $i$th TiO$_2$ atomic layer. For reference, the formation energy in the bulk is 5.18 eV.
Thickness dependence of sheet carrier density {#DFT-thickness}
---------------------------------------------
In experiment, the sheet carrier density was observed to display a thickness dependence, but a range of results exist. The sheet carrier density depends on both sample growth conditions and post-annealing [@Chen-Adv-2010]. In this section, we present first-principles calculations of the sheet carrier density for different LaAlO$_3$ film thicknesses. We study the $np$-type, $n$-type and QW systems, and discuss the similarities between these different interface systems.
We define the ‘sheet carrier density’ as the sum of all the conduction electrons per square unit cell in the Ti $d$-states of the SrTiO$_3$ substrate (for $n$-type and $np$-type interfaces) or in the Ti $d$-states of the single embedded TiO$_2$ layer (for the QW systems). Fig. \[fig:charge\] shows the sheet carrier density at different thicknesses for all three types of interface. The results of $n$-type interface are similar to what Son [*et al.*]{} recently obtained [@Son-PRB-2009]. Below the critical separation in each system, there are no mobile carriers since the system is insulating. Above the critical separation, the sheet carrier densities increase in all three cases, following a similar manner. In order to further reveal the similarities among these interface systems, we show in Fig. \[fig:nsvsE\] the relation of the sheet carrier density $\sigma_s$ versus the internal electric field $E$ through the LaAlO$_3$. The internal electric field is determined as follows. We calculate the macro-averaged smoothed total potential along the (001) direction. The internal electric field is the slope of the smoothed potential. Though the geometry details differ, the figure shows that the $(E, \sigma_s)$ relations are almost identical for all three interface types. This suggests that a single continuous model will be able to describe the behavior of all three interfaces. We discuss such a model in Section \[model-thickness\].
![\[fig:charge\] Comparison of the sheet carrier density computed with the continuous model (see Section \[model-thickness\]) and DFT simulations. The red, orange and green lines are the model calculations of the QW, the $n$-type and the $np$-type interfaces, respectively. The squares, triangles and circles are the DFT results of the QW, the $n$-type and the $np$-type interfaces, respectively.](charge.ps){width="10cm"}
![\[fig:nsvsE\] The sheet carrier density versus the internal electric field through the LaAlO$_3$ thin film. The red squares, orange triangles, and green circles are for the quantum well, the $n$-type, and the $np$-type interfaces, rspectively. The blue line is the result of the continuous model described in Section \[model-thickness\] .](cm.ps){width="10cm"}
We note that in our calculations of the sheet carrier density, we include all of the electrons in the conduction bands, regardless of whether they in fact contribute to the observed conductivity or not. We find that after the critical separation, the sheet carrier density increases continuously with the thickness of the LaAlO$_3$ film, asymptoting to 0.5 $e/a^2$. This behavior qualitatively agrees with some experiments [@Huijben-NatMat-2006; @Sing-PRL-2009], but conflicts with others [@Mannhart-Science-2006], in which the sheet carrier density remains almost constant (15 times smaller than 0.5 $e/a^2$) after the critical separation. It is possible that the low sheet carrier density in Ref. [@Mannhart-Science-2006] may be explained in terms of Anderson localization induced by disorder [@Satpathy-PRL-2008; @Son-PRB-2009].
External Field Effect {#DFT-external}
---------------------
We have shown that there is an insulating-to-metallic transition as a function of [LaAlO$_3$ ]{}thickness both in the SrTiO$_3$/LaAlO$_3$ interface and the quantum well (QW) systems. In this section, we examine the effects of inducing this transition at subcritical thicknesses via an applied electric field. We focus on the $n$-type interface and the QW system with 3 u.c. of LaAlO$_3$ (for both systems, the critical separation is 4 u.c. in the DFT simulations). To study the field effect, we apply a homogeneous electric field perpendicular to the interface. In order to avoid artificial effects from periodic boundary conditions (see Appendix \[appendix-PBC\]), we apply two electric fields of equal magnitude and opposite direction to each half of a symmetric simulation cell (see Fig. \[fig:fieldunitcell\]). Since the system is mirror symmetric, for simplicity we only focus on half of the simulation cell. At the $n$-type interface, the external electric field is parallel to the intrinsic electric field through the LaAlO$_3$ film, and in the QW systems, the direction is parallel to the field through the LaAlO$_3$ capping layer. Based on the polar catastrophe picture, such external electric fields will add to intrinsic electric fields and further reduce the energy gap, so that the insulating-to-metallic transition will occur at a subcritical separation.
We performed the simulations with different magnitudes of external electric field on both the $n$-type interface and QW systems. We need to elucidate a subtle point here: in the QW, the external electric field reduces the energy difference between the Ti $d$-states of the embedded TiO$_2$ layer and O $p$-states on the surface. This is the ‘Ti-O energy difference’ introduced in Section \[sec:quantumwell\]. Once this energy difference disappears, the insulating-to-metallic transition occurs. This energy difference is well defined because we can identify two Bloch states, one which characterizes the Ti $d$-states of the embedded TiO$_2$ layer and the other which characterizes the O $p$-states on the surface, and simply take their energy difference. At the $n$-type interface, the external electric field also reduces the energy difference between the Ti $d$-states of the first Ti atom at the interface and the O $p$-states on the surface. However, this energy difference is not easily identified because the external electric field also exists throughout the SrTiO$_3$ substrate, so the Bloch states that accept the transferred electrons are now composed of a mixture of all the Ti $d$-states in the SrTiO$_3$ thin film. For the purposes of this section, we simply extract the energy gap of this $n$-type interface system (between the O $p$-states on the surface and the lowest occupied Ti $d$-states) versus the external field. In Section \[model-external\], we will discuss more fully the relation of this energy gap and what would happen in an actual experiment.
In Fig. \[fig:externalfield\] we show the computed energy gap of the $n$-type interface and the Ti-O energy difference of the QW as a function of external electric field. As the figure shows, both monotonically decrease with increasing external electric field. Thus, the DFT simulations demonstrate how an external electric field can induce an insulating-to-metal transition.
Continuous model {#model}
================
In order to shed more light on the nature of the LaAlO$_3$/SrTiO$_3$ interface, we develop a simple model which approximates the LaAlO$_3$ film as a homogeneous continuous medium. This model shows that the polar and dielectric properties of LaAlO$_3$ largely determine the thickness dependence of the sheet carrier density and the external field effect behavior.
Thickness dependence of sheet carrier density {#model-thickness}
---------------------------------------------
In this section, we use model calculations, aided by DFT results, to quantitatively describe the insulating-to-metallic transition that occurs at the $n$-type and $np$-type interfaces and the quantum well (QW) systems. As detailed below, we find that the electric field dependence of $\epsilon_L$ is critical in obtaining quantitative accuracy.
![\[fig:average\] Schematic illustration of the continuous model. In the ionic limit, the LaAlO$_3$ thin film can be considered as a serial connection of capacitors. The electric field through LaAlO$_3$ is like an impulse and the resulting potential takes a stair shape. The dashed purple lines are the average electric field and the average potential.](average.ps){width="9cm"}
Before the insulating-to-metallic transition occurs, the alternating positively charged (LaO)$^+$ and negatively charged (AlO$_2$)$^-$ layers in the LaAlO$_3$ film can be idealized as a series of capacitors (see Fig. \[fig:average\]). As the figure illustrates, the thickness of each capacitor is only half of the unit cell of LaAlO$_3$ (there is zero voltage drop across the other half). Therefore the inner electric field through the LaAlO$_3$ behaves like an impulse and the resulting potential takes the shape of a stair function. If we approximate the LaAlO$_3$ thin film as a continuous homogeneous media and average this electric field, we can determine the average internal electric field:
$$\label{charge-1}\overline{E}=\frac{1}{2}\frac{4\pi
\sigma_0}{\epsilon_L}=\frac{4\pi \overline{\sigma}}{\epsilon_L},$$
where $\sigma_0=e/ a^2$ is an electron per two-dimensional unit cell, $\epsilon_L$ is the dielectric constant of LaAlO$_3$, and
$$\label{charge-2}\overline{\sigma}=\frac{\sigma_0}{2}=\frac{e}{2a^2}.$$
Once the insulating-to-metallic transition occurs, some electrons transfer across the interface to counteract the internal electric field through the LaAlO$_3$ thin film. The new average internal electric field through the LaAlO$_3$ film is then given by:
$$\label{charge-3}\overline{E}=\frac{4\pi(\overline{\sigma}-\sigma_s)}{\epsilon_L},$$
where $\sigma_s$ is the transferred electron density, or sheet carrier density. It is easy to recast Eq. (\[charge-3\]) into:
$$\label{charge-4} \sigma_s(\overline{E})=\overline{\sigma}-
\frac{\overline{E}\epsilon_L}{4\pi},$$
where $\overline{\sigma}=0.5e/a^2$. We also take into account that the dielectric constant $\epsilon_L$ depends on the internal electric field $\overline{E}$. The field dependence can be phenomenologically described by the Landau theory, and takes the following approximate form [@Devo-Phil-1949; @Antons-PRB-2005] (also see Appendix \[appendix-field\]):
$$\label{charge-5}
\epsilon_L(\overline{E}) \simeq \epsilon_0
\left(1+\left(\frac{\overline{E}}{\mathcal{E}_0}\right)^2\right)^{-1/3},$$
where $\epsilon_0$ is the dielectric constant of LaAlO$_3$ at vanishing electric field and $\mathcal{E}_0$ is a characteristic electric field. Eq. (\[charge-4\]) and Eq. (\[charge-5\]) establish a relation between the internal electric field $\overline{E}$ through the LaAlO$_3$ and the sheet carrier density $\sigma_s$, both of which can be calculated independently from DFT simulations. We also performed a separate slab calculation (see Appendix \[appendix-field-lao\] for details) and determined the parameters $\epsilon_0$ and $\mathcal{E}_0$ in Eq. (\[charge-5\]) as: $$\label{charge-6}\epsilon_0=40.95, \phantom{5} \mathcal{E}_0=0.15
\textrm{V/\AA}$$ Fig. \[fig:nsvsE\] compares $(E, \sigma_s)$ computed via DFT and the model. We can see from Fig. \[fig:nsvsE\] that $(E, \sigma_s)$ almost lies on the same curve in all three cases, which does not depend on the details of the structure. The continuous model shows a good agreement with the DFT results, and as the internal electric field decreases (i.e. the thickness of the LaAlO$_3$ increases), the continuous model becomes almost exact. This is expected because in the limit of an infinitely thick LaAlO$_3$ film, the interface and surface regions become neglegible. Our results show that even with only 7 u.c. of LaAlO$_3$ ([*i.e.*]{} the internal electric field $<$ 0.12 V/Å), the continuous model already works very well.
However, it is not easy to directly measure the internal electric field through the LaAlO$_3$ experimentally because the LaAlO$_3$ film is only a few unit cells thick. It is more useful to relate the sheet carrier density to the nominal thickness of the LaAlO$_3$ film ([*i.e.*]{} the number of LaAlO$_3$ unit cells grown on top of the SrTiO$_3$ substrate). Therefore, we need to find how the internal electric field through the LaAlO$_3$ depends on the LaAlO$_3$ thickness.
When the polar catastrophe takes place at the $n$-type interface, electrons leave the surface O-$p$ states and are transfered into the Ti-$d$ states. The charge transfer is halted once a common Fermi level is reached. Approximately, this means that the Ti-$d$ conduction band edge of the SrTiO$_3$ substrate is at the same energy as the O-$p$ surface valence band edge. Looking back at Fig. \[fig:interface\], we can see that there is a remanent field through the LaAlO$_3$ film and thus a potential difference. The existence of the electric field is also visible in Fig. \[fig:nsvsE\] from our calculations. The potential difference across the LaAlO$_3$ thin film, denoted by $K$, can be related to the remanent internal electric field by: $$\label{charge-7}e\overline{E}d= K$$ where $d$ is the thickness of LaAlO$_3$. In principle $K$ could depend on the thickness of LaAlO$_3$. But the detailed calculations show (see Fig. \[fig:K\]) that $K$ quickly saturates to a constant as the thickness $d$ gets larger. This could be understood because $K$ is essentially determined by the energies of the relevant electronic states and the band offsets, which are interface properties and have little dependence on the thickness of LaAlO$_3$. We therefore approximate $K$ as a constant. We illustrate later that this approximation does not significantly change the physics in the continuous model.
Inserting Eq. (\[charge-7\]) into Eq. (\[charge-4\]) and Eq. (\[charge-5\]), we eliminate $\overline{E}$ and obtain:
$$\label{charge-8}\sigma_s=\overline{\sigma}-\frac{K}{4\pi
ed}\epsilon_0\left(1+\left(\frac{K}{e\mathcal{E}_0d}\right)^2\right)^{-1/3}$$
Eq. (\[charge-8\]) highlights some qualitative features of the sheet carrier density. The critical separation is the smallest $d$ which makes $\sigma_s > 0$. Below the critical separation, the sheet carrier density is zero. Above the critical separation, the sheet carrier density gradually increases and then saturates to $\overline{\sigma}=0.5e/a^2$ in the limit of large $d$.
The parameter $K$ in Eq. (\[charge-8\]) is determined as follows. We simulate $n$-type, $np$-type interfaces and QW with different thicknesses and calculate the macro-averaged smoothed electric field through the LaAlO$_3$ at each thickness. In Fig. \[fig:K\], we show the potential difference across the LaAlO$_3$ thin film, defined by $e\overline{E}d$. We can see that in all three cases, $e\overline{E}d$ quickly saturates as $d$ gets larger. Since we do not have a microscopic model to describe how the potential difference $e\overline{E}d$ changes with the thickness $d$, we use the last data (corresponding to the largest $d$) to determine the values of $K$, which are highlighted by the dashed line in Fig. \[fig:K\]. In each case, we find that
$$\label{charge-10} K_{\rm QW}=3.16 \textrm{ eV}, \phantom{5} K_{n}=2.64
\textrm{ eV}, \phantom{5} K_{np}=2.40 \textrm{ eV}.$$
The $K$ in all three cases turn out to be different from each other and larger than the DFT-computed band gap of SrTiO$_3$ (1.85 eV). This is partly because the band offsets of these three cases are not the same. However, the small valence band offsets alone [@Albina-PRB-2007; @Pickett-PRB-2008] can not explain the big difference between the $K$ value and SrTiO$_3$ band gap. A more important factor is that we treat $\overline{E}$ as a uniform electric field throughout the LaAlO$_3$ thin film. However, close to the interface and surface, the internal electric field should be different from that in the middle of the LaAlO$_3$ film but difficult to average in those regions. Therefore we use an approximate relation ($K=e\overline{E}d$) to determine the potential difference, but as shown later, this approximation is good enough to reproduce the DFT-calculated thickness dependence of sheet carrier density.
Equipped with $K$, $\epsilon_0$ and $\mathcal{E}_0$, we can calculate the sheet carrier density at different thicknesses using Eq. (\[charge-8\]). We compare the sheet carrier densities calculated from the continuous model with those from the DFT simulations. The results are shown in Fig. \[fig:charge\]. The agreement is good in all three cases even though Eq. (\[charge-7\]) is not exact.
We would like to comment that recently Son *et al.* [@Son-PRB-2009] also calculated the thickness dependence of sheet carrier density and they made a similar model by assuming that the dielectric constant of LaAlO$_3$ has no field dependence and is simply constant. Rather than Eq.(\[charge-8\]), they had a simpler formula
$$\label{charge-11}\sigma_s=\overline{\sigma}\left(1-\frac{d_c}{d}\right)$$
where $\overline{\sigma}$ and $d_c$ are two constants determined by fitting. They found $\overline{\sigma}=0.455 e/a^2$, close to the ideal value $0.5 e/a^2$. Our analysis is different in that we fix $\overline{\sigma}=0.5 e/a^2$, which is necessitated by the polar catastrophe mechanism, but take into account that the dielectric constant of LaAlO$_3$ has a strong field dependence, as confirmed by the DFT calculations.
External electric field {#model-external}
-----------------------
In order to get a more quantitative description of the external field effect, we apply the continuous model established in the previous section. In principle, at the subcritical separation (3 u.c.), the continuous model should break down because the LaAlO$_3$ film is so thin that the interface and surface regions cannot be neglected and the LaAlO$_3$ is no longer uniform. However, we simplify the situation by assuming that LaAlO$_3$ is still a homogeneous media with the dielectric constant given by Eq. (\[charge-5\]) with the parameters Eq. (\[charge-6\]), but has an effective thickness $d_{eff}$. In order to simplify our notations, we use $E$ instead of $\overline{E}$ to denote the averaged macroscropic fields.
Based on the above argument, for the quantum well systems we have:
$$\label{field-1} \Delta=\Delta_0-(E^{L}-E^{L}_0)d^{L}_{eff},$$
where ‘$L$’ stands for LaAlO$_3$. $\Delta$ is the Ti-O energy difference at a given $E_{ext}$ and $E^{L}$ is the internal electric field. $\Delta_0$ and $E^{L}_0$ indicate the values of $\Delta$ and $E^{L}$, respectively, at vanishing external electric field:
$$\label{field-2} E^L=\frac{4\pi
\overline{\sigma}+E_{ext}}{\epsilon_L(E^L)}$$
and
$$\label{field-3} E^L_0=\frac{4\pi
\overline{\sigma}}{\epsilon_L(E^L_0)}$$
Equations (\[field-1\], \[field-2\], and \[field-3\]) establish the relation between $\Delta$ and $E_{ext}$. The effective thickness cannot be given $a$ $priori$ from the model. Instead, we fit the data and find $d^{L}_{eff}=5.9 \textrm{\AA} = 1.6$ u.c. A comparison of $(E_{ext}, \Delta)$ for the quantum well systems using DFT calculations and the model calculations is shown in Fig. \[fig:externalfield\].
For the $n$-type interface, we also need to account for the potential drop in the SrTiO$_3$ due to the external electric field. Therefore we have:
$$\label{field-4}
\Delta=\Delta_0-(E^{L}-E^{L}_0)d^{L}_{eff}-E^{S}d^{S}_{eff}$$
where ‘$S$’ stands for SrTiO$_3$. $\Delta$ is the energy gap of the system and $\Delta_0$ is the value of $\Delta$ at vanishing external electric field. Note that in the absence of an external electric field, the SrTiO$_3$ film is unpolarized in the continuous model. $E^{S}$ and $E_{ext}$ are related by:
$$\label{field-5} E^S=\frac{E_{ext}}{\epsilon_S(E^S)}$$
We use the Berry phase method [@Souza-PRL-2002] to calculate $\epsilon_S (E)$ and fit the data using Eq. (\[charge-5\]) (see Appendix \[appendix-field-sto\] for details). The results of the fit are:
$$\label{field-6}
\epsilon^{S}_0=309.6, \mathcal{E}^{S}_0=49.2 \textrm{V/}\mu\textrm{m}
=4.92 \times 10^{-3} \textrm{V/\AA}$$
We find that the values of $d^{L}_{eff}$ in different systems are different. Hence, in principle we have two fitting parameters in Eq. (\[field-4\]): $d^{L}_{eff}$ and $d^S_{eff}$. However, it turns out that Eq. (\[field-4\]) very insensitively depends on $d^{S}_{eff}$. This is expected because the dielectric constant of SrTiO$_3$ is at least 10 times larger than that of LaAlO$_3$ and the nominal thicknesses of LaAlO$_3$ and SrTiO$_3$ are both 3 u.c. in the DFT calculations (see Fig. \[fig:fieldunitcell\]), and thus the potential drop across the SrTiO$_3$ is much smaller than that across the LaAlO$_3$ film. Therefore we just use the real thickness of SrTiO$_3$ $d^{S}_{eff}=11.1$ Å (from the relaxed structure) and fit the data with $d^L_{eff}$. We find: $d^{L}_{eff}=7.5$ Å$\simeq 2$ u.c. Fig. \[fig:externalfield\] compares the values of $(E_{ext}, \Delta)$ of the $n$-type interface computed by DFT and the model, respectively.
Fig. \[fig:externalfield\] shows that the critical electric fields $E^c_{ext}$ extracted from the DFT calculations for the $n$-type interface and the QW are 0.79 and 1.15 V/Å, respectively. In the continuous model, the critical voltage across the whole sample is given by:
$$\label{external-field}V_{c}=
\frac{E^{c}_{ext}+4\pi\overline{\sigma}}{\epsilon_{\textrm{LAO}}}
d_{\textrm{LAO}}+\frac{E^{c}_{ext}}{\epsilon_{\textrm{STO}}}d_{\textrm{STO}}$$
where $d_{\textrm{LAO}}$ and $d_{\textrm{STO}}$ are the thickness of LaAlO$_3$ and SrTiO$_3$, respectively. $\overline{\sigma}=0.5 e/a^2$ and $E^{c}_{ext}$ is the critical external electric field. In order to estimate the critical voltage for the experimental setup, we need to use the thickness of the SrTiO$_3$ substrate in experiment.
Here we need to clarify some subtleties: in the presence of applied external field, DFT simulations do not realistically represent the spatial distribution of conduction electrons. Fig. \[fig:external-field\] illustrates the reason for the discrepancy. In the DFT simulations, the conduction electrons fill the lowest energy states. Since the “external” field bends the SrTiO$_3$ conduction bands, these states are located at the bottom surface of the SrTiO$_3$ substrate, as shown schematically in Fig. \[fig:external-field\]a. Although in principle the experimental situation is the same, realistically, electrons can only tunnel a few unit cells from the surface; thus they get trapped at the $n$-type interface, as illustrated in Fig. \[fig:external-field\]b. The trapping mechanism is the Ti-La interfacial hopping, which lowers the energy of the Ti atom closest to the interface and creates the energy barrier through which the electrons must tunnel. Hence, the real critical field has to be larger than the computed one, so that the energy of surface states equals that of the interface states (Fig. \[fig:external-field\]b). The critical external field $E_{ext}^c$ we obtain from DFT simulations is therefore only a lower bound.
Now we make a simple estimation of the lower bound: at low temperatures ($< 5$ K), the dielectric constant of SrTiO$_3$ can be as large as $2.5\times 10^4$ [@Saifi-PRB-1970] and the thickness of the SrTiO$_3$ substrate is typically $\simeq 1$ mm. Considering that $\epsilon_{\textrm{LAO}} \simeq 30$ and $d_{\textrm{LAO}}\simeq 10$ Å, the second term in Eq.(\[external-field\]) is dominant. Therefore, for the $n$-type interface, we have:
$$\label{external-field-ntype}
V^{n\textrm{-type}}_{c}\simeq\frac{E^{c}_{ext}}{\epsilon_{\textrm{STO}}}
d_{\textrm{STO}}\simeq 320~\textrm{V}$$
and for the QW,
$$\label{external-field-QW}
V^{\textrm{QW}}_{c}
\simeq\frac{E^{c}_{ext}}{\epsilon_{\textrm{STO}}}d_{\textrm{STO}}\simeq 480
~\textrm{V}$$
As these estimates are based on data from DFT simulations in which the band gap of SrTiO$_3$ is underestimated, they provide only lower bounds on the critical voltage. Experimentally, the critical voltage of the $n$-type interface with a 1 mm thick SrTiO$_3$ substrate has been measured to be $\simeq$ 70 V at low temperature [@Mannhart-Science-2006], which is much smaller than our estimates. The discrepancy may imply that the simple model of an ideally sharp interface is not sufficient to model the external field doping, due to additional details that play an important role at the $n$-type interface. Various types of atomic reconstructions, such as cation intermixtures at the interface [@Hwang-NatMat-2006] or oxygen vacancies on the surface [@Cen-NatMat-2008] may possibly trigger an insulating-to-metallic transition, leading to a smaller critical voltage.
![\[fig:external-field\] **a)** The spatial distribution of conduction electrons in the DFT simulations. The conduction electrons occupy the lowest energy states available, which are located at the bottom surface. **b)** The spatial distribution of conduction electrons in the actual experiment. The conduction electrons get trapped at the $n$-type interface by the tunneling barrier due to the self-consistent potential well formed at the interface that largely stems from the Ti-La hopping. The length of arrows illustrates the magnitude of external field.](external-field.ps){width="14cm"}
Conclusion
==========
We present a detailed study of ideal LaAlO$_3$/SrTiO$_3$ interfaces and a new class of quantum wells. In both systems, we confirm an intrinsic insulating-to-metallic transition. The observed transition can be triggered either by thickening the LaAlO$_3$ layers or by imposing an external electric field, both of which can be explained by the polar catastrophe mechanism. We show that for both the $n$-type interface and QW, the realistic critical separation (taking into account the underestimation of DFT band gap) is around 6 unit cells. We also show that, given a typical [SrTiO$_3$ ]{}substrate thickness of 1mm, the lower bound for the critical voltage necessary to induce an insulating-to-metallic transition is estimated to be $\sim 300$ V for the ideal $n$-type interface and $\sim 500$ V for the QW. In addition to theoretically demonstrating the observed physical properties, we provide a microscropic explanation of the observed binding of conduction electrons at the $n$-type interface, a phenomenon which can not be described by the polar catastrophe mechanism alone. We demonstrate that the large La-Ti hopping matrix element at the $n$-type interface, which is absent in both bulk constituents and is unique to the $n$-type interface, lowers the energy of the Ti atom at the interface relative to all other Ti sites, thus binding the electrons. Futhermore, we develop a continuous model that captures the essence of polar structure and dielectric properties and predicts the thickness dependence of sheet carrier densities.
We conclude by discussing some of the outstanding issues related to this system. Our understanding of the atomically sharp interfaces and QW suggests that the electronic properties of LaAlO$_3$/SrTiO$_3$ interfaces observed in experiments are unlikely to be solely a result of electronic reconstructions based on the polar catastrophe mechanism. For example, the theoretical critical separation, predicted by the polar catastrophe picture, is 6 u.c. [@Chen-PRB-2009; @Demkov-PRB-2008; @Pickett-PRL-2009; @Son-PRB-2009], which is larger than the experimental value (4 u.c.) [@Mannhart-Science-2006]. In addition, while the calculated thickness dependence of the sheet carrier density agrees qualitatively with some experiments [@Huijben-NatMat-2006; @Sing-PRL-2009] in that the sheet carrier density increases with LaAlO$_3$ thickness, it disagrees with other experiments [@Mannhart-Science-2006] which show an almost constant sheet carrier density above the critical separation. The sheet carrier density we calculate here includes all the conduction electrons. Whether some of them may be prone to the Anderson localization [@Satpathy-PRL-2008] is an issue beyond the scope of the DFT calculations. However, the possibility of multiple type of carriers [@Seo-2009] is an interesting question that warrants further study in both theory and experiment. Another remaining puzzle is that the polar catastrophe predicts the coexistence of electrons at the interface and holes one the surface, while experiments find that the surface region is insulating and only electron-like carriers are detected in the transport measurement [@Mannhart-Science-2006]. Recently, x-ray photoemssion measurement found that the polar field through LaAlO$_3$ is much smaller than the theoretical predictions [@Segal-PRB-2009]. It is possible that various atomic reconstructions, such as cation intermixting, nonstoichiometry and defects could also significantly affect the electronic structures of these interface systems and be partially responsible for the discrepancies mentioned above. While a detailed discussion of atomic reconstructions is beyond our present work, it serves as an interesting direction for future research.
Conduction electron and hole densities {#appendix-density}
======================================
In this appendix, we explain how we calculate the *conduction* electron and hole densities for the symmetric superlattices and stoichiometric interface systems. The basic tool we use is the local density of states (LDOS) for the system
$$\label{localdos}D(\textbf{r},E)=\sum_{n\textbf{k}}
|\psi_{n\textbf{k}}(\textbf{r})|^2\delta(E-\epsilon_{n\textbf{k}})$$
which we integrate over the appropriate energy range.
Double $n$-type and $p$-type superlattices
------------------------------------------
The symmetric nonstoichiometric double $n$-type ($p$-type) superlattice is shown in Fig. \[fig:symif\]a (Fig. \[fig:symif\]b). The corresponding electron (hole) denisty is presented in Fig. \[fig:chargespatial\]a (Fig. \[fig:chargespatial\]b). For the symmetric superlattices, there is no polar field and analysis of the density of states shows that there is an easily identifiable energy gap at all locations in the film; the Fermi level is either above the conduction band edge for the the $n$-type superlattice or below the conduction band for the $p$-type superlattice. For the transferred electron density (Fig. \[fig:chargespatial\]a), we integrate the LDOS from the middle of the band gap to the Fermi level,
$$\label{doublee}\eta(\textbf{r})=\int^{E_F}_{\cal{E}}D(\textbf{r},E)dE$$
where $\cal{E}$ is the energy in the middle of the band gap. We are in fact counting all the electrons in the conduction bands. For the transferred hole density (Fig. \[fig:chargespatial\]b), similarly we integrate the LDOS from the Fermi level to the middle of the band gap,
$$\label{doublee}\xi(\textbf{r})=\int^{\cal{E}}_{E_F}D(\textbf{r},E)dE$$
*i.e.* counting all the holes in the valence bands. The integrated LDOS $\eta(\textbf{r})$ and $\xi(\textbf{r})$ are then averaged over the $xy$ plane and finally plotted along the $z$ direction.
Stoichiometric $n$-type and $p$-type interface systems
------------------------------------------------------
The stoichiometric $n$-type ($p$-type) interface system is shown in Fig. \[fig:interface\]b (Fig. \[fig:interface\]c). The corresponding *conduction* electron (hole) density is shown in Fig. \[fig:chargespatial\]c (Fig. \[fig:chargespatial\]d). For the stoichiometric $n$-type and $p$-type interfaces, the polar field ensures that the bands edges in the LaAlO$_3$ are not flat and that the Fermi level will intersect the band edges of the SrTiO$_3$ as well as the surface of the LaAlO$_3$ film (see Fig. \[fig:interface\]). However, the local density of states in the SrTiO$_3$ film still clearly shows a band gap. Therefore, we can also compute the transferred charges at the interface using the above formula. The only complication is that in addition to showing the transferred charges in the interface region, $\eta(\textbf{r})$ and $\xi(\textbf{r})$ will necessarily have contributions localized at the surface region of LaAlO$_3$ which are not of direct interest when studying the interface alone; the choice of axis range in Fig. \[fig:chargespatial\]c and \[fig:chargespatial\]d effectively excludes this contribution. As a consistency check, we can also compute the transferred charge by computing the atomic projections of all Bloch states in the system and identifying all those states with the proper atomic character: Ti-$d$ or La-$d$ character and partial occupancy for the $n$-type interface, and O-$p$ character and partial occupancy for the $p$-type interface. One can then manually sum up these particular contributions to get the electron and hole distributions, and the results agree with the previous method in the relevant regions.
Periodic boundary condition effects {#appendix-PBC}
===================================
When we perform a slab calculation, the materials in the simulation are generally polarized either due to intrinsic polar fields ([ *e.g.*]{} LaAlO$_3$) or external electric fields. In DFT calculations, the periodic boundary conditions (PBC) imposed on the simulation cell influence the screening properties of the materials. We will show in the following two sections that the presence of PBCs does not significantly affect the polarization due to intrinsic polar fields in the system, but it does induce large artificial effects when one studies the response of the system to an external field, requiring a careful set up of the simulation and interpretation of the results.
Polarization due to intrinsic fields {#appendix-PBC-intrinsic}
------------------------------------
In this section, we illustrate that in the SrTiO$_3$/LaAlO$_3$ interface system, PBCs induce an artificial field through SrTiO$_3$, but we show that this effect is neglegible.
As shown in Fig. \[fig:PBCeffect\]a, the two boundary conditions of displacement $\mathbf{D}$ at each interface give:
$$\label{app-1}
D_{\rm LAO}-D_{\rm STO}=4\pi\sigma$$
$$\label{app-2}
D_{\rm STO}-D_{\rm V}=0$$
PBCs require that the total potential drop over the whole simulation cell is equal to zero:
$$\label{app-3}
E_{\rm STO}s+E_{\rm LAO}l+E_{\rm V}v=0$$
where $s$, $l$ and $v$ are the thicknesses of SrTiO$_3$, LaAlO$_3$ and vacuum, respectively. Considering that $E_i=D_i/\epsilon_i$ ($i$ is SrTiO$_3$, LaAlO$_3$ or vacuum), we can solve $D_{\rm STO}$ and $D_{\rm LAO}$ explicitly,
$$\label{app-4}
D_{\rm STO}=-4\pi\sigma\frac{l/\epsilon_{\rm LAO}}
{s/\epsilon_{\rm STO}+l/\epsilon_{\rm LAO}+v/\epsilon_{\rm V}}$$
$$\label{app-5}
D_{\rm LAO}=4\pi\sigma\frac{s/\epsilon_{\rm STO}+v}
{s/\epsilon_{\rm STO}+l/\epsilon_{\rm LAO}+v/\epsilon_{\rm V}}$$
The dielectric constants of SrTiO$_3$, LaAlO$_3$ and vacuum are roughly:
$$\label{app-6}
\epsilon_{\rm STO}\simeq 300 \gg \epsilon_{\rm LAO} \simeq 25 \gg
\epsilon_{\rm V}=1$$
and the thicknesses of SrTiO$_3$, LaAlO$_3$ and vacuum in the simulation cell are on the same order of magnitude (within a factor of two), therefore we have the following simplifications:
$$\label{app-7}
E_{\rm LAO}\simeq \frac{4\pi\sigma}{\epsilon_{\rm LAO}}$$
$$\label{app-8}
E_{\rm STO}\simeq -\frac{4\pi\sigma}
{\epsilon_{\rm LAO}\epsilon_{\rm STO}}\frac{l}{v}=-\frac{l}
{v\epsilon_{\rm STO}}E_{\rm LAO}\to |E_{\rm STO}|\ll |E_{\rm LAO}|$$
$$\label{app-9}
E_{\rm V}\simeq -\frac{4\pi\sigma}
{\epsilon_{\rm LAO}}\frac{l}{v}=-\frac{l}
{v}E_{\rm LAO}\to E_{\rm V}v \simeq -E_{\rm LAO}l$$
Eq. (\[app-8\]) and Eq. (\[app-9\]) show that though SrTiO$_3$ is artificially polarized due to the PBCs, the electric field through SrTiO$_3$ is much smaller than the intrinsic field through LaAlO$_3$ and the voltage built across the LaAlO$_3$ film mostly drops in the vacuum. Based on the above approximate dielectric constants, the field in SrTiO$_3$ is only $\sim 0.3\%$ of that in LaAlO$_3$. In addition, we can further reduce the field in SrTiO$_3$ by increasing the thickness of the vacuum $v$; the above relations give a quantitative measure of the (small) error for any finite value of $v$.
Polarization due to an applied external electric field {#appendix-PBC-extrinsic}
------------------------------------------------------
We illustrate in this section that when we use a slab calculation and a sawtooth potential to simulate an external electric field through nonpolar materials ([*e.g.*]{} SrTiO$_3$), PBCs will artificially undermine the screening and give rise to a significantly smaller dielectric constant.
Fig. \[fig:PBCeffect\]b shows a schematic of how an external electric field $E_{ext}$ along the $z$ direction is screened in a nonpolar material. There are two induced electric fields. The one in the material is the stardard depolarization field $E_{dep}$ and the other field, which we call the outside field $E_{out}$, is in the vacuum with the opposite direction. $E_{ext}$, $E_{dep}$ and $E_{out}$ are the $magnitudes$ of each field; their directions are explicitly shown in Fig. \[fig:PBCeffect\]b. We denote the size of the unit cell along the $z$ direction by $L$ and the thickness of the slab by $d$. The sawtooth potential automatically satisfies the periodic boundary condition (the reversed part is not shown in Fig. \[fig:PBCeffect\]b). Therefore, the induced electric fields required to satisfy the periodic boundary condition are given by:
$$\label{app-10}
E_{dep}d=E_{out}(L-d)$$
On the other hand, Gauss’s law gives:
$$\label{app-11}
4\pi \Sigma = E_{dep}+E_{out}$$
where the surface charge density $\Sigma$ is related to the polarization in the material by
$$\label{app-12}
\Sigma = \textbf{n}\cdot\textbf{P}=P$$
and the polarization $P$ is related to the total electric field $E_{tot}$ by
$$\label{app-13}
P=\chi E_{tot}$$
where $\chi$ is the permitivity. From Fig. \[fig:PBCeffect\]b, it is easy to see that
$$\label{app-14}
E_{tot}=E_{ext}-E_{dep}$$
Combining Eq. (\[app-10\]-\[app-14\]) gives
$$\label{app-15}
E_{tot}=\frac{E_{ext}}{1+4\pi\chi\left(1-\frac{d}{L}\right)}$$
Now we can identify the dielectric constant $\epsilon$
$$\label{app-16}
\epsilon=\frac{E_{ext}}{E_{tot}}=1+4\pi\chi\left(1-\frac{d}{L}\right)$$
Eq. (\[app-16\]) is different from the familiar formula $\epsilon=1+4\pi\chi$ in that we have an extra factor $(1-d/L)$. In typical slab calculations $(1-d/L) \sim 30\%$ and for high-$k$ materials ([*e.g.*]{} SrTiO$_3$) $\chi\gg 1$ so that $\epsilon$ is dominated by the second term in Eq. (\[app-16\]) and thus the reduction factor is a significant error.
A remedy
--------
The origin of the deviation of the dielectric constant from the correct value of $1+4\pi\chi$ is that the electric field induced by the bound charge $\pm \Sigma$ does not completely serve as the depolarization field, but is instead split into two parts, $E_{dep}$ and $E_{out}$. The presence of $E_{out}$ is purely due to the imposed periodic boundary condition, and is unphysical. However, in practical slab calculations it is too computationally expensive to make $d/L$ small.
In order to get rid of $E_{out}$, we need a simulation cell in which the depolarization field automatically satisfies periodic boundary conditions. Therefore, we use a mirror-symmetric simulation cell in which the external electric field is also mirror-symmetrically distributed. By symmetry, $E_{dep}$ automatically satisfies periodic boundary conditions and $E_{out}$ is guaranteed to be zero in the vacuum. Thus, the external electric field is correctly screened, but with a relatively modest increase in computational expense due to doubling the simulation cell along one direction.
Field dependence of dielectric constant {#appendix-field}
=======================================
The Landau theory phenomenologically describes the field dependence of the dielectric constant [@Devo-Phil-1949; @Antons-PRB-2005]. It is assumed that the free energy of the system $F(P,T)$ can be expanded in even powers of the polarization $P$:
$$\label{app-17}
F(P,T)=F_0+AP^2+BP^4+CP^6+...$$
where the coefficients $A,B,C...$ may depend on the temperature $T$. Keeping only terms in $F$ to the fourth order, we obtain:
$$\label{app-18}
E=\frac{\partial F}{\partial P}=2A P + 4BP^3$$
and the permitivity follows:
$$\label{app-19}
\frac{1}{\chi}=\frac{\partial E}{\partial P}=2A+12BP^2$$
Eq. (\[app-18\]) and Eq. (\[app-19\]) uniquely determine $\chi=\chi(P)$ and $P=P(E)$. The analytical solution to Eq. (\[app-18\]) $P=P(E)$ is complicated, but we can find a useful interpolation scheme [@Devo-Phil-1949; @Antons-PRB-2005] by noting that as $P$ is small, $\chi \to (2A)^{-1}$ and as $P$ is large, $\chi \propto P^{-2}$ and $E \propto P^3$, that is $\chi \propto E^{-2/3}$, so
$$\label{app-20}
\chi\simeq\chi_0\left(1+\left(\frac{E}{\mathcal{E}_0}\right)^2\right)^{-1/3}$$
For high-$k$ materials, we can also approximate $\epsilon \simeq 4\pi \chi \gg 1$, giving the final expression:
$$\label{app-21}
\epsilon\simeq\epsilon_0\left(1+\left(\frac{E}{\mathcal{E}_0}\right)^2\right)^{-1/3}$$
where $\epsilon_0$ and $\mathcal{E}_0$ are fitting parameters.
We note that the truncation of Eq. (\[app-17\]) is based on the assumption that $P$ is small (*i.e.* $E$ is small). However, the Landau theory itself does not give a characteristic polarization or electric field. Instead, we consider $\mathcal{E}_0$ as a characteristic electric field whose value is determined by the fitting. Thus, data points much larger than $\mathcal{E}_0$ should not be used in the fitting because Eq. (\[app-17\]) would break down. This is a self-consistent check.
LaAlO$_3$ {#appendix-field-lao}
---------
Before we calculate the dielectric constant $\epsilon_L$ of LaAlO$_3$, we need to elucidate a subtle point. The $\epsilon_L$ we calculate here is only well defined in the thin slab of LaAlO$_3$, not in the bulk. This is because we are in the region of very strong electric field ($\sim 0.7$ V/$a_{\rm LAO}$). With a few unit cells, the potential difference built across the slab will be larger than the band gap of LaAlO$_3$, and Zener tunneling will occur. Bulk LaAlO$_3$ can not accommodate such a large electric field and the Berry phase method [@Souza-PRL-2002] for calculating the dielectric constant breaks down in this regime. Instead, we resort to a mirror-symmetric slab calculation (see Fig. \[fig:LAO-slab\] and Appendix \[appendix-PBC\]). We turn on a mirror-symmetric external electric field $E_{ext}$ and calculate the macro-averaged electric field $E_{tot}$ through the material. The dielectric constant is then defined as:
$$\label{app-22}
\epsilon_L(E_{tot})=\frac{E_{ext}}{E_{tot}}$$
The raw data and the fitting curve using Eq. (\[app-21\]) are shown in Fig. \[fig:dielectric\]. The fitting results are:
$$\label{app-lao}\epsilon_0=40.95, \phantom{5} \mathcal{E}_0=0.15
\textrm{V/\AA}$$
SrTiO$_3$ {#appendix-field-sto}
---------
Since SrTiO$_3$ is a nonpolar material with a large dielectric constant, the typical internal electric field through SrTiO$_3$ is much smaller than that through LaAlO$_3$. The dielectric constant we are interested in can be defined in the bulk (based on the argument of metastable states) and calculated using the Berry phase method. In order to accurately determine the atom positions, we use $6\times6\times20$ $k$-point sampling and lower the force convergence threshold to 8 meV/ Å. We directly calculate the total polarization $P_{tot}$ (both ionic and electronic) in the unit cell at a given total electric field $E_{tot}$ and the dielectric constant follows straightforwardly:
$$\label{app-23}
\epsilon_S(E_{tot})=4\pi\frac{P_{tot}}{\Omega E_{tot}}+1$$
where $\Omega$ is the volume of unit cell. The raw data and fitting curve are shown in Fig. \[fig:dielectric\], with the fitting results:
$$\label{app-sto}\epsilon^{S}_0=309.6, \mathcal{E}^{S}_0=49.2
\textrm{V/}\mu\textrm{m}=4.92 \times 10^{-3} \textrm{V/\AA}$$
[^1]: See http://www.quantum-espresso.org
[^2]: Some atoms are fixed in order to simulate the bulk-like substrate. This is specified in each section.
[^3]: See the EPACS Document No. E-PRBMDO-79-R12912 for a detailed discussion of surface effects of TiO$_2$ termination.
[^4]: Detailed calculations show that the surface states have character of La $5d$, $6s$ and $6p$.
[^5]: We in fact perform the exact diagonalization and find that neglecting the hopping between adjacent Ti atoms and other higher order hopping matrix elements is an excellent approximation. More importantly, this leading order approximation described in the text shows that the Ti-La hopping is crucial in trapping the electrons at the $n$-type interface.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is proven that each indecomposable injective module over a valuation domain $R$ is polyserial if and only if each maximal immediate extension $\widehat{R}$ of $R$ is of finite rank over the completion $\widetilde{R}$ of $R$ in the $R$-topology. In this case, for each indecomposable injective module $E$, the following invariants are finite and equal: its Malcev rank, its Fleischer rank and its dual Goldie dimension. Similar results are obtained for chain rings satisfying some additional properties. It is also shown that each indecomposable injective module over one Krull-dimensional local Noetherian rings has finite Malcev rank. The preservation of Goldie dimension finiteness by localization is investigated too.'
address: 'Laboratoire de Mathématiques Nicolas Oresme, CNRS UMR 6139, Département de mathématiques et mécanique, Université de Caen Basse-Normandie 14032 Caen cedex, France'
author:
- François Couchot
title: Indecomposable injective modules of finite Malcev rank over local commutative rings
---
Introduction and preliminaries {#introduction-and-preliminaries .unnumbered}
==============================
In this paper all rings are associative and commutative with unity and all modules are unital. First we give some definitions.
An $R$-module $M$ is said to be **uniserial** if its set of submodules is totally ordered by inclusion and $R$ is a **chain ring**[^1] if it is uniserial as $R$-module. A chain domain is a valuation domain. In the sequel, if $R$ is a chain ring, we denote by $P$ its maximal ideal, $N$ its nilradical, $Z$ its set of zero-divisors ($Z$ is a prime ideal) and we put $Q=R_Z$. Recall that a chain ring $R$ is said to be [**Archimedean**]{} if $P$ is the sole non-zero prime ideal.
A module $M$ is said to be **finitely cogenerated** if its injective hull is a finite direct sum of injective hulls of simple modules. The **f.c. topology** on a module $M$ is the linear topology defined by taking as a basis of neighbourhoods of zero all submodules $G$ for which $M/G$ is finitely cogenerated (see [@Vam75]). This topology is always Hausdorff. We denote by $\widetilde{M}$ the completion of $M$ in its f.c. topology. When $R$ is a chain ring which is not a finitely cogenerated $R$-module, the f.c. topology on $R$ coincides with the $R$-topology which is defined by taking as a basis of neighbourhoods of zero all non-zero principal ideals. A chain ring $R$ is said to be **(almost) maximal** if $R/A$ is complete in its f.c. topology for any (non-zero) proper ideal $A$.
In 1959, Matlis proved that a valuation domain $R$ is almost maximal if and only if $Q/R$ is injective, and in this case, for each proper ideal $A$ of $R$, $E(R/A)\cong Q/A$, see [@Mat59 Theorem 4]. Since $Q$ is clearly uniserial and $Q/R\cong Q/rR$ for each non-zero element $r\in P$, we can also say that $R$ is almost maximal if and only if $E(R/rR)$ is uniserial, if and only if each indecomposable injective module is uniserial. This result was extended to any chain ring in 1971 by Gill, see [@Gil71 Theorem]: a chain ring $R$ is almost maximal if and only if $E(R/P)$ is uniserial, if and only if each indecomposable injective module is uniserial. By using [@Couch03 Proposition 14], if $R$ is a chain ring, it is easy to check that $E(R/rR)$ is uniserial if and only if so is $E(R/P)$. Let us observe that any indecomposable injective module is uniserial if and only if each finitely generated uniform module is cyclic.
If $M$ is a finitely generated module we denote by $\mathrm{gen}\ M$ its minimal number of generators. If $M$ is a module over a valuation domain $R$ the [**Fleischer rank**]{} of $M$, denoted by $\mathrm{Fr}\ M$, is defined to be the minimum rank of torsion-free modules having $M$ as an epimorphic image.
In the book “Modules over valuation domains” by Fuchs and Salce [@FuSa85 Proposition IX.3.1](1985), it is proven that $\mathrm{gen}\ M\leq\mathrm{Fr}\ E(R/P)$, for each finitely generated uniform module $M$ over a valuation domain $R$. However, it remains to give a characterization of valuation domains $R$ for which $\mathrm{Fr}\ E(R/P)$ is finite.
In 2005 [@Cou05 Proposition 2], if $R$ is an Archimedean chain ring, the author proved that there exists an integer $p>0$ such that $\mathrm{gen}\ M\leq p$ for each finitely generated uniform module $M$ if and only if $R$ is almost maximal, (i.e $p=1$).
An exact sequence $0 \rightarrow F \rightarrow E \rightarrow G \rightarrow 0$ is **pure** if it remains exact when tensoring it with any $R$-module. In this case we say that $F$ is a **pure** submodule of $E$. We say that a module $M$ is **polyserial** if it has a pure-composition series $$0=M_0\subset M_1\subset\dots\subset M_n=M,$$ i.e. $M_k$ is a pure submodule of $M$ and $M_k/M_{k-1}$ is a uniserial module for each $k=1,\dots, n$.
The **Malcev rank** of a module $M$ is defined as the cardinal number $$\mathrm{Mr}\ M=\mathrm{sup}\{\mathrm{gen}\ X\mid X\ \mathrm{finitely\ generated\ submodule\ of}\ M\}.$$ For each module $M$ over a valuation domain we have $\mathrm{Mr}\ M\leq\mathrm{Fr}\ M$.
An $R$-module $F$ is **pure-injective** if for every pure exact sequence $$0\rightarrow A\rightarrow B\rightarrow C\rightarrow 0$$ of $R$-modules, the following sequence $$0\rightarrow\mathrm{Hom}_R(C,F)
\rightarrow\mathrm{Hom}_R(B,F)\rightarrow\mathrm{Hom}_R(A,F)\rightarrow
0$$ is exact. An $R$-module $B$ is a **pure-essential extension** of a submodule $A$ if $A$ is a pure submodule of $B$ and, if for each submodule $K$ of $B$, either $K\cap A\ne 0$ or $(A+K)/K$ is not a pure submodule of $B/K$. We say that $B$ is a **pure-injective hull** of $A$ if $B$ is pure-injective and a pure-essential extension of $A$. By [@War69] or [@FuSa01 chapter XIII] each $R$-module $M$ has a pure-injective hull and any two pure-injective hulls of $M$ are isomorphic. In the sequel, for each $R$-module $M$, $\widehat{M}$ is its pure-injective hull.
In this paper we give a characterization of two classes of chain rings. The first is the class of chain rings $R$ for which each indecomposable injective module is polyserial (Theorem \[T:main\]). These rings are exactly the chain rings $R$ which satisfies the following two conditions:
1. $\mathrm{Mr}_{\widetilde{R}}\widehat{R}<\infty$;
2. each indecomposable injective module contains a pure uniserial submodule[^2].
The first condition is equivalent to the the following: there is a non-zero prime ideal $L$ such that $R_L$ is almost maximal and the valuation domain $R/L$ has a maximal immediate extension of finite rank which is equal to $\mathrm{Mr}_{\widetilde{R}}\widehat{R}$. These rings are almost maximal by stages, i.e. there exists a finite descending chain of prime ideals $(L_i)_{0\leq i\leq n}$, with $L_0=P$ such that $(R/L_{i+1})_{L_i}$ is almost maximal for $i=0,\dots,n-1$ and $R_{L_n}$ is maximal. Moreover, for each finitely generated uniform module $M$, $\mathrm{gen}\ M\leq\mathrm{Mr}_{\widetilde{R}}\widehat{R}$, and for each indecomposable injective module $E$, $\mathrm{Mr}_RE\leq\mathrm{Mr}_{\widetilde{R}}\widehat{R}$, the equalities hold for some $M$ and $E$. If $R$ is not a domain then $\mathrm{Mr}\ \widehat{R}<\infty$. A description of such chain rings $R$ is given, and this description is similar to the one of valuation domains with a maximal immediate extension of finite rank ([@Cou10 Theorem 10 and Proposition 11]).
The second class is the one of chain rings $R$ for which each localization of any $R$-module of finite Goldie dimension has finite Goldie dimension too. These rings are exactly the chain rings $R$ for which $R/L$ has a maximal immediate extension of finite rank for each non-zero prime ideal $L$. So, the first class is contained but strictly in the second one, and some examples are given.
It is also shown that the completion $\widetilde{R}$ of any chain ring $R$ in its f.c. topology is Gaussian, and $\widetilde{R}$ is a chain ring if and only if $R$ is either complete or a domain.
For each one Krull-dimensional local Noetherian ring $R$ it is proven that there exists a positive integer $m$ such that $\mathrm{Mr}\ E\leq m$ for every indecomposable injective $R$-module $E$. Moreover, for each integer $m>1$ we give an example of a one Krull-dimensional local Noetherian domain $D$ satisfying $\mathrm{gen}\ M\leq m$ for each finitely generated uniform $D$-module $M$. However, if $R$ is a chain ring with a such upper bound $m$ then $m$ is a prime power.
\[D:sharp\] Let $M$ be a non-zero module over a ring $R$. We set: $$M_{\sharp}=\{s\in R\mid \exists 0\ne x\in M\ \mathrm{such\ that}\
sx=0\}\quad\mathrm{and}\quad M^{\sharp}=\{s\in R\mid sM\subset M\}.$$ Then $R\setminus M_{\sharp}$ and $R\setminus M^{\sharp}$ are multiplicative subsets of $R$.
If $M$ is a module over a chain ring $R$ then $M_{\sharp}$ and $M^{\sharp}$ are prime ideals and they are called the [**bottom**]{} and the [**top prime ideal**]{}, respectively, associated with $M$.
We say that an $R$-module $E$ is **FP-injective** if $\mathrm{Ext}_R^1(F,E)=0,$ for every finitely presented $R$-module $F.$ A ring $R$ is called **self FP-injective** if it is FP-injective as $R$-module. Recall that a module $E$ is FP-injective if and only if it is a pure submodule of every overmodule.
If $L$ is a prime ideal of a chain ring $R$, as in [@FaZa86], we define the **total defect** at $L$, $d_R(L)$, the **completion defect** at $L$, $c_R(L)$, as the rank of the torsion-free $R/L$-module $\widehat{R/L}$ and the rank of the torsion-free $R/L$-module $\widetilde{R/L}$, respectively.
Relations between $\widehat{R}$ and $\widetilde{R}$ {#S:compl}
===================================================
Given a ring $R$, an $R$-module $M$ and $x\in M$, the **content ideal** $\mathrm{c}(x)$ of $x$ in $M$, is the intersection of all ideals $A$ for which $x\in AM$.
When $R$ is a chain ring, the **breadth ideal** $\mathrm{B}(x)$ of an element $x$ in $\widehat{R}$ is defined by $\mathrm{B}(x)=\mathrm{c}(x+R)$. So, $\mathrm{B}(x)=0$ if $x\in R$. Since $\widehat{R}=R+P\widehat{R}$ by [@Couc06 Proposition 1] then $\mathrm{B}(x)=\{r\in R\mid x\notin R+r\widehat{R}\}$ if $x\in\widehat{R}\setminus R$.
The following lemma will be often used in the sequel.
\[L:breadth\][@Couc06 Proposition 20 and Lemma 21] Let $R$ be a chain ring. Then:
1. $R/A$ is not complete in its f.c. topology if and only if there exists $x\in\widehat{R}\setminus R$ such that $A=\mathrm{B}(x)$;
2. if $x\in\widehat{R}$ and $x=r+ay$ for some $r,a\in R$ and $y\in\widehat{R}$, then $\mathrm{B}(y)=(\mathrm{B}(x):a)$.
\[P:compl\] Let $R$ be a chain ring. Then:
1. $\widehat{R}$ has a structure of $\widetilde{R}$-module which extends its structure of $R$-module;
2. $\widetilde{R}$ is isomorphic to the submodule of $\widehat{R}$ whose elements $x$ satisfy $\mathrm{B}(x)=0$;
3. for each non-zero prime ideal $L$ of $R$ there exists a prime ideal $L'$ of $\widetilde{R}$ such that $\widetilde{R}/L'\cong R/L$ and $L'\widehat{R}=L\widehat{R}$;
4. $\widetilde{R}/R$ is a $Q/Z$-vector space.
$(1)$. If $R$ is finitely cogenerated then $\widetilde{R}=R$. If not we have $\cap_{r\in R\setminus \{0\}}rP=0$. Let $a\in\widetilde{R}$ and $x\in\widehat{R}$. Let $(a_r+rP)_{r\in R\setminus \{0\}}$ be the family of cosets of $R$ which defines $a$. If $r\in sR$ then $(a_r-a_s)\in sP$, and it follows that $(a_rx-a_sx)\in sP\widehat{R}$. By [@Couc06 Proposition 4] the family $\mathcal{F}=(a_rx+rP\widehat{R})_{r\in R\setminus \{0\}}$ has a non-empty intersection. By [@Couc06 Lemma 19] $\cap_{r\in R\setminus \{0\}}rP\widehat{R}=0$, whence the intersection of the family $\mathcal{F}$ contains a unique element that we define to be $ax$. Now it is easy to complete the proof.
$(2)$. We do as in $(1)$ by taking $x=1$. So, for each $a\in \widetilde{R}$ corresponds a unique element $y\in\widehat{R}$ such that $\mathrm{B}(y)=0$. It is easy to check that we get a monomorphism from $\widetilde{R}$ into $\widehat{R}$.
$(3)$. We may assume that $R$ is not finitely cogenerated. Since each non-zero ideal is open in the f.c. topology of $R$, we have $\widetilde{R}\cong\varprojlim_{A\in\mathcal{I}}R/A$, where $\mathcal{I}$ is the set of non-zero ideals of $R$. So, there exists a surjection $\phi:\widetilde{R}\rightarrow R/L$. We put $L'=\ker\ \phi$. Let $0\ne a\in L'$ and $0\ne x\in\widehat{R}$, and let $(a_r+rR)_{r\in R\setminus \{0\}}$ be the family of cosets of $R$ which defines $a$. There exists $r\in L$ such that $a_r\in L\setminus rR$. We set $a'=a_r$. Let $s\in rR$. Since $(a_s-a')\in rR\subset a'R$ then $a_s=a'u_s$ for some $u_s\in R$. If $t\in sR$ then $a'(u_s-u_t)\in sR$, whence $(u_s-u_t)\in (sR:a')$. The family $(u_sx+(sR:a')\widehat{R})_{s\in Ra'\setminus \{0\}}$ has a non-empty intersection. Let $y$ an element of this intersection. Since $(a'y-a_sx)\in s\widehat{R}$ for each $s\in a'R\setminus \{0\}$, it follows that $(a'y-ax)\in\cap_{s\in a'R\setminus \{0\}}s\widehat{R}=0$. Hence $ax=a'y$.
$(4)$. Let $x\in\widetilde{R}$ and $s\in R\setminus Z$. Suppose that $sx=0$. By [@Couc06 Proposition 1] $x=r+rpy$ for some $r\in R$, $p\in P$ and $y\in\widehat{R}$. Since $R$ is a pure submodule of $\widehat{R}$, there exists $t\in R$ such that $sr(1+pt)=0$. We successively deduce that $sr=0$, $r=0$ and $x=0$. Now, suppose that $sx\in R$. Then there exists $t\in R$ such that $sx=st$, whence $x=t$. So, the multiplication by $s$ in $\widetilde{R}/R$ is injective. Since $\mathrm{B}(x)=0$, $x=a+sy$ for some $a\in R$ and $y\in\widehat{R}$. But $\mathrm{B}(y)=(0:s)=0$, so $y\in\widetilde{R}$. We conclude that the multiplication by $s$ in $\widetilde{R}/R$ is bijective. Now let $a\in Z$. Then $(0:a)$ contains a non-zero element $b$. From $\mathrm{B}(x)=0$ we deduce that $x=r+bz$ for some $r\in R$ and $z\in\widehat{R}$. It follows that $ax\in R$. The proof is complete.
A local ring $R$ is called **Gaussian**[^3] if, for any ideal $A$ generated by two elements $a,b,$ in $R$, the following two properties hold:
1. $A^2$ is generated by $a^2$ or $b^2$;
2. if $A^2$ is generated by $a^2$ and $ab=0$, then $b^2=0$.
\[T:compl\] Let $R$ be a chain ring. The following assertions hold:
1. $\widetilde{R}$ is a local Gaussian ring;
2. the following conditions are equivalent:
1. $\widetilde{R}$ is a chain ring;
2. $R$ is either complete or a domain;
3. $\widetilde{R}$ is a pure $R$-submodule of $\widehat{R}$;
4. $\widetilde{R}$ is a flat $R$-module.
$(1)$. $\widetilde{R}$ is local because it is the inverse limit of a system of local rings with local connecting homomorphisms.
We may assume that $R$ is not finitely cogenerated. Let $a$ and $b$ be two elements of $\widetilde{R}$. By [@Couc06 Proposition 1] there exist $a',b'\in R$ and $x,y\in 1+P\widehat{R}$ such that $a=a'x$ and $b=b'y$. We may assume that $b'=ra'$ for some $r\in R$. First suppose that $a'\notin Z$. By Lemma \[L:breadth\](2) $\mathrm{B}(x)=(0:a')=0$, whence $x\in\widetilde{R}$, and $\mathrm{B}(y)=(0:ra')=(0:r)$. It follows that $x$ is a unit of $\widetilde{R}$, $\mathrm{B}(ry)=0$, so $ry\in\widetilde{R}$ and $b=x^{-1}(ry)a$. If $ab=0$, it follows that $(ry)a^2=0$, whence $b^2=0$. Now, assume that $a'\in Z$. Let $0\ne t\in(0:a')\cap b'P$. Then $a=c+tz$ and $b=d+tw$ for some $c,d\in R$ and $z,w\in\widehat{R}$. So, we have $a=a'(1+px')=c+a't'z$ where $p,t'\in P$ and $x'\in\widehat{R}$. From $R$ pure submodule of $\widehat{R}$ we deduce that $c=ua'$ for some unit $u\in R$. In the same way $d=vb'$ for some unit $v\in R$. It follows that $a^2=u^2a'^2$, $ab=uvra'^2=u^{-1}vra^2$ and $b^2=v^2r^2a'^2=u^{-2}r^2v^2a^2$. If $ab=0$, then $ra'^2=0$, whence $b^2=0$.
$(2)$. It is well known that $(b)$ implies the other three conditions.
Assume that $R$ is neither complete nor a domain. Let $0\ne a\in\widetilde{R}$ and $0\ne r\in Z$. Since $\mathrm{B}(a)=0$ then $a=s+rx$ for some $s\in R$ and $x\in\widehat{R}$. It follows that $rx\in\widetilde{R}$. If $rx=ry$ then $\mathrm{B}(y)=(0:r)\ne 0$ by Lemma \[L:breadth\](2). So, $rx\notin r\widetilde{R}$ and $\widetilde{R}$ is not a pure submodule of $\widehat{R}$. Now suppose that $r\in\widetilde{R}rx$, whence $r=brx$ for some $b\in\widetilde{R}$. We have $\mathrm{B}(bx)=(0:r)$, so $bx\notin R+(0:r)\widehat{R}$. From $r(bx-1)=0$ and the flatness of $\widehat{R}$ we deduce that $bx-1=sy$ for some $s\in(0:r)$ and $y\in\widehat{R}$. We get a contradiction. So, $\widetilde{R}$ is not a chain ring. Hence, $(a)\Rightarrow (b)$ and $(c)\Rightarrow (b)$.
$(d)\Rightarrow (b)$. Since $R$ is a pure submodule of $\widetilde{R}$, $\widetilde{R}/R$ is flat. By Proposition \[P:compl\] it is a semisimple $Q$-module. It follows that either $\widetilde{R}/R=0$ or $Q$ is a field. We conclude that the condition $(b)$ holds.
\[P:complMr\] Let $R$ be a chain ring. Assume $(0:a)=Z$ for some $a\in R$. Then $\mathrm{Mr}\ \widetilde{R}=c_R(Z)$.
Let $R'=R/Z$. We shall prove that $\widetilde{R'}/R'$ and $\widetilde{R}/R$ are isomorphic. Since $\widetilde{R'}\subseteq\widehat{R'}\cong\widehat{R}/Z\widehat{R}$, $\widetilde{R'}/R'$ is isomorphic to the submodule of $\widehat{R}/(R+Z\widehat{R})$ whose elements $x+(R+Z\widehat{R})$ satisfy $\mathrm{B}(x)=Z$ (it is easy to check that $\mathrm{B}(x')=\mathrm{B}(x)$ if $x'\in x+(R+Z\widehat{R})$ and $x\ne R+Z\widehat{R}$). On the other hand, $\widetilde{R}\subseteq R+a\widehat{R}$, whence $\widetilde{R}/R$ is isomorphic to a submodule of $(R+a\widehat{R})/R$. For each $x\in\widehat{R}$ we put $\phi(x+(R+Z\widehat{R}))=ax+R$. It is easy to check that $\phi$ is a well defined epimorphism from $\widehat{R}/(R+Z\widehat{R})$ into $(R+a\widehat{R})/R$. If $ax\in R$, then $ax=ad$ for some $d\in R$ because $R$ is a pure submodule of $\widehat{R}$. From $a(x-d)=0$ and the flatness of $\widehat{R}$ we deduce that $(x-d)\in Z\widehat{R}$. So, $\phi$ is an isomorphism. By Lemma \[L:breadth\](2) $\mathrm{B}(x)=Z$ if and only if $\mathrm{B}(ax)=0$. Consequently, the restrition of $\phi$ to $\widetilde{R'}/R'$ is an isomorphism onto $\widetilde{R}/R$.
Polyserial injective modules {#S:poly}
============================
The following proposition is a slight generalization of [@Cou05 Proposition 2].
\[P:Arch\] Let $R$ be an Archimedean chain ring. Assume that there exists a non-zero injective module $E$ such that $E_{\sharp}=P$ and $\mathrm{Mr}\ E<\infty$. Then $R$ is almost maximal.
It is an immediate consequence of [@Cou05 Proposition 2] and its proof. The existence of an injective module $E$ with $\mathrm{Mr}\ E<\infty$ ($\nu(E)<\infty$) is used to show that $R$ is almost maximal.
We say that a module $M$ is **singly projective** if, for any cyclic submodule $G$, the inclusion map $G\rightarrow M$ factors through a free module $F$. The following theorem generalizes [@Cou10 Theorem 10 and Proposition 11]
\[T:InjHull\] Let $R$ be a chain ring. The following conditions are equivalent:
1. $\widehat{R}$ is a polyserial module;
2. $\mathrm{Mr}\ \widehat{R}<\infty$.
In this case there exists a finite family of prime ideals $$P=L_0\supset L_1\supset\dots\supset L_{m-1}\supset L_m\supseteq N$$ such that $(R/L_{k+1})_{L_k}$ is almost maximal, $\forall k,\ 0\leq k\leq m-1$, and $R_{L_m}$ is maximal.
Moreover,
- $\widehat{R}$ has a pure-composition series $\mathcal{S}$ $$0=F_0\subset R=F_1\subset\dots\subset F_m\subset F_{m+1}=\widehat{R}$$ where $F_{j+1}/F_j$ is a free $R_{L_j}$-module of finite rank, $\forall j,\ 0\leq j\leq m$;
- $\mathrm{Mr}\ \widehat{R}=d_R(L_m)=\prod_{k=1}^{k=m}c_R(L_k)$;
- $\widetilde{R}$ is polyserial and $\mathrm{Mr}\ \widetilde{R}=c_R(Z)$.
By [@FuSa01 Lemma XII.1.4] $(1)\Rightarrow (2)$.
$(2)\Rightarrow (1)$. When $R$ is a valuation domain each torsion-free module of finite rank is polyserial. So, we may assume that $R$ is not a domain.
First we will show that $R_N$ is maximal. Since each non-unit of $R_N$ is a zero-divisor, $R_N$ is self FP-injective. From [@Couc06 Propositions 1] it is easy to deduce that $\widehat{R}$ is singly projective. By [@Coucho07 Proposition 6] so that is $(\widehat{R})_N$ over $R_N$. By [@Coucho07 Propositions 3] $(\widehat{R})_N$ is FP-injective and by [@Couc06 Proposition 5] it is pure-injective, whence it is an injective module. It is easy to check that $\mathrm{Mr}\ (\widehat{R})_N<\infty$. By proposition \[P:Arch\] and [@Gil71 Proposition 1] we conclude that $R_N$ is maximal.
Now we shall build the pure composition series $\mathcal{S}$. By [@Couc06 Theorem 2.4.(2)] $\widehat{R/N}=\widehat{R}/N\widehat{R}$. Hence $\mathrm{rank}\ \widehat{R/N}=\mathrm{Mr}\ \widehat{R/N}<\infty$. We apply [@Cou10 Theorem 10 and Proposition 11] to $R/N$. There exists a finite family of prime ideals $$P=L_0\supset L_1\supset\dots\supset L_{m-1}\supset L_m\supseteq N$$ such that $(R/L_{k+1})_{L_k}$ is almost maximal, $\forall k,\ 0\leq k\leq m-1$, and $(R/N)_{L_m}$ is maximal. Moreover, $\widehat{R/N}$ has a pure-composition series $\mathcal{S'}$ $$0=F'_0\subset R/N=F'_1\subset\dots\subset F'_m\subset F'_{m+1}=\widehat{R/N}$$ where $F'_{j+1}/F'_j$ is a free $(R/N)_{L_j}$-module of finite rank, $\forall j,\ 0\leq j\leq m$. We proceed by induction on $j$. Obviously $F_1=R$. Suppose that $F_j$ is built and that $F_j/NF_j\cong F'_j$. If $M=\widehat{R}/F_j$ and $M'=\widehat{R/N}/F'_j$ then $M\cong M'$. So, $M^{\sharp}=L_j$, whence $M$ is a module over $R_{L_j}$. Moreover, $M/L_jM$ and $M'/L_jM'$ have the same rank $p_j$ over $(R/L_j)_{L_j}$ which is equal to the rank of $F'_{j+1}/F'_j$ over $(R/N)_{L_j}$. By [@Coucho07 Proposition 21] $M$ contains a free $R_{L_j}$-module $G$ of rank $p_j$ which is a pure submodule of $M$. Moreover, $G/NG\cong F'_{j+1}/F'_j$ and $L_j(M/G)=M/G$. Let $F_{j+1}$ be the inverse image of $G$ by the natural map $\widehat{R}\rightarrow M$. Hence $F_{j+1}/NF_{j+1}\cong F'_{j+1}$. Now, let $H=\widehat{R}/F_{m+1}$. Thus $H$ is flat and $NH=H$. By [@Coucho07 Proposition 19] $H$ is a module over $R_N$. It is obvious that $(F_{m+1})_N$ is a free module of finite rank over $R_N$. By [@Coucho07 Proposition 24] $(F_{m+1})_N$ is a pure-essential submodule of $(\widehat{R})_N$. But, since $R_N$ is maximal $(F_{m+1})_N$ is pure-injective. We deduce that $(F_{m+1})_N=(\widehat{R})_N$. So, $H=0$ and $F_{m+1}=\widehat{R}$. The maximality of $(R/N)_{L_m}$ and $R_N$ implies that $R_{L_m}$ is maximal if $L_m\ne N$ (see [@Couc06 Theorem 22]).
(b). We apply the last assertion of [@Cou10 Theorem 10] to $R/N$.
(c). We have $\mathrm{Mr}\ \widetilde{R}\leq\mathrm{Mr}\ \widehat{R}<\infty$. So, by Proposition \[P:compl\] $\widetilde{R}/R$ is a finite direct sum of modules isomorphic to $Q/Z$, whence $\widetilde{R}$ is polyserial. If $A$ is a non-zero proper ideal it is easy to check that $A^{\sharp}=\{s\in R\mid A\subset (A:s)\}$. So, if we take this definition of top prime ideal for each proper ideal of $R$ we have $0^{\sharp}=Z$. Then $Z$ is an element of the descending chain of prime ideals $(L_j)_{0\leq j\leq m}$ if $R$ is not complete in its f.c. topology. We shall show that there exists $a\in Z$ such that $Z=(0:a)$. By way of contradiction suppose that $Z$ is faithful. First assume that $N\subset Z$. If $t\in Z\setminus N$, then $0\ne (0:t)\subseteq N$. It follows that $N\subset Rt\subseteq (0:s)$ for some $0\ne s\in(0:t)$. By Lemma \[L:breadth\](1) there exists $x\in\widehat{R}\setminus R$ such that $\mathrm{B}(x)=0$ if $R$ is not complete. Then $x=r+sy$ for some $r\in R$ and $y\in\widehat{R}$ and $\mathrm{B}(y)=(0:s)$. Consequently, by using again Lemma \[L:breadth\](1) we deduce that $R/(0:s)$ is not complete. If we consider the valuation domain $R/N$, from the proof of [@Cou10 Proposition 11] we deduce that $(0:s)/N=bZ/N$ for some $b\in R$. It follows that $(0:s)=bZ$, whence $Z=(0:bs)$. Now, suppose that $Z=N$. If $Z$ is faithful, since $R_N$ is maximal, as in the proof of [@Cou10 Proposition 4] we prove that $R$ is complete. Hence, if $R$ is not complete, $Z$ is not faithful. We conclude by Proposition \[P:complMr\].
For each module $M$ we denote by $\mathcal{A}(M)$ its set of annihilator ideals, i.e. an ideal $A$ belongs to $\mathcal{A}(M)$ if there exists $0\ne x\in M$ such that $A=(0:x)$. If $E$ is an indecomposable injective module over a chain ring $R$, then, for any $A,\ B\in\mathcal{A}(E),\ A\subset B$ there exists $r\in R$ such that $A=rB$ and $B=(A:r)$.
Recall that a module $M$ has **Goldie dimension** $n$ (or $\mathrm{Gd}\ M=n$) if its injective hull is a direct sum of $n$ indecomposable injective modules.
\[T:main\] Let $R$ be a chain ring. Consider the following conditions:
1. there exists an indecomposable injective module $E$ such that $E_{\sharp}=P$ and $\mathrm{Mr}\ E<\infty$;
2. there exists a prime ideal $L$ such that $d_R(L)<\infty$ and $R_L$ is almost maximal;
3. $\mathrm{Mr}\ \widehat{R}<\infty$ if $R$ is not a domain, and $\widehat{R}$ is the extension of a reduced torsion-free module of finite rank with a divisible torsion-free module when $R$ is a domain;
4. the Malcev rank of $\widehat{R}$ over $\widetilde{R}$ is finite;
5. $\mathrm{Mr}\ E<\infty$ for each indecomposable injective module $E$;
6. there exists a positive integer $n$ such that $\mathrm{gen}\ M\leq n$ for each finitely generated uniform $R$-module $M$;
7. there exists a positive integer $n$ such that $\mathrm{gen}\ M\leq n\times\mathrm{Gd}\ M$ for each finitely generated $R$-module $M$;
8. each indecomposable injective module is polyserial;
9. there exists an indecomposable injective module $E$ such that $E_{\sharp}=P$ which is polyserial.
Then:
- the first seven conditions are equivalent and they are implied by the last two conditions. Moreover, if each indecomposable injective module contains a pure uniserial submodule then the nine conditions are equivalent.
- for each indecomposable injective module $E$, either $\mathrm{Mr}\ E=1$ if $J\subseteq L$ or $\mathrm{Mr}\ E=d_{R_J}(L_J)=d_R(L)/d_R(J)=\mathrm{Mr}_{\widetilde{R_J}}\widehat{R_J}$ if $L\subset J$, where $J=E_{\sharp}$ and $L$ is a prime ideal for which $R_L$ is almost maximal. Moreover, $\mathrm{Mr}\ E$ is the maximum of $\mathrm{gen}_R\ M$ where $M$ runs over all finitely generated $R$-submodules of uniform $R_J$-modules.
It is obvious that $(8)\Rightarrow (9)$, $(6)\Rightarrow (5)$, $(7)\Rightarrow (6)$ and $(5)\Rightarrow (1)$, and $(9)\Rightarrow (1)$ by [@Couc06 Proposition 13].
$(1)\Rightarrow (2)$. By [@Couch03 Corollary 28] $E$ is faithful or it is annihilated by a simple ideal if $P=Z$. So, for each non-zero prime ideal $J$ there exists $A\in\mathcal{A}(E)$ such that $A\subset J$. By [@Couch03 Lemma 26] $A^{\sharp}=E_{\sharp}=P$ and by [@Couc06 Proposition 1] $\widehat{R}/A\widehat{R}$ is an essential extension of $R/A$, whence it is isomorphic to a submodule of $E$. We deduce that $d_R(J)=\mathrm{Mr}\ \widehat{R}/J\widehat{R}\leq\mathrm{Mr}\ \widehat{R}/A\widehat{R}\leq\mathrm{Mr}\ E<\infty$. Let $p$ be the maximum of $d_R(J)$ where $J$ runs over all non-zero prime ideals of $R$ and let $L$ be the maximal prime ideal for which $d_R(L)=p$. By Theorem \[T:InjHull\] $(R/I)_L$ is maximal for each $I\in\mathcal{A}(E)$. We deduce that $R_L$ is almost maximal. Let us observe that $d_R(L)\leq \mathrm{Mr}\ E$.
$(2)\Rightarrow (3)$ and $(4)$. We do as in the proof of Theorem \[T:InjHull\]: from a pure composition series of $R/L$ we deduce a pure submodule $F$ of $\widehat{R}$ with $\mathrm{Mr}\ F=d_R(L)$ and if $H=\widehat{R}/F$, then $LH=H$. If $R$ is not a domain then, as in the proof of Theorem \[T:InjHull\], we show that $H=0$, whence $\widehat{R}$ is polyserial. It is easy to check that $\mathrm{Mr}_{\widetilde{R}}\widehat{R}\leq\mathrm{Mr}_R\widehat{R}$. If $R$ is a domain then, for each $a\in L$, $a\ne 0$, $(R/aR)_L$ is maximal. In the same way we get that $H=aH$, whence $H$ is divisible. Since $\widetilde{R}/L\widetilde{R}=R/L$ then $d_{\widetilde{R}}(L\widetilde{R})=d_R(L)$ and $(\widetilde{R})_L$ is maximal. So, $\widehat{R}$ is the extension of a torsion-free $\widetilde{R}$-module $F$ of rank $d_R(L)$ with a divisible torsion-free $\widetilde{R}$-module $H$. Since $(\widetilde{R})_L$ is maximal, $H=0$. So, $\mathrm{Mr}_{\widetilde{R}}\ \widehat{R} = d_R(L)$.
$(4)\Rightarrow (2)$. Let $J$ be a non-zero prime ideal of $R$. By Proposition \[P:compl\] there exists a prime ideal $J'$ of $\widetilde{R}$ such that $\widetilde{R}/J'=R/J$ and $J'\widehat{R}=J\widehat{R}$. So, $\mathrm{Mr}_R\ \widehat{R/J}\leq \mathrm{Mr}_{\widetilde{R}}\ \widehat{R}$. Now, we do as in $(1)\Rightarrow (2)$ to complete the proof.
$(3)\Rightarrow (6)$. Let $M$ be a finitely generated uniform module and $E$ its injective hull. Then $E$ is indecomposable. Let $J=E_{\sharp}$. Then $E$ is a module over $R_J$. If $J\subseteq L$ then $E$ is uniserial, so $\mathrm{Mr}\ E=1$. We may assume that $L\subset J$ and it is easy to check that $R_J$ also satisfies $(3)$. There exists $A\in\mathcal{A}(E)$ such that $M\subseteq (0:_EA)$. Let $E'=(0:_EA)$. By [@Couch03 Lemma 26] $A^{\sharp}=J$, so, by [@Couc06 Theorem 6] $E'\cong\widehat{R_J}/A\widehat{R_J}$. If $R$ is a domain, let $F$ be a pure reduced torsion-free $R_J$-submodule of finite rank of $\widehat{R_J}$ such that $\widehat{R_J}/F$ is divisible. Then $(\widehat{R_J}/F)\otimes_{R_J}(R/A)_J=0$. So, in this case $E'\cong (F/AF)_J$ and $\mathrm{Mr}\ E'\leq \mathrm{rank}\ F$. If $R$ is not a domain, then $\mathrm{Mr}\ E'\leq\mathrm{Mr}\ \widehat{R_J}$. We deduce that $\mathrm{gen}\ M\leq \mathrm{rank}\ F$ or $\mathrm{gen}\ M\leq \mathrm{Mr}\ \widehat{R_J}$.
$(6)\Rightarrow (7)$. Let $M$ be a finitely generated module and $E$ its injective hull. By [@FuSa85 Corollary IX.2.2] $\mathrm{Gd}\ M\leq\mathrm{gen}\ R$. So, $E=\oplus_{1\leq j\leq p}E_j$ where $E_j$ is indecomposable for $j=1,\dots,p$. Let $\pi_j: E\rightarrow E_j$ the natural projection and $M_j=\pi_j(M)$. Then $M$ is isomorphic to a submodule of $\oplus_{1\leq j\leq p}M_j$. By $(6)$ $\mathrm{gen}\ (\oplus_{1\leq j\leq p}M_j)\leq n\times p$. Since each finitely generated ideal is principal, we conclude that $\mathrm{gen}\ M\leq n\times p$ by [@WiWi75 Lemma 1.3].
$(3)\Rightarrow (8)$. Let $E$ be an indecomposable injective module. We assume that $E$ contains a pure uniserial submodule $U$. If $J=E_{\sharp}$, then $E\cong\widehat{R_J}\otimes_{R}U$ by [@Couc06 corollary 11.(4)]. If $R$ is not a domain, then $\widehat{R_J}$ is polyserial by Theorem \[T:InjHull\]. If $R$ is a domain, we may assume that $J\ne 0$. Let $F$ be a pure reduced torsion-free $R$-submodule of finite rank of $\widehat{R_J}$ such that $\widehat{R_J}/F$ is divisible. Then $(\widehat{R_J}/F)\otimes_{R}U=0$, $E\cong F\otimes_RU$ and we know that $F$ is polyserial. So, in the two cases, from a pure composition series of $\widehat{R_J}$ or $F$ with uniserial factors, we deduce a pure composition series of $E$ with uniserial factors. Hence $E$ is polyserial.
The second assertion is also proven.
\[L:pureunis\] Let $R$ be a chain ring and let $E$ be an indecomposable injective module such that $Z\subset E_{\sharp}$. Assume that $E$ contains a pure uniserial submodule. Then each indecomposable injective module $G$ for which $Z\subset G_{\sharp}$ contains a pure uniserial submodule.
After replacing $R$ by $R_{E_{\sharp}}$ we may assume that $E_{\sharp}=P$. First we shall prove that $E(R/Z)$ contains a pure uniserial submodule. If $Q$ is coherent, it is a consequence of [@Couch03 Corollary 22]. We assume that $Q$ is not coherent. So, $Z$ is flat by [@Couch03 Theorem 10]. By [@Cou06 Theorem 3] $E_Z$ is injective, $E_Z\ne 0$, and it contains a pure uniserial submodule $U$ and an injective hull of $U$. Let $A\in\mathcal{A}(E)$, $A\subset Z$. Since $A^{\sharp}=E_{\sharp}=P$, there exists $s\in P\setminus Z$ such that $A\subset (A:s)$. Let $t\in (A:s)\setminus A$. Then $Z\subset (A:t)$. So, $(A:t)_Z=Q$. It follows that $A_Z=tQ$. Hence $E(U)\cong E(Q/tQ)$. From [@Couc06 Proposition 14], we deduce that $E(R/Z)$ contains a pure uniserial submodule $V$. Let $x\in E(R/Z)$ such that $Z=(0:x)$. If $G$ is an indecomposable injective module such that $Z\subset G_{\sharp}$, $\mathcal{A}(G)$ contains a faithful ideal $B$. By [@Couch03 Proposition 6] $V/Bx$ is a pure uniserial submodule of $G$.
Let us observe that the condition $(8)$ of Theorem \[T:main\] implies that each indecomposable injective module contains a pure uniserial submodule.
\[P:polys\] Let $R$ be a chain ring and let $E$ be an indecomposable injective module such that $P=E_{\sharp}$. Assume that $E$ is polyserial. Then:
1. each indecomposable injective module $G$ for which $Z\subset G_{\sharp}$ is polyserial;
2. for each prime ideal $L\subseteq Z$, $E(R/L)$ and $E(R_L/aR_L)$ are polyserial, where $a\in L$ with $0\ne aR_L$.
$(1)$ holds by Lemma \[L:pureunis\] and Theorem \[T:main\].
$(2)$ holds by [@Couch03 Corollary 22] and Theorem \[T:main\].
If $R$ is a chain ring which is not a domain, satisfying $\mathrm{Mr}_{R}\widehat{R}<\infty$, then $\mathrm{Mr}_{R}\widehat{R}=\mathrm{Mr}_{\widetilde{R}}\widehat{R}$ even if $R\subset\widetilde{R}$.
Fleischer rank and dual Goldie dimension of indecomposable injective modules {#S:FldG}
============================================================================
\[R:Malcev\] If $M$ is a torsion-free module of finite rank over a valuation domain, it is easy to check that its Malcev rank is equal to its rank. So, if $M$ is a module over a chain ring $R$, then $\mathrm{Fr}\ M$ can be defined to be the minimum Malcev rank of flat modules having $M$ as an epimorphic image. Obviously $\mathrm{Mr}\ M\leq\mathrm{Fr}\ M$ for each module $M$.
\[P:IF\] Let $R$ be a chain ring and let $E$ be an indecomposable injective module such that $E_{\sharp}\subseteq Z$. Then $E$ is flat if $\mathcal{A}(E)\ne\{qQ\mid 0\ne q\in Z\}$.
If $\mathcal{A}(E)=\{rZ\mid r\in R\}$ then $E$ is flat by [@Couch03 Proposition 8]. So, we may assume that $A$ is not of the form $rZ$ if $A\in\mathcal{A}(E)$. By [@Couch03 Lemma 26] $A^{\sharp}=E_{\sharp}$ for each $A\in\mathcal{A}(E)$, so $A$ is an ideal of $Q$. It is easy to check that $(0:I)$ is also an ideal of $Q$ for each ideal $I$ of $R$. In the sequel we apply [@KlLe69 Proposition 1.3] to $Q$: $(0:(0:A))\ne A$ if and only if $A=qZ$ and $(0:(0:A))=qQ$ for some $q\in Z$. Let $r\in R$ and $x\in E$ such that $rx=0$. Then $r\in A$ where $A=(0:x)$. Since $rQ\subset A$, then $(0:A)\subset (0:r)$. Let $a\in (0:r)\setminus (0:A)$. It follows that $(0:a)\subseteq (0:(0:A))=A$. The injectivity of $E$ implies that there exists $y\in E$ such that $x=ay$. So, $E$ is flat.
Let $R$ be a chain ring. Assume that $d_R(L)<\infty$ and $R_L$ is almost maximal for a prime ideal $L$, and that $E(R/Z)$ contains a pure uniserial submodule $U$. Then $\mathrm{Mr}\ E=\mathrm{Fr}\ E$ for each indecomposable injective module $E$.
Let $E$ be an indecomposable injective module and $J=E_{\sharp}$. Since $\mathrm{Mr}\ E\leq\mathrm{Fr}\ E$ it is enough to show that $E$ is an epimorphic image of a flat module $G$ with $\mathrm{Mr}\ E=\mathrm{Mr}\ G$. First we assume that $J\subseteq Z$. If $Q$ is coherent then $E$ is flat. If $Q$ is not coherent and if $E\ncong E(Q/qQ)$, where $0\ne q\in Z$, then $E$ is flat by Proposition \[P:IF\]. If $E= E(Q/qQ)$, then by [@Couch03 Proposition 14] there exits an epimorphism $E(Q/Z)\rightarrow E$ whose kernel is a simple $Q$-module. It is easy to check that $\mathrm{Mr}\ E=\mathrm{Mr}\ E(Q/Z)$, and $E(Q/Z)$ is flat. Now, we assume that $Z\subset J$. In this case, $E\cong\widehat{R_J}\otimes_R(U/Ax)$ where $A$ is a faithful annihilator ideal of $E$ and $x\in U$ with $Z=(0:x)$. Moreover, $U$ is flat because so is $E(R/Z)$. Hence $E$ is an epimorphic image of $\widehat{R_J}\otimes_RU$ which is flat. If $R$ is not a domain then $\mathrm{Mr}\ E=\mathrm{Mr}\ \widehat{R_J}=\mathrm{Mr}\ \widehat{R_J}\otimes_RU$. If $R$ is a domain, by Theorem \[T:main\] $\widehat{R_J}$ contains a pure submodule $F$ of rank equal to $\mathrm{Mr}_{\widetilde{R}}\ \widehat{R_J}$ such that $\widehat{R_J}/F$ is a divisible module. In this case we can take $U=Q$ and $x=1$. Since $Q/A$ is a torsion module we have $E\cong F\otimes_RQ/A$. So, $E$ is a homomorphic image of $F\otimes_RQ$ and $\mathrm{Mr}\ E=\mathrm{Mr}_{\widetilde{R}}\ \widehat{R_J}=\mathrm{Mr}\ F\otimes_RQ$.
We say that a submodule $K$ of a module $M$ is **superfluous** if the equality $K+G=M$ holds only when $G=M$. A module $M$ is **co-uniform** if each of its proper submodules is superfluous. We say that $M$ has **dual Goldie dimension** $n$ (or $\mathrm{dG}\ M=n$) if there exists an epimorphism $\phi$ from $M$ into a direct sum of $n$ co-uniform modules such that $\ker\ \phi$ is superfluous.
\[P:MrdG\] Let $R$ be a chain ring. Then $\mathrm{dG}\ M\leq\mathrm{Mr}\ M$ for each $R$-module $M$.
Let $n$ a positive integer such that $n\leq\mathrm{dG}\ M$. Then there exists an epimorphism $\phi:M\rightarrow \oplus_{i=1}^n M_i$ where $M_i$ is a non-zero $R$-module for $i=1,\dots,n$. For each $i$, $1\leq i\leq n$, let $x_i$ be a non-zero element of $M_i$ and let $y_i\in M$ such that $x_i=\phi(y_i)$. If $\Sigma_{i=1}^na_iy_i=0$ where $a_i\in R$ for $i=1,\dots,n$, we successively deduce that $\Sigma_{i=1}^na_ix_i=0$, $a_ix_i=0$ and $a_i\in P$ for $i=1,\dots,n$. It follows that $\mathrm{Mr}\ M\geq n$ for each integer $n\leq\mathrm{dG}\ M$. So, $\mathrm{dG}\ M\leq\mathrm{Mr}\ M$.
\[P:dG\] Let $R$ be a chain ring. Suppose there exists a non-zero prime ideal $L$ such that $Z\subseteq L$, $d_R(L)<\infty$ and $R_L$ is almost maximal. Then each indecomposable injective module $E$ is polyserial and $\mathrm{Mr}\ E=\mathrm{dG}\ E$.
Let $H$ be the injective hull of $R/Z$. Then, since $H$ is an $R_Z$-module and $R_Z$ is almost maximal, $H$ is uniserial. Let $E$ be an indecomposable injective module and let $J=E_{\sharp}$. If $J\subseteq Z$ then $E$ is uniserial. If $Z\subset J$ we do as in the proof of Lemma \[L:pureunis\] to show that $E$ contains a pure uniserial submodule $U$. By Theorem \[T:main\] $E$ is a polyserial module. If $V$ is a uniserial factor of a pure composition series of $\widehat{R_J}$, then by Theorem \[T:InjHull\] $V\cong R_{L'}$ for some prime ideal $L'\supseteq L$. Il follows that $E_L\cong U_L^d$ where $d=d_{R_J}(L_J)=\mathrm{Mr}\ E$. Since $Z\subseteq L$, $E_L$ is a homomorphic image of $E$. So, $\mathrm{dG}\ E\geq d$. By Proposition \[P:MrdG\] $\mathrm{dG}\ E=d$.
We say that a chain ring is **strongly discrete** if $L^2\ne L$ for each non-zero prime ideal $L$.
\[P:discrete\] Let $R$ be a chain ring such that $Q$ is strongly discrete. Then each indecomposable injective module contains a pure uniserial submodule. For a such ring the nine conditions of Theorem \[T:main\] are equivalent.
Let $E$ be an indecomposable injective module and let $J=E_{\sharp}$. First assume that $J\subseteq Z$. Since $J^2\ne J$, $JR_J=aR_J$ where $a\in J\setminus J^2$. Recall that $E$ is a module over $R_J$. If $A\in\mathcal{A}(E)$ then $A^{\sharp}=J$ by [@Couch03 Lemma 26]. Since $JR_J$ is principal over $R_J$, so is $A$. Because $Z=sQ$ for some $s\in Z\setminus Z^2$, $(0:Z)=(0:s)\ne 0$. From [@Couch03 Corollary 22] we deduce that $E$ contains a pure uniserial submodule (which is isomorphic to $R_J$). If $Z\subset J$ we again use [@Couch03 Corollary 22] to conclude.
\[C:main\] Let $R$ be a valuation domain. Then the nine conditions of Theorem \[T:main\] are equivalent. Moreover, for each indecomposable injective module $E$, $\mathrm{Mr}\ E=\mathrm{Fr}\ E=\mathrm{dG}\ E=\mathrm{Mr}_{\widetilde{R_J}}\ \widehat{R_J}$ where $J=E_{\sharp}$.
It is possible to build examples of chain rings satisfying the nine equivalent conditions of Theorem \[T:main\] by using [@FaZa86 Example 6 and Theorem 8]. These examples are strongly discrete (and Henselian). If $R$ is a such example then $\mathrm{Mr}_{\widetilde{R}}\ \widehat{R}=p^m$, where $p$ is a prime number and $m$ a non-negative integer. By [@Vam90 Remark p.16] $\mathrm{Mr}_{\widetilde{R}}\ \widehat{R}$ is always a prime power.
Indecomposable injective modules over one Krull-dimensional local Noetherian rings {#S:Noe}
==================================================================================
From a result by Marie-Paule Malliavin we deduce Theorem \[T:Noe\]. If $M$ is a module of finite length, we denote by $\ell(M)$ its length.
\[T:Noe\] Let $R$ be a one Krull-dimensional local Noetherian ring. There exists a positive integer $n$ such that $\mathrm{Mr}\ E\leq n$ for each indecomposable injective $R$-module $E$.
By [@Mal66 Théorème 1.4.2] $\mathrm{Mr}\ R$ is finite. We put $m=\mathrm{Mr}\ R$. Let $E$ be an indecomposable injective module. Then there exists a prime ideal $L$ such that $E=E(R/L)$. First we assume that $L$ is a minimal prime. It follows that $E$ is a module of finite length over $R_L$ by [@Mat58 Theorem 3.11(2)] since $R_L$ is Artinian. In this case $E$ has a composition series whose factors are isomorphic to $R_L/LR_L$. It is easy to see that $\mathrm{Mr}\ R_L/LR_L=\mathrm{Mr}\ R/L\leq m$. Now, by induction on the length of $E$ over $R_L$ and by using [@FuSa01 Lemma XII.1.4] we get that $\mathrm{Mr}\ E<\infty$. Now we assume that $L=P$ the maximal ideal of $R$. Let $M$ be a finitely generated submodule of $E$ and $A=(0:M)$. Since $E$ is Artinian then $M$ is a module of finite length and $R/A$ is Artinian. By [@Cou81 Proposition 1.2] $M$ is injective over $R/A$ and $R/A=\mathrm{End}_R(M)$. Let $B$ be the ideal of $R$ such that $B/A$ is the socle of $R/A$. If $p=\ell(B/A)$ then $p=\mathrm{gen}\ B\leq m$. So, there is an exact sequence $0\rightarrow R/A\rightarrow M^p$ and by applying the functor $\mathrm{Hom}_{R/A}(-,M)$ to this sequence, we get that $M$ is a homomorphic image of $(R/A)^p$. So, $\mathrm{Mr}\ E\leq m$. Since the set of prime ideals of $R$ is finite the theorem is proven.
Let $R$ be a local ring of maximal ideal $P$ such that $P^2=0$. If $\mathrm{gen}\ P=n$ where $n>0$ it is easy to check that $\mathrm{Mr}\ E(R/P)\leq n$.
In the sequel, for each integer $n>1$ we shall give an example of a one Krull-dimensional local Noetherian domain $D$ whose all finitely generated uniform modules are generated by at most $n$ elements.
Consider the Noetherian domain $R$ defined in in the following way. Let $K$ be a field, $K[X,Y]$ the polynomial ring in two variables $X$ and $Y,$ and $f = Y^n-X^n(1+X).$ By considering that $f$ is a polynomial in one variable $Y$ with coefficients in $K[X],$ it follows from Eisenstein’s criterion that $f$ is irreducible. Then $D = \displaystyle{\frac{K[X,Y]}{fK[X,Y]}}$ is a domain. Let $x$ and $y$ be the images of $X$ and $Y$ in $D$ by the natural map and $P'$ the maximal ideal of $D$ generated by $\{x,y\}.$ Let $R=D_{P'}$ and $P=P'R$.
Then $\mathrm{Mr}\ R= n$ and $\mathrm{Mr}\ E= n$ for each indecomposable injective $R$-module $E$.
There are only two types of indecomposable injective modules: $E(R/P)$ and $Q$ the quotient field of $D$ and $R$. Let $M$ be a finitely generated submodule of $Q$. Then $M$ is isomorphic to an ideal of $R$. As a module over over $K[X]$ $D$ is generated by $n$ elements $1,y,y^2,\dots,y^{n-1}$. Since $K[X]$ is a principal ideal domain, by [@WiWi75 Lemma 1.3] each $K[X]$-submodule of $D$ is generated by at most $n$ elements. It follows that each ideal of $D$ and each ideal of $R$ is generated by at most $n$ elements. Let us observe that $\mathrm{gen}\ P^m=n$ for each $m\geq n-1$. So, $\mathrm{Mr}\ R=\mathrm{Mr}\ Q=n$. As in the proof of Theorem \[T:Noe\] we show that $\mathrm{Mr}\ E(R/P)\leq n$, and since $\mathrm{gen}\ P^{n-1}=n$ we have $\mathrm{gen}\ \mathrm{Hom}_R(R/P^n,E(R/P))=n$. The proof is now complete.
Goldie dimension and localization {#S:Gol}
=================================
At the beginning of this section $R$ is not necessarily a chain ring.
\[P:Gol\] Let $R$ be a ring satisfying one of the following two conditions:
1. $R_P$ is a one Krull-dimensional domain for each maximal ideal $P$;
2. $R_P$ is Noetherian for each maximal ideal $P$.
Then, $S^{-1}M$ has finite Goldie dimension for each $R$-module $M$ of finite Goldie dimension and for each multiplicative subset $S$ of $R$.
If $\mathrm{Gd}\ M<\infty$ then $M$ is a submodule of a finite direct sum of indecomposable injective modules $(E_i)_{1\leq i\leq n}$. It follows that $\mathrm{Gd}\ S^{-1}M<\infty$ if and only if $\mathrm{Gd}\ S^{-1}E_i<\infty$ for $i=1,\dots,n$. So, we may assume that $M$ is injective and indecomposable. On the other hand, since $\mathrm{End}_R(M)$ is a local ring, there exists a maximal ideal $P$ such that $M$ is a module over $R_P$. So, we may assume that $R$ is local of maximal ideal $P$.
If $R$ satisfies $(1)$ then $S=R\setminus \{0\}$. Either $M$ is torsion-free and $S^{-1}M=M$, or $M$ is torsion and $S^{-1}M=0$.
If $R$ satisfies $(2)$, we may assume that $M=E(R/P)$ and $S\cap P\ne\emptyset$. Let $\phi$ be the natural map $M\rightarrow S^{-1}M$. Since $M$ is artinian by [@Mat58 Corollary 3.4] then so is the image of $\phi$. It follows that $S^{-1}M$ is an essential extension of a semisimple module $X$. But, for each $s\in S\cap P$, $sX=0$. We conclude that $S^{-1}M=0$.
\[P:dGol\] Let $R$ be a zero Krull-dimensional ring. Then $\mathrm{dG}\ S^{-1}M<\infty$ for each module $M$ with $\mathrm{dG}\ M<\infty$ and for each multiplicative subset $S$ of $R$.
Since the natural maps $R\rightarrow S^{-1}R$ and $M\rightarrow S^{-1}M$ are surjective then $\mathrm{dG}\ S^{-1}M\leq\mathrm{dG}\ M$.
Let $R$ be a local two Krull-dimensional UFD, $p$ a prime element of $R$ and $S=\{p^n\mid n\in\mathbb{N}\}$. Then $\mathrm{dG}\ R=1$ and $\mathrm{dG}\ S^{-1}R=\infty$.
The first equality is obvious. Let $\Phi$ be the set of prime elements of $R$. If $P$ is the maximal ideal of $R$ then $P=\cup_{q\in\Phi} Rq$. So, $\Phi$ is not finite, else, by a classical lemma $P=Rq$ for some $q\in\Phi$ that is impossible. Let $n$ be a positive integer, let $q_1,\dots,q_n$ be $n$ distinct elements of $\Phi\setminus \{p\}$ and let $a=q_1\times\dots\times q_n$. By the chinese remainder theorem $S^{-1}R/aS^{-1}R\cong S^{-1}R/q_1S^{-1}R\times\dots\times S^{-1}R/q_nS^{-1}R$. So, $\mathrm{dG}\ S^{-1}R\geq n$ for each $n>0$.
\[T:main2\] Let $R$ be a chain ring. The following conditions are equivalent:
1. For each module $M$ of finite Goldie dimension and for each prime ideal $L$, $M_L$ has finite Goldie dimension;
2. for each prime ideal $L\ne 0$, $d_R(L)$ is finite.
$(1)\Rightarrow (2)$. By way of contradiction suppose there exists a non-zero ideal $L$ such that $d_R(L)=\infty$. Then, for each positive integer $n$, $\widehat{R}/L\widehat{R}$ contains a torsion-free $R/L$- module $F$ of rank $n$. Let $A$ be an ideal such that $A^{\sharp}=P$ and $A\subset L$. By [@Couc06 Proposition 1.(2)] $\widehat{R}/A\widehat{R}$ is isomorphic to a submodule of the injective hull $E$ of $R/A$. Since $F_L$ is a $(R/L)_L$-vector space of dimension $n$ contained in $(\widehat{R}/L\widehat{R})_L$, we deduce from [@Coucho07 Proposition 21], applied to $(\widehat{R}/A\widehat{R})_L$ that $E_L$ contains a free $(R/A)_L$-module of rank $n$. So, $\mathrm{Gd}\ E_L\geq n$ for each integer $n$.
$(2)\Rightarrow (1)$. It is sufficient to show that $\mathrm{Gd}\ E_L<\infty$ for each indecomposable injective module $E$ and each non-zero prime ideal $L$. Let $J=E_{\sharp}$. If $J\subseteq L$, then $E$ is a module over $R_J$, whence $E_L=E$. If $L\subset J$, since $d_{R_J}(L_J)\leq d_R(L)$, after replacing $R_J$ by $R$, we may assume that $J=P$. If $L=0$ (in the case where $R$ is a domain) then $E_L=0$. By [@Couch03 Corollary 28] $E$ is either faithful or annihilated by a simple ideal. So, if $L\ne 0$, there exists $A\in\mathcal{A}(E)$ such that $A\subset L$. We put $E'=(0:_EA)$. First we show that $E'_L$ is essential in $E_L$. Let $x\in E$ such that $\dfrac{x}{1}\notin E'_L$. Then $x\notin E'$. So, $(0:x)=rA$ where $r\in P$. It follows that $(0:rx)=A$. We conclude that $r\dfrac{x}{1}\in E'_L$ and $r\dfrac{x}{1}\ne 0$. Let $d=d_R(L)$. Then $(\widehat{R})_L/L(\widehat{R})_L\cong (R/L)_L^d$. Since $P/A$ is the set of zero-divisors of $R/A$ then $(R/A)_L$ is self FP-injective by [@Couch03 Theorem 11(2)]. From [@Couc06 Proposition 1] it is easy to deduce that $\widehat{R/A}$ is singly projective over $R/A$. By [@Coucho07 Proposition 6] so that is $(\widehat{R/A})_L$ over $(R/A)_L$. By [@Coucho07 Propositions 24 and 21] $E'_L$ (which is isomorphic to $(\widehat{R/A})_L$) contains an essential free $(R/A)_L$-submodule of rank $d$. We conclude that $\mathrm{Gd}\ E_L=d<\infty$.
\[C:main2\] Let $R$ be a chain ring and let $J$ be the intersection of all non-zero prime ideals. Then $J$ is prime ($=N$ if $R$ is not a domain) and the following assertions hold:
1. if $J\ne0$ then Goldie dimension finiteness is preserved by localization if and only if $d_R(J)$ is finite;
2. if $J=0$ and if $0$ is a non countable intersection of non-zero prime ideals then Goldie dimension finiteness is preserved by localization if and only if there exists a non-zero prime ideal $L$ such that $d_R(L)$ is finite and $R_L$ is almost maximal.
$(1)$ is an immediate consequence of Theorem \[T:main2\].
$(2)$. First we will show that there exists a positive integer $p$ such that $d_R(L)\leq p$ for each non-zero prime ideal $L$. By way of contradiction suppose there exists a non-zero prime ideal $L_n$ such that $d_R(L_n)\geq n$, for each integer $n>0$. Let $H=\cap_{n>0}L_n$. Then $H$ is a non-zero prime ideal and $d_R(H)\geq n$ for each integer $n>0$. We get a contradiction by Theorem \[T:main2\]. Let $p$ be the maximum of $d_R(I)$ where $I$ runs over all non-zero prime ideals of $R$ and let $L$ be the maximal prime ideal for which $d_R(L)=p$. If $L'$ is a non-zero prime ideal, $L'\subset L$, then $\mathrm{Mr}\ \widehat{R/L'}=p$. By Theorem \[T:InjHull\] $R_L/L'$ is maximal. We conclude that $R_L$ is almost maximal.
Let us observe that the following conditions:
1. each indecomposable injective $R$-module is polyserial;
2. the Goldie dimension finiteness is preserved by localization;
are equivalent if $R$ is a valuation domain such that $0$ is a non countable intersection of non-zero prime ideals. But, generally these two conditions are not equivalent. For instance, if $J\ne 0$, $d_R(J)<\infty$ and $R_J$ not almost maximal, where $J$ is the intersection of all non-zero prime ideals, then $R$ satisfies condition $(2)$ but not condition $(1)$. Another example of a chain ring satisfying condition $(2)$ but not condition $(1)$ is the following:
Let $R$ be a strongly valuation domain whose set of non-zero prime ideals is $\{L_n\mid n\in\mathbb{N}\}$ with $L_0=P$ and $L_{n+1}\subset L_n$ for each $n\in\mathbb{N}$. Moreover we assume that $c_R(L_n)=p$ for each $0\ne n\in\mathbb{N}$, where $p$ is a prime integer. A such ring $R$ exists by [@FaZa86 Theorem 8]. For each integer $n>0$, $d_R(L_n)=p^n$. So, condition $(2)$ is satisfied by Theorem \[T:main2\]. But condition $(1)$ doesn’t hold because there is no non-zero prime ideal $L$ with $R_L$ almost maximal.
[10]{}
F. Couchot. Les modules artiniens et leurs enveloppes quasi-injectives. In [*Séminaire d’algèbre Paul Dubreuil et Marie-Paule Malliavin*]{}, volume 867 of [*Lecture Notes in Math.*]{}, pages 380–395. Springer, (1981).
F. Couchot. Injective modules and fp-injective modules over valuation rings. , 267:359–376, (2003).
F. Couchot. Local rings of bounded module type are almost maximal valuation rings. , 33(8):2851–2855, (2005).
F. Couchot. Localization of injective modules over valuations rings. , 134(4):1013–1017, (2006).
F. Couchot. Pure-injective hulls of modules over valuation rings. , 207:63–76, (2006).
F. Couchot. Flat modules over valuation rings. , 211:235–247, (2007).
F. Couchot. Valuation domains with a maximal immediate extension of finite rank. , 323:32–41, (2010).
A. Facchini and P. Zanardo. Discrete valuation domains and ranks of their maximal extensions. , 75:143–156, (1986).
L. Fuchs and L. Salce. , volume 97 of [*Lecture Notes in Pure an Appl. Math.*]{} Marcel Dekker, New York, (1985).
L. Fuchs and L. Salce. . Number 84 in Mathematical Surveys and Monographs. American Mathematical Society, Providence, (2001).
D.T. Gill. Almost maximal valuation rings. , 4:140–146, (1971).
G.B. Klatt and L.S. Levy. Pre-self injectives rings. , 137:407–419, (1969).
M.P. Malliavin-Brameret. Largeur d’anneaux et de modules. , Mémoire 8, (1966).
E. Matlis. Injective modules over noetherian rings. , 8:511–528, (1958).
E. Matlis. Injective modules over [P]{}rüfer rings. , 15:57–69, (1959).
H. Tsang. . PhD thesis, University of Chicago, (1965).
P. Vámos. Classical [R]{}ings. , 34:114–129, (1975).
P. Vámos. Decomposition problems for modules over valuation domains. , 41:10–26, (1990).
R.B. Warfield. Purity and algebraic compactness for modules. , 28(3):689–719, (1969).
R. Wiegand and S. Wiegand. Finitely generated modules over [B]{}ezout rings. , 58(2):655–664, (1975).
[^1]: we prefer “chain ring ” to “valuation ring” to avoid confusion with “Manis valuation ring”.
[^2]: This condition holds for each valuation domain and other classes of chain rings but we don’t know if it is verified by any chain ring.
[^3]: this definition is equivalent to the usual one when $R$ is local, see [@Tsa65].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Auto-encoder is an important architecture to understand point clouds in an encoding and decoding procedure of self reconstruction. Current auto-encoder mainly focuses on the learning of global structure by global shape reconstruction, while ignoring the learning of local structures. To resolve this issue, we propose Local-to-Global auto-encoder (L2G-AE) to simultaneously learn the local and global structure of point clouds by local to global reconstruction. Specifically, L2G-AE employs an encoder to encode the geometry information of multiple scales in a local region at the same time. In addition, we introduce a novel hierarchical self-attention mechanism to highlight the important points, scales and regions at different levels in the information aggregation of the encoder. Simultaneously, L2G-AE employs a recurrent neural network (RNN) as decoder to reconstruct a sequence of scales in a local region, based on which the global point cloud is incrementally reconstructed. Our outperforming results in shape classification, retrieval and upsampling show that L2G-AE can understand point clouds better than state-of-the-art methods.'
author:
- Xinhai Liu
- Zhizhong Han
- Xin Wen
- 'Yu-Shen Liu'
- Matthias Zwicker
bibliography:
- 'reference.bib'
title: 'L2G Auto-encoder: Understanding Point Clouds by Local-to-Global Reconstruction with Hierarchical Self-Attention'
---
<ccs2012> <concept> <concept\_id>10010147.10010178.10010224</concept\_id> <concept\_desc>Computing methodologies Computer vision</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010147.10010178.10010224.10010240.10010242</concept\_id> <concept\_desc>Computing methodologies Shape representations</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002951.10003317</concept\_id> <concept\_desc>Information systems Information retrieval</concept\_desc> <concept\_significance>300</concept\_significance> </concept> </ccs2012>
{height="5cm" width="16cm"}
Introduction
============
In recent years, point clouds have attracted increasing attention due to the popularity of various depth sensors in different applications. Not only the traditional methods, deep neural networks have also been applied to point cloud analysis and understanding. However, it remains a challenge to directly learn from point clouds. Different from 2D images, point cloud is an irregular 3D data which makes it difficult to directly use traditional deep learning framework, e.g., traditional convolution neural network (CNN). The traditional CNN usually requires some fixed spatial distribution around each pixel so as to facilitate the convolution. One way to alleviate the problem is to voxelize a point cloud into voxels and then apply 3D Cov-Nets. However, because of the sparsity of point clouds, it leads to resolution-loss and explosive computation complexity, which sacrifices the representation accuracy.
To address above challenges, PointNet [@qi2016pointnet] has been proposed to directly learn shape representations from raw point sets. Along with the availability of directly learning from point clouds by deep learning models, auto-encoder (AE) has become an vital architecture of the involved neural networks. Current AE focuses on the learning of the global structure of point clouds in the encoding and decoding procedure. However, current AE structure is still limited by learning the local structure of point clouds, which tends to be an important piece of information for point cloud understanding.
To simultaneously learn global and local structure of point clouds, we propose a novel auto-encoder called Local-to-Global auto-encoder (L2G-AE). Different from traditional auto-encoder, L2G-AE leverages a local region reconstruction to learn the local structure of a point cloud, based on which the global shape is incrementally reconstructed for the learning of the global structure. Specifically, the encoder of L2G-AE can hierarchically encode the information at point, scale and region levels, where a novel hierarchical self-attention is introduced to highlight the important elements in each level. The encoder further aggregates all the information extracted from the point cloud into a global feature. In addition, L2G-AE employs a RNN-based decoder to decode the learned global feature into a sequence of scales in each local region. And based on scale features, the global point cloud is incrementally reconstructed. L2G-AE leverages this local to global reconstruction to facilitate the point cloud understanding, which finally enables local and global reconstruction losses to train L2G-AE.
Our key contributions are summarized as follows.
- We propose L2G-AE to enable the learning of global and local structures of point clouds in an auto-encoder architecture, where the local structure is very important in learning highly discriminative representations of point clouds.
- We propose hierarchical self-attention to highlight important elements in point, scale and region levels by learning the correlations among the elements in the same level.
- We introduce RNN as decoding layer in an auto-encoder architecture to employ more detailed self supervision, where the RNN takes the advantage of the ordered multi-scale areas in each local region.
Related Work
============
Point clouds is a fundamental type of 3D data format which is very close to the raw data of various 3D sensors. Recently, applications of learning directly on point clouds have received extensive attention, including shape completion [@Stutz2018CVPR], autonomous driving [@qi2017frustum], 3D object detection [@simon2018complex; @Yang2018pixor; @zhou2017voxelnet], recognition and classification [@qi2016pointnet; @NIPS2017_7095; @golovinskiy2009shape; @li2018pointcnn; @wang2018dynamic; @xu2018spidercnn; @shen2018mining; @xie2018attentional; @li2018so; @you2018pvnet], scene labeling [@NIPS2011_4226], upsampling [@yu2018pu; @2019arXiv181111286Y], dense labeling and segmentation [@Wang2018pointseg] , etc.
Due to the irregular property of point cloud and the inspiring performances of 2D CNNs on large-scale image repositories such as ImageNet [@deng2009imagenet], it is intuitive to rasterize point clouds into 3D voxels and then apply 3D CNNs. Some studies [@qi2017frustum; @zhou2017voxelnet; @HanCyber17a] represent each voxel with a binary value which indicates the occupation of this location in space. The main problem of voxel-based methods is the fast growth of neural network size and computation complexity with the increasing of spatial resolution. To alleviate this problem, some improvements [@li2016fpnn] have been proposed to explore the data sparsity of point clouds. However, when dealing with point clouds with huge number of points, the complexity of the neural network is still unacceptable.
Recently, deep neural networks work quite effectively on the raw 3D point clouds. Different from learning from readered views [@han2018seqviews2seqlabels; @han20193d2seqviews; @han20192seq2seq; @han2019view; @han2019parts4feature; @han20193dviewgraph] 2D meshes [@Zhizhong2016] or 3D voxels [@Zhizhong2016b; @han2017boscc; @han2018deep], PointNet [@qi2016pointnet] is the pioneer study which directly learns the representation for point clouds by computing features for each point individually and aggregating these features with max-pool operation. To capture the contextual information of local patterns inside point clouds, PointNet++ [@NIPS2017_7095] uses sampling and grouping operations to extract features from point clusters hierarchically. Similarly, several recent studies [@riegler2017octnet; @klokov2017escape] explores indexing structures, which divides the input point cloud into leaves, and then aggregates node features from leaves to the root. Inspired by the convolution operation, recent methods [@li2018pointcnn; @wang2018dynamic; @xu2018spidercnn] investigate well-designed CNN-like operations to aggregate points in local regions by building local connections with k-neareat-neighbors (kNN).
Capturing the context information inside local regions is very important for the discriminative ability of the learned point cloud representations. KC-Net [@shen2018mining] employs a kernel correlation layer and a graph pooling layer to capture the local patterns of point clouds. ShapeContextNet [@xie2018attentional] extends 2D Shape Context [@belongie2001shape] to the 3D, which divides a local region into small bins and aggregates the bin features. Point2Seqeuce [@liu2019point2sequence] employs an attention-based sequence to sequence architecture to encode the multi-scale area features inside local regions.
In order to alleviate the dependence on the labeled data, some studies have performed unsupervised learning for point clouds. FoldingNet [@yang2018foldingnet] proposes a folding operation to deform a canonical 2D grid onto the suface of a point cloud. 3D-PointCapsNet [@zhao2019pcn] employs a dynamic routing scheme in the reconstruction of input point clouds. However, it is difficult for these methods to capture the local patterns of point clouds. Similar to FoldingNet, PPF-FoldNet [@Deng2018ppffoldnet] also learns local descriptors on point cloud with a folding operation. LGAN [@Achlioptas2017latent] proposes an auto-encoder based on PointNet and extends the decoder module to the point cloud generation application with GAN. In this work, we propose a novel auto-encoder architecture to learn representations for point clouds. On the encoder side, an hierarchical self-attention mechanism is applied to embedding the correlation among features in each level. And on the decoder side, an interpolation layer and a RNN decoding layer are engaged to reconstruct multi-scale areas inside local regions. After building local areas, the global point cloud is generated by a fully-connected (FC) layer which acts as a down sampling function.
Method
======
Now we introduce the L2G-AE in detail, where the structure is illustrated in Figure \[fig:architecture\]. The input of the encoder is an unordered point set $\bm{P} = \{p_1, p_2, \cdots, p_N \}$ with $N$ ($N=1024$) points. Each point in the point set is composed of a 3D coordinate $(x,y,z)$. L2G-AE first establishes multi-scale areas $\bm{A}_t$ $(t \in [1,T])$ in each local region around the sampled points. Then, a hierarchical feature abstraction is enforced to obtain the global features of input point clouds with self-attentions. In the decoder, we simultaneously reconstruct local scale areas and global point clouds by hierarchical feature decoding. The output of L2G-AE is the reconstructed local areas $\bm{A}^{'}_t$ and the reconstructed $\bm{P}^{'}$ with same number of points to $\bm{P}$.
![A multi-scale example inside a local region of an airplane point cloud, where there are four scales areas $[\bm{A}_1, \bm{A}_2, \bm{A}_3, \bm{A}_4]$ with different colors around the centroid point (red).[]{data-label="fig:multiscale"}](ms.pdf){height="3.5cm" width="5cm"}
Multi-scale Establishment
-------------------------
To capture fine-grained local patterns of point clouds, we first establish multi-scale areas in each local region, which is similar to PointNet++ [@NIPS2017_7095] and Point2Sequence [@liu2019point2sequence]. Firstly, a subset $\{p_{i_1}, p_{i_2},\\ \cdots, p_{i_M}\}$ of the input points is selected as the centroid of local regions by iterative farthest point sampling (FPS). The latest point $p_{i_j}$ is always the farthest one from the rest points $\{p_{i_1}, p_{i_2}, \cdots, p_{i_{j-1}}\}$. Compared to other sampling method, such as random sampling, FPS can achieve a better coverage of the entire point cloud with the given same number of centroids. As shown in Figure \[fig:multiscale\], around each sampled centroid, $T$ different scale local areas are established continuously by kNN searching with $\{K_1, K_2, \cdots, K_T\}$ nearest points, respectively. An alternative searching method is ball query [@NIPS2017_7095] which selects all points with a radius around the centroid. However, it is difficult for ball query to ensure the information inside local regions, which is sensitive to the sparsity of the input point clouds.
![Self-attention module. The input of this module is a $D_1 \times D_2$ feature map and the output is another $D_1 \times (D_2 + C)$ feature map, where $C$ is a parameter.[]{data-label="fig:c_e"}](CE.pdf){height="4cm" width="8cm"}
Hierarchical Self-attention
---------------------------
In current work of learning on point clouds, Multi-Layer-Perceptron (MLP) layer is widely applied to integrate multiple features. Traditional MLP layer first abstracts each feature into higher dimension individually and then aggregates these features by a concise max pooling operation. However, these two simple operations can hardly encode the correlation between feature vectors in the feature space. Inspired by the self-attention machanism in [@zhang2018sa], the attention machanism is suitable for improving the traditional MLP by learning the correlation between features. In this work, we propose a self-attention module to make up the defects of the MLP layer with an attention mechanism. Here, self-attention refers to learn the correlation among features in the same level.
Different from the raw self-attention, we enforce a hierarchical feature extraction architecture with hierarchical self-attention in the encoder. There are three different levels inside the encoder, including point level, scale level, and region level. At each level, we introduce a self-attention module to learn self-attention weights by mining the correlations among the corresponding feature elements. Consequently, three self-attention modules are designed to propagate features from the lower level to the higher level. Supposed the input of the self-attention module is a feature map $\bm{x} \in \mathbb{R}^{D_1 \times D_2}$, where $D_1$, $D_2$ are the dimensions of the feature map. Therefore, $D_1$, $D_2$ are equal to $K_t$, $3$ in the point level, equal to $T$, $D$ in the scale level and equal to $M$, $D$ in the region level, respectively.
As depicted in Figure \[fig:c\_e\], the feature map $\bm{x}$ is first transformed into two feature spaces $\bm{f}$ and $\bm{g}$ to calculate the attention below, where $\bm{f}(\bm{x}) = \bm{W_f x}$, $\bm{g}(\bm{x}) = \bm{W_g x}$, $$\beta_{j,i} = \frac{exp(s_{ij})}{\sum_{i=1}^{D_1} exp(s_{ij})}, \text{where } s_{ij} = \bm{f}(\bm{x_i})^T\bm{g}(\bm{x_j}),$$ and $\beta_{j,i}$ evaluates the attention degree which the model pays to the $i^{th}$ location when synthesizing the $j^{th}$ feature vector. Then the attention result is $\bm{r} = (\bm{r_1}, \bm{r_2}, \cdots, \bm{r_j}, \cdots, \bm{r_{D_1}}) \in \mathbb{R}^{D_1 \times D_2}$, where $$\bm{r_j} = \sum_{i=1}^{D_1} \beta_{j,i}\bm{h}(\bm{x_i}), \text{where }\bm{h}(\bm{x_i}) = \bm{W_h x_i}.$$ In above formulation, $\bm{W_f},\bm{W_g},\bm{W_h} \in \mathbb{R}^{D_2 \times C}$ are learned weight matrices, which are implemented as $1\times1$ convolutions. We use $C = M / 8 $ in the experiments.
In addition, inspired by the skip link operation in ResNet[@he2016deep] and DenseNet [@huang2017densely], we further concatenate the result of the attention mechanism with the input feature matrix. Therefore, the final output of the self-attention module is given by $$\bm{o_i} = \bm{x_i} \oplus \bm{r_i},$$ where $\oplus$ is the concatenation operation. This allows the network to rely on the cues among the feature vectors.
To aggregate the features with correlation information, a MLP layer and a max pooling operation are employed to integrate the multiple features. In particular, the first self-attention module aggregates the points in a scale to a D-dimensional feature vector. The second one encodes the multi-scale features in a region into a D-dimensional feature. The final one integrates features of all local regions on a point cloud into a 1024-dimensional global feature. Therefore, the encoder hierarchically abstracts point features from the levels of point, scale and region to a global representation of the input point cloud.
Interpolation Layer
-------------------
The target of the decoder is to generate the points of the local areas and entire points. Previous approaches [@Achlioptas2017latent; @yang2018foldingnet; @Deng2018ppffoldnet] usually use simple fully-connected (FC) layers or MLP layers to build the decoder. However, the expressive ability of the decoder is largely limited without considering the relationship among features. In this work, we propose a progressive decoding way which can be regarded as a reverse process of the encoding. The first step is to generate local region features from the global feature. To propagate the global feature $\bm{g}$ to region features, a simple interpolation operation is first engaged in the decoder. The local region feature $\bm{l}_i$ is calculated by $$\bm{l}_i = \frac{c}{(p_i-p_0)^2}\bm{g}, i \in [1,M],$$ where $c$ ($c={10}^{-10}$) is a constant. Here, $p_0 = (0,0,0)$ is the centroid of the input point cloud after the normalization processing. And $p_i$ is the centroid point of the corresponding local region. By the simple interpolation operation, the spatial distribution information of local region can be integrated to facilitate the feature decoding. The interpolated local region features are then concatenated with skip linked local region features from the encoder. The concatenated features are passed through another MLP layer into a $M \times D$ feature matrix.
![The decoding process of the RNN layer.[]{data-label="fig:rnn"}](RNN.pdf){height="2cm" width="8cm"}
RNN Layer
---------
Given the feature of local regions, we want to decode the scale level features. Due to the multi-scale setting, the features of different scales in a local region can be regarded as a feature sequence with length $T$. As we all know that recurrent neural network [@hochreiter1997long] has shown excellent performances in processing sequential data. Thus, a RNN decoding layer is employed to generate the multi-scale area features. The decoding process is shown in Figure \[fig:rnn\]. We first replicate the local region feature $\bm{l}_i$ for $T$ times, and the replicated local region features are feed into the RNN layer by $$\bm{h}_t = f(\bm{h}_{t-1}, \bm{l}_i^t), t \in [1,T],$$ where $f$ is a non-linear activation function and $t$ is the index of RNN step. Therefore, the predicted $t^{th}$ area feature $\bm{a}_t$ can be calculated by $$\bm{a}_t = \bm{W}_\theta \bm{h}_t.$$ Here, $\bm{W}_d$ is a learnable weight matrix. To generate the points inside each local area, several FC layers are adopted to reconstruct the points. The local area $\bm{A}^{'}_t$ is reconstructed by $$\bm{A}^{'}_t = \bm{W}_{\theta_t}\bm{a}_t + b_{\theta_t},$$ where $\bm{W}_{\theta_t}$, $b_{\theta_t}$ are weights of the FC layer. Based on the reconstructed local areas, another FC layer is applied to incrementally reconstruct the entire point cloud. All reconstructed areas are concatenated and then passed through the FC layer by $$\bm{P} = \bm{W} [\bm{A}^{'}_1 \oplus \bm{A}^{'}_2 \oplus \cdots \oplus \bm{A}^{'}_T] + b.$$ Here, $\oplus$ represents the concatenation operation.
Loss Function
-------------
We propose a new loss function to train the network in an end-to-end fashion. There are two parts in the loss function, local scale reconstruction and global point cloud reconstruction, respectively. As mentioned earlier, we should encourage accurate reconstruction of local areas and the global point cloud at the same time. Suppose $\bm{A}_t$ is the $t^{th}$ scale area in the multi-scale establishment subsection, then, the local reconstruction error for $\bm{A}^{'}_t$ is measured by the well-known Chamfer distance, $$\begin{split}
L_{local} = d_{CH}(\bm{A}_t, \bm{A}^{'}_t) = \sum_{t=1}^T (\frac{1}{|\bm{A}_t|}\sum_{p_i \in \bm{A}_t} \min_{p^{'}_i \in \bm{A}^{'}_t} \lVert p_i - p^{'}_i \rVert_2 \\ + \frac{1}{|\bm{A}^{'}_t|}\sum_{p^{'}_i \in \bm{A}^{'}_t} \min_{p_i \in \bm{A}_t} \lVert p_i - p^{'}_i \rVert_2),
\end{split}$$ Similarly, let the input point set be $\bm{P}$ and the reconstructed point set be $\bm{P}^{'}$. The global reconstruction error can be denoted by $$\begin{split}
L_{global} = d_{CH}(\bm{P}, \bm{P}^{'}) = \frac{1}{|\bm{P}|}\sum_{p_i \in \bm{P}} \min_{p^{'}_i \in \bm{P}^{'}} \lVert p_i - p^{'}_i \rVert_2 \\
+ \frac{1}{|\bm{P}^{'}|}\sum_{p^{'}_i \in \bm{P}^{'}} \min_{p_i \in \bm{P}} \lVert p_i - p^{'}_i \rVert_2.
\end{split}$$ Altogether, the network is trained end-to-end by minimizing the following joint loss function $$L = L_{local} + \gamma L_{global},$$ where $\gamma$ ($\gamma=1$) is the proportion of two part errors.
Experiments
===========
In this section, we first investigate how some key parameters affect the performance of L2G-AE in the shape classification task on ModelNet10 [@wu20153d]. Then, an ablation study is done to show the effectiveness of each module in L2G-AE. Finally, we further evaluate the performances of L2G-AE in multiple applications including 3D shape classification, 3D shape retrieval and point cloud upsampling.
Network Configuration
---------------------
In L2G-AE, we first sample $M = 256$ points as the centroids of local regions by FPS. Then, around each centroid, a kNN searching algorithm selects $T=4$ scale areas with $[K_1=16, K_2=32, K_3=64, K_4=128]$ points inside each area. In the multi-level feature propagation process, we initialize the feature dimension $C = M/8 = 32$ and $D = 256$. The encoder learns a 1024-dimension global feature for the input point cloud through hierarchical feature extraction. Similarly, the decoder hierarchically reconstructs local scales and global point cloud. In the RNN decoding layer, we adopt LSTM as the default RNN cell with hidden state dimension $h = D = 256$. In the experiment, we train our network on a NVIDIA GTX 1080Ti GPU using ADAM optimizer with the initial learning rate of 0.0001 and batch size of 8. The learning rate is decreased by 0.3 for every 20 epochs.
Parameters
----------
All experiments on parameter comparison are evaluated under ModelNet10. ModelNet10 contains 4899 CAD models from 10 categories and is split into 3991 for training and 908 for testing. For each model, we adopt 1024 points which are uniformly sampled from mesh faces and are normalized into a unit ball before being fed into the network. During the training process, the loss function keeps decreasing and stabilizes around the 180th epoch. To acquire the accuracies on ModelNet10, we train a linear SVM from the global features obtained by the auto-encoder. Specifically, the OneVsRest strategy is adopted with the linearSVM function as the kernel.
We first explore the number of sampled points $M$ which determines the distribution of local regions inside point clouds. In the experiment, we keep the network settings as depicted in the network configuration and vary the number of sampled points $M$ from 128 to 320.
$M$ 128 192 256 320
--------- ------- ------- ----------- ------- --
Acc (%) 93.83 94.38 **95.37** 93.94
: The effects of the number of sampled points $M$ under ModelNet10.[]{data-label="table:sampled_points"}
The results are shown in Table \[table:sampled\_points\], where the instance accuracies on the benchmark of ModelNet10 have a tendency to rise first and then fall. This comparison implies that L2G-AE can effectively extract the contextual information in point clouds by multi-level feature propagation and $M=256$ is an optimal choice which can well cover input point clouds without excessive redundant.
![The reconstructed results with different sampled points, where the CD represents the Chamfer distance between ground-truth and the reconstructed point cloud.[]{data-label="fig:sample_re"}](sample.pdf){height="1.5cm" width="8cm"}
To learn the reconstructed results intuitively, Figure \[fig:sample\_re\] shows the reconstructed point clouds with different sampled points. According to Chamfer distances, L2G-AE can also reconstruct the input point cloud with the varying of sampled points.
With keeping the sampled points $M = 384$, we investigate the key parameter dimension $C$ inside the self-attention modules. To unify the parameter in self-attention module, we keep the same dimension $C$ in different semantic levels. We change the default $C = 32$ to 16 and 64, respectively. In Table \[table:c\], L2G-AE achieves the best performance when the feature dimension $C$ is 32.
$M$ 16 32 64
--------- ------- ----------- ------- --
Acc (%) 93.94 **95.37** 94.16
: The effects of the feature dimension $C$ of the self-attention module under ModelNet10.[]{data-label="table:c"}
Finally, we show the effects of feature dimension of local areas $D$ and the global feature $D_{global}$. The dimension is varied as shown in Table \[table:d\] and Table \[table:dglobal\]. Neither the biggest nor the smallest, L2G-AE gets better performances when $D$, $D_{global}$ are set to 256 and 1024 respectively. There is a trade-off between the network complexity and the expressive ability of our L2G-AE.
$D$ 128 256 512
--------- ------- ----------- ------- --
Acc (%) 93.72 **95.37** 93.28
: The effects of the local feature dimension $D$ on ModelNet10.[]{data-label="table:d"}
$D_{global}$ 512 1024 2048
-------------- ------- ----------- ------- --
Acc (%) 94.16 **95.37** 93.94
: The effects of the global feature dimension $D_{global}$ under ModelNet10.[]{data-label="table:dglobal"}
Ablation Study
--------------
To quantitatively evaluate the effect of the self-attention module, we show the performances of L2G-AE under four settings: with point level self-attention module only (PL), with area level self-attention module only (AL), with region level self-attention module only (RL), remove all self-attention modules (NSA) and with all self-attention modules (ASA). As shown in Table \[table:rsa\], the self-attention module is effective in learning highly discriminative representations of point clouds by capturing the correlation among feature vectors. The results with only one self-attention module outperform the results without any self-attention module. And we achieve the best performance when three self-attention modules work together. The performance of self-attentions is affected by the discriminative ability of features. At the area level, the features of areas in the same region are similar, since there are only four areas, which makes the self-attention at area level contribute the least among all three self-attentions. In contrast, at the point level and the region level, the features of points or regions change a lot, so these self-attentions contribute more. From our observation, the results of PL and RL are coincidentally equal in the experiments.
Metric PL AL RL NSA ASA
--------- ------- ------- ------- ------- -----------
Acc (%) 94.16 94.05 94.16 93.72 **95.37**
: The effects of the self-attention module on ModelNet10.[]{data-label="table:rsa"}
![The reconstruction results of L2G-AE with only the local loss and only the global loss.[]{data-label="fig:loss_compare"}](loss_compare.pdf){width="8cm" height="2cm"}
After exploring the self-attention module, we also discuss the contributions of the two loss functions $L_{local}$ and $L_{global}$. In Table \[table:loss\], the results with local loss only (Local), global loss only (Global) and two losses together (Local + Global) are listed. The local loss function is very important in capturing local patterns of point clouds. And the two loss functions together can further enhance the classification performances of our neural network. In addition, Figure \[fig:loss\_compare\] shows the reconstruction results of our L2G-AE with only local loss and only global loss, respectively. From the results of the reconstructed point clouds, L2G-AE can reconstruct the input point cloud with only part of the joint loss function. In particular, the local reconstructed result in Figure \[fig:loss\_compare\] is a dense point cloud.
Metric Local Global Local+Global
--------- ------- -------- --------------
Acc (%) 94.71 92.84 **95.37**
: The effects of the two loss functions $L_{local}$ and $L_{global}$ on ModelNet10.[]{data-label="table:loss"}
Methods Supervised MN40 MN10
------------------ ------------ ----------- ----------- --
PointNet Yes 89.20 -
PointNet++ Yes 90.70 -
ShapeContextNet Yes 90.00 -
Kd-Net Yes 91.80 94.00
KC-Net Yes 91.00 94.4
PointCNN Yes 92.20 -
DGCNN Yes 92.20 -
SO-Net Yes 90.90 94.1
Point2Sequence Yes 92.60 95.30
MAP-VAE No 90.15 94.82
LGAN No 85.70 95.30
LGAN(MN40) No 87.27 92.18
FoldingNet No 88.40 94.40
FoldingNet(MN40) No 84.36 91.85
Our No **90.64** **95.37**
: The comparison of classification accuracy (%) under ModelNet10 and ModelNet40.[]{data-label="table:compare"}
Classification
--------------
In this subsection, we evaluate the performance of L2G-AE under ModelNet10 and ModelNet40 benchmarks, where ModelNet40 contains 12, 311 CAD models which is split into 9, 843 for training and 2, 468 for testing. Table \[table:compare\] compares L2G-AE with state-of-the-art methods in the shape classification task on ModelNet10 and ModelNet40. The compared methods include PointNet [@qi2016pointnet], PointNet++ [@NIPS2017_7095], ShapeContextNet [@xie2018attentional], KD-Net [@klokov2017escape], KC-Net [@shen2018mining], PointCNN [@li2018pointcnn], DGCNN [@wang2018dynamic], SO-Net [@li2018so], Point2Sequence [@liu2019point2sequence], MAP-VAE [@Zhizhong2019mapvae], LGAN [@Achlioptas2017latent] and FoldingNet [@yang2018foldingnet].
L2G-AE significantly outperforms all the unsupervised competitors under ModelNet10 and ModelNet40, respectively. In particular, L2G-AE achieves accuracy $95.37\%$ which is even higher than other methods of supervision under ModelNet10. Although the results of LGAN [@Achlioptas2017latent] and FoldingNet [@yang2018foldingnet] also show good performances under ModelNet10 and ModelNet40. This is because these methods are trained under a version of ShapeNet55 that contains more than 57,000 3D shapes. However, this version of ShapeNet55 dataset is not avaiable for public download from the official website. Therefore, we train all these methods under ModelNet40 for the fair comparison.
Methods LGAN FoldingNet Our
--------- ------- ------------ -----------
Acc (%) 49.94 53.42 **67.81**
: The comparison of retrieval in terms of under ModelNet10.[]{data-label="table:retrieval"}
![The comparison of PR curves for retrieval under ModelNet10.[]{data-label="fig:pr"}](pr.pdf){height="3cm" width="4cm"}
$10^{-3}$ bathtub bed chair desk dresser monitor n.stand sofa table toilet
----------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- --
PU **1.01** **1.12** **0.82** **1.22** 1.55 **1.19** **1.77** **1.13** **0.69** **1.39**
EC 1.43 1.81 1.80 1.30 1.43 2.04 1.88 1.79 1.00 1.72
Our 1.74 1.46 1.58 2.08 **1.40** 1.61 1.86 1.67 1.86 2.10
Retrieval
---------
L2G-AE is further evaluated in the shape retrieval task under ModelNet10 and compared with some other unsupervised methods of learning on point clouds. The compared results include two state-of-the-art unsupervised methods for point clouds, i.e., LGAN [@Achlioptas2017latent] and FoldingNet [@yang2018foldingnet]. The target of shape retrieval is to obtain the relevant information of a inquiry from a collection. In these experiments, the 3D shapes in the test set are used as quires to retrieve the rest shapes in the same set, and mean Average Precision (mAP) is used as a metric.
As shown in Table \[table:retrieval\], our results outperform all the compared results under ModelNet10. It shows that L2G-AE can be effect in improving the performance of unsupervised shape retrieval on point clouds. Their PR curves under ModelNet10 are also compared in Figure \[fig:pr\] which intuitively shows the performances of these three methods.
Unsupervised Upsampling for Point Clouds
----------------------------------------
Benefit from the design of local to global reconstruction, it is competent for our L2G-AE to be applied in the unsupervised point cloud upsampling application. In the local reconstruction, a dense point cloud is obtained by reconstructing multiple local scales with overlapping. Therefore, it is convenient to produce the upsampling results by downsampling from the dense local reconstructed results using some unsupervised methods, such as random sampling or farthest point sampling. As far as we know, L2G-AE is the first method which performs point cloud upsampling with deep neural networks in an unsupervised manner. To evaluate the performance of L2G-AE, We compare our method on relatively sparse (625 points) inputs with state-of-the-art supervised point cloud upsampling methods, including PU-Net [@yu2018pu] and EC-Net [@yu2018ecnet]. The target of upsampling is to generate a dense point clouds with 10000 points. For PU-Net and EC-Net, the $16 \times$ results (10000 points) are obtained from inputs (625 points) in a supervised manner. Differently, L2G-AE first obtains the local reconstruction results and then downsamples them to 10000 points.
![Some upsampled results of L2G-AE.[]{data-label="fig:upsampled"}](upsampling.pdf){width="8cm" height="6cm"}
As shown in Table \[table:upsampling\], mean Chamfer Distance (mCD) is used as a metric for quantitative comparison with PU-Net (PU) and EC-Net (EC) under ModelNet10. Although the results of PU-Net and EC-Net are better than “Our” in some classes under ModelNet10, the most likely reason is that the ground-truth is not visible to L2G-AE in the training. In addition, the input point cloud with 625 points contains very limited information. Figure \[fig:upsampled\] shows some upsamled results of our L2G-AE.
![Some reconstructed examples of L2G-AE.[]{data-label="fig:reconstruct"}](reconstruct.pdf){height="7cm" width="7cm"}
![Some examples of the attention in the region level, where each subfigure represents a 3D object.[]{data-label="fig:attetnion"}](attention_rl.pdf){height="10cm" width="7cm"}
Visualization
-------------
In this section,we will show some important visualization results of L2G-AE. Firstly, some reconstructed point clouds by L2G-AE are listed with the ground-truths as shown in Figure \[fig:reconstruct\]. From the results, the reconstructed point clouds of L2G-AE are consistent with the ground-truths.
Then, some visualizations of the attention map inside self-attention modules are engaged to show the effect of attentions in the hierarchical feature abstraction. There are three self-attention modules in the encoder, and we first visualize the attention map inside the local region level. For intuitively understanding, we directly attach the attention values to the centroids of local regions and then show these centroids. By summing attention map by column in the region level, the attention value of each centroid is caculated. For example, a $256 \times 256$ attention map is translated to a 256-dimension attention vector, when the number of sampled centroids is 256. Then, both the size and the color of centroids are associated with the attention values. Therefore, the centroids with lighter colors and larger sizes indicate larger attention values. As depicted in Figure \[fig:attetnion\], we show some examples of the region level attention. Figure \[fig:attetnion\] shows that the self-attention in the region level tends to on special local regions at conspicuous locations such as edges, corners or protruding parts.
![Some examples of the attention in the scale level. The abscissa represents the 4 scales $[s_1,s_2,s_3,s_4]$ around each centroid in a point cloud and the ordinate indicates the index of 256 centroids, where each subfigure represents a 3D object.[]{data-label="fig:sl_attention"}](sl_attention.pdf){width="7cm" height="6cm"}
![Some examples of the attention in the point level, where the four subfigures in each row represent the four scales of a local region.[]{data-label="fig:pl_attention"}](pl_attention.pdf){width="7cm" height="6cm"}
Similarly, we also show some examples of the scale level attention in Figure \[fig:sl\_attention\] and the point level attention in Figure \[fig:pl\_attention\]. In Figure \[fig:sl\_attention\], each image shows the 4 scale attention values around 256 sampled centroids of a point cloud. And the color indicates the value of attention, where large attention value corresponds to a bright color such as yellow. The results indicate that the network tends to focus on the $4^{th}$ scale which contains more information of local structures. In Figure \[fig:pl\_attention\], each row represents the 4 scale areas around a centroid. In different scale areas, the network concern on different points inside the areas to capture the local patterns in the local region.
Conclusions
===========
In this paper, we propose a novel local to global Auto-encoder framework for point cloud understanding in the shape classification, retrieval and point cloud upsampling tasks. In the encoder, a self-attention mechanism is employed to explore the correlation among features in the same level. In contrast, an interpolation layer and RNN decoding layer successfully reconstruct local scales and global point clouds hierarchically. Experimental results show that our method achieves competitive performances with state-of-the-art methods.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Anonymous Author(s)'
- |
[Carl Yang$^*$, Jieyu Zhang$^*$, Jiawei Han]{}\
*University of Illinois, Urbana Champaign, 201 N Goodwin Ave, Urbana, Illinois 61801, USA\
{jiyang3, jieyuz2, hanj}@illinois.edu*
bibliography:
- 'carlyang.bib'
title: Neural Embedding Propagation on Heterogeneous Networks
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An exact arithmetic, memory efficient direct solution method for finite element method (FEM) computations is outlined. Unlike conventional black-box or low-rank direct solvers that are opaque to the underlying physical problem, the proposed method leverages physical insights at every stage of the development through a new symmetric domain decomposition method (DDM) with one set of Lagrange multipliers. Comparisons with state-of-the-art exact direct solvers on electrically large problems suggest up to 10 times less memory and better run-time complexity while maintaining the same accuracy.'
author:
-
title: 'Direct Domain Decomposition Method (D$^3$M) for Finite Element Electromagnetic Computations'
---
at (current page.south) ;
This work has been presented at the 2016 International Conference on Electromagnetics in Advanced Applications. The parallel version of this work has been presented at the 2019 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting \[\[ref:moshfegh2019\]\]
Introduction
============
Modern computational electromagnetics FEM tools often rely on memory efficient iterative solvers such as Multigrid \[\[ref:zhu2006\]\] and DDM \[\[ref:toselli2005\]\] that may experience convergence difficulties near resonances or multi-scale problems, and lose efficiency at multiple excitation runs. Contrary, direct solvers such as MUMPS \[\[ref:amestoy2001\]\] or PARDISO are reliable but scale unfavorably and are hard to parallelize. Thus, recent trends in direct solvers \[\[ref:ying2014\]\] strive to reduce workload by leveraging low-rank approximations at the cost of accuracy and possibly reliability. Yet, these solvers are opaque to important underlying physics, leaving room for further improvements.
To achieve an efficient $exact$ direct solver, one must start from scratch and attempt to leverage deep physical and numerical insights that may require re-formulating the BVP and FEM assembling, in addition to the symbolic and numeric factorization stages. This is critical not only to avoid internal resonances at all intermediate factorization separators, but also to produce numerically efficient matrix structures i.e. reduced size, block-wise sparse symmetric matrices.
This work achieves all these via a direct DDM (D$^3$M) framework. A set of auxiliary variables is used to cast a decomposed BVP that, after an initial reduction/elimination step, leads to an auxiliary blocked matrix that is suitable for factorization. To attain maximal performance, this matrix is factored with a special blocked LDL$^T$ method with restricted Bunch-Kaufman pivoting \[\[ref:bunch1977\]\].
The accuracy and performance of the proposed D$^3$M has been verified and tested in 3D scattering problems by perfect electric conductor (PEC) plates and dielectric spheres of progressively larger electrical sizes. The proposed D$^3$M solver requires $3\!\times\! - \!10\times$ less memory than MUMPS mainly due to the choice of structured separators and the absence of delayed pivots attributed to the interior resonance free formulation. An initial serial implementation of D$^3$M was up to two times slower than MUMPS for small problems but becomes competitive on problems larger than one million unknowns.
Theory
======
Consider a computational domain $\Omega$ decomposed into $N$ non-overlapping subdomain $\Omega_i$. For example, a decomposed problem with four domains is shown in Fig. \[fig:2Ddecomp\]. The decomposed BVP reads, $find \left({\bf E}_{*}, \pmb{\lambda} \right) \in \left\{{\bf V}_{*}, \pmb{\Lambda}\right\} such \mbox{ } that$: $$\nonumber
\left\{
\begin{array}{ll}
\nabla\times\frac{1}{\mu_r}\nabla\times {\bf E}_{*} -k^2\epsilon_r {\bf E}_{*} = -j
k\eta{\bf J}^{imp}, \hspace{15pt} \mbox{in } \, \Omega = \bigcup \Omega_i\\
\gamma_{\times}\!\!\left(\! \nabla\!\!\times\!\!{\bf E} \right) \!- \!jk \gamma_{t}\left( {\bf E}\! \right) \! =\!\gamma_{\times}\!\left( \nabla\!\times\!{\bf E}^{inc} \right) \!- \!jk \gamma_{t}\left( {\bf E}^{inc}\! \right)\!, \hspace{1pt} \! \mbox{on } \, \partial\Omega \\
{\bf R}^{(i)}_{ij}\left[\mathbf{j} - \alpha \mathbf{e}\right]-{\bf R}^{(\mathcal{N}(i))}_{ij}\left[-\mathbf{j} - \alpha \mathbf{e}\right] = {\bf 0}, \! \hspace{15pt}\! \mbox{on } \, {\cal S}, \forall i \,\&\, \{ij\} \\
{\bf R}^{(i)}_{ij}{\bf e}-{\bf R}^{(\mathcal{N}(i))}_{ij}{\bf e} = {\bf 0}, \hspace{80pt}\! \mbox{on } \, {\cal S}, \forall i\,\&\, \{ij\}
\end{array}
\right.$$ where $\mathbf{e} = \gamma_t(\mathbf{E}|_\mathcal{S})$, and $\mathbf{j} = \gamma_\times(\nabla \!\times\! \mathbf{E}|_\mathcal{S})$. ${\bf R}^{(i)}_{ij}$ is the restriction operator from domain $i$ to interface $ij$, and $\mathcal{N}(i)$ denotes the neighbor of domain $i$.\
![2D decomposition problem with four domains.[]{data-label="fig:2Ddecomp"}](./generic-model-4domains-lambda.pdf){width="2.5in"}
After transforming to $\pmb{\lambda}=\bf{j} \pm \mathbf{e}$, casting the variational problem and expanding trial and testing function spaces gives
$$\!\!\left[\!\! \begin{array}{cc} \!\!\mathbf{A}\!\! & \!\!\mathbf{D}\!\! \\
\!\!\mathbf{D}^{T}\!\! & \!\!\mathbf{0}\!\! \\
\end{array}
\!\!\right]\!\!\!\left[\!\!
\begin{array}{c}
\!\mathbf{E}\!\!\!\\
\!\pmb{\lambda}\!\!\!\\
\end{array}
\!\!\right]\!\!=\!\!\left[\!\!\!
\begin{array}{c}
{\bf f}\\
{\bf 0}\\
\end{array}
\!\!\!\right]
\!{\ext@arrow 0359\Rightarrowfill@{}{Elim.\mathbf{E}}}\!
\!\underbrace{\left(\sum\limits_{i=1}^{N_I} \mathbf{D}^T_i (\mathbf{A}_i)^{-1} \mathbf{D}_i\!\right)}_\text{\bf{K}} \! \pmb{\lambda} = \mathbf{g},$$ where $\mathbf{A}$ is the diagonal blocked matrix of $\mathbf{A}_i, i = 1,2,...,N$ where $\mathbf{A}_i$ is the FEM-ABC matrix for domain $i$ with loss or gain at the interfaces. $\mathbf{D}$ is a blocked matrix of sparse matrices $\mathbf{D}_i, i = 1,2,...,N_I$ which maps the primal space to LM space. $N_I$ is the number of interfaces.
The reduced matrix ${\bf K}$ is symmetric block-wise sparse but indefinite. Hence LDL$^T$ factorization with symmetric partial pivoting a.k.a Bunch-Kaufmann LDL$^T$ \[\[ref:bunch1977\]\] can be used to save memory and CPU time. Since ${\bf K}$ is a block-wise sparse matrix, we have modified the Bunch-Kaufman LDL$\!^T$ factorization to its block restricted partial pivoted form. Each block in ${\bf K}$ corresponds to a super-node of typical order $n > 300$. Therefore, D$^3$M consistently operates at the maximum performance region of Level 3 BLAS. The main steps of the proposed D$^3$M are:
1. Generate dense domain matrices $\!\mathbf{K}_D^{(d)}\!$,
2. Assemble the block-wise sparse reduced matrix ${\bf K}$,
3. Reorder the clique graph of ${\bf K}$,
4. Symbolic factorize the reordered clique graph,
5. Factorize ${\bf K}$ with restricted B-K pivoting block LDL$\!^T$ factorization (see algorithm 1),
6. Forward/Backward substitute the reduced system ${\bf K}$ for auxiliary unknowns,
7. Recover primal unknowns.
The clique graph of blocked sparse matrix ${\bf K}$ is reordered using METIS (same as MUMPS). Assuming that the clique graph has $l$ levels, the algorithm of block LDL$^T$ factorization is given in Algorithm 1. A multifrontal version of block LDL$^T$ can be used to further speed computations.
Dense LDL$^T$ factorize ${\bf K}_{jj}$ Dense Triangular solve $\mathbf{L}_{jj} \mathbf{X}_{ij}^T \!\!=\!\! \mathbf{K}_{ij}^T,$ for $\mathbf{X}_{ij} \!\!=\!\! \mathbf{P}_{ii} \mathbf{L}_{ij} \mathbf{D}_{jj}$ Compute work variable $\mathbf{W}_{ij} =\mathbf{X}_{ij} \mathbf{D}_{ii}^{-1}$ Dense update $\mathbf{K}_{ik} \leftarrow \mathbf{A}_{ik} \!-\! \mathbf{X}_{ij}\mathbf{W}_{ij}^T$
Numerical Results
=================
First, the scattering of progressively larger PEC plates (from $3\lambda \!\times\! 3\lambda$ up to $19\lambda \!\times\! 19\lambda$) are considered. Computational complexity of factorization time and memory for these problems using the proposed D$^3$M and MUMPS are shown in Fig. \[fig:pec\_plate\]. It is noted that a 3M unknown problem is solved with only 10 GB of RAM at full double precision accuracy.
Next, the scattering of progressively larger dielectric spheres is considered. Computational complexity of factorization time and memory for these problems using the proposed D$^3$M and MUMPS are shown in Fig. \[fig:sphere\]. Again, the proposed method uses more than 2.25 times less memory and surprisingly better time complexity than MUMPS. The relative residual error ($\|{\bf A} \bar{x} - f \|_{inf} / \|f\|_{inf}$) of all runs using the proposed D$^3$M is around $10^{-12}\!\!-\!\!10^{-13}$ which was the same as MUMPS.
[![Computational complexity for progressively larger PEC plates using the proposed D$^3$M and MUMPS: Factorization time (top); Factorization memory (bottom).[]{data-label="fig:pec_plate"}](./PEC_Plate.pdf "fig:"){width="2.5in"}]{}
[![Computational complexity for progressively larger dielectric spheres using the proposed D$^3$M and MUMPS: Factorization time (top); Factorization memory (bottom).[]{data-label="fig:sphere"}](./Sphere.pdf "fig:"){width="2.5in"}]{}
[1]{}
J. Moshfegh, D. G. Makris, and M. N. Vouvakis,“Parallel Direct Domain Decomposition Methods (D$^3$M) for Finite Elements." *2019 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting*, pp. 777–778. \[ref:moshfegh2019\]
Y. Zhu and A. C. Cangellaris, Multigrid finite element methods for electromagnetic field modeling. *John Wiley & Sons*, Vol. 28, 2006. \[ref:zhu2006\]
A. Toselli and O. Widlund, *Domain decomposition methods: algorithms and theory*, Vol. 3. Berlin: Springer, 2005. \[ref:toselli2005\]
P. R. Amestoy, I. S. Duff, J. Y. L’Excellent, and J. Koster, “A fully asynchronous multifrontal solver using distributed dynamic scheduling," *SIAM Journal on Matrix Analysis and Applications* Vol. 23, No. 1, pp.15-41, 2001. \[ref:amestoy2001\]
P. G. Schmitz, and L. Ying, “A fast nested dissection solver for Cartesian 3D elliptic problems using hierarchical matrices," *Journal of Computational Physics* Vol. 258, pp.227-245, 2014. \[ref:ying2014\]
J. R. Bunch, L. Kaufman, “Some stable methods for calculating inertia and solving symmetric linear systems," *Math. Comp.*, Vol. 31, No. 137, pp. 163–179, 1977. \[ref:bunch1977\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We introduce the idea of Assur graphs, a concept originally developed and exclusively employed in the literature of the kinematics community. The paper translates the terminology, questions, methods and conjectures from the kinematics terminology for one degree of freedom linkages to the terminology of Assur graphs as graphs with special properties in rigidity theory. Exploiting recent works in combinatorial rigidity theory we provide mathematical characterizations of these graphs derived from ‘minimal’ linkages. With these characterizations, we confirm a series of conjectures posed by Offer Shai, and offer techniques and algorithms to be exploited further in future work.'
author:
- 'Brigitte Servatius[^1]'
- 'Offer Shai[^2]'
- 'Walter Whiteley [^3]'
bibliography:
- 'assurcombArX.bib'
title: Combinatorial Characterization of the Assur Graphs from Engineering
---
Introduction
============
Working in the theory of mechanical linkages, the concept of ‘Assur groups’ was developed by Leonid Assur (1878-1920), a professor at the Saint-Petersburg Polytechnical Institute. In 1914 he published a treatise (reprinted in [@Assur]) entitled [*Investigation of plane bar mechanisms with lower pairs from the viewpoint of their structure and classification*]{}. In the kinematics literature it is common to introduce ‘Assur groups’ (selected groups of links) as special minimal structures of links and joints with zero mobility, from which it is not possible to obtain a simpler substructure of the same mobility [@PK]. Initially Assur’s paper did not receive much attention, but in 1930 the well known kinematician I.I. Artobolevskiĭ, a member of the Russian academy of sciences, adopted Assur’s approach and employed it in his widely used book [@Arto]. From that time on Assur groups are widely employed in Russia and other eastern European countries, while their use in the west is not as common. However, from time to time Assur groups are reported in research papers for diverse applications such as: position analysis of mechanisms [@Mitsi]; finding dead-center positions of planar linkages [@PK] and others.
The mechanical engineering terminology for linkages (kinematics) and their standard counting techniques are introduced via an example in the next section. Central to Assur’s method is the decomposition of complex linkages into fundamental, minimal pieces whose analyses could then be merged to give an overall analysis. Many of these approaches for Assur groups were developed from a range of examples, analyzed geometrically and combinatorially, but never defined with mathematical rigor.
In parallel, rigidity of bar and joint structures as well as motions of related mechanisms have been studied for several centuries by structural engineers and mathematicians. Recently (since 1970) a focused development of a mathematical theory using combinatorial tools was successful in many applications. For example for planar graphs there is a simple geometric duality theory, which, if applied to mechanisms and frameworks yields a relation between statics and kinematics: any locked planar mechanism is dual to an unstable planar isostatic framework (determinate truss) [@Shai; @CW2].
The purpose of this paper is twofold. First, we want to draw together the vocabulary and questions of mechanical engineering with the rigidity theory terminologies of engineering and mathematics. Second, we want to apply the mathematical tools of rigidity theory, including the connections between statics and kinematics, to give precision and new insights into the decomposition and analysis of mechanical linkages.
The mathematical tools we need are briefly sketched with references provided in §2.3-2.7. Our main result is the description of Assur graphs (our term for Assur groups) in Engineering terms (§2.1,2.2) and its reformulation in mathematical terms. We show that our mathematical reformulation allows us in a natural way to embed Assur’s techniques in the theory of frameworks (§3) and bring the results back to linkages. In the process we veriy several conjectured characterizations presented by Offer Shai in his talk concerning the generation of Assur graphs and the decomposition of linkages into Assur graphs, at the 2006 Vienna Workshop on Rigidity and Flexibility §3.1. We also give algorithmic processes for decomposing general linkages into Assur graphs, as well as for generating all Assur graphs (§3.2-3.3).
In a second paper [@SSW2], we will apply the geometric theory of bar-and-joint framework rigidity in the plane to explore additional properties and characterizations of Assur graphs. This exploration includes singular (stressed) positions of the frameworks, explored using reciprocal diagrams, and the introduction of ‘drivers’, which appear in passing in the initial example in the next section.
Preliminaries
=============
In the first two sub-sections we present the mechanical engineering vocabulary, problems and approaches through an example. These offer the background and the motivation for the concepts of the paper, but do not yet give the formal mathematical definitions. In the remaining five sub-sections we give the framework basics needed to mathematically describe these approaches.
Linkages and Assur graphs
--------------------------
A linkage is a mechanism consisting of rigid bodies, the [*links*]{}, held together by joints. Since we only consider linkages in the plane, all our joints are pin joints, or pins. A complex linkage may be efficiently studied by decomposing it into simple pieces, the Assur groups. (Engineers use the term group to mean a specified set of links. In mathematics the word group is used for an algebraic structure, but most of the tools we will use come from graph theory, so the word graph seems more natural and we will use it as soon as we start our mathematical section). We introduce these ideas via an example.
Figure \[tracfig01a\] depicts an excavator attached with a linkage system. In the following, we illustrate how the schematic drawing of this system is constructed and how it is decomposed into Assur groups.
![The excavator with its kinematic system.\[tracfig01a\]](tracfig01a)
In order to get a uniform scheme, termed [*structural scheme*]{}, it is common to represent all the connections between the links by revolute joints as appears in Figure \[tracfig02b\]. Here joints $0_1$ and $0_3$ attach the excavator to the vehicle (fixed ground) and these special joints are marked with a small hatched triangle, and are called [*pinned joints*]{}. All other joints are called [*inner*]{} joints. A link which can be altered (e.g. by changing its length) is called a [*driving link*]{}. A driving link can be thought of as driving or changing the distance between its endpoints like the pistons in our excavator example, which may be modeled in the structural scheme by a rotation of an inserted link 1.
![The unified structural scheme of the kinematic system of the excavator.\[tracfig02b\]](tracfig02b "fig:")![The unified structural scheme of the kinematic system of the excavator.\[tracfig02b\]](tracfig03 "fig:")
Once the engineering system is represented in the structural scheme, to start the analysis, all the driving links are deleted and replaced by pinned joints (mathematically speaking the driving links are contracted and their endpoints identified). In the current example, links 1 and 4 are deleted and joints A and D are pinned. Then, the system is decomposed into three Assur groups, each consisting of two links, one inner joint and two pinned joints. In the literature the Assur groups of this type are referred to as dyads [@Norton]. The order of the decomposition is important. If an inner joint of a group, $G_1$, becomes a pinned joint in group $G_2$, then $G_1$ should precede $G_2$.
In our example (see Figure \[tracfig04\]), the unique order of decomposition is: Dyad$_{1} = \{2,3\}$; Dyad$_{2}=\{5,6\}$; Dyad$_{3}=\{7,8\}$.
![Assur group decomposition of the structural scheme of the kinematic system of the excavator.\[tracfig04\]](tracfig04)
Degree of freedom of a mechanism: Grübler’s equation
----------------------------------------------------
The degree of freedom (DOF) of a linkage is the number of independent coordinates or measurements required to define its position.
In mechanical engineering [@Norton], Grübler’s equation relates the (least number of internal) degrees of freedom $F$ of a linkage mechanism to the number $L$ of links and the number $J$ of joints in the mechanism. In the plane, if $J_i$ is the number of joints from which $i$ links emanate, $i\geq 2$ then $$\label{Grubler1}
F =3(L-1)-2\sum (i-1)J_i$$
In the example above $L=9$ because the fixed ground is considered a link, and $\sum (i-1)J_i= 11$, because the revolute joint $E$ is counted twice as it pins links 5, 6, and 7. By Grübler’s equation we get $F=3(9-1)-2\cdot11=2$, which is correct as there are two driving links (two distances being controled). If these two driving links are removed and their ends pinned (identified) as in the example analysis, we have only 7 links left and the number of revolute joints is now 9, so $F=0$. This is another indication that the drivers work independently.
Note that Grübler’s equation only gives a lower bound on the degree of freedom and there are many cases where the actual DOF is larger than the predicted one [@Norton]. If a linkage contains a sub-collection of links pinned in such a way among themselves that the Grübler count is negative for the sub-collection, then the predicted DOF for the linkage might be smaller than the actual DOF (see Figure \[miscount\](a)). This situation can be detected and corrected by combinatorial means as we will describe in Laman’s Theorem, see §2.5 Theorem \[OverviewTheorem\].
Counting techniques, however, cannot detect special geometries, e.g. parallelism or symmetry of links, which also might lead to a false Grübler DOF prediction. We will examine this type of geometric singularity of a combinatorially correct graph in [@SSW2].
![In some cases Grübler’s equation provides a false answer (a), due to an overcounted subgraph. In others (b) it correctly predicts a pinned isostatic framework (determinate truss). \[miscount\]](henny03)
Frameworks
----------
For a linkage in which all the links are bars, with revolute joints at the two endpoints of the bar, we can rewrite Grübler’s equation in terms of graph theory, by introducing a graph whose vertices, $V$, are the joints and whose edges, $E$, are the bars. With $V_i$ denoting the set of vertices of valence $i$, Grübler’s equation becomes $$F=3(|E|-1)-2\sum (i-1)J_i=3(|E|-1)-2\sum i|V_i| +2\sum |V_i|$$ $$=
3|E|-3-4|E|+2|V|=2|V|-3 -|E|.$$ So if the edges of a graph embedded in the plane are interpreted as rigid bars and the vertices as revolute joints, the graph needs to have at least $2|V|-3$ edges in order to have no internal degrees of freedom, an observation made already by Maxwell. The count $2|V|-3$ will be central to the rest of the paper.
By a [*framework*]{} we mean a graph $G = (V,E)$ together with a configuration ${\mathbf{p}}$ of $V$ into Euclidean space, for our purpose the Euclidean plane. We will always assume that the two ends of an edge (a bar) are distinct points). A motion of the framework is a displacement of the vertices which preserves the distance between adjacent vertices, and a framework is [*rigid*]{} if the only motions which it admits arise from congruences.
Let us assume that the location ${\mathbf{p}}_{i}$ of a vertex is a continuous function of time, so that we can differentiate with respect to time. If we consider the initial velocities, ${\mathbf{p}}'_{i}$, of the endpoints ${\mathbf{p}}_{i}$ of a single edge $(i,j)$ under a continuous motion of a framework, then, to avoid compressing or extending the edge, it must be true that the components of those velocities in the direction parallel to the edge are equal, i.e. $$\label{InfEqn}
({\mathbf{p}}_i - {\mathbf{p}}_j)\cdot({\mathbf{p}}'_i - {\mathbf{p}}'_j) = 0.$$ A function assigning vectors to each vertex of the framework such that equation \[InfEqn\] is satisfied at each edge is called an [*[first-order ]{}motion.*]{} If the only [first-order ]{}motions are [*trivial*]{}, that is, they arise from [first-order ]{}translations or [first-order ]{}rotations of ${\mathbb{R}}^2$, then we say that the framework is [first-order ]{}rigid in the plane. [First-order ]{}rigidity implies rigidity, see for example [@Connelly].
In our excavator example not all links are bars. In the structural scheme link 8 is modeled by a bar because it contains only two pins, while link 3, which contains 5 pins appears as a “body". We can replace such a body by a rigid subframework on these 5 vertices (or more). In general, any linkage consisting of rigid bodies held together by pin joints can be modeled as a framework by replacing the bodies with rigid frameworks.
The rigidity matrix
-------------------
Any graph $G$ can be considered a subgraph of the complete graph $K_n$ on the vertex set $V = \{1,\ldots, n\}$, where $n$ is large enough. Let ${\mathbf{p}}$ be a fixed [*configuration*]{} (embedding) of $V$ into ${\mathbb{R}}^2$.
Equation \[InfEqn\] defines a system of linear equations, indexed by the edges $(i,j)$, in the variables for the unknown velocities ${\mathbf{p}}'_i$. The matrix $R({\mathbf{p}})$ of this system is a real $n(n-1)/2$ by $2n$ matrix and is called the [*rigidity matrix*]{}. As an example, we write out coordinates of ${\mathbf{p}}$ and of the rigidity matrix $R({\mathbf{p}})$, in the case $n=4$. $${\mathbf{p}} = ({\mathbf{p}}_{1},{\mathbf{p}}_{2},{\mathbf{p}}_{3},{\mathbf{p}}_{4}) =
(p_{11},p_{12},p_{21},p_{22},p_{31},p_{32},p_{41},p_{42});$$ $$\left[\begin{array}{cccccccc}
_{p_{11}-p_{21}} & _{p_{12}-p_{22}} & _{p_{21}-p_{11}} & _{p_{22}-p_{12}} &
_{0} & _{0} & _{0} & _{0} \\
_{p_{11}-p_{31}} & _{p_{12}-p_{32}} & _{0} & _{0} &
_{p_{31}-p_{11}} & _{p_{32}-p_{12}} & _{0} & _{0} \\
_{p_{11}-p_{41}} & _{p_{12}-p_{42}} & _{0} & _{0} &
_{0} & _{0} & _{p_{41}-p_{11}} & _{p_{42}-p_{12}} \\
_{0} & _{0} & _{p_{21}-p_{31}} & _{p_{22}-p_{32}} &
_{p_{31}-p_{21}} & _{p_{32}-p_{22}} & _{0} & _{0} \\
_{0} & _{0} & _{p_{21}-p_{41}} & _{p_{22}-p_{42}} &
_{0} & _{0} & _{p_{41}-p_{21}} & _{p_{42}-p_{22}} \\
_{0} & _{0} & _{0} & _{0} &
_{p_{31}-p_{41}} & _{p_{32}-p_{42}} & _{p_{41}-p_{31}} & _{p_{42}-p_{32}}
\end{array}\right]$$
A framework $(V,E,{\mathbf{p}})$ is infinitesimally rigid (in dimension $2$) if and only if the submatrix of $R({\mathbf{p}})$ consisting of the rows corresponding to $E$ has rank $2n-3$. We say that the vertex set $V$ is in generic position if the determinant of any submatrix of $R({\mathbf{p}})$ is zero only if it is identically equal to zero in the variables ${\mathbf{p}}'_i$. For a generically embedded vertex set, linear dependence of the rows of $R({\mathbf{p}})$ is determined by the graph induced by the edge set under consideration. The rigidity properties of a graph are the same for any generic embedding. A graph $G$ on $n$ vertices is [*generically rigid*]{} if the rank $\rho$ of its rigidity matrix $R_G ({\mathbf{p}})$ is $2n-3$, where $R_G ({\mathbf{p}})$ is the submatrix of $R({\mathbf{p}})$ containing all rows corresponding to the edges of $G$, for a generic embedding of $V$. The (generic) DOF of $G$ is defined to be $ 2n-3-\rho$.
Results for the plane
---------------------
Linear dependence of the rows of the rigidity matrix defines a matroid on the set of rows and for generic configurations we speak about [*independent edge sets*]{} rather than independent rows of $R({\mathbf{p}})$. For a generic embedding of $n$ vertices into ${\mathbb{R}}^2$ we call the matroid on the complete graph obtained from $R({\mathbf{p}})$ the generic rigidity matroid in dimension $2$ on $n$ vertices, $\mathfrak{R}_2(n)$.
The following theorem characterizes $\mathfrak{R}_2(n)$.
\[OverviewTheorem\]
[*(Laman [@Laman])*]{} The independent sets of $\mathfrak{R}_2(n)$ are those sets of edges which satisfy Laman’s condition: $$\label{LamansIneq}
|F| \leq 2|V(F)| - 3
\mbox{~for all $F \subseteq E, F \not = \emptyset $};$$
Laman’s Theorem was proved in 1970 and it was this theorem that promoted the use of matroids to attack rigidity questions. There are many equivalent axiom systems known for matroids. These can be used to reveal structural properties of various types and their relationships. The fact that matroids are exactly those structures for which independent sets can be constructed greedily has important algorithmic conseqences.
From the count condition in the inequalities (\[LamansIneq\]) for independent edge sets it is straight forward to deduce count conditions for [*bases*]{} of $\mathfrak{R}_2(n)$ (edge sets inducing minimally rigid or isostatic graphs), as well as for minimally dependent sets, or [*circuits*]{}, which will play a fundamental role in our analysis and will be treated in the next section.
Independent sets of $\mathfrak{R}_2(n)$ may be constructed inductively [@TW]. Given an independent edge set $E$ in $\mathfrak{R}_2(n)$, we can extend $E$ by new edges provided that the inequalities \[LamansIneq\] are not violated. Starting with an independent set (e.g. a single edge):
1. We can attach a new vertex $v$ by two new edges $x=(v,u)$ and $y=(v,w)$ to the subgraph of $G$ induced by $E$ and $E\cup\{x,y\}$ is independent, see Figure \[extendfig01\]a. This is also called [*$2$-valent vertex addition*]{}.
2. Similarly, we can attach a new vertex $v$ by three new edges to the endpoints of an edge $e \in E$ plus any other vertex in the subgraph of $G$ induced by $E$, and ${E\setminus e}\cup\{x,y,z\}$ is independent, see Figure \[extendfig01\]b. This operation is called [*edge-split*]{}, because the new vertex $v$ is thought of as splitting the edge $e$.
![Building up an independent set of edges by: $2$-valent vertex (a) and edge-split (b)\[extendfig01\]](extendfig01)
These [*Henneberg techniques*]{} developed in [@TW] have become standard in rigidity theory, see also [@GSS], and when we resort to “the usual arguments” within some of the proofs to come, we have these standard proof techniques in mind. For further reference we state the following well known result.
Any independent set in $\mathfrak{R}_2(n)$ can be obtained from a single edge by a sequence of $2$-valent vertex additions and edge-splits.\[Henneberg\]
Figure \[henny01\] illustrates a sequence described in Theorem \[Henneberg\].
![A Henneberg sequence constructing an isostatic graph.\[henny01\]](henny02)
Rigidity circuits
-----------------
Minimally dependent sets, or circuits, in $\mathfrak{R}_2(n)$ are edge sets $C$ satisfying $|C|=2|V(C)|-2$ and every proper non-empty subset of $C$ satisfies inequality (\[LamansIneq\]). Note that these circuits, called [*rigidity circuits*]{}, always have an even number of edges. We will, as is commonly done, not distinguish between edge sets and the graphs they induce.
Similarly to the inductive constructions of independent sets (see Figure \[extendfig01\]), all circuits in $\mathfrak{R}_2(n)$ can be constructed from a tetrahedron (the complete graph on four vertices) by two simple operations, see [@Berg-Jordan], namely [*edge-split* ]{} as in Figure \[extendfig01\]b, and [*2-sum*]{}, where the 2-sum of two (disjoint) graphs is obtained by “gluing” the graphs along an edge and removing the glued edge, see Figure \[twosum\].
![2-sums taken along the lined up edge pairs combine circuits into a larger circuit. \[twosum\]](twosum)
Any circuit in $\mathfrak{R}_2(n)$ can be obtained from $K_4$ by a sequence edge-splits and 2-sums.\[BJ\]
Isostatic Pinned Framework
--------------------------
Given a framework associated with a linkage, we are interested in its internal motions, not the trivial ones, so following the mechanical engineers we pin the framework by prescribing, for example, the coordinates of the endpoints of an edge, or in general by fixing the position of the vertices of some rigid subgraph, see Figure \[pinsub01\]. We call these vertices with fixed positions [*pinned*]{}, the others [*inner*]{}. (Inner vertices are sometimes called [*free*]{} or [*unpinned*]{} in the literature.) Edges among pinned vertices are irrelevant to the analysis of a pinned framework. We will denote a pinned graph by $G(I,P;E)$, where $I$ is the set of inner vertices, $P$ is the set of pinned vertices, and $E$ is the set of edges, where each edge has at least one endpoint in $I$.
A pinned graph $G(I,P;E)$ is said to satisfy the [*Pinned Framework Conditions*]{} if $|E|=2|I|$ and for all subgraphs $G'(I', P'; E')$ the following conditions hold:
1. $|E'|\leq 2|I'|$ if $|P'| \geq 2$,
2. $|E'|\leq 2|I'| -1$ if $|P'|=1$ , and
3. $|E'|\leq 2|I'|-3$ if $P' = \emptyset$.
We call a pinned graph $G(I,P;E)$ [*pinned isostatic*]{} if $E=2|I|$ and $\widetilde{G}= G\cup K_{P}$ is rigid as an unpinned graph, where $K_{P}$ is a complete graph on a vertex set containing all pins (but no inner vertices). In other words, we “replace” the pinned vertex set by a complete graph containing the pins and call $G(I,P;E)$ isostatic, if choosing any basis in that replacement produces an (unpinned) isostatic graph.
A pinned graph $G(I,P;E)$ realized in the plane, with ${\mathbf{P}}$ for the pins, and ${\mathbf{p}}$ for all the vertices, is a [*pinned framework*]{}. A pinned framework is [*rigid*]{} if the matrix $R_{\widetilde{G}}$ has rank $2|I|$, where $R_{\widetilde{G}}$ is the rigidity matrix of $\widetilde{G}$ with the columns corresponding to the vertex set of $K_p$ removed, [*independent*]{} if the rows of $R_{\widetilde{G}}$ corresponding to $E$ are independent, and [*isostatic*]{}, if it is rigid and independent. The vertices $I$ of a pinned framework are in [*generic position*]{} if any submatrix of the rigidity matrix is zero only if it is identically equal to zero with the coordinates of the inner vertices as variables. The coordinates of the pins are prescribed constants.
Figure \[pinsub01\] shows an example of a pinned isostatic $G$ and a corresponding basis ${\widehat{G}}$ of $\widetilde{G}$.
![Framework (a) is pinned isostatic because Framework (b) is isostatic.\[pinsub01\]](pinsub01)
It is common in engineering to choose pins in advance and their placement ${\mathbf{P}}$ might not be generic, in fact not even in [*general position*]{} (no three points collinear), as it is sometimes necessary to have all pins on a line. The following result shows that this is not a problem.
\[PinThm\] Given a pinned graph $G(I,P;E)$, the following are equivalent:
\(i) There exists an isostatic realization of $G$.
\(ii) The Pinned Framework Conditions are satisfied.
\(iii) For all placements ${\mathbf{P}}$ of $P$ with at least two distinct locations and all generic positions of $I$ the resulting pinned framework is isostatic.
Let $\widetilde{G}= G\cup K_{P'}$ $P'\supseteqq P$, and $F$ a maximal independent edge set of $K_{P'}$. Then by Theorem \[OverviewTheorem\] we deduce that $E\cup F$ is isostatic if and only if the Pinned Framework Conditions are satisfied, so $(i)\Leftrightarrow (ii)$
In order to show $(ii)\Rightarrow (iii)$ we first show that we can extend $F$ to $F\cup E$ by a Henneberg sequence of 2-valent vertex additions and edge-splits (see Figure \[extendfig01\]). To this end we de-construct $\widetilde{G}$ first by removing inner vertices as follows. Assume that $|I|>1$. Since $|E|=2|I|$ and $I$ spans at most $2|I|-3$ edges, we must have at least three edges joining the set of inner vertices $I$ to the pinned vertices $P$. Therefore, if we sum over the valence of inner vertices, and denote the the set of edges with both endpoints in $I$ by $E_i$, the ones with one endpoint in $I$ by $E_p$, we obtain $\sum val(i) = 2|E_i|
+|E_p| = 2|I|+|E_i| \leq 4|I|-3$. So there will be some inner vertices of valence 2 or 3. If at this stage there is some vertex of degree $2$, we can just remove it, to create a smaller graph with the same isostatic count. If there is some vertex of degree $3$, then by the usual arguments [@GSS; @TW], it can be removed, and replaced by a new edge joining two of its neighbors, which were not yet joined in a remaining rigid subgraph (e.g. not both pinned), to create a smaller subgraph with the isostatic count. This produces a reverse sequence of smaller and smaller isostatic graphs until we have no inner vertices.
To obtain a realization, place $P$ in an arbitrary position ${\mathbf{P}}$ with at least two distinct vertices. Create an isostatic graph whose vertex set contains $P$, by, for example, ordering the vertices in $P$ with distinct positions arbitrarily, $P=\{ p_1, p_2, \ldots \}$, adding edges between consecutive vertices and attaching an extra vertex, ${p}_0$ by edges $(p_0,p_i)$. This graph is clearly isostatic, provided the point ${\mathbf{p}}_0$ is not placed on the line through ${\mathbf{p}}_i$ and ${\mathbf{p}}_{i+1}$ for any $i$, because it has the correct edge count and is rigid since it consists of a string of non-collinear triangles.
To complete the proof, we work back up the sequence of subgraphs we created in the de- construction process. We assume the current graph is realized as isostatic. When the next graph is created by adding a 2-valent vertex, then adding such a vertex in any position except on the line joining its two attachments gives a new isostatic realization.
When the next graph is created by an edge-split, note that at least one of the neighbors of the new vertex is inner, so this added inner vertex can be placed in a generic position ensuring that the three new attachments are not collinear. Therefore, by the usual arguments [@GSS; @TW], this insertion is also isostatic when placed along the line of the bar being removed, and therefore also when placed in any generic position. Since (iii) trivially implies (i), the proof is complete.
A pinned graph $G(I, P;E)$ satisfying the Pinned Framework Conditions must have at least two pins and in every isostatic realization of $G$ there must be at least two distinct pin locations. Placing all pins in the same location never yields an isostatic framework, but we can make an important observation about the DOF of such a “pin collapsed” framework.
\[contractThm\] Let $G(I, P; E)$ be a pinned graph satisfying the Pinned Framework Conditions. Identifying the pinned vertices to one vertex $p^{*}$ yields a graph $G^{*}(V,E)$, $V=I\bigcup \{p^{*}\}$ and the DOF of $G^{*}$ is one less than the number of rigidity circuits contained in $G^{*}$.
Since $|E|=2|I|=2|V|-2$, $G^{*}$ contains too many edges to be isostatic. If $G^{*}$ is rigid, it is overbraced by exactly one edge, hence contains exactly one rigidity circuit. If $G_p$ is not rigid, each of the rigidity circuits in $G^{*}$ must contain $p^{*}$. If two rigidity circuits intersected in a vertex other than $p^{*}$, the union of their edge sets, together with the pinned subgraph would violate the Pinned Subgraph Conditions. Therefore all circuits in $G^{*}$ have exactly the vertex $p$ in common. Removing exactly one edge from each circuit yields a basis for $G^{*}$ in $\mathfrak{R}(G^{*})$, establishing the desired connection between the DOF and the number of circuits.
Characterizations of Assur graphs
=================================
We start with two citations from the mechanical engineering literature as motivation for our combinatorial conditions. The following definition appears in [@YV]: “An Assur group is obtained from a kinematic chain of zero mobility by suppressing one or more links, at the condition that there is no simpler group inside". In [@SS] we find: “An element of an Assur group is a kinematic chain with free or unpaired joints on the links which when connected to a stationary link will have zero DOF. A basic rigid chain is a chain of zero DOF and whose subchains all have DOF greater than zero. In other words an element of an Assur group is a basic rigid chain with one of its links deleted".
These descriptions from the engineering literature are not definitions in the mathematical sense, but rather use ‘minimality’ informally, as in the original work of Assur. We are now ready to give a formal definition by confirming a series of equivalent combinatorial characterizations of Assur graphs. These statements are new, and (iii) and (iv) come from the conjectures offered by Offer Shai at the workshop.
Basic Characterization of Assur graphs
--------------------------------------
\[CharacterThm\] Assume $G= (I,P; E)$ is a pinned isostatic graph. Then the following are equivalent:
\(i) $G= (I,P; E)$ is minimal as a pinned isostatic graph: that is for all proper subsets of vertices $I'\cup P'$, $I' \cup P'$ induces a pinned subgraph $G' = (I'\cup P', E')$ with $|E'| \leq 2|I'| -1$.
\(ii) If the set $P$ is contracted to a single vertex $p^*$, inducing the unpinned graph $G^*$ with edge set $E$, then $G^*$ is a rigidity circuit.
\(iii) Either the graph has a single inner vertex of degree $2$ or each time we delete a vertex, the resulting pinned graph has a motion of all inner vertices (in generic position).
\(iv) Deletion of any edge from $G$ results in a pinned graph that has a motion of all inner vertices (in generic position).
\(i) implies (iv) If we delete an edge, there must be a motion by the count. If there is a set of inner vertices that are not moving, in generic position, then these vertices, and their edges to the pinned vertices, must form a proper isostatic pinned subgraph contradicting condition (i).
\(iv) implies (iii) Removing an edge with an endpoint of valence $2$ produces a graph with a pendant edge. This must be the only inner vertex, since any other inner vertices are not moving, contradicting (iv). Since removing a single edge results in a motion of all inner vertices, removing all edges incident with one particular vertex results in a framework with a motion on all the remaining vertices.
Conversely, (iii) implies (i) If the graph contains a minimal proper pinned subgraph, then removing any vertex outside of this subgraph will produce a motion at most in the vertices outside of the subgraph. This contradicts (iii).
\(i) is equivalent to (ii) If $G= (V,P; E)$ is a pinned isostatic graph, then identifying the vertices in $P$ to a single vertex $p*$ yields a graph $G^{*} = (V^{*}, E^{*})$ with $|E^{*}| = 2(|V|+1) = |V^{*}|-2$, so $G^{*}$ is dependent and if the minimality condition in (i) is satisfied, it must be a rigidity circuit. Conversely, if $G^{*}$ is a rigidity circuit, we can pick an arbitrary vertex of $G^{*}$ and call it $p^{*}$. Splitting $p^{*}$ into a vertex set $P$, $|P|\geq 2$, (where $P$ may have as many vertices as the valence of $p^{*}$ in $G^{*}$ allows) and specifying for each edge with endpoint $p^{*}$ a new endpoint from $P$ so that no isolated vertices are left, yields an isostatic pinned framework satisfying the minimality condition.
This theorem provides a rigorous mathematical definition: an [*Assur graph*]{} is a pinned graph satisfying one of the four equivalent conditions in Theorem \[CharacterThm\].
Condition (i) is a refinement of the Grübler count (\[Grubler1\]), in a form which is now necessary and sufficient.
Condition (ii) translates the minimality condition to minimal dependence in $\mathfrak{R}_2 (n)$ and thus serves as a purely combinatorial description of Assur graphs and may be checked by fast algorithms [@JacobsHendrickson; @pebblegame].
Conditions (iii) and (iv) are similar in nature. Condition (iii) provides the engineer with a quicker check for the Assur property for smaller graphs than (iv), since there are fewer vertices than edges to delete. However, condition (iv) tells the engineer that a driver inserted for an arbitrary edge will (generically) move all inner vertices. We will expand on this property in [@SSW2]
![Assur graphs\[assurwordfig01\]](assurwordfig01)
![Corresponding circuits for Assur graphs\[assurwordfig03\]](assurwordfig03)
Some examples of Assur graphs are drawn in Figure \[assurwordfig01\] and their corresponding rigidity circuits in Figure \[assurwordfig03\].
Decomposition of general isostatic frameworks
----------------------------------------------
We now show that a general isostatic framework can be decomposed into a partially ordered set of Assur graphs. The given framework can be re-assembled from these pieces by a basic linkage composition. Figure \[assurwordfig02\] shows isostatic pinned frameworks and Figure \[assurwordfig05\] indicates their decomposition.
![Decomposable - not Assur graphs\[assurwordfig02\]](assurwordfig02)
![The first step of a decomposition for isostatic frameworks in \[assurwordfig02\] - with identified subcircuit(s).\[assurwordfig04\]](assurwordfig04)
Given two linkages as pinned frameworks $H= (W,Q;F)$ and $G= (V,P; E)$ and an injective map $C: Q \rightarrow V\cup P$, the [*linkage composition*]{} $C(H,G)$ is the linkage obtained from $H$ and $W$ by identifying the pins $Q$ of $H$ with their images $C(Q)$.
Given two pinned isostatic graphs $H= (W,Q,F)$ , $G= (V,P; E)$, the composition $C(H,G)$ creates the new composite pinned graph: $C=(V\cup W, P, E\cup F)$ which is also isostatic.
By the counts, we have $|F|=2|W|$, and $|E|=2|V|$, so $|E\cup F| = 2|W \cup V|$. A similar analysis of the subgraphs confirms the isostatic status.
![Recomposing the pinned isostatic graphs in Figure \[assurwordfig02\] from their Assur components.\[assurwordfig05\]](assurwordfig05)
Under this operation, the Assur graphs will be the minimal, indecomposable graphs. We can show that every pinned isostatic graph $G$ is a unique composition of Assur graphs, which we will call the [*Assur components*]{} of $G$.
A pinned graph is isostatic if and only if it decomposes \[POTheorem\] into Assur components. The decomposition into Assur components is unique.
Take the isostatic pinned framework, and identify the ground pins. This is now a dependent graph. Using properties of $\mathfrak{R}_2 (n)$, see [@CMW], we can identify minimal dependent subgraphs - which, by Theorems \[contractThm\] and \[CharacterThm\], are Assur components after the pins are separated. These are the initial components. When all of these initial components are contracted in step two, we seek additional Assur components. We iterate the process until only the ground is left.
The decomposition process described in the proof of Theorem \[POTheorem\] naturally induces a partial order on the Assur components of an isostatic graph: component $A\leq B$ if $B$ occurs at a higher level, and $B$ has at least one vertex of $A$ as a pinned vertex. The algorithm for decomposing the graph guarantees that $A\leq B$ means that $B$ occurs at a later stage than $A$. This partial order can be represented in an [*Assur scheme*]{} as in Figure \[assurscheme\]. This partial order, with the identifications needed for linkage composition, can be used to re-assemble the graph from its Assur components.
![An isostatics pinned framework a) has a unique decomposition into Assur graphs b) which is represented by a partial order or Assur scheme c).\[assurscheme\]](AssurDecomp)
Replacing any edge in an isostatic framework produces a 1 DOF linkage. The decomposition described in Theorem \[POTheorem\] permits the analysis of this linkage in layers. In fact, we can place drivers in each Assur component to obtain linkages with several degrees of freedom and their complex behavior can be simply described by analyzing the individual Assur components. This process of adding drivers is studied in more detail in [@SSW2].
Generating Assur graphs
------------------------
We summarize inductive techniques to generate all Assur graphs. Engineers find such techniques of interest to generate basic building blocks for synthesizing new linkages.
![An Edge-Split takes an Assur graph with at least four vertices to an Assur graph\[assurwordfig06\]](assurwordfig06)
![2-sum of Assur graphs gives a new Assur graph with a removed pin.\[palo01Fig\]](palo01)
The dyad is the only Assur graph on three vertices. There is no Assur graph on four vertices. An Assur graph, whose corresponding rigidity circuit is $K_4$ is called a [*basic*]{} Assur graph.
To generate all Assur graphs (on five or more vertices) we use Theorem \[CharacterThm\](ii) together with Theorem \[BJ\] to generate all rigidity circuits. To get from a rigidity circuit $C$ to an Assur graph, we choose a vertex $p^*$ of $C$ and split it into two or more pins (as in the proof of Theorem \[CharacterThm\]). The choice of $p^*$, the splitting of $p^*$ into a set $P$ of pins ($2\leq |P|\leq val(p^*)$), and choosing for each edge incident to $p^*$ an endpoint from $P$ allows us to construct several Assur graphs from one rigidity circuit, see Figure \[PinRearrange\]. We say that $G(I, P; E)$ and $G'(I, P', E)$ are related by [*pin rearrangement*]{} if $G^* =G'^*$ (see Figure \[PinRearrange\]).
![Pin rearrangement (maintaining fact of at least two pins) \[PinRearrange\]](assurwordfig08)
The operations of edge-split and 2-sum, which were used to generate rigidity circuits inductively, can also be used directly on Assur graphs to generate new Assur graphs from old, see Figures \[assurwordfig06\] and \[palo01Fig\]. In particular, the operation of 2-sum may be of practical value if, for space reasons for example, a pinned vertex is to be eliminated, see Figure \[palo01Fig\].
All Assur graphs on 5 or more vertices can be obtained from basic Assur graphs by a sequence of edge-splits, pin-rearrangements and $2$-sums of smaller Assur graphs.
Since mechanical engineers might want to have additional tools readily available for generating Assur graphs, one can seek additional operations under which the class of Assur graphs is closed. Vertex-split (creating two vertices of degree at least three) is another operation which takes a rigidity circuit to a rigidity circuit, and therefore takes an Assur graph with at least three vertices to a larger Assur graph (Figure \[assurwordfig10\]). This is, in a specific sense, the dual operation to edge-split [@CMW]. More generally, [@CMW] explores a number of additional operations for generating larger circuits from smaller circuits. Each of these processes will take an Assur graph to a larger Assur graph.
![Vertex split taking an Assur graph to an Assur graph. \[assurwordfig10\]](assurwordfig10)
The inductive constructions for Assur graphs can be used to provide a visual [*certificate sequence*]{} for an Assur graph. If we are given a sequence of edge-splits and 2-sums starting from a dyad and ending with $G$, see Figure \[palo02Fig\], it is trivial to verify that $G$ is an Assur graph. We constructed such a sequence in the proof of Theorem \[PinThm\]. It is well known, see [@TW], that there are exponential algorithms to produce such a certificate. However, there are other fast algorithms, for example the so called pebble games, see [@JacobsHendrickson; @pebblegame] to detect all the rigidity properties of graphs that can be adapted to verify the Assur property.
![Certificate sequence for the final Assur graph.\[palo02Fig\]](palo02W)
Concluding comments
===================
The paper introduces, for the first time, the concept of Assur graphs, in the rigorous mathematical terminology of rigidity theory. This work paves a new channel for cooperation between the communities in kinematics and in rigidity theory. An example for such channel is the material appearing in §2.7 showing how to transform determinate trusses used by the kinematicians into isostatic frameworks, widely employed by the rigidity theory and structural engineering communities.
At this point, it is hard to predict all the practical applications that are to benefit from this new relation between the disciplines. Nevertheless, we anticipate practical results from the use of rigidity theory in mechanisms as introduced in the paper. Examples of such results, include using rigid circuits from rigidity theory to find the proper decomposition sequence of pinned isostatic framework into Assur components (Section 3.2) and generation of Assur graphs by applying two known operations to their corresponding rigidity circuits (Section 3.3).
It is expected that new opportunities will be opened up, for example, for mechanical engineers to comprehend topics in rigidity theory that are used today in many disciplines, including biology, communications and more. Mechanical engineers in the west may be motivated to use the Assur graphs (Assur groups) concept as it is widely applied in eastern Europe and Russia.
Decomposing a larger linkage into Assur graphs and analyzing these pieces one at a time is an effective way to analyze the overall motion, working in layers. This paper has given a precise mathematical foundation for the Assur method as well as a proof of its correctness and completeness.
We have followed standard engineering practice and developed the theory in the language of bar-and-joint frameworks, but of course there is no need to replace a larger link (rigid body) with an isostatic bar-and-joint sub-framework to apply the counting techniques. It is simply convenient to do so in order to streamline notation and graphics. All of our results may be reworded in terms of body and bar structures or body and pin frameworks. This presentation would be much closer to the original example in Figures 1 and 2 and the counts of §2.2.
To extend this type of analysis to 3D linkages, the lack of good characterization of isostatic bar-and-joint frameworks in $3$D is an initial obstacle. However, if we focus on body-and-bar or body-and-hinge structures (the analog of body-and-bar and body-and-pin frameworks in the plane) then the generic DOF of these structures is computable by analogous counting techniques [@WW2; @Wh], and all our combinatorial methods will carry over sucessfully.
[^1]: Mathematics Department, Worcester Polytechnic Institute. [email protected]
[^2]: Faculty of Engineering, Tel-Aviv University. [email protected]
[^3]: Department of Mathematics and Statistics, York University Toronto, ON, Canada. [email protected]. Work supported in part by a grant from NSERC Canada
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A barrier certificate is an inductive invariant function which can be used for the safety verification of a hybrid system. Safety verification based on barrier certificate has the benefit of avoiding explicit computation of the exact reachable set which is usually intractable for nonlinear hybrid systems. In this paper, we propose a new barrier certificate condition, called *Exponential Condition*, for the safety verification of semi-algebraic hybrid systems. The most important benefit of *Exponential Condition* is that it has a lower conservativeness than the existing convex condition and meanwhile it possesses the property of convexity. On the one hand, a less conservative barrier certificate forms a tighter over-approximation for the reachable set and hence is able to verify critical safety properties. On the other hand, the property of convexity guarantees its solvability by semidefinite programming method. Some examples are presented to illustrate the effectiveness and practicality of our method.
invariant, barrier certificate, safety verification, hybrid system, nonlinear system, sum of squares
author:
- Hui Kong
- Fei He
- Xiaoyu Song
- 'William N. N. Hung'
- Ming Gu
title: 'Exponential-Condition-Based Barrier Certificate Generation for Safety Verification of Hybrid Systems[^1]'
---
Introduction
============
Hybrid systems [@henzinger1996theory], [@alur1995algorithmic] are models for those systems with interacting discrete and continuous dynamics. Embedded systems are often modeled as hybrid systems due to their involvement of both digital control software and analog plants. In recent years, as embedded systems are becoming ubiquitous, more and more researchers are devoted to the theory of hybrid systems. Reachability problems or safety verification problems are among the most challenging problems in verifying hybrid systems. The aim of safety verification is to decide that starting from an initial set, whether a continuous system or hybrid system can reach an unsafe set. For this purpose, many methods have been proposed for various hybrid systems with different features.
Deductive methods based on inductive invariant play an important role in safety verification of hybrid systems. An inductive invariant of a hybrid system is an invariant $\varphi$ that holds at the initial states of the system, and is preserved by all discrete and continuous transitions. A safety property is an invariant $\psi$ (usually not inductive) that holds in all reachable states of the system. The standard technique for proving a given property $\psi$ is to generate an inductive invariant $\varphi$ that implies $\psi$. Therefore, the problem of safety verification is converted to the problem of inductive invariant generation and hence avoid the reachability computation of the hybrid system. The key points in generating inductive invariant for hybrid systems is how to define an inductive condition that is the least conservative and how to efficiently compute the inductive invariant that satisfies the inductive condition. Usually, these two aspects contradicts with each other, that is, an inductive condition with sufficiently low conservativeness often encounters the computability or complexity problem. For different class of hybrid systems, various inductive invariants and computational methods have been proposed.
Some methods were primarily proposed for constructing inductive invariant for linear hybrid systems [@jirstrand1998invariant], [@rodriguez2005generating]. In recent years, however, researchers concentrate more and more on nonlinear hybrid systems, especially on algebraic or semi-algebraic hybrid systems (i.e. those systems whose vector fields are polynomials and whose set descriptions are polynomial equalities or inequalities), as they have a higher universality. In [@sankaranarayanan2004constructing], [@sankaranarayanan2010automatic], Sankaranarayanan et al. presented a computational method based on the theory of ideal over polynomial ring and quantifier elimination for automatically generating algebraic invariants for algebraic hybrid systems. Similarly, Tiwari et al. proposed in [@tiwari2004nonlinear] a technique based on the theory of ideal over polynomial ring to generate the inductive invariant for nonlinear polynomial systems. In [@prajna2007framework], [@prajna2004safety], S. Prajna et al. proposed a new inductive invariant called *Barrier Certificate* for verifying the safety of semialgebraic hybrid systems and the computational method they applied is the technique of sum-of-squares decomposition of semidefinite polynomials. In [@sloth2012compositional], C. Sloth et al. proposed a new *Barrier Certificate* for a special class of hybrid systems which can be modeled as an interconnection of subsystems. In [@platzer2008computing], A. Platzer et al. proposed the concept of *Differential Invariant* which is a boolean combination of multiple polynomial inequalities for verifying semialgebraic hybrid systems. In [@gulwani2008constraint], S. Gulwani et al. proposed an inductive invariant similar to *Differential Invariant* except that they defined a different inductive condition and they used SMT solver to solve the inductive invariant. In [@taly2009deductive], A. Taly et al. discussed the soundness and completeness of several existing invariant condition and presented several simpler and practical invariant condition that are sound and relatively complete for different classes of inductive invariants. In [@taly2011synthesizing], A. Taly et al. proposed to use inductive controlled invariant to synthesize multi-modal continuous dynamical systems satisfying a specified safety property.
In this paper, we propose a new barrier certificate (called *Exponential Condition*) for the safety verification of semialgebraic hybrid systems. A barrier certificate is a special class of inductive invariant for the safety verification of hybrid systems: a function $\varphi(x)$ which maps all the states in the reachable set to non-positive reals and all the states in the unsafe set to positive reals. Given a dynamical system $S$ with dynamics $\dot{x}=f(x)$ with initial set $Init$, to prove a safety property $P$ (we use $X_u$ to denote the unsafe set) is satisfied by $S$, the basic idea of *Exponential Condition* is to identify a function $\varphi(x)$ such that 1) $\varphi(x)\leq 0$ for any point $x \in Init$, 2) $\varphi(x)>0$ for any point $x \in X_u$, and 3) $\mathcal{L}_f\varphi(x) \leq \lambda\varphi(x)$, where $\mathcal{L}_f\varphi(x)=\frac{\partial{\varphi}}{\partial{x}} f(x)$ is the Lie derivative of $\varphi$ with respect to the vector field $f$ and $\lambda$ is any negative constant real value. The first condition and the third condition together guarantee that $\varphi(x) \leq 0$ for any point $x$ in the reachable set $R$, which implies that $R \cap X_u=\emptyset$. Therefore, we can assert that the safety property $P$ is satisfied by the system $M$ as long as we can find a function $\varphi(x)$ satisfying the above condition. The above condition can be extended to semialgebraic hybrid systems naturally. The idea is to identify a set of functions $\{\varphi_i(x)\}$, one for each mode of the hybrid system, which not only satisfy the above condition but also satisfy an additional sign-preserving constraint for each discrete transition.
The most important benefit of *Exponential Condition* is that it is less conservative than *Convex Condition* [@prajna2007framework] and *Differential Invariant* [@platzer2008computing], where the Lie derivative of $\varphi(x)$ is required to satisfy that $\mathcal{L}_f\varphi(x)\leq 0$ (a stronger condition than $\mathcal{L}_f\varphi(x) \leq \lambda\varphi(x)$), and meanwhile, it possesses the property of convexity as well. On the one hand, a less conservative inductive invariant forms a tighter over-approximation for the reachable set and hence is able to verify critical safety properties (i.e., the unsafe region is very close to reachable region). On the other hand, a convex inductive invariant condition can be solved efficiently by semidefinite programming method, which is widely used for computing Lyapunov functions in the stability analysis of nonlinear systems. In fact, there exist some other less conservative inductive invariants than *Exponential Condition*, such as [@prajna2007framework], [@gulwani2008constraint], [@taly2009deductive], however, these inductive conditions are not convex and thus cannot be solved by semidefinite programming method. Instead, they are usually solved by quantifier elimination and SMT solver, which usually has a much higher computational complexity than semidefinite programming method.
Given a semialgebraic hybrid system, we choose a set of polynomials of bounded degree with unknown coefficients as the candidate inductive invariant, and then we obtain a set of positive semidefinite polynomials (i.e. $P(x)\geq 0$) according to *Exponential Condition*. Therefore, the generation of barrier certificate based on *Exponential Condition* can be transformed to the problem of sum-of-squares programming of positive semidefinite polynomials [@sturm1999using], [@prajna2005sostools]. Based on our theory, we develop an algorithm for generating the inductive invariant satisfying *Exponential Condition*. Experiments on both nonlinear systems and hybrid systems show the effectiveness and practicality of our method.
The remainder of this paper is organized as follows. Section \[sec:Preliminaries\] introduces the preliminaries of our method. Section \[sec:CertCondition\] presents the barrier certificate conditions for continuous systems and hybrid systems. Section \[sec:computemethod\] introduces the computational method we use to construct barrier certificates according to the barrier certificate conditions. Section \[sec:examples\] gives some examples to demonstrate the application of our method to the safety verification of continuous and hybrid systems. Finally, we conclude our work in Section \[sec:conclusion\].
Preliminaries {#sec:Preliminaries}
=============
In this paper, we adopt the model proposed in [@carloni2006languages] as our modeling framework. Many other models for hybrid system can be found in [@maler1992prom], [@lygeros1999controllers], [@alur1995algorithmic].
A continuous system is specified by a differential equation $$\label{sec:ContinuousSys}
\dot{x} = f(x)$$ where $x \in \mathbb{R}^n$ and $f$ is a Lipschitz continuous vector function from $\mathbb{R}^n$ to $\mathbb{R}^n$. Note that the Lipschitz continuity guarantees the existence and uniqueness of the solution $x(t)$ to the system . A hybrid system can then be defined as:
**(Hybrid System)** A hybrid system is a tuple $\mathcal{H} = \langle L,X,E,R,G,$ $I,F\rangle$, where
- $L$ is a finite set of locations (or modes);
- $X \subseteq \mathbb{R}^n$ is the continuous state space. The hybrid state space of the system is denoted by $\mathcal{X} = L \times X$ and a state is denoted by $(l,x) \in \mathcal{X}$;
- $E \subseteq L \times L$ is a set of discrete transitions;
- $G:E \mapsto 2^X$ is a guard mapping over discrete transitions;
- $R:E \times X \mapsto 2^X$ is a reset mapping over discrete transitions;
- $I : L \mapsto 2^X$ is an invariant mapping;
- $F : L \mapsto (X\mapsto X)$ is a vector field mapping which assigns to each location $l$ a vector field $f$.
The transition and dynamic structure of the hybrid system defines a set of trajectories. A trajectory is a sequence starting from a state $(l_0,x_0) \in \mathcal{X}_0$, where $\mathcal{X}_0 \subseteq \mathcal{X}$ is an initial set, and consisting of a series of interleaved continuous flows and discrete transitions. During the continuous flows, the system evolves following the vector field $F(l)$ at some location $l\in L$ until the invariant condition $I(l)$ is violated. At some state $(l,x)$, if there is a discrete transition $(l,l')\in E$ such that $(l,x) \in G(l,l')$ (we write $G(l,l')$ for $G((l,l'))$), then the discrete transition can be taken and the system state can be reset to $R(l,l',x)$. The problem of safety verification of a hybrid system is to prove that the hybrid system cannot reach an unsafe set $\mathcal{X}_u$ from an initial set $\mathcal{X}_0$.
An important concept used in this paper is the Lie derivative. In our context, the Lie derivative evaluates the change of a scalar function $\varphi(x)$ along the flow of a vector field $f(x)=(f_1(x),\cdots,f_n(x))$. Formally, $$\mathcal{L}_f \varphi \triangleq \frac{\partial \varphi}{\partial x} f(x) = \sum_{i=1}^n \frac{\partial \varphi}{\partial x_i} f_i(x)$$
Some other notations that are used in this paper are presented here. $\mathbb{R}$ denotes the real number field. $\mathcal{C}^1(\mathbb{R}^n)$ denotes the space of 1-time continuously differentiable functions mapping $X \subseteq \mathbb{R}^n$ to $\mathbb{R}$. $\mathbb{R}[x]$ denotes the polynomial ring in $x$ over the real number field and $\mathbb{R}[x]^m$ denotes the $m$-dimensional polynomial vector space over $\mathbb{R}[x]$. $M^T$ denotes the transpose of the matrix $M$.
Conditions for Constructing Barrier Certificates {#sec:CertCondition}
================================================
Barrier Certificate Condition for Continuous Systems {#contSys}
----------------------------------------------------
Given a continuous system $S$, an initial set $X_0$ and an unsafe set $X_u$, a barrier certificate is a real-valued function $\varphi(x)$ of states satisfying that $\varphi(x)\leq 0$ for any point $x$ in the reachable set $R$ and $\varphi(x)>0$ for any point $x$ in the unsafe set $X_u$ (called *General Constraint* hereafter). Therefore, if there exists such a function $\varphi(x)$, we can assert that $R \cap X_u = \emptyset$, that is, the system can not reach a state in the unsafe set from the initial set. However, the exact reachable set $R$ is not computable for most hybrid systems, we cannot decide directly whether $\varphi(x)\leq 0$ holds for all the points in $R$. Therefore, various alternative inductive conditions that are equivalent to or sufficient for *General Constraint* are proposed. In what follows, we present a new barrier certificate which is a sufficient condition for *General Constraint*.
Consider a continuous system $\mathbb{C}$ specified by the differential equation , we assume that $X_0 (\subseteq X)$, $X_u$ are the initial set and the unsafe set respectively. Then, we have the following theorem as a barrier certificate condition.
\[theorem1\] Given the continuous system and the corresponding sets $X$, $X_0$ and $X_u$, for any given $\lambda \in \mathbb{R}$, if there exists a barrier certificate, i.e, a real-valued function $\varphi(x)\in \mathcal{C}^1(\mathbb{R}^n)$ satisfying the following formulae: $$\begin{aligned}
&\forall x \in X_0: \varphi(x) \leq 0\label{cond11}\\
&\forall x \in X: \mathcal{L}_f\varphi(x) - \lambda \varphi(x) \leq 0\label{cond12}\\
&\forall x \in X_u: \varphi(x) > 0\label{cond13}\end{aligned}$$ then the safety property is satisfied by the system .
Suppose $x_0 \in X_0$ and $x(t)$ be the corresponding particular solution of the system . We aim to prove that for any function $\varphi(x(t))$ satisfying the formulae –, the following formula holds: $$\forall \zeta \geq 0: \varphi(x(\zeta)) \leq 0.$$ Let $g(x) = \mathcal{L}_f\varphi(x) - \lambda\varphi(x)$, then by $$\forall x \in X: g(x) \leq 0\label{gx}$$ Since $\frac{d\varphi(x(t))}{dt} = \frac{\partial{\varphi}}{\partial{x}} \frac{dx}{dt}=\frac{\partial{\varphi}}{\partial{x}} f(x)=\mathcal{L}_f\varphi(x)$, we have the differential equation about $\varphi(x(t))$ $$\label{eqcomb}
\left\{\begin{array}{ll}
\frac{d\varphi(x(t))}{dt} - \lambda \varphi(x(t)) - g(x(t)) = 0\\
\varphi(x(0)) = \varphi(x_0)\\
\end{array}\right.$$ By solving the differential equation , we have following the solution: $$\label{odesolution}
\varphi(x(t)) = (\int_{0}^{t}{(g(x(\tau)) e^{-\lambda \tau}d\tau}+\varphi(x_0)) e^{\lambda t}.$$ By , we have $$\int_{0}^{t}{(g(x(\tau)) e^{-\lambda \tau}d\tau} \leq 0.\label{intgx}$$ then by and $\varphi(x_0)\leq 0$, we finally have $$\varphi(x(t)) \leq \varphi(x_0) e^{\lambda t} \leq 0.\label{bxtineq}$$ Hence, for any $\zeta\geq 0$, $\varphi(x(\zeta)) \leq 0$ holds.
The formulae and ensure that the barrier separates the initial set $X_0$ from the unsafe set $X_u$, and the formula ensures that system trajectories cannot escape from inside of the barrier. These conditions together imply that $\varphi(x) \leq 0$ is an inductive invariant of the system .
From another point of view, the semi-algebraic set $\{x\in \mathbb{R}^n| \varphi(x)\leq 0\}$ forms an over-approximation for the reachable set of the system , and the zero level set of the function $\varphi(x)$ (i.e., $\{x\in \mathbb{R}^n|\varphi(x)=0\}$) forms the boundary of the over-approximation. In order to be less conservative, we hope the boundary of the over-approximation encloses the reachable set $\{x(t)| x(0) \in X_0, \dot{x}= f(x), t \in \mathbb{R}_+\}$ as tightly as possible, in other words, to make the upper-bound of $\varphi(x(t))$ approach zero as closely as possible. According to the above proof (i.e., ), the scope over which the function $\varphi(x(t))$ can range depends closely on the value of the parameter $\lambda$: the less value the $\lambda$ is, the closer the upper-bound of the scope that $\varphi(x(t))$ can reach is to zero (see \[fig:barrieronlambda\]).
![Dependency of Barrier Certificate Condition on $\lambda$. As the value of $\lambda$ decreases (e.g. from $1/4$ to $-3$), the upper-bound of the value of $\varphi(x(t))$ approaches to zero, which means the barrier certificate condition becomes less conservative[]{data-label="fig:barrieronlambda"}](Upperbound-fig.pdf)
Roughly speaking, the values of $\lambda$ are divided into three classes according to the conservativeness of the barrier certificate condition:
- $\lambda = 0$. In this case, the formula is degenerated to $\frac{\partial{\varphi}}{\partial{x}} f(x) \leq 0$, which is the case of *Convex Condition*. This condition implies that the value of $\varphi(x(t))$ will never get close to zero over time $t$. Thus, the condition is very conservative.
- $\lambda <0$. In this case, we know that 1) $\varphi(x(t)) \leq \varphi(x_0) e^{\lambda t}\leq 0$, and 2) $\frac{\partial{\varphi}}{\partial{x}} f(x) \leq \lambda \varphi(x)\geq 0$. These two inequalities together imply that the value of $\varphi(x(t))$ can increase over the time $t$ but never get across the upper bound $0$, provided that $\varphi(x(0))\leq 0$ at the beginning.
- $\lambda >0$. In this case, $\frac{\partial{\varphi}}{\partial{x}} f(x) \leq \lambda \varphi(x)\leq 0$, which means that the value of $\varphi(x(t))$ get far away from $0$. Apparently, the condition is much more conservative than the first case.
Therefore, as long as we let $\lambda <0$, we can get less conservative barrier certificate conditions than *Convex Condition*. Note that *Exponential Condition* is convex as well and its convexity can be easily proved by verifying that for any two functions $\varphi_1(x)$ and $\varphi_2(x)$ satisfying the formulae – and any $\theta$ with $0\leq \theta \leq 1$, $\varphi(x)= \theta \varphi_1(x) + (1-\theta) \varphi_2(x)$ satisfies the formulae – as well. Based on this fact, we can convert the problem of constructing barrier certificate into the problem of convex optimization which we will discuss in Section \[sec:computemethod\].
In addition, as a generalization of *Convex Condition*, *Differential Invariant* is basically as conservative as *Convex Condition*. Here we present informally an explanation on this point. The differences in their definitions include mainly two aspects:
1. invariant template: *Convex Condition* employs a single inequality $p(x)\leq 0$ as the invariant template while *Differential Invariant* employs a conjunction $\bigwedge_{i=1}^m q_i(x)\rhd_i r_i(x)$, where $\rhd_i$ denotes a connective in $\{=,\geq,>,\leq,<\}$.
2. inductive condition: *Convex Condition* employs $\mathcal{L}_f(p)\leq 0$ as the inductive condition while *Differential Invariant* employs the conjunction $\bigwedge_{i=1}^m \mathcal{L}_f q_i\rhd_i \mathcal{L}_f r_i$, which results from applying the Lie derivative to each of the conjuncts in the invariant template respectively.
Note that each conjunct of a *Differential Invariant* is still an inductive invariant by itself, which is named *Sub-Differential-Invariant* here. Based on the above definition, we can easily prove that every *Sub-Differential-Invariant* $q_i(x)\rhd_i r_i(x)$ satisfies *Convex Condition*. For example, suppose we have a *Sub-Differential-Invariant* $q_i(x) > r_i(x)$ and the corresponding inductive condition $\mathcal{L}_f q_i>\mathcal{L}_f r_i$, let $p(x)=r_i(x) - q_i(x)$, then we can obtain an equivalent inductive invariant $p(x)<0$ and the corresponding inductive condition $\mathcal{L}_f p=\mathcal{L}_f q_i - \mathcal{L}_f r_i<0$, which implies $p(x)\leq 0$ and $\mathcal{L}_f p \leq 0$ hold. Therefore, the *Sub-Differential-Invariant* $q_i(x)>r_i(x)$ satisfies *Convex Condition*. Similarly, all the other cases of $q_i(x)\rhd_i r_i(x)$ can be proved to satisfy *Convex Condition*. Hence, *Sub-Differential-Invariant* is no less conservative than *Convex Condition*. By taking a conjunction of multiple *Sub-Differential Invariant*s, *Differential Invariant* actually enhances the ability to over-approximate complex-shaped reachable sets. However, this does not overcome the drawback that no trajectory of the system can move towards the boundary of the over-approximation formed by a *Differential Invariant*. Therefore, in this sense, we say that *Differential Invariant* is basically as conservative as *Convex Condition* and consequently is more conservative than *Exponential Condition*.
In the following subsection, we extend the barrier certificate condition for continuous systems to hybrid systems.
Barrier Certificate Condition for Hybrid Systems
------------------------------------------------
Different from the barrier certificate for a continuous system, the barrier certificate for a hybrid system consists of a set of functions $\{\varphi_l(x)| l \in L\}$, each of which corresponds to a discrete location of the system and forms a barrier between the reachable set and the unsafe set at that individual location. For each function $\varphi_l(x)$ at location $l$, in addition to defining constraints for the continuous flows, the barrier certificate conditions have to take into account all the discrete transitions starting from location $l$ to make the overall barrier certificate an inductive invariant. Formally, we define the barrier certificate condition for hybrid systems as the following theorem.
\[theorem2\] Given the hybrid system $\mathcal{H}=\langle L,X,$ $E,R,G,I,F\rangle$, the initial set $\mathcal{X}_0$ and the unsafe set $\mathcal{X}_u$ of $\mathcal{H}$, then, for any given set of constant real numbers $S_{\lambda}=\{\lambda_l\in \mathbb{R}| l\in L\}$ and any given set of constant non-negative real numbers $S_{\gamma}=\{\gamma_{ll'} \in \mathbb{R}_+| (l,l') \in E\}$, if there exists a set of functions $\{\varphi_l(x)| \varphi_l(x) \in \mathcal{C}^1(\mathbb{R}^n),l\in L\}$ such that, for all $l\in L$ and $(l,l')\in E$, the following conditions hold: $$\begin{aligned}
&\forall x \in Init(l): \varphi_l(x) \leq 0\label{cond21}\\
&\forall x \in I(l): \mathcal{L}_{f_l}\varphi_l(x) - \lambda_l \varphi_l(x) \leq 0\label{cond22}\\
&\forall x \in G(l,l'), \forall x' \in R((l,l'),x): \gamma_{ll'} \varphi_l(x)-\varphi_{l'}(x') \geq 0\label{cond23}\\
&\forall x \in Unsafe(l): \varphi_l(x) > 0\label{cond24}\end{aligned}$$ where $Init(l)$ and $Unsafe(l)$ denote respectively the initial set and the unsafe set at location $l$, then the safety property is satisfied by $\mathcal{H}$.
To prove this theorem, it is sufficient to prove that given any trajectory, say $\pi$, of the system $\mathcal{H}$, it cannot reach an unsafe state. Suppose the infinite time interval $\mathbb{R}_+$ associated with $\pi$ is divided into an infinite sequence of continuous time subintervals, i.e., $\mathbb{R}_+ = \bigcup_{n=0}^\infty I_n$, where $I_n = \{t\in \mathbb{R}_+| t_n \leq t \leq t_{n+1}\}$ is the time interval that the system spent at location $\rho(I_n)$ (where $\rho(I_n)$ returns the location corresponding to $I_n$), we define the trajectory as $\pi = \{x_{\rho(I_n)}(t)| t \in I_n, n \in \mathds{N}\}$, where $x_{\rho(I_0)}(t_0) \in Init(\rho(I_0))$. Then, our objective is to prove the following assertion: $$\label{assertion1}
\forall n \in \mathds{N}: \forall t\in I_n:\varphi_{\rho(I_n)}(x_{\rho(I_n)}(t))\leq 0.$$ The basic proof idea is by induction.
*Basis:* $n=0$. According to Theorem \[theorem1\], it’s obvious that $$\forall t\in I_0:\varphi_{\rho(I_0)}(x_{\rho(I_0)}(t))\leq 0$$
*Induction:* $n=k$. Assume for some $k$, $$\forall n \in [0,k]:\forall t\in I_n:\varphi_{\rho(I_n)}(x_{\rho(I_n)}(t))\leq 0$$ we mean to prove that $$\forall t\in I_{k+1}:\varphi_{\rho(I_{k+1})}(x_{\rho(I_{k+1})}(t))\leq 0$$
Case 1. (Discrete Transition) By the inductive assumption, we know that $$\forall t\in I_k:\varphi_{\rho(I_k)}(x_{\rho(I_k)}(t))\leq 0$$ hence $$\forall t\in I_k:x(t)\in G(\rho(I_k),\rho(I_{k+1})) \implies \varphi_{\rho(I_k)}(x(t)) \leq 0$$ According to condition , we know that $\varphi_{\rho(I_{k+1})}(x_{\rho(I_{k+1})}(t_{k+1})) \leq 0$.
Case 2. (Continuous Transition) According to Case 1 and condition , we can conclude that $\forall t \in I_{k+1}:\varphi_{\rho(I_{k+1})}(x_{\rho(I_{k+1})}(t))\leq 0$ by Theorem \[theorem1\].
By induction, we know that the assertion holds. Therefore, the safety property is guaranteed.
Informally, the formulae , and together ensure that at each location $l \in L$, the system never evolves into an unsafe state continuously. The formula ensures that the system never jumps from a safe state to an unsafe state discretely. By induction, the formulae – together guarantee the safety of the system.
The selection of the parameter set $S_\lambda$ is essential to the conservativeness of the barrier certificate conditions. As discussed in Subsection \[contSys\], by setting all the elements of $S_\lambda$ to $0$, we can derive *Convex Condition* for hybrid systems. However, *Convex Condition* is too restrictive to be useful for hybrid systems. For example, see the hybrid system in \[fig:convexcondexamp1\], there is a reset operation $x = x_r$ (which is often the case) at the transition $(l_2,l_1)$. Assume there exists a barrier certificate $\{\varphi_{l_1}(x),\varphi_{l_2}(x)\}$ if we set all the elements of $S_\lambda$ to $0$ and (without loss of generality) set all the elements of $S_\gamma$ to $1$, then for any trajectory containing at least two times of the transition $(l_2,l_1)$, one at time instant $t_1$ and another at $t_2$, $t_1 < t_2$, respectively, we can assert that $\varphi_{l_1}(x_{l_1t_1}) > \varphi_{l_1}(x_{l_1t_2})$ according to Theorem \[theorem2\], this contradicts with $x_{l_1t_1} = x_{l_1t_2} = x_r$, that is, the barrier certificate satisfying *Convex Condition* does not exist no matter what the unsafe set is.
![A hybrid system without barrier certificate satisfying *Convex Condition*.[]{data-label="fig:convexcondexamp1"}](convexcondexample.pdf)
Therefore, in order to make the barrier certificate condition less conservative, we try to choose negative values for $\lambda_l \in S_\lambda$ and theoretically: the less, the better. However, in practice, the optimal domain for $\lambda$ may depend on the specific computational method. For example, the interval $[-1,0)$ appears to be optimal and not too sensitive in-between for the semidefinite programming method used in this paper.
The selection of $S_\gamma$ is relatively simple. We usually set all of its elements to $1$ except for the discrete jumps with a reset operation that is independent of the pre-state of the jump, for which we usually set $\gamma_{ll'}$ to $0$.
Construction Method for Barrier Certificate {#sec:computemethod}
===========================================
Constructing inductive invariants for general hybrid systems is very hard. Fortunately, for some existing inductive conditions, several computational methods are available for semialgebraic hybrid systems. The most representative methods include the fixed-point method based on saturation [@platzer2008computing], the constraint-solving methods based on semidefinite programming [@prajna2004safety] and quantifier elimination [@gulwani2008constraint] and the Grobner-bases method [@tiwari2004nonlinear], [@sankaranarayanan2010automatic]. Similar to *Convex Condition*, *Exponential Condition* defines a convex set of barrier certificate functions as well and hence can be solved by semidefinite programming method supposing the hybrid system is semialgebraic and the barrier certificate function $\varphi(x)$ is a polynomial.
In our computational method, a barrier certificate is assumed to be a set $\Phi=\{\varphi_l(x)| l\in L\}$ of multivariate polynomials of fixed degrees with a set of unknown real coefficients. According to the constraint inequalities in Theorem \[theorem1\] or Theorem \[theorem2\], we can obtain a set of positive semidefinite (*PSD*) polynomials $Q = \{Q_i| Q_i(x) \geq 0, deg(Q_i)= 2n, x \in \mathbb{R}^n, n \in \mathds{N}\}$, where $deg(\cdot)$ returns the degree of a polynomial. Note that a polynomial $Q(x)$ of degree $2k$ is said to be *PSD* *if and only if* $Q(x)\geq 0$ for all $x\in \mathbb{R}^n$. Thus, our objective is to find a set of real-valued coefficients for $\varphi_l\in \Phi$ to make all the $Q_i\in Q$ be *PSD*.
A famous sufficient condition for a polynomial $P(x)$ of degree $2k$ to be *PSD* is that it is a sum-of-squares (*SOS*) $P(x) = \sum q_i(x)^2$ for some polynomials $q_i(x)$ of degree $k$ or less [@lasserre2007sufficient]. Furthermore, it is equivalent to that $P(x)$ has a positive semidefinite quadratic form, i.e., $P(x) = v(x)Mv(x)^T$, where $v(x)$ is a vector of monomials with respect to $x$ of degree $k$ or less and $M$ is a real symmetric *PSD* matrix with the coefficients of $P(x)$ as its entries. Therefore, the problem of finding a *PSD* polynomial $P(x)$ can be converted to the problem of solving a linear matrix inequality (LMI) $M\succeq 0$ [@boyd1994linear], which can be solved by semidefinite programming [@parrilo2003semidefinite].
In our work, we extend SOSTOOLS based on the theory in this paper to implement an algorithm for discovering barrier certificate automatically.
Sum-of-squares Transformation for Continuous System {#subsec:continuousmethod}
---------------------------------------------------
In order to be solvable for the barrier certificate condition by SOS programming, we need to restate it with multivariate polynomials. In this context, we assume that all the state sets involved in the condition are semialgebraic, that is, they can be written as $\{x \in \mathbb{R}^n|P_1(x) \geq 0,...,P_m(x)\geq 0, P_i(x)\in \mathbb{R}[x], 1\leq i\leq m\}$). For convenience, we write it compactly as $\{x \in \mathbb{R}^n|\mathcal{P}(x) \geq 0, \mathcal{P}(x) \in \mathbb{R}[x]^m\}$, where $\mathcal{P}(x)=(P_1(x),P_2(x),...,P_m(x))$. In addition, each dimension of the vector field $f(x)$ and the barrier certificate function $\varphi(x)$ are all polynomials in $\mathbb{R}[x]$. Based on the previous assumption, we present the sum-of-squares transformation of *Exponential Condition* for continuous systems as the following corollary.
\[corollary1\] Given the continuous polynomial system and the initial set $X_0 = \{x\in \mathbb{R}^n|I_0(x)\geq 0, I_0(x) \in \mathbb{R}[x]^r\}$ and the unsafe set $X_u = \{x\in \mathbb{R}^n|U(x) \geq 0, U(x) \in \mathbb{R}[x]^s\}$, where $r$ and $s$ are the dimensions of the polynomial vector spaces, for any $\lambda \in \mathbb{R}$ and any real number $\epsilon>0$, if there exists a polynomial function $\varphi(x)\in \mathbb{R}[x]$ and two *SOS* polynomial vectors (i.e., every element of the vector is a *SOS* polynomial) $\mu(x) \in \mathbb{R}[x]^r$ and $\eta(x)\in \mathbb{R}[x]^s$ satisfying that the following polynomials $$\begin{aligned}
&-\varphi(x) - \mu(x) I_0(x) \label{corocond1}\\
&-\mathcal{L}_f\varphi(x) + \lambda \varphi(x) \label{corocond2}\\
&\varphi(x) - \eta(x) U(x) - \epsilon \label{corocond3}\end{aligned}$$ are all *SOS*s, then the safety property is satisfied by the system .
It is sufficient to prove that any $\varphi(x)$ satisfying – also satisfies –. By , we have $-\varphi(x) - \mu(x) I_0(x)\geq 0$, that is, $\varphi(x) \leq -\mu(x) I_0(x)$. Because for any $x \in X_0$, $-\mu(x) I_0(x)\leq 0$, this means $\varphi(x) \leq 0$. Similarly, we can derive from . By , it’s easy to prove that $\varphi(x)-\epsilon \geq 0$ holds for any $x \in X_u$. Since $\epsilon$ is greater than $0$, then the condition holds. Therefore, the system is safe.
Since the polynomials – are required to be *SOS*s, each of them can be transformed to a positive semidefinite quadratic form $v(x)M_iv(x)^T$, where $M_i$ is a real symmetric *PSD* matrix with the coefficients of $\varphi(x)$, $\mu(x)$ and $\eta(x)$ as its variables. As a result, we obtain a set of *LMI*s $\{M_i \succeq 0\}$ which can be solved by semidefinite programming.
\[algorithm1\]
[ Initialize. Set $\Lambda$ to a set of negative values between $-1$ and $0$; Set $\epsilon$ to a small positive value; Set $dMin$ and $dMax$ to positive integer respectively Pick $\lambda$ and $d$. For each $\lambda \in \Lambda$ and for each $d$ from $dMin$ to $dMax$, perform step \[loopstart\]–\[loopend\] until a barrier certificate is found Decide the degree of $\mu(x)$ and $\eta(x)$ according to $d$. To be *SOS*s for both and , at least one of the degrees of $\mu(x) I_0(x)$ and $\eta(x) U(x)$ is greater than or equal to the degree of $\varphi(x)$\[loopstart\] Generate complete polynomials $\varphi(x)$, $\mu(x)$ and $\eta(x)$ of specified degree with unknown coefficient variables Eliminate the monomials of odd top degrees in –, $\mu(x)$ and $\eta(x)$, respectively. To be a *SOS*, a polynomial has to be of even degree. Concretely, let the coefficients of the monomials to be eliminated be zero to get equations about coefficient variables and then reduce the number of coefficient variables by solving the equations and substituting free variables for non-free variables in all the related polynomials\[eliminatemon\] Perform the *SOS* programming on the positive semidefinite constraints – and $\mu(x)$, $\eta(x)$\[programming\] Check if a feasible solution is found, if not found, continue with a new loop; else, check if the solution can indeed enable the corresponding polynomials to be *SOS*s, if so, return $\varphi(x)$; else, for all the polynomials in the programming, eliminate all the monomials whose coefficients have too small absolute values(usually less than $10^{-5}$) by using the same method as step \[eliminatemon\], then go to step \[programming\] unless an empty polynomial is produced\[loopend\]]{}
We use *Algorithm* \[algorithm1\] to compute the desired barrier certificate. In the algorithm, we first choose a small set of negative values $\Lambda$ as a candidate set for $\lambda$ and an integer interval $[dMin,dMax]$ as a candidate set for degree $d$ of $\varphi(x)$. Then, we attempt to find a barrier certificate satisfying the conditions – for a fixed pair of $\lambda$ and $d$ until such one is found. Theoretically, according to the analysis about the dependence of conservativeness of barrier certificate on the value of $\lambda$, we should set $\lambda$ to as small negative value as possible. However, experiments show that too small negative numbers for $\lambda$ often lead the semidefinite programming function to numerical problems. In practice, the negative values in the interval $[-1,0)$ are good enough for $\lambda$ to verify very critical safety properties. Note that the principle for step \[loopstart\] in *Algorithm 1* is that if $\varphi(x)$ has a dominating degree in both polynomials, there couldn’t exist a solution that make both polynomials be *SOS*s because $-\varphi(x)$ and $\varphi(x)$ occur in and simultaneously. The motive for eliminating the monomials with small coefficients in step \[loopend\] is from the observation that those monomials are usually the cause of the failed *SOS* decomposition for the polynomials when the semidefinite programming function gives a seemingly feasible solution.
The idea for constructing barrier certificates for continuous systems can be easily extended to hybrid systems. We describe it in the following subsection.
Sum-of-squares Transformation for Hybrid System {#subsec:hybridmethod}
-----------------------------------------------
Similar to continuous system, in order to be solvable by semidefinite programming, we need to limit the hybrid system model in Section \[sec:Preliminaries\] to semialgebraic hybrid system.
Consider the hybrid system $\mathbb{H} = \langle L,X,E,R,G,I,F\rangle$, where the mappings $F,R,G,I$ of $\mathbb{H}$ are defined with respect to polynomial inequalities as follows:
- $F:l\mapsto~f_l(x)$
- $G:(l,l')\mapsto\{x\in\mathbb{R}^n|G_{ll'}(x)\geq~0, G_{ll'}(x)\in\mathbb{R}[x]^{p_{ll'}}\}$
- $R:(l,l',x)\mapsto~\{x'\in\mathbb{R}^n|R_{ll'x}(x')\geq~0, R_{ll'x}(x')\in\mathbb{R}[x]^{q_{ll'}}\}$
- $I: l\mapsto\{x\in\mathbb{R}^n|I_l(x)\geq~0, I_l(x)\in\mathbb{R}[x]^{r_l}\}$
and the mappings of the initial set and the unsafe set are defined as follows:
- $\operatorname{Init}:l\mapsto\{x\in\mathbb{R}^n| \operatorname{Init}_l(x)\geq 0, \operatorname{Init}_l(x)\in\mathbb{R}[x]^{s_l}\}$
- $\operatorname{Unsafe}:l\mapsto\{x\in\mathbb{R}^n| \operatorname{Unsafe}_l(x)\geq 0, \operatorname{Unsafe}_l(x)\in\mathbb{R}[x]^{t_l}\}$
where $p_{ll'}$, $q_{ll'}$, $r_l$, $s_l$ and $t_l$ is the dimension of polynomial vector space. Then we have the following corollary for constructing barrier certificate for the semialgebraic hybrid system $\mathbb{H}$.
\[corollary3\] Let the hybrid system $\mathbb{H}$ and the initial state set mapping $Init$ and the unsafe state set mapping $Unsafe$ be defined as the above. Then, for any given set of constant real numbers $S_{\lambda}=\{\lambda_l\in\mathbb{R}| l\in~L\}$ and any given set of constant non-negative real numbers $S_{\gamma}=\{\gamma_{ll'}\in\mathbb{R}_+| (l,l')\in~E\}$ ,and any given small real number $\epsilon>0$, if there exists a set of polynomial functions $\{\varphi_l(x)\in\mathbb{R}[x]| l\in~L\}$ and five sets of *SOS* polynomial vectors $\{\mu_l(x)\in\mathbb{R}[x]^{s_l}| l\in~L\}$, $\{\theta_l(x)\in\mathbb{R}[x]^{r_l}| l\in~L\}$, $\{\kappa_{ll'}(x) \in \mathbb{R}[x]^{p_{ll'}}| (l,l') \in E\}$ , $\{\sigma_{ll'}(x)\in\mathbb{R}[x]^{q_{ll'}}| (l,l')\in~E\}$ and $\{\eta_l(x)\in\mathbb{R}[x]^{t_l}| l\in~L\}$, such that the polynomials $$\begin{aligned}
&\varphi_l(x) - \mu_l(x) \operatorname{Init}_l(x)\label{contCond1}\\
&\lambda_l \varphi_l(x) - \mathcal{L}_{f_l}\varphi_l(x)- \theta_l(x) I_l(x)\label{contCond2}\\
&\gamma_{ll'} \varphi_l(x) - \varphi_{l'}(x') -\kappa_{ll'}(x) G_{ll'}(x)-\sigma_{ll'}(x') R_{ll'x}(x') \label{contCond3}\\
&\varphi_l(x) - \epsilon - \eta_l(x) \operatorname{Unsafe}_l(x)\label{contCond4}\end{aligned}$$ are *SOS*s for all $l\in L$ and $(l,l')\in E$, then the safety property is satisfied by the system $\mathbb{H}$.
Similar to Corollary \[corollary1\], it’s easy to prove that any set of polynomials $\{\varphi_l(x)\}$ satisfying – also satisfies –, hence the hybrid system $\mathbb{H}$ is safe.
The algorithm for computing the barrier certificates for hybrid systems is similar to the algorithm for continuous systems except that it needs to take into account the constraint for the discrete transitions. We do not elaborate on it here any more. Note that the strategy for the selection of $\lambda$’s for continuous system applies here as well and we only need to set all the elements of $S_\gamma$ to $1$ except for the discrete transition whose post-state is independent of the pre-state, where we set $\gamma_{ll'}$ to $0$ to reduce the computational complexity.
Examples {#sec:examples}
========
Example 1 {#subsec:example1}
---------
Consider the two-dimensional system (from [@khalil2002nonlinear] page 315) $$\begin{gathered}
\label{system1}
\begin{bmatrix}\dot{x_1} \\ \dot{x_2}\end{bmatrix} = \begin{bmatrix} x_2\\ -x_1 + \frac{1}{3}x_1^3 - x_2\end{bmatrix}\end{gathered}$$ with $\mathcal{X} = \mathbb{R}^2$, we want to verify that starting from the initial set $X_0 = \{x \in \mathbb{R}^2| (x_1-1.5)^2 + x_2^2 \leq 0.25\}$, the system will never evolve into the unsafe set $X_u = \{x \in \mathbb{R}^2| (x_1 + 1)^2 + (x_2 + 1)^2 \leq 0.16\}$. We attempted to use both the method based on *Convex Condition* proposed in [@prajna2007framework] and the method based on *Exponential Condition* in this paper to find the barrier certificates with a degree ranging from $2$ to $10$. (Note that in [@gulwani2008constraint], [@taly2009deductive], the inductive invariants are not sufficient in general according to [@taly2011synthesizing] and hence cannot be applied to our examples. The work of \[19\] applies only to a very special class of hybrid systems which is not applicable to our examples either.) During this process, all the programming polynomials are complete polynomials automatically generated (instead of the non-complete polynomials consisting of painstakingly chosen terms) and all the computations are performed in the same environment. The result of the experiment is listed in Table \[tbl:comparison\]. The first column is the degree of the barrier certificate to be found, the second column is the amount of time spent by the method based on *Convex Condition*, and the rest columns are the amount of time spent by the method based on *Exponential Condition* for different value of $\lambda$. Note that the symbol $\times$ in the table indicates that the method failed to find a barrier certificate with the corresponding degree either because the semidefinite programming function found no feasible solution or because it ran into a numerical problem.
\[tbl:comparison\]
-------------- ---------- ------------------------ ------------------------ --------------
of
$\varphi(x)$ $\lambda=\frac{-1}{8}$ $\lambda=\frac{-1}{4}$ $\lambda=-1$
2 $\times$ 0.4867 0.4836 0.2496
3 $\times$ 0.5444 0.6224 0.4976
4 0.4368 0.4103 0.4072 0.3853
5 $\times$ 0.4321 0.4103 0.3947
6 $\times$ 0.3214 0.3011 0.2714
7 $\times$ 0.9563 0.9532 0.9453
8 $\times$ 0.9188 0.8970 0.7893
9 $\times$ 1.4944 1.4149 1.5132
10 $\times$ 1.4336 1.3931 1.3650
-------------- ---------- ------------------------ ------------------------ --------------
: Computing results for *Convex Condition* and *Exponential Condition*. *Exponential Condition* shows much stronger capability in finding barrier certificates.
As shown in Table \[tbl:comparison\], the method based on *Convex Condition* succeeded only in one case ($Degree = 4$) due to the conservativeness of *Convex Condition*. Comparably, our method found all the barrier certificates of the specified degrees ranging from $2$ to $10$. Especially, the lowest degree of barrier certificate we found is quadratic: $\varphi(x) = -.86153-.87278x_1-1.1358x_2-.23944x_1^2-.5866x_1x_2$ with $\mu(x)=0.75965$ and $\eta(x)=0.73845$ when $\lambda$ is set to $-1$. The phase portrait of the system and the zero level set of $\varphi(x)$ are shown in \[fig:example1\]. Note that being able to find a lower degree of barrier certificates is essential in reducing the computational complexity.
In addition, we can see from Table \[tbl:comparison\] that the runtime of *Exponential Condition*-based method decreases with the value of $\lambda$ for each fixed degree except for $Degree=3,9$, this observation can greatly evidence our theoretical result about $\lambda$ selection: the less, the better.
Example 2 {#subsec:example3}
---------
In this example, we consider a hybrid system with two discrete locations (from [@prajna2004safety]). The discrete transition diagram of the system is shown in \[fig:example3\] and the vector fields describing the continuous behaviors are as follows: $$\begin{aligned}
&f_1(x)= \begin{bmatrix}x_2 \\ -x_1+ x_3 \\ x_1+(2x_2+3x_3)(1+x_3^2)\end{bmatrix}, f_2(x)= \begin{bmatrix}x_2 \\ -x_1+x_3 \\ -x_1-2x_2-3x_3\end{bmatrix}\end{aligned}$$ At the beginning, the system is initialized at some point in $X_0 = \{x \in \mathbb{R}^3 | x_1^2 + x_2^2 + x_3^2 \leq 0.01\}$ and then it starts to evolve following the vector fields $f_1(x)$ at location 1(NO CONTROL mode). When the system reaches some point in the guard set $G(1,2)=\{x\in \mathbb{R}^3|0.99 \leq x_1^2 + 0.01x_2^2 + 0.01x_3^3 \leq 1.01\}$, it can jump to location 2 (CONTROL mode) nondeterministically without performing any reset operation (i.e., $R(1,2,x)=G(1,2)$). At location 2, the system will operate following the vector field $f_2(x)$, which means that a controller will take over to prevent $x_1$ from getting too big. As the system enters the guard set $G(2,1) = \{x \in \mathbb{R}^3 | 0.03 \leq x_1^2 + x_2^2 + x_3^2 \leq 0.05\}$, it will jump back to location 1 nondeterministically again without reset operation (i.e., $R(2,1,x)=G(2,1)$). Different from the experiment in [@prajna2004safety], where the objective is to verify that $|x_1| < 5.0$ in CONTROL mode, our objective is to verify that $x_1$ will stay in a much more restrictive domain in CONTROL mode: $|x_1| < 3.2$.
We define the unsafe set as $\operatorname{Unsafe}(1)=\emptyset$ and $\operatorname{Unsafe}(2)=\{x\in \mathbb{R}^3| 3.2 \leq |x_1| \leq 10\}$, which is sufficient to prove $|x_1|< 3.2$ in CONTROL mode. Similarly, we tried to use both the method in this paper and the method in [@prajna2007framework] to compute the barrier certificate. By setting $\lambda_1=\lambda_2=-\frac{1}{5}$ and $\gamma_{12}=\gamma_{21}=1$, our method found a pair of quartic barrier certificate functions: $\phi_1(x)$ and $\phi_2(x)$, whose zero level set is shown in \[fig:example31\_phi1\] and \[fig:example31\_phi2\] respectively. As you can see, at each location $l=1,2$, the zero level set of $\phi_l(x)$ forms the boundary of the over-approximation $\phi_l(x) \leq 0$ (denoting the points within the pipe) for the reachable set at location $l$. On the one hand, the hybrid system starts from and evolves within the corresponding over-approximation and jumps back and forth between the two over-approximations. On the other hand, the unsafe set does not intersect the over-approximation formed by $\phi_2(x)\leq 0$ (see \[fig:example31\_phi2\_uns\]). Therefore, the safety of the system is guaranteed. However, using the method in [@prajna2007framework], we cannot compute the barrier certificate, which means it cannot verify the system.
Conclusion {#sec:conclusion}
==========
In this paper, we propose a new barrier certificate condition (called *Exponential Condition*) for the safety verification of continuous systems and hybrid systems. Our barrier certificate condition is parameterized by a real number $\lambda$ and the conservativeness of the barrier certificate condition depends closely on the value of $\lambda$: the less value the $\lambda$ is, the less conservative the barrier certificate condition is. Specifically, *Convex Condition* is just the special case of *Exponential Condition* with $\lambda=0$. Therefore, we can obtain the barrier certificate condition that is less conservative than *Convex Condition* as long as we set $\lambda$ to a negative value. The most important benefit of *Exponential Condition* is that it possesses a relatively low conservativeness as well as the convexity and hence can be solved efficiently by semidefinite programming method.
Based on our method, we are able to construct polynomial barrier certificate to verify very critical safety property for semialgebraic continuous systems and hybrid systems. The experiments on a continuous system and a hybrid system show the effectiveness and practicality of our method.
[10]{}
R. Alur, C. Courcoubetis, N. Halbwachs, T. Henzinger, P. Ho, X. Nicollin, A. Olivero, J. Sifakis, and S. Yovine. The algorithmic analysis of hybrid systems. , 138(1):3–34, 1995.
S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. , volume 15. Society for Industrial Mathematics, 1994.
L. Carloni, R. Passerone, and A. Pinto. Languages and tools for hybrid systems design. , 1(1-2), 2006.
S. Gulwani and A. Tiwari. Constraint-based approach for analysis of hybrid systems. In [*Computer Aided Verification*]{}, pages 190–203. Springer, 2008.
T. Henzinger. The theory of hybrid automata. In [*Proc. IEEE Symp. Logic in Computer Science (LICS)*]{}, pages 278–292, 1996.
M. Jirstrand. Invariant sets for a class of hybrid systems. In [*Proc. IEEE Conference on Decision and Control*]{}, volume 4, pages 3699–3704, 1998.
H. K. Khalil. . Prentice Hall, 3rd edition, 2001.
J. Lasserre. Sufficient conditions for a real polynomial to be a sum of squares. , 89(5):390–398, 2007.
J. Lygeros, C. Tomlin, and S. Sastry. Controllers for reachability specifications for hybrid systems. , 35(3):349–370, 1999.
O. Maler, Z. Manna, and A. Pnueli. Prom timed to hybrid systems. In [*Real-time: theory in practice*]{}, pages 447–484. Springer, 1992.
P. Parrilo. Semidefinite programming relaxations for semialgebraic problems. , 96(2):293–320, 2003.
A. Platzer and E. Clarke. Computing differential invariants of hybrid systems as fixedpoints. In [*Computer Aided Verification*]{}, pages 176–189. Springer, 2008.
S. Prajna and A. Jadbabaie. Safety verification of hybrid systems using barrier certificates. , pages 271–274, 2004.
S. Prajna, A. Jadbabaie, and G. Pappas. A framework for worst-case and stochastic safety verification using barrier certificates. , 52(8):1415–1428, 2007.
S. Prajna, A. Papachristodoulou, P. Seiler, and P. Parrilo. Sostools and its control applications. , pages 580–580, 2005.
E. Rodr[í]{}guez-Carbonell and A. Tiwari. Generating polynomial invariants for hybrid systems. , pages 590–605, 2005.
S. Sankaranarayanan. Automatic invariant generation for hybrid systems using ideal fixed points. In [*Proc. ACM International Conference on Hybrid Systems: Computation and Control*]{}, pages 221–230, 2010.
S. Sankaranarayanan, H. Sipma, and Z. Manna. Constructing invariants for hybrid systems. , pages 69–77, 2004.
C. Sloth, G. Pappas, and R. Wisniewski. Compositional safety analysis using barrier certificates. In [*Proc. ACM international conference on Hybrid Systems: Computation and Control*]{}, pages 15–24, 2012.
J. Sturm. Using sedumi 1.02, a matlab toolbox for optimization over symmetric cones. , 11(1-4):625–653, 1999.
A. Taly, S. Gulwani, and A. Tiwari. Synthesizing switching logic using constraint solving. , 13(6):519–535, 2011.
A. Taly and A. Tiwari. Deductive verification of continuous dynamical systems. In [*FSTTCS*]{}, volume 4, pages 383–394, 2009.
A. Tiwari and G. Khanna. Nonlinear systems: Approximating reach sets. , pages 171–174, 2004.
[^1]: This work was supported by the Chinese National 973 Plan under grant No. 2010CB328003, the NSF of China under grants No. 61272001, 60903030, 91218302, the Chinese National Key Technology R&D Program under grant No. SQ2012BAJY4052, and the Tsinghua University Initiative Scientific Research Program.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The LHC has delivered already 10 ${\rm fb}^{-1}$ of proton proton collisions at a centre-of-mass energy of 7-8 TeV. With this data set, ATLAS and CMS have discovered a new boson at a mass of about 125 GeV and have searched for new physics at the TeV scale.'
author:
- Gigi Rolandi
bibliography:
- 'pa-bib.bib'
title: 'LHC Results - Highlights'
---
Introduction
============
The LHC [@Evans:2008zzb] performs well above expectations with a peak luminosity of $4\, 10^{33}\, {\rm cm}^{-2} {\rm s}^{-1}$ from collisions of two 3.5 TeV proton beams in 2011 and two 4 TeV beams in 2012 . In 2011, the LHC delivered about 6 ${\rm fb}^{-1}$ to ATLAS [@Aad:2008zzm] and CMS [@ref:CMS]. With this large integrated luminosity, it is possible to search effectively for the Standard Model (SM) Higgs boson and to probe the existence of new physics at the TeV scale. This luminosity has been reached with an inter bunch spacing of 50 ns and more than $1.2\;10^{11}$ protons per bunch. At these currents the average number of inelastic interactions per bunch crossing (pile-up) is about 15, hence posing new challenges to trigger, event reconstruction, and quality of reconstructed physics objects.
New algorithms have been designed to mitigate the effect of the pileup. In spite of the difficult experimental conditions, ATLAS and CMS have been able to calibrate quickly their data and to deliver new physics results shortly after the start of data taking. They are producing physics papers at a rate of about 100 papers per year per experiment, probing the Standard Model and searching for new physics. The main physics messages of the analyses of the data collected up to now are:
- the Standard Model is still in excellent shape;
- a new boson [@Aad:2012gk; @Chatrchyan:2012gu] has been found with properties compatible with those predicted for the SM Higgs Boson;
- no sign of new physics has been found yet.
In this lecture, I will concentrate on two topics : search for Supersymmetry and search for the Higgs boson.
Supersymmetry
=============
Supersymmetry (SUSY) [@ref:SUSY-1; @ref:SUSY0; @ref:SUSY1; @ref:SUSY2; @ref:SUSY3; @ref:SUSY4; @ref:hierarchy1; @ref:hierarchy2] is a well motivated extension of the SM. It introduces a large number of new particles with the same quantum numbers as their SM partners, but differing by half a unit of spin.
With R-parity conservation [@Farrar:1978xj], the supersymmetric particles, such as squarks and gluinos, are produced in pairs and decay to the lightest, stable supersymmetric particle (LSP). If the LSP is neutral and weakly interacting, a typical signature is a final state of multi-jets and possibly leptons accompanied by large Missing Transverse Energy (MET).
The cross sections for producing SUSY particles are shown in Fig. \[fig:prospino\] as a function of the mass of the particles. In quark and gluons collisions it is easy to produce coloured objects like gluinos and squarks, which decay typically to jets and MET, while the cross sections for Electroweak productions are smaller and the mass reach substantially reduced. These “ewkinos” decays typically produce many leptons and MET.
![Cross sections for producing SUSY particles in pp collisions at $\sqrt{s}=$ 7 TeV computed with PROSPINO[@Beenakker:1996ch; @Beenakker:1997ut; @Beenakker:1999xh][]{data-label="fig:prospino"}](prospino_xsections.png){width="40.00000%"}
Both ATLAS [@Aad201267; @ATLASJetMET; @ATLAS2011; @Aad:2012hm; @Aad:2012rz] and CMS [@RA2paper35pb-1; @RAZORpaper35pb-1; @AlphaTPaper35pb-1; @AlphaTPaper1fb-1; @Chatrchyan:2012jx] presented many hadronic SUSY searches on the 7 TeV data based directly or indirectly on MET. These searches can be interpreted in many ways. Simplified versions of SUSY, with a drastic reduction of the more than 100 parameter space like CMSSM [@ref:CMSSM] or mSugra [@ref:MSUGRA], are excluded for gluinos and squarks below about 1 TeV and are now cornered. The searches can also be interpreted in terms of simplified models [@Alves:2011wf] where a single decay chain is considered with the assumption that the branching fractions along this chain are 100%.
Figure \[fig:sms\] shows two examples. The models assumed here are i) gluino pair production when squarks are much heavier than gluinos and the gluinos decay to two light quarks and a neutralino (left) and ii) squark gluino associated production with gluino decaying into a quark pair and a neutralino and the squark decaying into quark neutralino. The neutralino here is assumed to be massless (right).
[cc]{} ![Upper limit [@Chatrchyan:2012jx] on the cross section for gluino pair production when squarks are much heavier than gluinos and the gluinos decay to two light quarks and a neutralino (left). Upper limit [@Aad:2012rz] on squark-gluino associated production with the gluino decaying into a quark pair and a neutralino and the squark decaying into quark neutralino. The neutralino here is assumed to be massless(right).[]{data-label="fig:sms"}](SMS_mt2_1.png "fig:"){width="48.00000%"} & ![Upper limit [@Chatrchyan:2012jx] on the cross section for gluino pair production when squarks are much heavier than gluinos and the gluinos decay to two light quarks and a neutralino (left). Upper limit [@Aad:2012rz] on squark-gluino associated production with the gluino decaying into a quark pair and a neutralino and the squark decaying into quark neutralino. The neutralino here is assumed to be massless(right).[]{data-label="fig:sms"}](SMS_ATLAS.png "fig:"){width="42.00000%"}
While generic SUSY production at the scale of about 1 TeV is not compatible with data, there is still room for natural models [@Barbieri:1987fn; @Papucci:2011wy] where gluinos and third-generation squarks are below 1 TeV. Both ATLAS and CMS have done specific searches for the third-generation squarks. Examples are the search for same-sign lepton pairs, b jets and MET by CMS [@Chatrchyan:2012sa] - motivated by final states including four top or two top and two W as shown in Fig. \[fig:sms1\] (a) and (b) - and the search for three b-jets and MET by ATLAS [@Aad:2012pq] addressing models shown in Fig \[fig:sms1\] (c) and (d). Broadly speaking, these searches exclude gluinos of 900 GeV for third generation squarks lighter than 300-400 GeV.
[cccc]{} ![Simplified models used for interpretations of searches for production of third generation squarks [@Chatrchyan:2012sa; @Aad:2012pq][]{data-label="fig:sms1"}](SMS1.png "fig:"){width="22.00000%"} & ![Simplified models used for interpretations of searches for production of third generation squarks [@Chatrchyan:2012sa; @Aad:2012pq][]{data-label="fig:sms1"}](SMS2.png "fig:"){width="25.00000%"} & ![Simplified models used for interpretations of searches for production of third generation squarks [@Chatrchyan:2012sa; @Aad:2012pq][]{data-label="fig:sms1"}](SMS3.png "fig:"){width="22.00000%"} & ![Simplified models used for interpretations of searches for production of third generation squarks [@Chatrchyan:2012sa; @Aad:2012pq][]{data-label="fig:sms1"}](SMS4.png "fig:"){width="22.00000%"}\
(a)&(b)&(c)&(d)
Both ATLAS and CMS have also performed a model-independent search for Weakly Interacting Massive Particles (WIMP) production [@Aad:2011xw; @Chatrchyan:2012me] triggering on events with a monojet and MET. Here the jet is produced by initial state radiation of one of the interacting partons and the two WIMPs escape undetected leading to spectacular events like the one shown in Fig. \[fig:cms\_wimp\]: the detector is empty with the exception of a single high energy jet. In the SM these events are produced by a high transverse momentum Z decaying into a neutrino pair. This background can be effectively measured from similar events where the Z decays into a muon pair.
[r]{}[0.5]{}
{width="40.00000%"}
The measured cross section is compatible with the SM background and limits can be set on generic WIMP production. For dark matter models, the observed limit on the cross section depends on the mass of the dark matter particle and the nature of its interaction with the SM particles. The limits on the effective contact interaction scale as a function of the wimp mass can be translated into a limit on the dark matter-nucleon scattering cross section[@bib:RoniHarnik]. These limits can be compared with the constraints from direct and indirect detection experiments. The LHC limits [@Chatrchyan:2012me] are competitive with those from direct WIMP search for the spin-dependent interaction for WIMP mass below few hundreds GeV, and also for the spin-independent interaction for WIMP mass below 10 GeV .
Search for the Higgs boson
==========================
Three weeks after I gave this lecture in Anjoux, ATLAS [@Aad:2012gk] and CMS [@Chatrchyan:2012gu] announced the observation of a new boson compatible with the SM Higgs boson in a CERN seminar. In this section, with the agreement of the organizers of the school, I describe the observation of the boson.
The search for the Higgs boson (H) and the justification of the electroweak symmetry breaking [@Englert:1964et; @Higgs:1964pj] was one of the main reasons for the construction of the Large Hadron Collider. This search was therefore a high priority analysis for ATLAS and CMS. In the SM the cross section for H production in proton proton collision at 7 TeV is about 17 pb [@LHCHiggsCrossSectionWorkingGroup:2011ti] for ${\rm m}_{\rm H} \simeq 125$ GeV and 30% higher at 8 TeV. In about 10 fb$^{-1}$ of integrated luminosity, shared about equally between the two energies, some 200,000 Higgs bosons are produced in each experiment. It is very difficult however to separate this signal from the very large background of SM processes, especially the hadronic final states, and specific searches with leptons or photons in the final state are performed. Table \[tab:TH1\] lists the search channels together with the branching ratio expected for a SM Higgs boson.
channel Branching Fraction
---------------------------------------------------------------------- --------------------
${\rm H }\rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $ $1.2\; 10^{-4}$
${\rm H } \rightarrow \gamma \gamma $ $ 2.3\; 10^{-3}$
${\rm H } \rightarrow {\rm WW} \rightarrow \ell \nu \; \ell' \nu' $ $1.0\; 10^{-2}$
${\rm H } \rightarrow \tau \tau $ $6.0\; 10^{-2}$
${\rm H } \rightarrow {\rm bb} $ $5.8\; 10^{-1}$
: Channels used in the search for the Higgs boson and the branching ratios expected in the SM for ${\rm m}_{\rm H} =125$ GeV. Here $\ell$ and $\ell'$ indicate an electron or a muon.[]{data-label="tab:TH1"}
A value of ${\rm m}_{\rm H}\simeq 125$ GeV is smaller than the sum of the masses of the vector boson (V) pairs, in the decays ${\rm H} \rightarrow {\rm VV}$ one or both V are off mass-shell. In ${\rm H}\rightarrow 4\ell$ and ${\rm H} \rightarrow \gamma \gamma $ the H mass is reconstructed with high resolution (1-2%) with the precise measurements of the momenta of leptons and photons. Because the width of the SM Higgs boson for ${\rm m}_{\rm H} \simeq 125$ GeV is a few MeV, one expects to see a narrow peak dominated by the instrumental resolution. The other channels have worse mass resolution because of the missing neutrinos in the W and $\tau$ decays or because the b-jets are reconstructed with some 10% resolution. The search in the bb final state is performed in the associate production ${\rm VH}\rightarrow {\rm Vbb}$ where the V decays into leptons that provide the trigger and reduce the overwhelming hadronic background.
[r]{}[0.5]{}
{width="40.00000%"}
Figure \[fig:expected\] displays the expected discovery potential for a SM Higgs boson in the CMS experiment as a function of ${\rm m}_{\rm H}$ (ATLAS have similar figures). The probability that the background can produce a fluctuation greater than the potential excess produced in data by a SM Higgs boson (the so called local p-value) is estimated less than $10^{-8}$ i.e. more than five standard deviations. The most sensitive channels are those where a narrow peak can be observed: ${\rm H}\rightarrow 4\ell$ and ${\rm H}\rightarrow \gamma \gamma$. The WW channel is also quite sensitive.
${\rm H} \rightarrow {\rm WW} \rightarrow \ell \nu \; \ell' \nu' $
------------------------------------------------------------------
The decay mode ${\rm H} \rightarrow {\rm WW} \rightarrow \ell \nu \; \ell' \nu' $ is the main search channel for a SM Higgs boson with mass above the WW threshold of 160 GeV. With good experimental control of the MET and very tight lepton identification, it is possible to reject large part of the reducible background and extend the sensitivity down to ${\rm m}_{\rm H} \simeq 120$ GeV. The signature is two isolated opposite-sign charged leptons and large MET caused by the two undetected neutrinos. The most sensitive channel is when the two leptons have opposite flavour and there are no extra jets in the event. Here the main background is the irreducible non resonant WW production and the reducible W+jet production when the jet fakes a lepton. The other channels with same flavour leptons or with associated jets have larger backgrounds from Drell-Yan and top quark decay respectively and contribute less than 20% to the sensitivity. The yields of the largest backgrounds are estimated from control regions. One important variable to separate the signal from the irreducible background is the angle between the two leptons. Due to spin correlations, this variable is small for W pairs from the spin-0 H decay and large for the WW non resonant production [@Dittmar:1996ss]. In ATLAS the transverse mass of the MET vector and the di-lepton system - shown in Fig. \[fig:ww\] a - is used to test for the presence of a signal for all jet multiplicities. In CMS, the signal is separated from the background with kinematical and topological requirements optimized for each mass hipothesis; one of the most sensitive variables is the dilepton invariant mass shown in Fig. \[fig:ww\] b.
[cc]{} ![ ${\rm H} \rightarrow {\rm WW}$ analysis. a) ATLAS, Distribution of the transverse mass, in the 0-jet and 1-jet analyses with opposite flavour, for events satisfying all selection criteria. The expected signal for ${\rm m}_{\rm H}$ = 125 GeV is shown added to the background prediction. The hashed area indicates the total uncertainty on the background prediction. b) CMS, Distribution of dilepton invariant mass for the zero-jet opposite flavour category at 8 TeV after the full selection, except for the selection on $m_{\ell \ell}$ itself. The signal expected from a Higgs boson with a mass ${\rm m}_{\rm H}$ = 125 GeV is shown added to the background.[]{data-label="fig:ww"}](ATLAS_ww_mt.png "fig:"){width="48.00000%"} & ![ ${\rm H} \rightarrow {\rm WW}$ analysis. a) ATLAS, Distribution of the transverse mass, in the 0-jet and 1-jet analyses with opposite flavour, for events satisfying all selection criteria. The expected signal for ${\rm m}_{\rm H}$ = 125 GeV is shown added to the background prediction. The hashed area indicates the total uncertainty on the background prediction. b) CMS, Distribution of dilepton invariant mass for the zero-jet opposite flavour category at 8 TeV after the full selection, except for the selection on $m_{\ell \ell}$ itself. The signal expected from a Higgs boson with a mass ${\rm m}_{\rm H}$ = 125 GeV is shown added to the background.[]{data-label="fig:ww"}](CMS_ww_mll.png "fig:"){width="48.00000%"}
\
(a)&(b)
In the most sensitive channel, leptons with opposite flavour and zero jets, CMS observe 158 events, estimate a background of $124 \pm 12$ events and expect $24\pm 5 $ events for a SM Higgs boson of 125 GeV. Similarly ATLAS observe 185 events, estimate a background of $142 \pm 16$ events and expect $20\pm 4$ events for a SM Higgs boson of 125 GeV.
${\rm H} \rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $
---------------------------------------------------------------------
The decay mode ${\rm H} \rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $ , the so called golden channel, is characterized by a small signal yield over a flat irreducible background of direct ZZ production. Because the signal yield is small, it is important to maximize the efficiency lowering as much as possible the threshold on the lepton transverse momenta: at ${\rm m}_{\rm H}= 125$ GeV the average transverse momentum of the softest lepton is about 7 GeV. Also the lepton identification is relaxed in order to maximize the efficiency. The reducible background is evaluated with control regions and is small, in spite of the relaxed identification criteria, thanks to the presence of four leptons in the final state.
In the ATLAS analysis the thresholds for muons and electrons are 6 and 7 GeV. In CMS, they are 5 and 7 GeV. The events are selected pairing opposite charge and same flavour leptons and with requirements on the invariant masses of the pairs. A recovery of the final state radiation photons is also used in CMS . The expected background yield of the irreducible background is estimated using the MC simulation normalized to the theoretical cross section for ZZ production. The identification efficiency, the energy scale and the energy resolution are measured using large samples of Z , Y and J/$\psi$ decaying into two leptons.
[cc]{} ![ ${\rm H} \rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $. Distribution of the four-lepton invariant mass for the selected candidates, compared to the background expectation. The signal expectation for a SM Higgs boson with ${\rm m}_{\rm H}=125$ GeV is also shown. (a) ATLAS, (b) CMS; the inset shows the distribution after selection of events with KD > 0.5, as described later in the text.[]{data-label="fig:4leptons"}](ATLAS_4l_mass.png "fig:"){width="48.00000%"} & ![ ${\rm H} \rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $. Distribution of the four-lepton invariant mass for the selected candidates, compared to the background expectation. The signal expectation for a SM Higgs boson with ${\rm m}_{\rm H}=125$ GeV is also shown. (a) ATLAS, (b) CMS; the inset shows the distribution after selection of events with KD > 0.5, as described later in the text.[]{data-label="fig:4leptons"}](CMS_4l_mass.png "fig:"){width="48.00000%"}
\
(a)&(b)
Figure \[fig:4leptons\] shows the invariant mass distribution of the selected events compared to the estimated background. The peak at about 90 GeV is the decay of the Z into four leptons, where a lepton pair is radiated by one of the leptons originating from the Z decay. In this process the distribution of the lowest momentum lepton is softer than in the Higgs boson decays. The Z peak has a different yield in the two experiments because of the larger efficiency in CMS for low momentum leptons. The reducible background in CMS is indeed much smaller than the reducible one, in ATLAS they are comparable for ${\rm m}_{\rm H}=125$ GeV.
[cc]{} ![ ${\rm H}\rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $. CMS. The distribution of events selected in the four lepton channels for the kinematic discriminant, KD, versus four lepton mass. Events are marked by symbols. The horizontal error bars indicate the estimated mass resolution. In the left plot the colours show the expected background; in the right plot the colours show the event density expected from a SM Higgs boson of 125 GeV (both in arbitrary units).[]{data-label="fig:mela"}](CMS_mela_bkg.png "fig:"){width="48.00000%"} & ![ ${\rm H}\rightarrow {\rm ZZ} \rightarrow \ell \ell \; \ell' \ell' $. CMS. The distribution of events selected in the four lepton channels for the kinematic discriminant, KD, versus four lepton mass. Events are marked by symbols. The horizontal error bars indicate the estimated mass resolution. In the left plot the colours show the expected background; in the right plot the colours show the event density expected from a SM Higgs boson of 125 GeV (both in arbitrary units).[]{data-label="fig:mela"}](CMS_mela_signal.png "fig:"){width="48.00000%"}
In the four-lepton mass region 121.5-130.5 GeV CMS observe 9 events, estimate a background of $3.8 \pm 0.5$ events and expect $7.5\pm 0.9 $ events for a SM Higgs boson of 125 GeV. In the four-lepton mass region 120-130 GeV ATLAS observe 13 events, estimate a background of $4.8 \pm 0.2$ events and expect $5.3\pm 0.5$ events for a SM Higgs boson of 125 GeV.
The scalar nature of the Higgs boson provides important discriminating power between the signal and the irreducible background. The kinematics of the ${\rm ZZ} \rightarrow \ell \ell \; \ell' \ell'$ process is fully described by five angles and the invariant masses of the two lepton pairs [@Cabibbo:1965zz; @Gao:2010qx; @DeRujula:2010ys] for a fixed invariant mass of the four-lepton system. In CMS, a kinematic discriminant (KD) is constructed based on the probability ratio of the signal and background hypotheses as described in Ref. [@Chatrchyan:2012sn]. Figure \[fig:mela\] shows the the distribution of four leptons invariant mass versus KD for the selected events. The discriminant KD takes large values for signal like events and small values for background like events.
A clustering of events is observed with a high value of the kinematic discriminant at ${\rm m}_{\rm H} \simeq 125$ GeV where the background expectation is low, corresponding to the excess seen in the one-dimensional mass distribution.
${\rm H} \rightarrow \gamma \gamma $
------------------------------------
The decay mode ${\rm H} \rightarrow \gamma \gamma $ is characterized by a narrow peak in the diphoton invariant mass distribution above a large irreducible background from QCD production of two photons and a reducible background where one - and in few cases two - reconstructed photons originate from misidentification of jets. The resolution in the invariant mass of the two photons varies on event by event basis, depending on the properties of the reconstructed photons and of the overall event properties (e.g. number of reconstructed vertices). In order to improve the sensitivity of the analysis the events are categorized in exclusive sets with varying signal purity. The events with two reconstructed jets belong to separate categories to exploit the better signal to background ratio in the H production via Vector Boson Fusion (VBF) where two jets originating from the two scattered quarks are expected with large difference in rapidity.
[cc]{} ![ ${\rm H} \rightarrow \gamma \gamma $. The distributions of the invariant mass of diphoton candidates after all selections are shown together with the result of a fit to the data of the sum of a signal component and a background component. (left) ATLAS. The panel in the middle (c) shows the weighted sample with the weights are explained in the text. The signal is fixed to ${\rm m}_{\rm H} = 126.5$ GeV and the background is described by a fourth-order Bernstein polynomial . The residuals of the weighted data with respect to the fitted background component are displayed in (d). The panel on the top (b) shows the residuals of the un-weighted data with respect to their fitted background. (right) CMS. The panel shows the weighted sample where the background is fitted with a fifth order polynomial . The coloured bands represent the $\pm1$ and $\pm2$ deviation uncertainties on the background estimate. The inset shows the central part of the unweighted invariant mass distribution []{data-label="fig:mgg"}](ATLAS_mgg.png "fig:"){width="48.00000%"} & ![ ${\rm H} \rightarrow \gamma \gamma $. The distributions of the invariant mass of diphoton candidates after all selections are shown together with the result of a fit to the data of the sum of a signal component and a background component. (left) ATLAS. The panel in the middle (c) shows the weighted sample with the weights are explained in the text. The signal is fixed to ${\rm m}_{\rm H} = 126.5$ GeV and the background is described by a fourth-order Bernstein polynomial . The residuals of the weighted data with respect to the fitted background component are displayed in (d). The panel on the top (b) shows the residuals of the un-weighted data with respect to their fitted background. (right) CMS. The panel shows the weighted sample where the background is fitted with a fifth order polynomial . The coloured bands represent the $\pm1$ and $\pm2$ deviation uncertainties on the background estimate. The inset shows the central part of the unweighted invariant mass distribution []{data-label="fig:mgg"}](CMS_mgg.png "fig:"){width="48.00000%"}
The diphoton invariant mass is reconstructed from the energies measured by the calorimeter and the position of the primary vertex. In ATLAS, the primary vertex of the hard interaction is identified exploiting the directions of flight of the photons as determined with the longitudinal segmentation of the electromagnetic calorimeter. The CMS calorimeter has no pointing information and the vertex is identified from the kinematic properties of the tracks associated with that vertex and their correlation with the diphoton kinematics.
In CMS, a multivariate regression algorithm is used to extract the photon energy and a photon-by-photon estimate of the uncertainty in that measurement. This information is used in a boosted decision tree (BDT) together with photon and vertex quality variables and kinematic variables. The BDT is trained to separate signal and background events and its output is used to assign the events in an optimal way to four categories with different signal-to-background ratio.
In ATLAS, the non-VBF events are separated in nine categories defined by the rapidity of the photons, the component of the diphoton transverse orthogonal to the axis defined by the difference between the two photon momenta [@Ackerstaff:1997rc; @Vesterinen:2008hx] and the presence or not of converted photons.
The background in each category is estimated from data by fitting the measured diphoton mass spectrum with a background model with free shape and normalization. This background model is chosen by requiring that the potential residual bias is smaller than 20% of the statistical accuracy of the fit. The statistical analysis of the data is done with an unbinned likelihood function in each category.
The distribution of the invariant mass of the diphoton candidates summed an all categories is shown in Fig. \[fig:mgg\]. In order to exploit the different sensitivity of each category, the events are weighted with category-dependent factors reflecting the different signal-to-background ratios. In ATLAS, the weights are defined as ln(1 + S/B) where S is 90% of the expected signal and B is the integral of the background fit in the window containing S. In CMS, the weights proportional to S/(S + B), where S and B are defined \[almost\] as in ATLAS and the weights are normalized such that the integral of the weighted signal model matches the number of signal events given by the best fit. An excess near 125 GeV appears clearly in both the weighted and unweighted distributions.
Statistical analysis
--------------------
A common statistical procedure [@LHC-HCG-Report] for the interpretation of the SM Higgs boson searches has been developed by ATLAS and CMS. Data and background predictions are compared to the expected SM Higgs boson signal and the ratio $\mu$ between the measured and predicted signal strength is evaluated. The background-only hypothesis corresponds to $\mu=0$ while a positive value of $\mu$ significantly different from zero indicates the presence of a signal. The signal expected for a SM Higgs boson is $\mu=1$. At each mass, possible values of $\mu$ are tested with a test statistics based on the profile likelihood ratio [@Cowan:2010st] to extracts the information on the signal strength from a full likelihood fit to the data. The likelihood function includes all the parameters describing the data and all the parameters that describe the systematic uncertainties and their correlations.
The interpretation strategy is based on the modified frequentist criterion [@Junk:1999kv; @Read:2002hq] CLs. A value of $\mu$ is excluded at 95% CL when CLs is less than 5%. For each mass, the value of $\mu$ excluded at 95% CL, $\mu_{95}$, is computed. In practice a scan in steps of a fraction of the mass resolution is done for each channel. Figure \[fig:muexclusion\] shows the value of $\mu_{95}$ as a function of the Higgs boson mass for the $\gamma\gamma$ and $4\ell$ channels.
[c]{} ![ Top: $\gamma\gamma$ channel. Bottom: $4\ell$ channel. Left: ATLAS. Right: CMS. The expected exclusion limit $\mu_{95}$ computed in the background only hypothesis is shown together with its 1$\sigma$ (green) and 2$\sigma$ (yellow) bands as a function of the Higgs boson mass. The black line shows the observed $\mu_{95}$.[]{data-label="fig:muexclusion"}](gg_excluded.png "fig:"){width="80.00000%"}\
![ Top: $\gamma\gamma$ channel. Bottom: $4\ell$ channel. Left: ATLAS. Right: CMS. The expected exclusion limit $\mu_{95}$ computed in the background only hypothesis is shown together with its 1$\sigma$ (green) and 2$\sigma$ (yellow) bands as a function of the Higgs boson mass. The black line shows the observed $\mu_{95}$.[]{data-label="fig:muexclusion"}](4l_excluded.png "fig:"){width="80.00000%"}
In this plot, nearby mass points are correlated with a correlation length that is given by the mass resolution of about 1-2%. The data follow roughly the expected value of $\mu_{95}$ with the exception of the region at ${\rm m}_{\rm H}\simeq 125$ GeV where a peak exceeding the 2$\sigma$ band is observed in all distributions.
The significance of the excess is quantified by the probability for a background fluctuation to be at least as large as the observed excess. This local p-value is shown in Fig \[fig:pvalues\] for the combination of channels presented by ATLAS and CMS. The combination assumes the SM branching fractions.
[cc]{} ![Expected and observed local p-value as function of the Higgs boson mass for the combination of the channels of each experiment. Left: ATLAS, the band shows the $\pm 1\sigma$ statistical fluctuation of the expected p-value. Right: CMS the observed p-values of the 7 TeV and 8 TeV data sets are also shown. []{data-label="fig:pvalues"}](ATLAS_pvalue.png "fig:"){width="48.00000%"} & ![Expected and observed local p-value as function of the Higgs boson mass for the combination of the channels of each experiment. Left: ATLAS, the band shows the $\pm 1\sigma$ statistical fluctuation of the expected p-value. Right: CMS the observed p-values of the 7 TeV and 8 TeV data sets are also shown. []{data-label="fig:pvalues"}](CMS_pvalue.png "fig:"){width="36.00000%"}
At ${\rm m}_{\rm H}\simeq125$ GeV ATLAS expect $5\sigma$ and observe $6\sigma$ while CMS expect $6\sigma$ and observe $5\sigma$. A clear signal is established in each experiment. The combination of the two experiments has a global significance in large excess of the $5\sigma$ value that is the value conventionally required for claiming an observation. Since this new resonance decays to two photons, it must be a boson with spin different from 1 [@Landau; @Yang]. Its mass can be measured by fitting the signal strength as a function of the mass in the most sensitive channels: $\gamma\gamma$ and $4\ell$. This fit allows the signal strength in each channel to float independently in order to reduce the model dependence. As a result, CMS quotes a mass of $125.3\pm0.4\pm0.5$ GeV and ATLAS quotes a mass of $126.0\pm0.4\pm0.4$ GeV. In both experiments, the main systematic error comes from the energy scales of electrons and photons evaluated from a comparison of data and simulation at the Z peak.
[ccc]{} ![Measurements of the signal strength parameter $\mu$ for for the individual channels.The horizontal bars indicate the $\pm 1 \sigma$ uncertainties on $\mu$ for individual modes; they include both statistical and systematic uncertainties. (a): ATLAS assume ${\rm m}_{\rm H}=126$ GeV. (b): CMS assume ${\rm m}_{\rm H}=125.5$ GeV, the vertical band shows the overall $\mu$ value from a global fit assuming the SM branching fractions. (c) Comparison of the ATLAS and CMS measurements shown in (a) and (b). []{data-label="fig:muchannels"}](ATLAS_muchannels.png "fig:"){width="28.00000%"} & ![Measurements of the signal strength parameter $\mu$ for for the individual channels.The horizontal bars indicate the $\pm 1 \sigma$ uncertainties on $\mu$ for individual modes; they include both statistical and systematic uncertainties. (a): ATLAS assume ${\rm m}_{\rm H}=126$ GeV. (b): CMS assume ${\rm m}_{\rm H}=125.5$ GeV, the vertical band shows the overall $\mu$ value from a global fit assuming the SM branching fractions. (c) Comparison of the ATLAS and CMS measurements shown in (a) and (b). []{data-label="fig:muchannels"}](CMS_muchannels.png "fig:"){width="25.00000%"} & ![Measurements of the signal strength parameter $\mu$ for for the individual channels.The horizontal bars indicate the $\pm 1 \sigma$ uncertainties on $\mu$ for individual modes; they include both statistical and systematic uncertainties. (a): ATLAS assume ${\rm m}_{\rm H}=126$ GeV. (b): CMS assume ${\rm m}_{\rm H}=125.5$ GeV, the vertical band shows the overall $\mu$ value from a global fit assuming the SM branching fractions. (c) Comparison of the ATLAS and CMS measurements shown in (a) and (b). []{data-label="fig:muchannels"}](comb_muchannels.png "fig:"){width="28.00000%"}
\
(a)&(b)&(c)
The best-fit signal strength is fit for each search channel independently at the measured value of the mass of the boson. The results of these fits are shown in Fig \[fig:muchannels\]. There is large consistency between the measurements of the two experiments. The bb and $\tau\tau$ channels, with low sensitivity, are compatible with $\mu=0$ and with $\mu=1$. The WW and ZZ channels have strengths close to $\mu=1$ in both experiments, with ATLAS consistently larger than CMS. The strength of the $\gamma\gamma$ channel is larger than 1 in both experiments: ATLAS measures $\mu=1.8\pm0.5$ and CMS $\mu=1.6\pm0.4$ . The average of the two values gives $\mu=1.7\pm0.3$, some two sigmas above the SM expectation.
[cc]{} ![Best-fit signal strength as a function of the Higgs boson mass hypothesis for the full combination of the 2011 and 2012 data. The band shows the $\pm 1 \sigma$ uncertainty. Left: ATLAS. Right: CMS. []{data-label="fig:mu"}](ATLAS_mu.png "fig:"){width="48.00000%"} & ![Best-fit signal strength as a function of the Higgs boson mass hypothesis for the full combination of the 2011 and 2012 data. The band shows the $\pm 1 \sigma$ uncertainty. Left: ATLAS. Right: CMS. []{data-label="fig:mu"}](CMS_mu.png "fig:"){width="38.00000%"}
An important quantity is the best fit value of the signal strength $\mu$ as a function of the Higgs boson mass for the combination of all search channels in each experiment. The combination assumes the SM branching fractions. This quantity is shown in Fig. \[fig:mu\]. The signal strength is compatible with the SM expectation: ATLAS measures a slight excess $\mu=1.4\pm0.3$ and CMS measures $\mu=0.87\pm0.23$. The average gives $\mu=1.1\pm0.2$.
Conclusions
===========
The Standard Model has passed the first scrutiny by LHC, which probed for the first time the TeV scale with data collected in 2011. The Higgs particle was the main missing block of the SM. The new boson found by ATLAS and CMS in the range of masses preferred by the precision electroweak tests is a spectacular confirmation of the SM framework. Still the SM leaves too many open questions to be considered a complete description of Nature.
Supersymmetry, considered as one of the most natural extensions of the SM, has been tested already with 2011 data. Direct searches exclude constrained SUSY models. The room for natural supersymmetry is quite restricted but scenarios with light stops or sbottoms are still open.
We look forward to the analysis of the new data collected by LHC in 2012 and to the higher energy run that will start after the energy upgrade to 13.5 TeV in 2015.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider nonsymmetric hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of commutative nonsymmetric association schemes. First, we give a characterization of the eigenmatrix of a commutative nonsymmetric association scheme of class $3$ whose Bose–Mesner algebra contains a nonsymmetric hermitian complex Hadamard matrix, and show that such a complex Hadamard matrix is necessarily a Butson-type complex Hadamard matrix whose entries are $4$-th roots of unity. We also give nonsymmetric association schemes $\mathfrak{X}$ of class $6$ on Galois rings of characteristic 4, and classify hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of $\mathfrak{X}$. It is shown that such a matrix is again necessarily a Butson-type complex Hadamard matrix whose entries are $4$-th roots of unity.'
address:
- 'Kobe Gakuin University, Kobe, 650-8586, Japan'
- 'Tohoku University, Sendai, 980-8579, Japan'
author:
- Takuya Ikuta
- Akihiro Munemasa
date: 'December 17, 2017'
title: |
Butson-type complex Hadamard matrices\
and association schemes on Galois rings of characteristic 4
---
[^1]
Introduction {#sec.Intro}
============
A complex Hadamard matrix is a square matrix $W$ of order $n$ which satisfies $W\overline{W}^\top= nI$ and all of whose entries are complex numbers of absolute value $1$. A complex Hadamard matrix is said to be Butson-type, if all of its entries are roots of unity. In our earlier work [@MI], we proposed a method to classify symmetric complex Hadamard matrices belonging to the Bose–Mesner algebra of a symmetric association scheme. In this paper, we propose an analogous method to classify nonsymmetric hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of a commutative nonsymmetric association scheme. First we consider nonsymmetric hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of a commutative nonsymmetric association scheme of class $3$, and then we show that the first eigenmatrix of such a scheme is characterized as $$\label{d3P1}
\begin{pmatrix}
1&a(2a-1)&a(2a-1)&2a-1\\
1&ai&-ai&-1\\
1&-ai&ai&-1\\
1&-a&-a&2a-1
\end{pmatrix},$$ where $a$ is a positive integer and $i=\sqrt{-1}$. Moreover, we show that such a complex Hadamard matrix is necessarily a Butson-type complex Hadamard matrix whose entries are $4$-th roots of unity. An association scheme with the first eigenmatrix (\[d3P1\]) is a nonsymmetric amorphous association scheme belonging to $L_{2a;1}$, according to [@IMY]. An example of such an association scheme was constructed from a Galois ring of characteristic $4$ in [@IMY Theorem 9]. Galois rings have been used to construct certain association schemes. Yamada [@Y] used Galois rings of characteristic $4$ to construct distance-regular digraphs, which are nonsymmetric association schemes of class $3$. This construction was generalized in [@IMY] to produce amorphous association schemes. Ma [@Ma] gave association schemes of class $3$ on Galois rings of characteristic $4$, which do not come from fusions in any amorphous association schemes. Moreover, certain properties of association schemes obtained from Galois rings of odd characteristic have been investigated in [@EP].
Our second result of this paper is a construction of nonsymmetric association schemes $\mathfrak{X}$ of class $6$ on Galois rings of characteristic 4, whose first eigenmatrix is given by $$\label{P6}
(p_{i,j})_{\substack{0\leq i\leq 6 \\ 0\leq j\leq 6}}=\begin{pmatrix}
1&2b(b-1)&2b(b-1)&b&b&b-1&b\\
1&bi&-bi&0&0&-1&0\\
1&-bi&bi&0&0&-1&0\\
1&0&0&bi&-bi&b-1&-b \\
1&0&0&-bi&bi&b-1&-b \\
1&-2b&-2b&b&b&b-1&b \\
1&0&0&-b&-b&b-1&b
\end{pmatrix},$$ where $b$ is a power of $4$. We also classify hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of $\mathfrak{X}$. We show that such a matrix is necessarily a Butson-type matrix whose entries are $4$-th roots of unity.
Association schemes and complex Hadamard matrices {#sec.AS}
=================================================
In this section, we consider hermitian matrices belonging to the Bose–Mesner algebra of a commutative association scheme. Assuming that all entries are complex numbers with absolute value $1$, we find conditions under which such a matrix is a complex Hadamard matrix. We refer the reader to [@BI; @BCN] for undefined terminology in the theory of association schemes.
Let $X$ be a finite set with $n$ elements, and let $\mathfrak{X}=(X,\{R_i\}_{i=0}^d)$ be a commutative association scheme with the first eigenmatrix $P=(P_{i,j})$. We let $\mathfrak{A}$ denote the Bose–Mesner algebra spanned by the adjacency matrices $A_0,A_1,\ldots,A_d$ of $\mathfrak{X}$. Then the adjacency matrices are expressed as $$\label{eq:pij}
A_j=\sum_{i=0}^d P_{i,j}E_i \quad (j=0,1,\ldots,d),$$ where $E_0=\frac{1}{n}J,E_1,\ldots,E_d$ are the primitive idempotents of $\mathfrak{A}$. Let $w_0=1$ and $w_1,\ldots,w_d$ be complex numbers with absolute value $1$. Set $$\label{eq:W}
W=\sum_{j=0}^d w_jA_j\in\mathfrak{A}.$$ Define $$\label{eq:gamma}
\gamma_k=\sum_{j=0}^d w_jP_{k,j} \quad (k=0,1,\ldots,d).$$ By , and we have $$\label{eq:WE}
W=\sum_{k=0}^d \gamma_kE_k.$$ Let $X_j$ [($1\leq j\leq d$)]{} be indeterminates. For $k=1,2,\dots,d$, let $e_k$ be the polynomial defined by $$\label{eq:ek}
e_k=1+2\left(\sum_{j=1}^d P_{k,j}X_j+
\sum_{1\leq j_1<j_2\leq d}P_{k,j_1}P_{k,j_2}X_{j_1}X_{j_2}\right)
+\sum_{j=1}^dP_{k,j}^2X_j^2-n.$$ Then we have the following.
\[lem:equiv\] Assume that the matrix $W$ given in is a hermitian matrix. Then the following statements are equivalent:
- $W$ is a complex Hadamard matrix,
- $\gamma_k^2=n$ for $k=1,\ldots,d$,
- $(w_j)_{1\leq j\leq d}$ is a common zero of $e_k$ [($k=1,\ldots,d$)]{}.
By we have $W\overline{W}^T=W^2=\sum_{k=0}^d\gamma_k^2E_k$. Therefore, (i) implies (ii). To prove the converse, it is enough to show that $\gamma_0^2=n$. Since $$\begin{aligned}
(W^2)_{j,j}&=\sum_{k=1}^nW_{j,k}\overline{W_{j,k}}\\
&=n,\end{aligned}$$ the diagonal entries of $W^2$ are all $n$. Thus, $$\begin{aligned}
n^2&=\operatorname{Tr}W^2\\
&=\sum_{k=0}^d\gamma_k^2\operatorname{Tr}E_k\\
&=\gamma_0^2+\sum_{k=1}^dn\operatorname{Tr}E_k\\
&=\gamma_0^2+n\operatorname{Tr}(I-E_0)\\
&=\gamma_0^2+n(n-1).\end{aligned}$$ Hence $\gamma_0^2=n$. By we have $$\gamma_k^2=1+2\left(\sum_{j=1}^d P_{k,j}w_j+
\sum_{1\leq j_1<j_2\leq d}P_{k,j_1}P_{k,j_2}w_{j_1}w_{j_2}\right)
+\sum_{j=1}^dP_{k,j}^2w_j^2$$ for $k=1,\dots,d$. Therefore, the equivalence of (ii) and (iii) follows.
Hermitian complex Hadamard matrices and nonsymmetric association schemes of class $3$ {#sec.3-class}
=====================================================================================
In this section, we classify nonsymmetric hermitian complex Hadamard matrices belonging to the Bose–Mesner algebra of a nonsymmetric commutative association scheme $\mathfrak{X}=(X,\{R_i\}_{i=0}^3)$ of class $3$, where $R_1^\top=R_2$, $R_3$ symmetric. We use the following result.
\[lem:song\] Let $$\begin{pmatrix}
1&\frac{k_1}{2}&\frac{k_1}{2}&k_2\\
1&\frac12(r+bi)&\frac12(r-bi)&-(r+1)\\
1&\frac12(r-bi)&\frac12(r+bi)&-(r+1)\\
1&\frac{s}{2}&\frac{s}{2}&-(s+1)
\end{pmatrix}$$ be the first eigenmatrix of $\mathfrak{X}$, where $r,s$ are integers and $i^2=-1$. Then one of the following holds.
- $(r,s,b^2)=(0,-(k_2+1),\frac{k_1(k_2+1)}{k_2})$,
- $(r,s,b^2)=(-(k_2+1),0,(1+k_2)(1+k_1+k_2))$,
- $(r,s,b^2)=(-1,k_1,k_1+1)$.
Note that, if we put $$\label{eq:0930}
(k_1,k_2)=(2a(2a-1),2a-1),$$ in part (i) of Lemma \[lem:song\], we have the matrix (\[d3P1\]).
Let $\mathfrak{A}$ be the Bose–Mesner algebra of $\mathfrak{X}$ which is the linear span of the adjacency matrices $A_0,A_1,A_2,A_3$ of $\mathfrak{X}$, where $A_1^\top=A_2$, $A_3$ symmetric. Let $w_0=1$ and $w_1,w_2,w_3$ be complex numbers of absolute value $1$. Set $$\label{eq:w0W}
W=w_0A_0+w_1A_1+w_2A_2+w_3A_3\in\mathfrak{A}.$$ Then our first main theorem is the following.
\[thm:main0925\] Assume that the matrix is a hermitian complex Hadamard matrix and not a real Hadamard matrix. Then $\mathfrak{X}$ is a nonsymmetric association scheme whose unique nontrivial symmetric relation consists of $2a$ cliques of size $2a$, and the first eigenmatrix of $\mathfrak{X}$ is given by . Moreover, $w_1=\pm i$ and $w_3=1$.
Suppose that the matrix $W$ is a complex Hadamard matrix. Without loss of generality, we may assume $w_0=1$ in (\[eq:w0W\]). Since $W$ is a hermitian matrix, we have $w_1w_2=1$ and $w_3=\pm1$. Since $W$ is not a real Hadamard matrix, we have $w_1-w_2\not=0$. By Lemma \[lem:equiv\], $(w_i)_{i=1}^3$ is a common zero of the polynomials $e_k$ ($k=1,2,3$) defined in (\[eq:ek\]). Since $$e_1-e_2=bi(X_1-X_2)(r(X_1+X_2-2X_3)-2X_3+2),$$ we have $$\label{eq:1-1}
r(w_1+w_2-2w_3)-2w_3+2=0.$$
First assume that $w_3=1$. Then by (\[eq:1-1\]) we have $r(w_1+w_2-2)=0$. Hence $r=0$. Therefore we have case (i) of Lemma \[lem:song\]. Then $$\label{eq:nk1}
(k_1,k_2)=\left(\frac{(s+1)b^2}{s},-(s+1)\right).$$ After specializing $X_3=1$, we have $$e_3-e_1=\frac{1}{4}((s-bi)X_1+(s+bi)X_2-2s)((s+bi)X_1+(s-bi)X_2-2s).$$ Hence $((s-bi)w_1+(s+bi)w_2-2s)((s+bi)w_1+(s-bi)w_2-2s)=0$. Then by $w_2=w_1^{-1}$ we have $$\label{eq:1-2}
(w_1-1)^2((s-bi)w_1-(s+bi))((s+bi)w_1-(s-bi))=0.$$ Put $w=(s+2bi)/(s-2bi)$. By $w_1\not=1$ we have $w_1\in\{w,\overline{w}\}$ by (\[eq:1-2\]). We may assume $w_1=w$. After specializing $X_1=w$, $X_2=\overline{w}$ and $X_3=1$, we have $$e_1=-(1+k_1+k_2)-\frac{4b^4s^2}{(b^2+s^2)^2}.$$ Hence $$\label{eq:1-3}
k_1+k_2=-1-\frac{4b^4s^2}{(b^2+s^2)^2}.$$ Then by (\[eq:nk1\]), (\[eq:1-3\]) we have $$(b-s)(b+s)((s+1)b^4-(s-2)s^2b^2+s^4)=0.$$ Since $b>0$ and $s=-(k_2+1)<-1$, we have $b-s\not=0$. First assume that $s=-b$. Then, by putting $b=2a$, we have $(k_1,k_2)=(2a(2a-1),2a-1)$ and $w_1=-i$. Therefore we have the first eigenmatrix (\[d3P1\]). Next assume that $$\label{eq:0921}
(s+1)b^4-(s-2)s^2b^2+s^4=0.$$ Then the discriminant of (\[eq:0921\]) as a quadratic equation in $b^2$ is $s^4((s-4)^2-16)$. Since $b^2$ is rational, $(s-4)^2-16$ is a square. This implies $s\in\{0,-1\}$. This contradicts $s=-(k_2+1)<-1$.
Next assume that $w_3=-1$. Then by (\[eq:1-1\]) we have $r(w_1+w_2+2)+4=0$. Put $g(x)=rx^2+2(r+2)x+r$. By $w_2=w_1^{-1}$ we have $g(w_1)=0$. Since $w_1\not\in{\mathbb{R}}$, we have $r+1<0$. Thus we have case (ii) of Lemma \[lem:song\]. After specializing $X_1=w_1$, $X_2=w_1^{-1}$, $X_3=-1$ and $s=0$, we have $$e_1-e_3=\frac{(w_1+1)f(w_1)}{4w_1^2},$$ where $f(x)=((r+bi)x+r-bi)((r+bi)x^2+2(r+4)x+r-bi)$. Since $w_1\not=-1$, we have $f(w_1)=0$. Since $$\begin{aligned}
f(x)&=\frac{(r+bi)((r+bi)rx+r(r+4)-(3r+4)bi)}{r^2}g(x)\\
&\quad-\frac{4((r+1)b^2+r^2)}{r^2}((r+4)x-r)\end{aligned}$$ and $(r+4)w_1-r\not=0$, we have $$\label{eq:0918}
(r+1)b^2+r^2=0.$$ Substituting $(r,b^2)=(-k_2+1,(1+k_1+k_2)(k_2+1))$ in (\[eq:0918\]), we have $(k_2+1)(k_2^2+k_1k_2-1)=0$. This is a contradiction.
In the next section, we will construct a nonsymmetric association scheme $\mathfrak{X}$ of class $6$ on a Galois ring of characteristic $4$, with the first eigenmatrix . An association scheme with the first eigenmatrix (\[d3P1\]) can be obtained by fusing some relations of $\mathfrak{X}$.
Association schemes on Galois rings {#sec.Galoisring}
===================================
For the reminder of this section, we let $e\geq 3$ be an odd positive integer. We refer the reader to [@Y] for basic theory of Galois rings. Let $F=\text{GF}(2)$ be the prime field of characteristic $2$ and $K=\text{GF}(2^e)$ be an extension of degree $e$. We identify $K$ with $F[x]/(\varphi(x))$, where $\varphi(x)$ is a primitive polynomial of degree $e$ over $F$, and we denote by $\zeta$ a root of $\varphi(x)$ in $K$. Let ${\mathcal{A}}={\mathbb{Z}}/4{\mathbb{Z}}$. There exists a monic polynomial $\Phi(x)$ over ${\mathcal{A}}$ such that $\Phi(x)\equiv \varphi(x)$ mod $2{\mathcal{A}}[x]$ and $\Phi(x)$ divides $x^{2^e-1}-1$ in ${\mathcal{A}}[x]$ (see [@CS Theorem 1]). The ring ${\mathcal{R}}={\mathcal{A}}[x]/(\Phi(x))$ is called a Galois ring, and it is a local ring with maximal ideal ${\mathcal{P}}=2{\mathcal{R}}$ and residue field ${\mathcal{R}}/{\mathcal{P}}\cong K$.
Let $\xi$ be the image of $x$ in ${\mathcal{R}}$, so that $\xi+{\mathcal{P}}$ is mapped to $\zeta$ under the isomorphism ${\mathcal{R}}/{\mathcal{P}}\cong K$. Then ${\mathcal{R}}={\mathcal{A}}[\xi]$ and $\xi$ has order $2^e-1$ in ${\mathcal{R}}$. Let ${\mathcal{T}}$ be the cyclic group of order $2^e-1$ generated by $\xi$. Since ${\mathcal{T}}$ is mapped bijectively onto $K\setminus\{0\}$ under the natural homomorphism ${\mathcal{R}}\rightarrow K$, each element $\alpha\in {\mathcal{R}}$ is uniquely expressed as $$\alpha=\alpha_0+2\alpha_1, \quad \alpha_0,\alpha_1\in{\mathcal{T}}\cup\{0\}.$$ The unit group ${\mathcal{R}}^{\ast}$ is the direct product of ${\mathcal{T}}$ and the principal unit group $\mathcal{E}=1+{\mathcal{P}}$. Let ${\mathcal{T}}_\delta=\{\xi^j\mid 0\leq j\leq 2^e-2,\;\mathrm{Tr}(\zeta^j)=\delta\}$ ($\delta=0,1$). We set ${\mathcal{P}}_0=2{\mathcal{T}}_0\cup\{0\}$ and $H=1+{\mathcal{P}}_0$. Thus ${\mathcal{P}}_0$ and $H$ are subgroups of index $2$ in the additive group ${\mathcal{P}}$ and the multiplicative group $\mathcal{E}$, respectively. Setting $b=2^{e-1}$, we have $b=|{\mathcal{P}}_0|=|H|$.
The mapping $\xi\mapsto\xi^2$ can be extended to a ring automorphism $f$ of ${\mathcal{R}}$ which fixes ${\mathcal{A}}$ elementwise. The ring automorphism $f$ is called the [*Frobenius automorphism*]{}. For $\alpha=\alpha_0+2\alpha_1$ ($\alpha_0,\alpha_1\in{\mathcal{T}}\cup\{0\}$), we have $$\alpha^f=\alpha_0^2+2\alpha_1^2,$$ and the trace of $\alpha\in{\mathcal{R}}$ is defined by $${\mathcal{S}}(\alpha)=\alpha+\alpha^f+\cdots+\alpha^{f^{e-1}}.$$ Note that ${\mathcal{S}}$ is an ${\mathcal{A}}$-linear mapping from ${\mathcal{R}}$ to ${\mathcal{A}}$. The additive characters of ${\mathcal{R}}$ are given by $\alpha\mapsto\chi(\alpha\beta)$ ($\beta\in{\mathcal{R}}$), where $$\chi(\alpha)=i^{e{\mathcal{S}}(\alpha)},$$ and $i^2=-1$. Define $$\lambda_{\alpha}(A)=\sum_{\beta\in A}\chi(\alpha\beta),$$ for $\alpha\in{\mathcal{R}}$ and $A\subset{\mathcal{R}}$. Set $\lambda=\lambda_1$.
Observe $$\begin{aligned}
\mathcal{E}&=H\cup(-H), \nonumber \\
\alpha{\mathcal{T}}&={\mathcal{P}}\setminus\{0\} \quad \text{for any $\alpha\in{\mathcal{P}}\setminus\{0\}$}, \label{eq:aT} \\
\mathrm{Tr}(\zeta^j)&\equiv {\mathcal{S}}(\xi^j)\bmod{2}. \nonumber\end{aligned}$$ Since $e$ is odd, we obtain $$\begin{aligned}
\chi(\alpha)&=
\begin{cases}
1 & \ \text{if} \ \alpha\in{\mathcal{P}}_0, \\
-1 & \ \text{if} \ \alpha\in{\mathcal{P}}\setminus{\mathcal{P}}_0,
\end{cases}\label{chival} \\
\chi(\alpha)&=i \quad \text{for} \ \alpha\in H. \label{chiH}\end{aligned}$$
\[lem:aP0\] For any $\alpha\in{\mathcal{R}}^{\ast}\setminus\mathcal{E}$ we have the following. $$\begin{aligned}
\lambda(\alpha\mathcal{P}_0)&=0, \label{eq:aP0} \\
\lambda(\alpha H)&=0. \label{eq:aH}\end{aligned}$$
Note that the correspondence $2\xi^j\mapsto \zeta^j$ extends to an isomorphism from ${\mathcal{P}}$ to $K$ as additive groups. Under this isomorphism, ${\mathcal{P}}_0$ is mapped to $\mathrm{Tr}^{-1}(0)$. Since $\alpha\notin\mathcal{E}$, the image $a=\alpha+{\mathcal{P}}$ of $\alpha$ in $K={\mathcal{R}}/{\mathcal{P}}$ is not $1$. By the non-degeneracy of the trace function on $K$, we have $a\mathrm{Tr}^{-1}(0)\neq\mathrm{Tr}^{-1}(0)$. This implies $\alpha\mathcal{P}_0\neq\mathcal{P}_0$. Then $$A=\alpha{\mathcal{P}}_0\cap{\mathcal{P}}_0$$ is a subgroup of index $4$ in the additive group ${\mathcal{P}}$. By we have $$\chi(\beta)=
\begin{cases}
1 & \text{if} \quad \beta\in A, \\
-1 & \text{if} \quad \beta\in\alpha{\mathcal{P}}_0\setminus A.
\end{cases}$$ Therefore $\lambda(\alpha{\mathcal{P}}_0)=0$. Since $\alpha H=\alpha+\alpha{\mathcal{P}}_0$, we have $\lambda(\alpha H)=\lambda(\alpha)\lambda(\alpha{\mathcal{P}}_0)
=0$.
Recall $b=|{\mathcal{P}}_0|=|H|$.
\[lem:lamH\] We have the following.
- $\lambda(H)=bi$,
- $\lambda({\mathcal{T}}H)=bi$.
\(i) Obvious from .
\(ii) Since ${\mathcal{T}}\setminus\{1\}\subset{\mathcal{R}}^{\ast}\setminus\mathcal{E}$, we have $\lambda(({\mathcal{T}}\setminus\{1\})H)=0$ by . Then the result follows from (i).
\[lem:lamaH\] We have $$\lambda_{\alpha}(H)=
\begin{cases}
b & \text{if} \quad \alpha\in{\mathcal{P}}_0, \\
-b & \text{if} \quad \alpha\in{\mathcal{P}}\setminus{\mathcal{P}}_0.
\end{cases}$$
Let $\alpha\in\mathcal{P}$. Then $\alpha H=\{\alpha\}$. Hence $\lambda_{\alpha}(H)=|H|\chi(\alpha)$, and the result follows from .
\[lem:lamaTH\] For any $\alpha\in{\mathcal{P}}\setminus\{0\}$ we have $$\lambda_{\alpha}({\mathcal{T}}H)=-b.$$
By we have $$\begin{aligned}
\lambda_{\alpha}({\mathcal{T}}H)&=\sum_{\beta\in{\mathcal{P}}\setminus\{0\}}\lambda_{\beta}(H) \\
&=\sum_{\beta\in{\mathcal{P}}}\lambda_{\beta}(H)-\lambda_0(H) \\
&=-\lambda_0(H) && (\text{by Lemma~\ref{lem:lamaH}}) \\
&=-b.\end{aligned}$$
Then our second main theorem is the following.
\[thm:1\] Let $$\begin{aligned}
S_0&=\{0\}, \\
S_1&=({\mathcal{T}}\setminus\{1\})H, \\
S_2&=({\mathcal{T}}\setminus\{1\})(-H), \\
S_3&=H, \\
S_4&=-H, \\
S_5&={\mathcal{P}}_0\setminus\{0\}, \\
S_6&={\mathcal{P}}\setminus{\mathcal{P}}_0.\end{aligned}$$ Then $\lambda_{\alpha}(S_j)$ is constant for any $\alpha\in S_i$ [($i=0,1,\ldots,6$)]{}. Put $p_{i,j}=\lambda_{\alpha}(S_j)$ for $\alpha\in S_i$ [($i,j=0,1,\ldots,6$)]{}. Then $P=(p_{i,j})_{\substack{0\leq i\leq 6 \\ 0\leq j\leq 6}}$ is given by [(\[P6\])]{}. In particular, if we set $$R_j=\{(\alpha,\beta)\in{\mathcal{R}}\times{\mathcal{R}}\mid \alpha-\beta\in S_j \} \quad (j=0,1,\ldots,6),$$ then $\mathfrak{X}=({\mathcal{R}},\{R_i\}_{i=0}^6)$ is a nonsymmetric association scheme of class $6$ with the first eigenmatrix .
It is easy to check that $|S_1|=|S_2|=2b(b-1)$, $|S_3|=|S_4|=b$, $|S_5|=b-1$, and $|S_6|=b$. So we have $p_{0,j}$ ($j=0,1,\ldots,6$) as in . It is trivial that $\lambda_{\alpha}(S_0)=1$ for any $\alpha\in S_0\cup S_1\cup \cdots\cup S_6$. Hence $p_{j,0}=1$ for $j=0,1,\ldots,6$ as in .
We claim that $$\begin{aligned}
\lambda_{\alpha}(S_1)&=\begin{cases}
bi & \text{if} \quad \alpha\in S_1,\\
0 & \text{if} \quad \alpha\in S_3,\\
-2b & \text{if} \quad \alpha\in S_5,\\
0 & \text{if} \quad \alpha\in S_6,
\end{cases} \label{case:1}
\displaybreak[0]\\
\lambda_{\alpha}(S_3)&=\begin{cases}
0 & \text{if} \quad \alpha\in S_1,\\
bi & \text{if} \quad \alpha\in S_3,\\
b & \text{if} \quad \alpha\in S_5,\\
-b & \text{if} \quad \alpha\in S_6,
\end{cases} \label{case:2}
\displaybreak[0]\\
\lambda_{\alpha}(S_5)&=\begin{cases}
-1 & \text{if} \quad \alpha\in S_1,\\
b-1 & \text{if} \quad \alpha\in S_3,\\
b-1 & \text{if} \quad \alpha\in S_5,\\
b-1 & \text{if} \quad \alpha\in S_6.
\end{cases} \label{case:3}\end{aligned}$$
First we prove . Assume that $\alpha\in S_1$. Then $\lambda_{\alpha}({\mathcal{T}}H)=\lambda({\mathcal{T}}H)=bi$ by Lemma \[lem:lamH\] (ii), and $\lambda_{\alpha}(H)=\lambda(\alpha H)$. Since $S_1\subset{\mathcal{R}}^{\ast}\setminus\mathcal{E}$, we have $\lambda(\alpha H)=0$ by . Hence $\lambda_{\alpha}(S_1)=bi$. Next assume that $\alpha\in S_3$. Similarly, we have $\lambda_{\alpha}({\mathcal{T}}H)=\lambda({\mathcal{T}}H)$ and $\lambda_{\alpha}(H)=\lambda(H)$. By Lemma \[lem:lamH\] (i), (ii) we have $\lambda_{\alpha}(S_1)=0$. Next assume that $\alpha\in S_5\cup S_6$. Then by Lemma \[lem:lamaH\] and Lemma \[lem:lamaTH\], we have $\lambda_{\alpha}(S_1)=-2b$ or $0$ depending on $\alpha\in S_5$ or $\alpha\in S_6$. This completes the proof of .
Next we prove . Assume that $\alpha\in S_1$. Then $\lambda_{\alpha}(S_3)=\lambda(\alpha S_3)$. Since $S_1\subset{\mathcal{R}}^{\ast}\setminus\mathcal{E}$, we have $\lambda(\alpha S_3)=0$ by . Next assume that $\alpha\in S_3$. Then $\lambda_{\alpha}(S_3)=\lambda(S_3)=bi$ by Lemma \[lem:lamH\] (i). Next assume that $\alpha\in S_5\cup S_6$. Then by Lemma \[lem:lamaH\], $\lambda_{\alpha}(S_3)=b$ or $-b$ depending on $\alpha\in S_5$ or $\alpha\in S_6$. This completes the proof of .
Finally we prove . Assume that $\alpha\in S_1$. Since $S_1\subset{\mathcal{R}}^{\ast}\setminus\mathcal{E}$, we have $\lambda_{\alpha}(S_5)=\lambda(\alpha({\mathcal{P}}_0\setminus\{0\}))=-1$ by . Next assume that $\alpha\in S_3$. Since $\alpha\mathcal{P}_0={\mathcal{P}}_0$, we have $\lambda_{\alpha}(S_5)
=\lambda({\mathcal{P}}_0\setminus\{0\})=|S_5|=b-1$. Next assume that $\alpha\in S_5\cup S_6$. Then $\alpha({\mathcal{P}}_0\setminus\{0\})=\{0\}$. Hence $\lambda_{\alpha}(S_5)=|S_5|=b-1$. This completes the proof of .
From – we have $p_{1,1}$, $p_{3,1}$, $p_{5,1}$, $p_{6,1}$, $p_{1,3}$, $p_{3,3}$, $p_{5,3}$, $p_{6,3}$, $p_{1,5}$, $p_{3,5}$, $p_{5,5}$, and $p_{6,5}$ as in . Since $\{S_j\}_{j=0}^6$ is a partition of ${\mathcal{R}}$, we have $\sum_{j=0}^6\lambda_{\alpha}(S_j)=0$ for $\alpha\not=0$. Thus, it is enough to check that $\lambda_{\alpha}(S_j)$ is a constant independent of $\alpha\in S_i$ for $i=1,\ldots,6$ and $j=1,\ldots,5$. Since $S_2=-S_1$, we have $p_{2,j}=\overline{p_{1,j}}$ and $p_{j,2}=\overline{p_{j,1}}$ for $j=1,\ldots,6$. Since $S_4=-S_3$, we have $p_{4,j}=\overline{p_{3,j}}$ and $p_{j,4}=\overline{p_{j,3}}$ for $j=1,\ldots,6$. Therefore we find all $p_{i,j}$ as in from –.
Let $\mathfrak{X}=(X,\{R_i\}_{i=0}^d)$ be a commutative association scheme. A partition $\Lambda_0, \Lambda_1, \ldots, \Lambda_e$ of the index set $\{0,1,\ldots,d\}$ of the association scheme $\mathfrak{X}$ is said to be [*admissible*]{} if $\Lambda_0=\{0\}$, $\Lambda_i\not=\emptyset$ $(1\leq i\leq e)$ and $\Lambda_i'=\Lambda_j$ for some $j$ ($1\leq j\leq e$), where $\Lambda_i'=\{\alpha' \mid \alpha\in\Lambda_i\}$, $R_{\alpha'}=\{(x,y)\mid (y,x)\in R_{\alpha}\}$. Let $R_{\Lambda_i}=\bigcup_{\alpha\in\Lambda_i}R_{\alpha}$. If $\mathfrak{Y}=(X,\{R_{\Lambda_i}\}_{i=0}^e)$ becomes an association scheme, then it is called a [*fusion*]{} scheme of $\mathfrak{X}$.
[**Bannai–Muzychuk criterion**]{} ([@B; @Muzychuk]). Let $\mathfrak{X}$ be a commutative association scheme with the first eigenmatrix $P$. Let $\{\Lambda_j\}_{j=0}^e$ be an admissible partition of the index set $\{0,1,\ldots,d\}$. Then $\{\Lambda_j\}_{j=0}^e$ gives rise to a fusion scheme $\mathfrak{Y}$ if and only if there exists a partition $\{\Delta_i\}_{i=0}^e$ of $\{0,1,\ldots,d\}$ with $\Delta_0=\{0\}$ such that each $(\Delta_i,\Lambda_j)$ block of the first eigenmatrix $P$ has a constant row sum. Moreover, the constant row sum of the $(\Delta_i,\Lambda_j)$ block is the $(i,j)$ entry of the first eigenmatrix of $\mathfrak{Y}$.
Let $\mathfrak{X}$ be an association scheme given in Theorem \[thm:1\]. Fusion schemes of $\mathfrak{X}$ with at least three classes are listed in Table \[table\].
fused relations class nonsymmetirc or symmetric
------------------ --------------------------- ------- ---------------------------
$\mathfrak{X}_1$ $\{1,2\}$ $5$ nonsymmetric
$\mathfrak{X}_2$ $\{3,4\}$ $5$ nonsymmetric
$\mathfrak{X}_3$ $\{1,2\},\{3,4\}$ $4$ symmetric
$\mathfrak{X}_4$ $\{3,4,6\}$ $4$ nonsymmetric
$\mathfrak{X}_5$ $\{1,2\},\{3,4\},\{5,6\}$ $3$ symmetric
$\mathfrak{X}_6$ $\{1,2,3,4\}$ $3$ symmetric
$\mathfrak{X}_7$ $\{1,3\},\{2,4\},\{5,6\}$ $3$ nonsymmetric
$\mathfrak{X}_8$ $\{1,4\},\{2,3\},\{5,6\}$ $3$ nonsymmetric
: Fusion schemes of $\mathfrak{X}$[]{data-label="table"}
The first eigenmatrix of $\mathfrak{X}_7$ and $\mathfrak{X}_8$ in Table \[table\] are the matrix (\[d3P1\]) by putting $a=b$. Put $b=4^p$. We can verify that $\mathfrak{X}_7$ and $\mathfrak{X}_8$ for $p=3$ are isomorphic, and $\mathfrak{X}_7$ and $\mathfrak{X}_8$ for $p=5$ are not isomorphic. The computation needed to verify these facts was done with the help of Magma [@magma].
Complex Hadamard matrices and Galois rings {#sec:5}
==========================================
We continue to use the notation introduced in Section \[sec.Galoisring\]. Let $\{A_j\}_{j=0}^6$ be the set of the adjacency matrices of $\mathfrak{X}$. Then $A_1^T=A_2$, $A_3^T=A_4$, and $A_5,A_6$ symmetric. We call the algebra $\mathfrak{A}=\langle A_0,A_1,\ldots,A_6\rangle$ the Bose–Mesner algebra of $\mathfrak{X}$. Our next theorem gives a classification of hermitian complex Hadamard matrices belonging to $\mathfrak{A}$.
\[thm:2\] Let $w_0=1$ and $w_j$ [(]{}$1\leq j\leq6$[)]{} be complex numbers of absolute value $1$. Set $$W=\sum_{j=0}^6w_jA_j\in\mathfrak{A},$$ and assume that $W$ is hermitian. Then, $W$ is a complex Hadamard matrix if and only if $$\begin{aligned}
W=&A_0+\epsilon_1i(A_1-A_2)+\epsilon_2i(A_3-A_4)+A_5+A_6, \quad \text{or} \label{eq:W1}\\
W=&A_0+\epsilon_1i(A_1-A_2)+\epsilon_2(A_3+A_4)+A_5-A_6,\label{eq:W2}\end{aligned}$$ for some $\epsilon_1,\epsilon_2\in\{\pm1\}$.
Recall that $\{A_j\}_{j=0}^6$ is the set of the adjacency matrices of $\mathfrak{X}$ given in Theorem \[thm:1\]. Notice that $A_1^T=A_2$, $A_3^T=A_4$, while $A_5$ and $A_6$ are symmetric. Since $W$ is hermitian, we have $w_1w_2=1$ and $w_3w_4=1$.
Suppose that the matrix $W$ is a complex Hadamard matrix. Then $(w_i)_{i=1}^6$ is a common zero of the polynomials $e_k$ ($k=1,\ldots,6$) defined in . Since $$\label{01-7}
e_2-e_1=4ib(X_1-X_2)(X_5-1),$$ we have $w_1=w_2$ or $w_5=1$.
Suppose first that $w_1=w_2$. After specializing $X_1=X_2$ and $X_1^2=1$, we have $$e_1=(X_5-(2b+1))(X_5+2b-1).$$ Then $|w_5|=|2b\pm1|\geq 2b-1>1$, which is a contradiction. Therefore, we must have $w_5=1$. After specializing $X_5=1$, we have $$e_1=-b^2(X_1-X_2-2i)(X_1-X_2+2i).$$ Hence $w_1=-w_2\in\{\pm i\}$. Moreover, after specializing $X_5=1$, $X_1=-X_2$ and $X_1^2=-1$, we have $$\begin{aligned}
e_4-e_3&=4ib^2(X_6-1)(X_3-X_4),\label{01-8}\\
e_5-e_6&=4b^2(X_6+1)(X_3+X_4).\label{01-9}\end{aligned}$$ If $w_6=1$, then $w_4=-w_3\in\{\pm i\}$ by . Therefore we have . If $w_3=w_4$, then $w_3\in\{\pm1\}$ and $w_6=-1$ by . Therefore we have .
Conversely, assume that $W$ is one of the matrices , . We show that $W$ is a complex Hadamard matrix. By Lemma \[lem:equiv\], it suffices to show that $(w_i)_{i=1}^6$ is common zero of the polynomials , and it is easy to do this.
We note that the matrix belongs to the Bose–Mesner algebra of $\mathfrak{X}_7$ or $\mathfrak{X}_8$ in Table \[table\], depending on $\epsilon_1=\epsilon_2$ or $\epsilon_1=-\epsilon_2$. Also, the matrix belongs to the Bose–Mesner algebra of $\mathfrak{X}_2$ or $\mathfrak{X}_4$ in Table \[table\], depending on $\epsilon_2=1$ or $\epsilon_2=-1$. Therefore, no proper fusion scheme of $\mathfrak{X}$ contains all the matrices and in its Bose–Mesner algebra $\mathfrak{A}$. We also note that Ma [@Ma] considered association schemes which are invariant under the multiplication by ${\mathcal{T}}$. The only association schemes in Table \[table\] which are invariant under the multiplication by ${\mathcal{T}}$ is $\mathfrak{X}_7$, and it has the first eigenmatrix described by [@Ma Theorem 7].
[9]{} E. Bannai, [*Subschemes of some association schemes*]{}, J. Algebra 144 (1991), 167–188.
E. Bannai and T. Ito, [*Algebraic Combinatorics I: Association Schemes*]{}, Benjamin/Cummings, Menlo Park, 1984.
W. Bosma, J. Cannon, and C. Playoust, *The Magma algebra system. I. The user language*, J. Symbolic Comput., 24 (1997), 235–265.
A. E. Brouwer, A. M. Cohen, and A. Neumaier, [*Distance-Regular Graphs*]{}, Springer-Verlag, Berlin, Heidelberg, 1989.
A. R. Calderbank and N. J. A. Sloane, *Modular and $p$-adic cyclic codes*, Des. Codes Cryptogr., 6 (1995), 21–35.
S. Evdokimov and I. Ponomarenko, *Normal cyclotomic schemes over a finite commutative ring*, St. Petersburg Math. J., 19 (6) (2008), 911–929.
T. Ikuta and A. Munemasa, *Complex Hadamard matrices contained in a Bose–Mesner Algebra*, Spec. Matrices, 3 (2015), 91–110.
T. Ito, A. Munemasa, and M. Yamada, *Amorphous association schemes over the Galois rings of characteristic $4$*, European J. Combin., 12 (1991), 513–526.
J. Ma, *Three-class association schemes on Galois rings in characteristic $4$*, Graphs and Combinatorics, 23 (2007), 73–86.
M. E. Muzychuk, *V-rings of permutation groups with invariant metric*, Ph.D. thesis, Kiev State University, 1987.
S. Y. Song, *Class 3 association schemes whose symmetrizations hace two classes*, J. Combin. Theory, Ser. A, 70 (1995) 1–29.
M. Yamada, *Distance-regular digraphs of girth $4$ over an extension ring of ${\mathbb{Z}}/4{\mathbb{Z}}$*, Graphs and Combinatorics, 6 (1990), 381–394.
[^1]: The work of A.M. was supported by JSPS KAKENHI grant number 17K05155.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '3D object detection from raw and sparse point clouds has been far less treated to date, compared with its 2D counterpart. In this paper, we propose a novel framework called FVNet for 3D front-view proposal generation and object detection from point clouds. It consists of two stages: generation of front-view proposals and estimation of 3D bounding box parameters. Instead of generating proposals from camera images or bird’s-eye-view maps, we first project point clouds onto a cylindrical surface to generate front-view feature maps which retains rich information. We then introduce a proposal generation network to predict 3D region proposals from the generated maps and further extrude objects of interest from the whole point cloud. Finally, we present another network to extract the point-wise features from the extruded object points and regress the final 3D bounding box parameters in the canonical coordinates. Our framework achieves real-time performance with 12ms per point cloud sample. Extensive experiments on the 3D detection benchmark KITTI show that the proposed architecture outperforms state-of-the-art techniques which take either camera images or point clouds as input, in terms of accuracy and inference time.'
author:
-
bibliography:
- 'references.bib'
title: |
FVNet: 3D Front-View Proposal Generation for\
Real-Time Object Detection from Point Clouds
---
*3D object detection, Point clouds, Real-time.*
Acknowledgment {#acknowledgment .unnumbered}
==============
Thanks to the National Natural Science Foundation of China (No. 61972157), Science and Technology Commission of Shanghai Municipality Program (No. 18D1205903) and the National Social Science Foundation of China (No. 18ZD22) for funding.
| {
"pile_set_name": "ArXiv"
} |
0.6cm
\#1 \#1\#2
UCB-PTH-12/07\
1.3cm
[**The Static Quantum Multiverse**]{}
0.7cm
[Yasunori Nomura]{}
0.4cm
[*Berkeley Center for Theoretical Physics, Department of Physics,\
University of California, Berkeley, CA 94720, USA*]{}
0.1cm
[*Theoretical Physics Group, Lawrence Berkeley National Laboratory, CA 94720, USA*]{}
0.8cm
Introduction {#sec:intro}
============
The goal of fundamental physics is to find a prescription in which (potentially) testable predictions can be made and, through it, to learn how nature works at the most fundamental level. In the present way physics is formulated, this can be done in three steps:
- [*“Theory”*]{} — We must specify the fundamental structure of the theory, which consists of the following two parts:
- [*Kinematics*]{} — We must understand what is a “state” which somehow represents the status of a physical system. We must also understand how it is related to the observed reality. For example, in conventional quantum mechanics a state is a ray in Hilbert space, which is related to reality through the Born rule, while in classical mechanics a state is a point in classical phase space (so is directly observable).
- [*Dynamics*]{} — We must know a (set of) fundamental law(s) the states obey. In quantum mechanics it is the Schrödinger equation, $i \frac{d}{dt} \ket{\Psi(t)} = H \ket{\Psi(t)}$, while in classical mechanics it is the Newton equation, $m\, \ddot{\bf x}(t) = {\bf F}$.
- [*“System”*]{} — We then need to specify a system we consider, which again consists of two parts:
- [*Kinematics*]{} — We need to know the kinematical structure of the system. In quantum mechanics this corresponds to specifying the Hilbert space, which is characterized by its dimension and operators acting on its elements. In classical mechanics, it is given by the dimension of the phase space.
- [*Dynamics*]{} — We also need to specify dynamics of the system. In the examples in (i-2), we need to give the forms of $H$ and ${\bf F}$, respectively.
- [*“Selection Conditions”*]{} — Even if (i) and (ii) are known, we still need to provide “selection conditions” on a state. Usually, they are given in the form of boundary conditions, for example as the knowledge one already has, e.g. $\ket{\Psi(0)}$ and $\{ {\bf x}(0),
\dot{\bf x}(0) \}$, before making predictions on something unknown, e.g. $\ket{\Psi(t)}$ and $\{ {\bf x}(t), \dot{\bf x}(t) \}$ for $t > 0$.
To understand the ultimate structure of nature, we would want to do the above in the context of cosmology, and see whether the resulting predictions are consistent with what we observe. In this respect, physics of eternal inflation—which occurs under rather general circumstances [@Guth:1982pn]—has caused tremendous confusions in recent years. A major problem has been the so-called measure problem: even if we know the initial state and its subsequent evolution, we cannot define (even probabilistic) predictions unambiguously.[^1] This occurs because in eternal inflation anything that can happen will happen infinitely many times, so it apparently leads to arbitrariness in predictions, associated with how these infinities are regularized [@Guth:2000ka]. Such an arbitrariness would prevent us from making well-defined predictions, so it seemed that to define the theory we needed to specify the exact way of regulating spacetime, where the infinities occur. This would be quite uncomfortable, since then the theory requires a specification of a (ad hoc) regularization prescription [*beyond the basic principles of quantum mechanics and relativity*]{}.
Recently, a framework that addresses this problem has been proposed in Refs. [@Nomura:2011dt; @Nomura:2011rb], which allows for an intrinsically quantum mechanical treatment of the eternally inflating multiverse (see Ref. [@Nomura-review] for a review directed to a wide audience). In this framework, physics is described in [*a fixed reference (local Lorentz) frame*]{} associated with a fixed reference point $p$, with spacetime existing only within its (stretched) apparent horizon. An essential point is that the principles of quantum mechanics constrain the space of states ${\cal H}_{\rm QG}$ [@Nomura:2011rb] in such a way that the problem of infinity does not arise. Namely, the correct identification of (ii-1) avoids the problem, without changing (i) from that of usual unitary quantum mechanics. A state representing the multiverse $\ket{\Psi(t)}$ “evolves” deterministically and unitarily in ${\cal H}_{\rm QG}$, following the laws of quantum mechanics: $i \frac{d}{dt} \ket{\Psi(t)} = H \ket{\Psi(t)}$. Here, $t$ is an auxiliary parameter introduced to describe the “evolution” of the state, and need not be directly related to physical time we observe. Once the state $\ket{\Psi(t)}$ is known, physical predictions can be obtained through the (extended) Born rule [@Nomura:2011dt; @Nomura:2011rb] without suffering from an infinity or ambiguity. This framework makes it possible that once a boundary condition on the state, e.g. $\ket{\Psi(t)}$ at some $t = t_0$, is given (element (iii)) and the explicit form of $H$ acting on ${\cal H}_{\rm QG}$ is understood, e.g. by studying string theory (element (ii-2)), then unambiguous predictions are obtained for any physical questions one asks. While the framework does not achieve all of (i)–(iii), it does eliminate the ambiguity associated with the measure problem and provides a setting in which the remaining issues can be discussed.
In this paper we consider the issue of (iii) in the quantum mechanical framework of the multiverse described above. We take the following hypothesis: $$\ovalbox{Hypothesis I: The laws of quantum mechanics are not violated.}
\label{eq:hypo-1}$$ This—in particular the fact that the evolution of a quantum state is deterministic and unitary—implies that the multiverse state exists all the way from $t = -\infty$ to $+\infty$. Namely, the multiverse does not have a beginning or end. (For recent discussions on the beginning of the eternally inflating multiverse, see, e.g., Refs. [@Mithani:2012ii].) There are three potential issues in this picture:
- [*Uniqueness*]{} — What is the selection condition imposed on the multiverse state, on which physical predictions will depend? In particular, what is the principle determining it?
- [*Well-definedness*]{} — The (extended) Born rule formula in general involves $t$ integrals, which would now run from $-\infty$ to $+\infty$. Will this give well-defined probabilities?
- [*Consistency*]{} — Are the resulting predictions consistent with observation? In particular, are they consistent with the observed arrow of time, even if there is no beginning or end?
In this paper we argue that consistency with observation excludes the possibility that the selection condition is determined purely in ${\cal H}_{\rm QG}$, without referring to an operator algebra. In particular, this excludes the possibility that the multiverse is in the maximally mixed state in ${\cal H}_{\rm QG}$. We then propose that the multiverse state must satisfy the following simple criterion: $$\ovalbox{Hypothesis II: Physical predictions do not depend on the
reference frame one chooses.}
\label{eq:hypo-2}$$ We show that this requirement leads to the condition $$\frac{d}{dt}\ket{\Psi(t)} = 0
\qquad\Leftrightarrow\qquad
H \ket{\Psi(t)} = 0,
\label{eq:static}$$ where we have taken $t$ to be the proper time at $p$; namely, we find that [*the multiverse state must be static!*]{} We will argue that despite its naive appearance, this does not contradict observation, including the fact that we observe that time flows in a definite direction. It simply gives constraints on the structure of $H$, on which we will allow for making arbitrary assumptions, given that its explicit form is not available under current theoretical technology. We will also argue that the hypothesis leads to unique and well-defined predictions for any physical questions, once one knows the explicit form of $H$ (element (ii-2) listed at the beginning). Specifically, any physical question can be phrased in the form: given what we know $A$ about a state, what is the probability for it to be consistent also with $B$? And the relevant probability is given by $$P(B|A) = \frac{\bra{\Psi} {\cal O}_{A \cap B} \ket{\Psi}}
{\bra{\Psi} {\cal O}_A \ket{\Psi}},
\label{eq:prob-final}$$ where $\ket{\Psi} \equiv \ket{\Psi(0)}$, and ${\cal O}_X$ is the operator projecting onto states consistent with condition $X$.
There are two comments. First, given Hypothesis I, Hypothesis II arises as a consequence of general covariance (and its suitable extension to the quantum regime) if we assume that the multiverse is in a zero-eigenvalue eigenstate of global energy and boost operators. This condition, therefore, provides another, more technical way of stating Hypothesis II. Second, without knowledge of the ultimate structure of $H$ in quantum gravity, the scenario presented here is not the only option available within the framework of Refs. [@Nomura:2011dt; @Nomura:2011rb], although it seems to be the most natural possibility. For example, one might imagine that the multiverse has a “beginning,” and evolves only thereafter. (This violates both Hypotheses I and II.) The framework of Refs. [@Nomura:2011dt; @Nomura:2011rb] itself may still be applied in such a case.
The organization of this paper is as follows. In the next section, we review the framework of the quantum multiverse given in Refs. [@Nomura:2011dt; @Nomura:2011rb], and discuss the issue of selection conditions in that context. In Section \[sec:obs\], we reconsider what the arrow of time is. We emphasize that the observed flow of time does not necessarily mean that the state is actually evolving. In Section \[sec:max-mixed\], we explore the possibility that the selection condition is expressed in ${\cal H}_{\rm QG}$ without referring to any quantum operator. We find that this forces the multiverse to be in the maximally mixed state in ${\cal H}_{\rm QG}$, which is observationally excluded. In Section \[sec:stat\], we present our main scenario in which the multiverse state is determined by the two hypotheses described above. We find this implies that the multiverse state must be static, and discuss how it can be realized in the cosmological context. We also see that the scenario arises as a consequence of quantum mechanics and general covariance if we assume that the multiverse is in a zero-eigenvalue eigenstate of global energy and boost operators. In Section \[sec:consistency\], we discuss the consistency of the scenario with observation, specifically the observed arrow of time. In Section \[sec:discuss\], we provide our final discussions. We draw a close analogy of the present scenario with the case of the hydrogen atom, underscoring the intrinsically quantum nature of the scenario.
Framework—the Quantum Multiverse {#sec:framework}
================================
In this section we review the framework of Refs. [@Nomura:2011dt; @Nomura:2011rb], describing the quantum multiverse. We also discuss the issue of selection conditions in making predictions within this framework.
The Hilbert space {#subsec:Hilbert}
-----------------
The framework is based on the principles of quantum mechanics. In particular, we formulate it using Hamiltonian (canonical) quantum mechanics, although the equivalent Lagrangian (path integral) formulation should also be possible. We take the Schrödinger picture throughout.
Recall that to do Hamiltonian quantum mechanics, we need to fix all gauge redundancies. Since these redundancies include coordinate transformations in a theory with gravity, states must be defined [*as viewed from a fixed (local Lorentz) reference frame*]{} associated with a fixed reference point $p$. Moreover, to avoid violation of the principles of quantum mechanics, they must represent only spacetime regions within the (stretched) apparent horizons of $p$, as suggested first in the study of black hole physics [@Susskind:1993if]. Together with the states associated with spacetime singularities, these states form the Hilbert space for quantum gravity ${\cal H}_{\rm QG}$.
The construction of ${\cal H}_{\rm QG}$ can proceed analogously to the usual Fock space construction in quantum field theory. For a set of fixed semi-classical geometries ${\cal M} = \{ {\cal M}_i \}$ having the same apparent horizon $\partial {\cal M}$, the Hilbert space is given by $${\cal H}_{\cal M} = {\cal H}_{{\cal M}, {\rm bulk}}
\otimes {\cal H}_{{\cal M}, {\rm horizon}},
\label{eq:ST-H_M}$$ where ${\cal H}_{{\cal M}, {\rm bulk}}$ and ${\cal H}_{{\cal M}, {\rm
horizon}}$ represent Hilbert space factors associated with the degrees of freedom inside and on the horizon $\partial {\cal M}$. The dimensions of these factors are both $\exp({\cal A}_{\partial {\cal M}}/4)$, where ${\cal A}_{\partial {\cal M}}$ is the area of the horizon in Planck units: $${\rm dim}\,{\cal H}_{\cal M}
= {\rm dim}\,{\cal H}_{{\cal M}, {\rm bulk}} \times
{\rm dim}\,{\cal H}_{{\cal M}, {\rm horizon}}
= \exp\left(\frac{{\cal A}_{\partial {\cal M}}}{2}\right),
\label{eq:H_M-dimension}$$ consistently with the holographic principle [@'tHooft:1993gx]. The full Hilbert space for dynamical spacetime is then given by the direct sum of the Hilbert spaces for different ${\cal M}$’s $${\cal H} = \bigoplus_{\cal M} {\cal H}_{\cal M}.
\label{eq:ST-H}$$ In addition, the complete Hilbert space for quantum gravity must contain “intrinsically quantum mechanical” states, associated with spacetime singularities [@Nomura:2011rb]: $${\cal H}_{\rm QG} = {\cal H} \oplus {\cal H}_{\rm sing},
\label{eq:QG-H}$$ where ${\cal H}_{\rm sing}$ represents the Hilbert space for the singularity states. The evolution of the multiverse state $\ket{\Psi(t)}$, which represents the entire multiverse, is deterministic and unitary in ${\cal H}_{\rm QG}$, but not in ${\cal H}_{\cal M}$ or ${\cal H}$.
The dimension of the complete Hilbert space ${\cal H}_{\rm QG}$ is infinite, as the dimensions of Hilbert subspaces associated with stable Minkowski space and spacetime singularities are infinite: $${\rm dim}\,{\cal H}_{\rm Minkowski} = \infty,
\qquad
{\rm dim}\,{\cal H}_{\rm sing} = \infty.
\label{eq:dim-inf}$$ This implies, by the second law of thermodynamics, that a [*generic*]{} multiverse state in ${\cal H}_{\rm QG}$ will evolve at large $t$ into a superposition of terms corresponding to supersymmetric Minkowski space or spacetime singularity: $$\ket{\Psi(t)}
\,\,\stackrel{t \rightarrow \infty}{\longrightarrow}\,\,
\sum_i a_i(t) \ket{\mbox{supersymmetric Minkowski space $i$}}
\,+\, \sum_j b_j(t) \ket{\mbox{singularity state $j$}},
\label{eq:asympt}$$ where we have assumed that the only absolutely stable Minkowski vacua are supersymmetric ones, as suggested by the string landscape picture [@Bousso:2000xa].
Note that an infinite number of states exist only in a Hilbert subspace associated with a spacetime singularity or a Minkowski space [*in which the area of the apparent horizon diverges ${\cal A}_{\partial {\cal M}}
= \infty$*]{}. In particular, the number of states associated with a fixed Friedmann-Robertson-Walker (FRW) time in a Minkowski bubble is finite for any finite energy density $\rho$, since the area of the apparent horizon is given by ${\cal A}_{\partial {\cal M}} = 3/2\rho$ (with $\rho$ in Planck units) [@Bousso:2002ju-FRW], so that ${\cal A}_{\partial {\cal M}} < \infty$ for $\rho > 0$.
The (extended) Born rule {#subsec:Born}
------------------------
For a given multiverse state $\ket{\Psi(t)}$, physical predictions can be obtained following the rules of quantum mechanics. An important point is that the “time” parameter $t$ here is simply an auxiliary parameter introduced to describe the “evolution” of the state. The physical information is only in [*correlations*]{} between events; specifically, time evolution of a physical quantity $X$ is nothing more than a correlation between $X$ and a quantity that can play the role of time, such as the location of the hands of a clock or the average temperature of the cosmic microwave background in our universe. A particularly useful choice for $t$ is the proper time at $p$, which we will assume for the rest of the paper.
Any physical question can then be phrased as: given what we know $A$ about a state, what is the probability for that state to be consistent also with condition $B$? In the context of the multiverse, this probability is given by [@Nomura:2011dt] $$P(B|A) = \frac{\int\!dt \bra{\Psi(0)} U(0,t)\,
{\cal O}_{A \cap B}\, U(t,0) \ket{\Psi(0)}}
{\int\!dt \bra{\Psi(0)} U(0,t)\, {\cal O}_A\, U(t,0) \ket{\Psi(0)}},
\label{eq:prob}$$ where $U(t_1,t_2) = e^{-iH(t_1-t_2)}$ is the “time evolution” operator with $H$ being the Hamiltonian of the entire system for a fixed “time” parameterization $t$ (here the proper time at $p$), and ${\cal O}_X$ is the operator projecting onto states consistent with condition $X$. Note that since we have already fixed a reference frame, conditions $A$ and $B$ in general must involve specifications of ranges of location and velocity in which a physical object must be with respect to the reference point $p$.
As we will discuss in more detail in Section \[sec:obs\], the formula in Eq. (\[eq:prob\]) can be used to answer any physical questions including those about dynamical evolution of a system, despite the fact that conditions $A$ and $B$ both act at the same moment $t$. We therefore base all our discussions on Eq. (\[eq:prob\]) in this paper. (For a different formula that can be used more easily in many practical contexts, see Ref. [@Nomura:2011rb].) The $t$ integrals in the equation run over the entire region under consideration. Suppose, for example, that we know the universe/multiverse is in a particular, e.g. eternally inflating, state $\ket{\Psi(0)}$ at $t = 0$, and want to predict what happens in $t > 0$. In this case, the integrals must be taken from $t = 0$ to $\infty$, since condition $A$ may be satisfied at any value of $t > 0$ in some component of $\ket{\Psi(t)}$. Note that despite the integrals running to $\infty$ the resulting probability is well-defined, because Eq. (\[eq:asympt\]) prohibits an event from occurring infinitely many times with a finite probability, which would cause divergences.
The issue of selection conditions {#subsec:s-c}
---------------------------------
What kind of predictions does the framework described above allow us to make? While the framework addresses the issues of infinity and the ambiguity associated with it (i.e. the measure problem as defined here), it is certainly not complete. In particular, ...
- [*“Unspecified System”*]{} — We did not identify the system [ *explicitly*]{}. Specifically, the complete theory of quantum gravity is not known, so that we do not know the form of $H$, especially the part acting on the horizon degrees of freedom. This particular issue can be bypassed if we focus only on questions addressed at the semi-classical level. Even then, however, current technology does not give us the explicit form of $H$, e.g. the structure of the string landscape.
- [*“Selection Conditions”*]{} — Predictions in general depend on the selection condition we impose on $\ket{\Psi(t)}$ (even if we know $H$ explicitly). For example, in the situation considered at the end of the previous subsection, they depend on the initial condition $\ket{\Psi(0)}$.
These limitations may still allow us to make certain predictions, possibly with some assumption on the dynamics of the system. First of all, if we are interested in a system localized in a small region compared with the horizon scale, then we can make predictions on the evolution of the system (i.e. correlation with a physical quantity that plays the role of time) using prior information about the system—indeed, one can show that Eq. (\[eq:prob\]) is reduced to the standard Born rule in such a case. Second, if we are interested in quantities whose distributions in $H$ are reasonably inferred in an anthropically allowed range, then we can predict the probability distribution of these quantities seen by a typical observer, under the assumption that the selection condition provides a statistically uniform prior [@Weinberg:1987dv]. This is, for example, the case if we are interested in the probability distribution of the cosmological constant one observes [@Larsen:2011mi].
However, if we want to answer general “multiversal” questions, e.g. if we want to predict the probability distribution of the structure of the low-energy Lagrangian found by an intellectual observer in the multiverse, then we would need to address both (a) and (b) above. (What the intellectual observer means can be specified explicitly by condition $A$.) For (a), one could hope that future progress, e.g. in string theory, might provide us (at least the relevant information on) the form of $H$ in ${\cal H}_{\rm QG}$. But what about (b)?
There are at least three aspects which make this problem substantial:
- One might speculate that a physical theory only allows for relating a given initial state to another final state, which is indeed the case in conventional Newtonian and quantum mechanics. In the present context, this implies that to make general predictions, we need to know the state $\ket{\Psi(t)}$ explicitly for some $t$. This is, however, impossible to do observationally! Quantum mechanics does not allow us to know the exact state [*including us, the observer*]{}. Moreover, $\ket{\Psi(t)}$ is the quantum state for the whole multiverse, so it in general contains terms representing different semi-classical universes than what we live in.
- General predictions in the multiverse, therefore, will be possible only if we have a [*theoretical*]{} input on the selection condition of $\ket{\Psi(t)}$. Suppose it takes the form of a specific “initial condition,” $\ket{\Psi(0)}$. Then, the predictions depend on $\ket{\Psi(0)}$, so that, unless we have a separate theory of the initial condition, the uniqueness of (even statistical) predictions will be lost.
- Imagine that there is, indeed, a theory of the initial condition giving a particular state $\ket{\Psi(0)}$, and that the framework described in Sections \[subsec:Hilbert\] and \[subsec:Born\] applies only to $t > 0$. In this case, the laws of quantum mechanics, especially deterministic and unitary evolution of the state, is violated at $t = 0$. While this is possible, it would be more comfortable if fundamental principles, such as those of quantum mechanics, do not have an “exception” like this.
In the rest of the paper, we will address the problem of selection conditions, i.e. issue (b), from the viewpoint of extrapolating the principles of quantum mechanics to the maximum extent possible. By postulating a certain simple criterion, and requiring consistency with observation, we will arrive at the picture that the multiverse state must, in fact, be static. This provides a strong selection of the possible states. The observed flow of time arises from the structures of $H$ in ${\cal H}_{QG}$, and not because of a $t$ dependence of $\ket{\Psi(t)}$.
The Observational “Data” {#sec:obs}
========================
Any selection condition imposed on the multiverse state must not lead to results inconsistent with observation, if it is to do with nature. The basic observational fact in our universe is that we see time flow in a definite direction, and predictions of a theory must not contradict it. As we will see, this seemingly weak requirement, in fact, provides a powerful tool to determine the selection condition. Here, we carefully consider what the observed flow of time actually means in the context of the quantum multiverse.
What is the arrow of time? {#subsec:arrow-time}
--------------------------
What does the fact that we see time flow really mean? At the most elementary level, it just means that the memory state of my (or your) brain is consistent with the hypothesis that it is generated by an environment whose coarse-grained entropy evolves from lower to higher values. The point is that the states consistent with such a hypothesis are very special ones among all the possible states the brain can take. What the fundamental theory must explain is why my brain is in one of these highly exceptional states.
![Suppose you know that there are a half of a chair and of a room in the first half of the scene (the upper picture). In a regular ordered world, you expect the second half of the scene contains the other half of the chair and the room, possibly with some other things (the lower left picture). On the other hand, the number of such states is much smaller than that of states in which the second half contains random, disordered configurations (the lower right picture).[]{data-label="fig:chair"}](chair.epsi){width="16cm"}
To illustrate the basic idea further, let us consider a more corporeal example of a chair in a room. Suppose you are looking at only a half of the scene and find a half of a chair and of a room there; see the upper picture in Fig. \[fig:chair\]. What would you expect to be in the other half? In the ordered world we live in, we expect to see the other half of the chair and the room, possibly with some other things such as a painting on the wall, as depicted in the lower left picture in Fig. \[fig:chair\]. However, any such configurations are extremely rare among all the possible configurations physically allowed and consistent with the first half of the scene. The vast majority of these general configurations correspond to the ones in which the other half of the scene is completely disordered, as depicted in the lower right picture in Fig. \[fig:chair\]. The arrow of time refers to the fact that we always find ordered configurations (as in the lower left picture) rather than disordered ones (as in the lower right picture) in any similar situation, i.e. not only for a chair in a room but also for other objects. Such ordered configurations can be naturally expected if the entire system is evolved from a state having a much lower coarse-grained entropy; otherwise, we would expect disordered ones since the number of states corresponding to disordered configurations is much larger than that corresponding to ordered ones.
In the context of the multiverse, the fact that we live in our universe and see the arrow of time tells us two things:
- A typical observer among all the “conscious” observers in the multiverse (including fluke, Boltzmann brain observers [@Dyson:2002pf]) must live in a universe consistent with our current knowledge, i.e. a universe whose low energy physics is described by the standard model of particle physics and cosmology.
- When we ask any conditional probability $P(B|A)$ within our universe, i.e. when precondition $A$ is chosen such that it selects a situation in our universe (e.g. my brain state), the answer should be dominated by one that arises from a low coarse-grained entropy state through evolution.
These two are the only things we definitely know from observation about the structure of the multiverse; for example, the arrow of time may not exist in other universes, i.e. the probabilities may be dominated by disordered configurations in those universes. What we must require is that the theory must (at least) be compatible with these two conditions.
The above discussion shows that the following two statements are literally [*equivalent*]{} as concepts: “An observer sees the arrow of time” and “There is no Boltzmann brain problem.” This is consistent with the picture presented recently by Bousso [@Bousso:2011aa], who analyzed the arrow of time in the context of the evolving multiverse in the landscape. Historically, the argument like the one here was first used to exclude the possibility that our universe, which has a positive cosmological constant, is absolutely stable [@Dyson:2002pf]. It was also argued in Ref. [@Nomura:2011rb] that it excludes the possibility that the multiverse is a closed, finite system if it has a [*generic*]{} initial condition in ${\cal H}_{\rm QG}$. This possibility, however, is allowed if the selection condition imposed on the entire multiverse state is special, as is the case in the scenario considered in this paper.
In summary, a selection condition imposed on the multiverse state must be such that the resulting probabilities are consistent with conditions (A) and (B) listed above. In particular, this leads to the following corollary:
- Any selection condition on $\ket{\Psi(t)}$ that leads to an (almost) equal probability for all the possible states in ${\cal H}_{\rm QG}$ corresponding to our universe is observationally excluded.
This is because such a scenario would lead to the probabilities being dominated by disordered configurations in our universe, contradicting observation. This condition will play an important role in rejecting a possible selection condition in Section \[sec:max-mixed\].
Is the multiverse really evolving? {#subsec:evolve}
----------------------------------
The consideration given above also illuminates the following question: is the multiverse really evolving? The answer is: it need not. In order to be consistent with the observed arrow of time, it is only necessary that the probabilities [*in our universe*]{} are dominated by configurations that are consistent with the hypothesis that the system has evolved from a lower coarse-grained entropy state. This, however, does not necessarily mean that the multiverse state $\ket{\Psi(t)}$ is actually evolving in $t$. It simply says that the probabilities obtained from $\ket{\Psi(t)}$ should be consistent with the hypothesis that our universe has evolved from a lower entropy state.
One might think that we actually “witnessed” that the state evolved as we came into being and grew. The interpretation of this fact, however, needs care—all we know is that our memory states are such that they are [*consistent with*]{} those obtained by interacting with environments that evolve from lower to higher entropy states. Similarly, we usually consider that our universe has evolved from the early big-bang, but all we really know is that the current state of the universe is consistent with the hypothesis that it has evolved from a lower entropy, big-bang state. As we have seen in the previous subsection, what these observations are really telling us is that in our universe different parts of physical configurations are correlated in certain (very) special ways. They do not mean that the multiverse state $\ket{\Psi(t)}$ [*must*]{} be evolving.
The question of whether a physical system is viewed as evolving or not, therefore, can be determined by asking questions about a “current” configuration, i.e. configuration at a fixed value of $t$. If the configuration is consistent with the hypothesis that the system has evolved from a lower entropy state, then we [*interpret*]{} it as the system evolving—it is not necessary that the state itself is actually changing with $t$. To do such a determination, it is enough to use the formula of Eq. (\[eq:prob\]), in which conditions $A$ and $B$ act at the same moment. In fact, in quantum mechanics, when we obtain information about a system we do that indirectly by observing imprints in the environment left by the system [@q-Darwinism], so this is almost exactly what we do in reality when we study the “history” of a system.
Summarizing, the observed flow of time does not require that the multiverse state is actually changing with $t$. It simply requires that the resulting probabilities satisfy the two conditions described in the previous subsection: (A) and (B). The probability formula in which conditions $A$ and $B$ both act at the same moment can be used to answer any physical questions, including those about a system that we interpret as dynamically evolving.
Selection Conditions and Operators {#sec:max-mixed}
==================================
We now start exploring possible selection conditions that can be imposed on the multiverse state. As stated in the introduction, we consider that the laws of quantum mechanics are not violated (Hypothesis I), which forces the multiverse state to exist for all values of $t$: from $-\infty$ to $+\infty$. This implies that, once a selection condition is given at a particular moment, which we take as $t = 0$, then the state is uniquely determined by solving the Schrödinger equation both forward and backward in $t$.
In this section, we ask the following question: can the selection condition be given in Hilbert space ${\cal H}_{\rm QG}$ without referring to any quantum operator? If this is possible, then it would imply that the form of the selection condition, written purely in terms of quantum states, must be basis independent, since we cannot specify a basis without knowledge of operators and how they act on elements in the Hilbert space. (Note that Hilbert space itself does not contain any physical information except for its dimension, i.e. any complex Hilbert spaces having the same dimension are identical with each other.) We will see that there is only one possible selection condition satisfying this criterion, and that it is observationally excluded. We will therefore learn that the expression for the selection condition in ${\cal H}_{\rm QG}$ must involve some information about the quantum operators.
The selection condition without an operator {#subsec:state-MM}
-------------------------------------------
Suppose that the multiverse is in a pure state, and that the selection condition at $t = 0$ is given by $$\ket{\Psi(0)} = \sum_i c_i \ket{\alpha_i},
\label{eq:bc-pure}$$ where $\ket{\alpha_i}$ represents a complete, orthonormal basis for the elements in ${\cal H}_{\rm QG}$, and $c_i$ are fixed coefficients characterizing the selection condition. Can the expression in Eq. (\[eq:bc-pure\])—[*including the values of $c_i$*]{}—be basis independent?
Consider that we perform an arbitrary basis change $$\ket{\alpha_i} = \sum_j U_{ij} \ket{\alpha'_j},
\label{eq:basis-change}$$ where $U_{ij}$ is an arbitrary unitary matrix. In the new basis, the expression in Eq. (\[eq:bc-pure\]) is written as $\ket{\Psi(0)} =
\sum_i c'_i \ket{\alpha'_i}$, where the new coefficients $c'_i$ are given by $c'_i = \sum_j c_j U_{ji}$. In order for the form of the selection condition to be basis independent, we need to have $$c_i = c'_i = \sum_j c_j U_{ji}
\label{eq:pure-indep}$$ [*for an arbitrary $U_{ij}$*]{}. This condition cannot be satisfied unless $c_i = 0$ for all $i$. Therefore, it is not possible to write a selection condition without referring to any quantum operator if the multiverse state is pure.
Suppose now that the multiverse is in an intrinsically mixed state, which takes the form $$\rho(0) = \sum_{i,j} d_{ij} \ket{\alpha_i} \bra{\alpha_j}
\label{eq:bc-mixed}$$ at $t = 0$, where $d_{ij}$ is a positive semi-definite Hermitian matrix. The basis change in Eq. (\[eq:basis-change\]) then leads to $\rho(0)
= \sum_i d'_{ij} \ket{\alpha'_i} \bra{\alpha'_j}$, where the new coefficients are given by $d'_{ij} = \sum_{k,l} U_{ik} d_{kl} U^*_{jl}$. In order for the selection condition to be basis independent, we must have $$d_{ij} = d'_{ij} = \sum_{k,l} U_{ik} d_{kl} U^*_{jl}
\label{eq:mixed-indep}$$ for an arbitrary $U_{ij}$. This has the unique solution (up to the overall coefficient): $$d_{ij} \propto \delta_{ij}.
\label{eq:d-indep}$$ We thus find that the requirement is satisfied if the multiverse state is specified by $$\rho(0) \propto \sum_i \ket{\alpha_i} \bra{\alpha_i},
\label{eq:rho-MM}$$ namely if the multiverse is in the maximally mixed state in ${\cal H}_{\rm QG}$ at $t = 0$.
Can the multiverse be in the maximally mixed state? {#subsec:max-ig}
---------------------------------------------------
Once the selection condition is given by Eq. (\[eq:rho-MM\]), the multiverse state $\rho(t)$ for arbitrary $t$ can be obtained using the evolution equation $$\rho(t) = U(t,0)\, \rho(0)\, U(0,t).
\label{eq:mixed-evol}$$ Since $\rho(0)$ is proportional to the unit matrix in ${\cal H}_{\rm QG}$, however, this gives $$\rho(t) = \rho(0),
\label{eq:mixed-0}$$ i.e. the multiverse is in the maximally mixed state at all times.
Equations (\[eq:rho-MM\]) and (\[eq:mixed-0\]) imply that all the possible states in ${\cal H}_{\rm QG}$ corresponding to our universe are equally probable. This is exactly the possibility that is observationally excluded by corollary ($\ast$) in Section \[subsec:arrow-time\]. Since we have arrived at this conclusion only by assuming that the selection condition is written without referring to a quantum operator in ${\cal H}_{\rm QG}$, we learn that the condition must in fact involve a quantum operator. The significance of this result lies in the fact that in quantum mechanics, operators are the objects that contain information about the system—the condition imposed on the multiverse state must reflect the structure of the system.
The Static Quantum Multiverse {#sec:stat}
=============================
What operators can be used in the condition imposed on the multiverse state? Since the multiverse contains many universes in which low energy physical laws differ, they cannot be “vacuum specific” operators. In this section, we identify candidate operators—those generating reference frame changes and that generating evolution.
We then impose the requirement that physical predictions are independent of a reference frame one chooses to describe the multiverse (Hypothesis II in the introduction). We will see that this implies that the multiverse state is independent of $t$, i.e. it must be static. As discussed in Section \[subsec:evolve\], this does not necessarily contradict observation. (The consistency with the observed flow of time will be discussed further in Section \[sec:consistency\].) We will also see that with Hypothesis I, Hypothesis II can be viewed as a consequence of requiring that the multiverse is in an eigenstate of global energy and boost operators with zero eigenvalues.
Reference frame changes {#subsec:ref-change}
-----------------------
Recall that in the framework of Refs. [@Nomura:2011dt; @Nomura:2011rb], quantum states allowing for spacetime interpretation, i.e. elements of ${\cal H} \subset {\cal H}_{\rm QG}$, represent only the spacetime regions inside and on the (stretched) apparent horizons as viewed from a fixed reference frame associated with a fixed reference point $p$. What happens if we change the reference frame?
Consider a state representing a configuration in de Sitter space. If we perform a spatial translation, which is equivalent to shifting the location of $p$, then it will necessarily mix the degrees of freedom inside and on the horizon because the state is defined only in the restricted spacetime region. This is precisely the phenomenon we call the observer dependence of the horizon: (some of) the degrees of freedom associated with internal space for one observer are described as those associated with the horizon by another. Next, consider a state which will later form a black hole, with $p$ staying outside of the black hole horizon. Such a state will not contain the spacetime region inside the black hole horizon because it will be outside $p$’s horizon. Now, imagine that we change the reference frame by performing a boost at an early time so that $p$ will be inside the black hole horizon at late times. In this new frame, the state at late times [*does*]{} contain the spacetime region inside the black hole horizon, although now it does [*not*]{} contain Hawking radiation quanta escaping to the future null infinity, which were included in the state before performing the reference frame change. This is exactly the phenomenon of black hole complementarity [@Susskind:1993if]. The present framework, therefore, allows us to understand the two phenomena described above in a unified manner as special cases of general reference frame changes [@Nomura:2011rb]; in particular, the concept of spacetime depends on the reference frame.
As any symmetry transformation, reference frame changes must be represented by unitary transformations acting on Hilbert space ${\cal H}_{\rm QG}$. What is the set of generators representing these transformations, and what is the algebra they satisfy?
In the limit $G_N \rightarrow 0$, the set of transformations associated with the reference frame changes and a shift of the origin of $t$ (time translation) is reduced to the standard Poincaré transformations, which is analogous to the fact that the standard Poincaré group is reduced to the Galilean group in the limit $c \rightarrow
\infty$ [@Nomura:2011rb]. Here, $G_N$ and $c$ are Newton’s constant and the speed of light, respectively. In the case of the reduction associated with $c \rightarrow \infty$, the structure of infinitesimal transformations changes. This is seen clearly in the Poincaré algebra: $$\begin{array}{c}
[ J_{[ij]}, J_{[kl]} ]
= i \left( \delta_{ik} J_{[jl]} - \delta_{il} J_{[jk]}
- \delta_{jk} J_{[il]} + \delta_{jl} J_{[ik]} \right),
\\[10pt]
[ J_{[ij]}, K_k ] = i \left( \delta_{ik} K_j - \delta_{jk} K_i \right),
\qquad
[ K_i, K_j ] = - \frac{i}{c^2} J_{[ij]},
\\[10pt]
[ J_{[ij]}, P_k ] = i \left( \delta_{ik} P_j - \delta_{jk} P_i \right),
\qquad
[ K_i, P_j ] = \frac{i}{c^2} \delta_{ij} H,
\qquad
[ P_i, P_j ] = 0,
\\[10pt]
[ J_{[ij]}, H ] = [ P_i, H ] = [ H, H ] = 0,
\qquad
[ K_i, H ] = i P_i,
\end{array}
\label{eq:Poincare-alg}$$ where $J_{[ij]}$, $K_i$, and $P_i$ are the generators of spatial rotations, boosts, and spatial translations, respectively, and we have exhibited $c$ explicitly. This algebra is reduced to a different algebra, i.e. that of the Galilean group, as $c \rightarrow \infty$: $$\begin{array}{c}
[ J_{[ij]}, J_{[kl]} ]
= i \left( \delta_{ik} J_{[jl]} - \delta_{il} J_{[jk]}
- \delta_{jk} J_{[il]} + \delta_{jl} J_{[ik]} \right),
\\[10pt]
[ J_{[ij]}, K_k ] = i \left( \delta_{ik} K_j - \delta_{jk} K_i \right),
\qquad
[ K_i, K_j ] = 0,
\\[10pt]
[ J_{[ij]}, P_k ] = i \left( \delta_{ik} P_j - \delta_{jk} P_i \right),
\qquad
[ K_i, P_j ] = i \delta_{ij} M,
\qquad
[ P_i, P_j ] = 0,
\\[10pt]
[ J_{[ij]}, H ] = [ P_i, H ] = [ H, H ] = 0,
\qquad
[ K_i, H ] = i P_i,
\end{array}
\label{eq:Galilean-alg}$$ where we have rescaled $H \rightarrow c^2 M + H$ to allow for the possibility that the original $H$ has a constant piece that goes as $c^2$. Can the algebra corresponding to the reference frame changes and time translation have extra terms beyond Eq. (\[eq:Poincare-alg\]) that disappears in the limit $G_N \rightarrow 0$?
One can immediately see that it cannot. The generators of the reference frame changes consist of $J_{[ij]}$, $K_i$, and $P_i$, while that of time translation is $H$. Taking natural units, the mass dimensions of these generators are $[J_{[ij]}] = [K_i] = 0$ and $[P_i] = [H] = 1$, while that of Newton’s constant is $[G_N] = -d+2$, where $d$ is the number of spacetime dimensions. It is then easy to find that for $d \geq 4$, where gravity is dynamical, there is no term one can add to the commutators in Eq. (\[eq:Poincare-alg\]) that is linear in generators and has a positive integer power of $G_N$.[^2] The algebra for the reference frame changes and time translation, therefore, is the same as that of the Poincaré transformations in Eq. (\[eq:Poincare-alg\]). The effect of nonzero $G_N$ appears as the reduction of the Hilbert space, but not in the transformation generators of the Poincaré group.
Selecting the multiverse state {#subsec:selec}
------------------------------
Let us now require that predictions do not depend on the reference frame one chooses to describe the multiverse (Hypothesis II). Physically, this implies that there is neither absolute center nor the frame of absolute rest in the multiverse.
Formally, our requirement can be stated as follows. Suppose we want to make physical predictions using projection operators ${\cal O}_X$, e.g. $X = A$, $A \cap B$, and so on. The relevant matrix elements are then $\bra{\Psi(t)} {\cal O}_X \ket{\Psi(t)}$. Now, consider a multiverse state as viewed from a different reference frame: $\ket{\Psi'(t)} = S \ket{\Psi(t)}$, where $S$ is the unitary operator representing the corresponding reference frame change. Our requirement is then $$\bra{\Psi(t)} {\cal O}_X \ket{\Psi(t)}
= \bra{\Psi'(t)} {\cal O}_X \ket{\Psi'(t)}
\label{eq:req}$$ [*for arbitrary $S$ and ${\cal O}_X$*]{}. Note that the operator in the right-hand side is [*not*]{} ${\cal O}'_X = S {\cal O}_X S^\dagger$, but the same ${\cal O}_X$ as in the left-hand side. This equation, therefore, has a nontrivial physical content, imposing constraints on the multiverse state. (If we had ${\cal O}'_X$ in the right-hand side, then the equation would simply represent a basis change, and thus would be trivial.)
In order to satisfy Eq. (\[eq:req\]), the multiverse state must satisfy $S \ket{\Psi(t)} \propto \ket{\Psi(t)}$, so that it must be a simultaneous eigenstate of operators $J_{[ij]}$, $K_i$ and $P_i$.[^3] One can then easily see from Eq. (\[eq:Poincare-alg\]) that this requires that the multiverse state is also an eigenstate of $H$, and that the eigenvalues under $J_{[ij]}$, $K_i$, $P_i$, and $H$ are all zero. The fact that the multiverse state is an eigenstate of $H$ with zero eigenvalue means that $$\frac{d}{dt}\ket{\Psi(t)} = 0,
\label{eq:d-dt_Psi}$$ i.e. the multiverse state is static! We can therefore write it simply as $\ket{\Psi} \equiv \ket{\Psi(t)} = \ket{\Psi(0)}$. The conditions coming from Hypothesis II can then be summarized as $$J_{[ij]} \ket{\Psi} = K_i \ket{\Psi} = P_i \ket{\Psi} = H \ket{\Psi} = 0.
\label{eq:Psi-cond}$$ This provides selection conditions for the multiverse state.
In fact, given Hypothesis I, the conditions in Eq. (\[eq:Psi-cond\]) follow from a standard procedure of quantizing a system with redundancies [@Dirac-book], if we assume that the multiverse state is invariant under the action of global energy and boost operators. In this procedure, any gauge redundancy, including general coordinate transformations, appears as a supplementary condition imposed on quantum states, which eliminates unphysical degrees of freedom from the states. Starting from a consistent, general covariant quantum theory of gravity (which is presumably string theory), the states are subject to a huge number of supplementary conditions, some of which will be used to reduce the number of degrees of freedom from that implied by local field theory to that suggested by the holographic principle, as in Eq. (\[eq:H\_M-dimension\]). (This implies that the number of constraints is much larger than that of the standard constraints associated with classical general coordinate transformations [@DeWitt:1967yk].) In this bigger (more redundant) picture, the framework of Refs. [@Nomura:2011dt; @Nomura:2011rb] corresponds to the scheme in which all the gauge redundancies are explicitly fixed, except for the ones associated with the reference frame changes. These residual redundancies, i.e. those of the reference frame changes, must then have their own supplementary conditions imposed on the states living in ${\cal H}_{\rm QG}$.
To illustrate this in a simple example, let us consider a spacetime that admits rectilinear coordinates $x_i$ in a constant $t$ hypersurface. In terms of Hamiltonian and momentum densities, ${\cal H}(x)$ and ${\cal P}_i(x)$, the Hilbert space ${\cal H}_{\rm QG}$ then corresponds to the space of states in which the constraints of the form $$\begin{aligned}
&& \int\! x_i x_j {\cal H}(x)\, d^3x \ket{\Psi}
= \int\! x_i x_j x_k {\cal H}(x)\, d^3x \ket{\Psi} = \cdots
\nonumber\\
&& \qquad = \int\! x_i x_j {\cal P}_k (x)\, d^3x \ket{\Psi}
= \int\! x_i x_j x_k {\cal P}_l(x)\, d^3x \ket{\Psi} = \cdots = 0,
\label{eq:const-1}\end{aligned}$$ as well as those associated with holography and complementarity, are [*already*]{} imposed; namely, the states in ${\cal H}_{\rm QG}$ satisfy these constraints by construction. On the other hand, the constraints of the form $$\int\! {\cal H}(x)\, d^3x \ket{\Psi}
= \int\! x_i {\cal H}(x)\, d^3x \ket{\Psi}
= \int\! {\cal P}_i (x)\, d^3x \ket{\Psi}
= \int\! x_i {\cal P}_j(x)\, d^3x \ket{\Psi} = 0
\label{eq:const-2}$$ are [*not*]{} imposed to obtain ${\cal H}_{\rm QG}$, so they must still be imposed on the states in ${\cal H}_{\rm QG}$. Now, the generators of time translation and the reference frame changes are given by $$\begin{array}{c}
H = \int\! {\cal H}(x)\, d^3x + \epsilon,
\qquad
P_i = \int\! {\cal P}_i(x)\, d^3x + p_i,
\\[10pt]
K_i = \int\! x_i {\cal H}(x)\, d^3x + k_i,
\qquad
J_{[ij]} = \int\! (x_i {\cal P}_j(x) - x_j {\cal P}_i(x)) d^3x + j_{[ij]},
\end{array}$$ where we have included global energy $\epsilon$ and momentum $p_i$ operators (and the corresponding quantities in $K_i$ and $J_{[ij]}$) that represent possible contributions from surface terms. Such terms can indeed arise in asymptotically Minkowski space, and play the role of what we consider the total energy and momentum of the system [@Arnowitt:1962hi].
Note that it is the effect of global energy $\epsilon$ that allows for any evolution of states in $t$ in quantum gravity, because $$\ket{\Psi(t_1)} = e^{-iH(t_1-t_2)} \ket{\Psi(t_2)}
= e^{-i\epsilon (t_1-t_2)} \ket{\Psi(t_2)},
\label{eq:gen-evol}$$ so unless $\ket{\Psi(t)}$ is a superposition of terms that give different values of $\epsilon$, the state is stationary. In this picture, our Hypothesis II corresponds to the assumption that the multiverse is an eigenstate of $\epsilon$ and $k_i$ with vanishing eigenvalues: $$\epsilon \ket{\Psi} = k_i \ket{\Psi} = 0,
\label{eq:zero-ep}$$ in which case we immediately see that $\ket{\Psi}$ also has zero eigenvalues under $p_i$ and $j_{[ij]}$, and that Eq. (\[eq:Psi-cond\]) follows from the constraints in Eq. (\[eq:const-2\]) (and vice versa). An important point is that for a state in ${\cal H}_{\rm QG}$, the surface terms reside on the (stretched) apparent horizon, so that Eq. (\[eq:zero-ep\]) is the assumption about the structure of the theory on this surface. This is in the intrinsically quantum gravitational regime, over which we currently do not have good theoretical control.
The selection of possible multiverse states, therefore, is boiled down to solving the infinite-dimensional matrix equations in Eq. (\[eq:Psi-cond\]). Here, we assume that there is no other selection condition, i.e. Eq. (\[eq:Psi-cond\]) is enough to fully select the system. (We assume that other supplementary conditions, e.g. those associated with standard gauge symmetries, are already taken care of. Also, since all the redundancies associated with gravity other than those corresponding to the reference frame changes are supposed to be fixed in the present framework [@Nomura:2011dt], there are no more conditions arising from considerations of gravity.) We look for solutions to Eq. (\[eq:Psi-cond\]) of the form $$\ket{\Psi} = \sum_i c_i \ket{\alpha_i},
\qquad
\sum_i |c_i|^2 < \infty,
\label{eq:sol-cond}$$ where $\ket{\alpha_i}$ represents a complete, orthonormal basis in ${\cal H}_{\rm QG}$, so that the sums of $i$ run to infinity; see Eq. (\[eq:dim-inf\]). The normalizability condition here is imposed for the following (usual) reason. Suppose there are normalizable solutions $\ket{\Psi_I}$ ($I = 1,\cdots,N$) satisfying Eq. (\[eq:sol-cond\]), as well as non-normalizable solutions $\ket{\Psi_I}$ ($I = N+1,\cdots,K$). The non-normalizable solutions will have coefficients which strongly diverge as the dimensions of corresponding Hilbert subspaces ${\cal H}_{\cal M}$ become large. This is because the process transforming an element of ${\cal H}_{\cal M}$ to that of ${\cal H}_{\cal M'}$ with ${\rm dim}\,{\cal H}_{\cal M'} < {\rm dim}\,{\cal H}_{\cal M}$ becomes highly suppressed as ${\rm dim}\,{\cal H}_{\cal M}$ gets large (because of Eq. (\[eq:dim-inf\])). Let us now imagine regulating the sums of $i$ as $\sum_i \rightarrow \sum_{i=1}^{n}$, in which case we can normalize all the solutions so that $\inner{\Psi_I}{\Psi_J} = \delta_{IJ}$ for $I,J = 1,\cdots,K$. We can then consider state $\rho = \sum_{I,J=1}^{K}
d_{IJ} \ket{\Psi_I} \bra{\Psi_J}$ with arbitrary finite positive semi-definite Hermitian matrix $d_{IJ}$, and calculate probabilities arising from $\rho$ using a projection operator that selects (a finite number of) configurations compatible with some condition $X$: ${\cal O}_X
= \sum_{i \in X} \ket{\alpha_i} \bra{\alpha_i}$. The resulting probabilities are the same as those arising from $\rho' = \sum_{I,J=1}^{N}
d_{IJ} \ket{\Psi_I} \bra{\Psi_J}$, i.e. the state obtained by eliminating all the non-normalizable solutions from $\rho$, up to terms disappearing for $n \rightarrow \infty$. Therefore, the non-normalizable solutions can all be dropped from physical considerations.
The Hilbert space relevant for the multiverse ${\cal H}_{\rm Multiverse}$, then, is spanned by the normalizable solutions to Eq. (\[eq:Psi-cond\]), and so is much smaller than ${\cal H}_{\rm QG}$: $${\cal H}_{\rm Multiverse} \subset {\cal H}_{\rm QG},
\qquad
{\rm dim}\,{\cal H}_{\rm Multiverse} \ll {\rm dim}\,{\cal H}_{\rm QG}.
\label{eq:H_multiverse}$$ We note that this situation is analogous to usual quantum mechanical systems, e.g. a hydrogen atom. In the hydrogen atom, the state factor corresponding to a radial wavefunction $c(r)$ can be written as $\ket{\psi}
= \int_0^\infty\!dr\, c(r) \ket{r}$. The only states relevant to physics of the hydrogen atom are those satisfying the normalizability condition $\int_0^\infty\!dr\, |c(r)|^2 < \infty$ in the Hilbert space spanned by $\ket{r}$. The other, non-normalizable solutions (which behave as $\ln c(r) \sim r$ at large $r$) are irrelevant. The situation in the quantum multiverse is similar. The non-normalizable solutions have infinitely strong supports in supersymmetric Minkowski vacua or singularity worlds, which have infinite-dimensional Hilbert spaces. These solutions, therefore, are irrelevant in making predictions in a “realistic world,” i.e. in a universe that has nonzero free energy. The only relevant states are those that are normalizable in the Hilbert space of quantum gravity, ${\cal H}_{\rm QG}$. For a schematic drawing of this analogy, see Fig. \[fig:analogy\].
![A schematic depiction of the analogy between the hydrogen atom and the quantum multiverse. In the case of the hydrogen atom, the only relevant states are those that satisfy the Schrödinger equation and are normalizable in the Hilbert space spanned by $\ket{r}$ (solid line); the non-normalizable modes are irrelevant (dashed line). In the quantum multiverse, the relevant states are those that satisfy Eq. (\[eq:Psi-cond\]) and are normalizable in Hilbert space ${\cal H}_{\rm QG}$ (solid line); the non-normalizable modes, which have diverging coefficients for supersymmetric Minkowski or singularity states, are irrelevant (dashed line).[]{data-label="fig:analogy"}](analogy.epsi){width="17cm"}
The static multiverse states in [${\cal H}_{\rm QG}$]{} {#subsec:static}
-------------------------------------------------------
We now discuss how our conditions Eqs. (\[eq:d-dt\_Psi\], \[eq:sol-cond\]) can be compatible with Eq. (\[eq:asympt\]), which says that a generic multiverse state in ${\cal H}_{\rm QG}$ will evolve into a superposition of supersymmetric Minkowski and singularity states as $t \rightarrow
\infty$. In order for Eq. (\[eq:d-dt\_Psi\]) to be satisfied, the coefficients $c_i$ of all the terms in $\ket{\Psi(t)} = \ket{\Psi}$ must be constant when expanded in components $\ket{\alpha_i}$. In a basis in which $\ket{\alpha_i}$ in ${\cal H}$ have well-defined semi-classical configurations, the evolution operator $\exp(-iHt)$ (and thus $H$ as well) is not diagonal. Therefore, the processes in Eq. (\[eq:asympt\]) will occur for [*generic*]{} $\ket{\Psi(t)}$, but they must exactly be canceled by some “inverse processes” in $\ket{\Psi}$. In particular, in order for the normalization condition in Eq. (\[eq:sol-cond\]) to be satisfied, this must occur before the state is dissipated into infinite-dimensional Hilbert space.
Let us consider a physical configuration in a Minkowski universe in which there is a bubble wall surrounding us, which, however, is contracting toward us rather than expanding away. Such a configuration, which is exactly the time reversal of a usual expanding bubble configuration, is physically allowed, as the fundamental equation of the theory is symmetric under $t \rightarrow -t$. Usually, we do not consider this kind of configuration as it is only an exponentially small subset of all the configurations allowed by the theory; in particular, there is only an exponentially small probability for forming such a configuration starting from a generic, e.g. thermal, state. We are, however, now considering very special states, i.e. the states that satisfy Eqs. (\[eq:d-dt\_Psi\], \[eq:sol-cond\]), and in these states such configurations could balance the “loss” of semi-classically unstable states in Eq. (\[eq:asympt\]). For example, the entire multiverse state is so “fine-tuned” that a reheating that occurs in a Minkowski universe produces exactly the configuration that puts the system back to (a superposition of) states in unstable vacua. Similar processes must also occur for singularities. Note that since these processes are exponentially suppressed under normal circumstances, they are invisible in the usual semi-classical analysis.
The states given by Eq. (\[eq:sol-cond\]) are the ones in which all these and other processes are balanced.[^4] Since the inverse processes are unlikely to occur at the zero density, these states will explore only a finite-dimensional portion of Minkowski vacua (see the discussion at the end of Section \[subsec:Hilbert\]). The number of independent states, therefore, may well be finite: ${\rm dim}\,{\cal H}_{\rm Multiverse} < \infty$, which we will assume to be the case. Note that the sizes of various elements in $H$ represented as a matrix acting on ${\cal H}_{\rm QG}$ differ significantly; in fact, they are expected to differ exponentially, or even double-exponentially, as some of the processes are highly suppressed. This implies that the values of $|c_i|$’s in Eq. (\[eq:sol-cond\]) will also vary significantly. The resulting states $\ket{\Psi}$, therefore, are not excluded by corollary ($\ast$) in Section \[subsec:arrow-time\]. The structure of $\ket{\Psi}$, and its consistency with observation, will be discussed further in Section \[sec:consistency\].
Predictions in the static quantum multiverse {#subsec:pred}
--------------------------------------------
The number of independent normalizable solutions to Eq. (\[eq:Psi-cond\]) will depend on the structure of the multiverse, i.e. issue (a) in Section \[subsec:s-c\]. In particular, the existence of a solution requires $H$ to take a certain special form (so that it has at least one normalizable, zero-eigenvalue eigenvector), which we assume to be the case. Suppose there are $N$ such solutions $\ket{\Psi_I}$ ($I =
1,\cdots,N = {\rm dim}\,{\cal H}_{\rm Multiverse} < \infty$). How can the physical predictions be made?
If $N = 1$, the multiverse state is simply $\ket{\Psi} \equiv
\ket{\Psi_1}$. The probabilities are then given by the generalized Born rule, Eq. (\[eq:prob\]), but now without the $t$ integrals. (They simply give a constant factor $\int_{-\infty}^{+\infty} dt$, which cancels between the numerator and denominator.) The final formula is given by Eq. (\[eq:prob-final\]), which we reproduce here: $$P(B|A) = \frac{\bra{\Psi} {\cal O}_{A \cap B} \ket{\Psi}}
{\bra{\Psi} {\cal O}_A \ket{\Psi}}.
\nonumber$$ As discussed in Section \[subsec:evolve\], this formula can be used to answer any physical questions, including those about a system that we view as dynamically evolving.
In the case that $N > 1$, any multiverse states of the form $\ket{\Psi}
= \sum_{I=1}^N c_I \ket{\Psi_I}$ or $\rho = \sum_{I,J=1}^{N} d_{IJ}
\ket{\Psi_I} \bra{\Psi_J}$ are allowed. In the absence of more information (or selection conditions), it is natural to assume that the multiverse is in the maximally mixed state $$\rho = \frac{1}{N} \sum_{I=1}^N \ket{\Psi_I} \bra{\Psi_I},
\label{eq:multiverse-mixed}$$ where we have taken $\ket{\Psi_I}$’s to be orthonormal. This state is invariant under the basis change $\ket{\Psi_I} \rightarrow U_{IJ}
\ket{\Psi_J}$, and is reduced to $\ket{\Psi} = \ket{\Psi_1}$ for $N = 1$. The probabilities are given by the mixed-state version of Eq. (\[eq:prob-final\]): $$P(B|A) = \frac{{\rm Tr}\left[ \rho\, {\cal O}_{A \cap B} \right]}
{{\rm Tr}\left[ \rho\, {\cal O}_A \right]}.
\label{eq:prob-final-mixed}$$ Note that Eq. (\[eq:multiverse-mixed\]), i.e. the maximally mixed state in ${\cal H}_{\rm Multiverse}$, is different from Eq. (\[eq:rho-MM\]), i.e. the maximally mixed state in ${\cal H}_{\rm QG}$, in which the sum runs over all the possible states in ${\cal H}_{\rm QG}$ including the ones that do not satisfy Eq. (\[eq:Psi-cond\]). The state in Eq. (\[eq:multiverse-mixed\]), therefore, is not excluded by corollary ($\ast$) in Section \[subsec:arrow-time\].
Consistency with Observation {#sec:consistency}
============================
In this section we discuss the consistency of the present scenario with observation, specifically the observed arrow of time. Our approach here will be to allow for making assumptions on the structures of $H$ and ${\cal H}_{\rm QG}$ (unless they are inconsistent with what we already know about string theory), and to see if the scenario is consistent. We do not claim that all of these assumptions are absolutely necessary—our purpose here is to argue that, despite its naive appearance, the scenario is not excluded by observation. More detailed analysis/modeling of the landscape will be left for future work.
The structure of [${\cal H}_{\rm QG}$]{} {#subsec:structure}
----------------------------------------
Solutions to Eq. (\[eq:Psi-cond\]) depend on the structure of ${\cal H}_{\rm QG}$ as well as the form of $H$ (and other operators). Here we assume that ${\cal H}_{\rm QG}$ contains only “cosmologically relevant” states. The minimally required set of ${\cal H}_{\cal M}$’s that must be included in ${\cal H}$, i.e. in the right-hand side of Eq. (\[eq:ST-H\]), will then be those of FRW universes corresponding to all the possible vacua in the theory (and their straightforward generalizations, e.g. those of FRW universes with black holes). Not all spacetime must be contained in ${\cal H}$; for example, ${\cal H}$ need not contain a stable anti-de Sitter space without a singularity, which might only be a mathematical idealization because it does not arise through dynamical evolution in the FRW universes.
For each vacuum $I$ of the theory, the number of states associated with an FRW universe in $I$ is estimated as $${\cal N}_I \,\,=
\sum_{n=\exp({\cal A}_{I,{\rm min}}/2)}^{\exp({\cal A}_{I,{\rm max}}/2)}
\!\!\!\! n
\,\,\simeq\,\,
\frac{1}{2}\, e^{{\cal A}_{I,{\rm max}}},
\label{eq:N_I}$$ where ${\cal A}_{I,{\rm min}}$ and ${\cal A}_{I,{\rm max}}$ are the minimum and maximum areas of the apparent horizon in this universe, and we have used ${\cal A}_{I,{\rm max}} \gg {\cal A}_{I,{\rm min}}$ in the last equation. While possible deformations of the apparent horizon, e.g. by the existence of black holes, can have corrections to the explicit expression, we expect that the above estimate gives a qualitatively correct result: $\ln {\cal N}_I \approx O({\cal A}_{I,{\rm max}})$. The area ${\cal A}_{I,{\rm max}}$ is given by the inverse of the absolute value of the vacuum energy density (in Planck units) ${\cal A}_{I,{\rm max}}
\sim 1/|\rho_{\Lambda,I}|$, since in a de Sitter universe the apparent horizon approaches the event horizon at late times, while in an anti-de Sitter universe it has the maximum area when $p$ hits the singularity at $t \sim 1/|\rho_{\Lambda,I}|^{1/2}$. We therefore find $$\ln {\cal N}_I \sim \frac{1}{|\rho_{\Lambda,I}|}.
\label{eq:ln-N_I}$$ This implies that the number of states associated with a vacuum with $\rho_{\Lambda,I} \neq 0$ is finite.
The arrow of time in the static multiverse {#subsec:arrow}
------------------------------------------
We now consider a solution to the equation $H \ket{\Psi} = 0$, a part of the conditions in Eq. (\[eq:Psi-cond\]). We can view this equation as requiring that $\ket{\Psi}$ is in a stationary state in ${\cal H}_{\rm QG}$. (In fact, the equation is stronger than that, since the eigenvalue of $H$ must be zero.) In particular, it implies that the probability current creating states in vacuum $I$ must be balanced with that destroying those for each $I$ (in fact, each state in $I$). At the semi-classical level, this condition is impossible to satisfy for terminal vacua. As discussed in Section \[subsec:static\], however, our state is special, obtained after solving the “quantization condition” $H \ket{\Psi} = 0$, so that it can also be satisfied for these vacua.
Let us now consider vacuum $J$ that can support any observer, either an ordinary observer or a Boltzmann brain. We will argue that the arrow of time is predicted if the following three conditions are met for all possible $J$’s:
- Transitions to states in $J$ from those in other vacua are mainly through the states having low coarse-grained entropies in $J$, i.e. elements of ${\cal H}_{\cal M}$ with $\ln {\rm dim}\,{\cal H}_{\cal M}
\ll {\cal A}_{J,{\rm max}}$.
- Subsequent evolution in vacuum $J$ produces ordinary observers with probability $\epsilon_J$, which may be suppressed exponentially but not double-exponentially.
- The rate of producing Boltzmann brains $\Gamma_{{\rm BB},J}$ in vacuum $J$, which is double-exponentially suppressed (see, e.g. [@Bousso:2008hz]), is smaller than the decay rate $\Gamma_J$ of the vacuum itself.
Namely, if the structure of $H$ is such that it satisfies all these conditions, then the scenario is compatible with observation. (The “transitions” and “evolution” here, of course, refer to the apparent ones in $\ket{\Psi}$, which is in itself static.)
To see this, let us consider the distribution of the size of the coefficients $|c^J_i|$ of various terms in $\ket{\Psi}$ corresponding to the states in vacuum $J$, $\ket{\alpha^J_i}$. For this purpose, we define the quantity $P^J_\tau$ corresponding to the probability for a universe to be at FRW time $t_{\rm FRW}$ between $\tau$ and $\tau+d\tau$: $$P^J_\tau\, d\tau
= \sum_{i | \tau < t_{\rm FRW} < \tau+d\tau}\!\!\!\! |c^J_i|^2,$$ where $t_{\rm FRW}$ should be specified by physical configurations in $\ket{\alpha^J_i}$. The distribution of $P^J_\tau$ then follows from the definition of $\Gamma_J$: $$P^J_\tau = P^J_0 e^{-\Gamma_J\, \tau},
\label{eq:PJtau}$$ where we have assumed that the transitions to states in $J$ occur at $\tau = 0$ either through Coleman-De Luccia [@Coleman:1980aw] or Hawking-Moss [@Hawking:1981fz] processes (or their inverses), although our conclusion is insensitive to this assumption. Note that in these cases it is indeed natural to expect that states just after the transitions are the ones having low coarse-grained entropies, i.e. in ${\cal H}_{\cal M}$ with $\ln {\rm dim}\,{\cal H}_{\cal M}
\ll {\cal A}_{J,{\rm max}}$, because both the start and end points of the Coleman-De Luccia tunneling in field space are away from local minima (if the false vacuum has a positive vacuum energy), and the Hawking-Moss transition is a thermal process occurring through the field climbing up the potential barrier [@Weinberg:2006pc].
Now, the definitions of $\epsilon_J$ and $\Gamma_{{\rm BB},J}$ in (II) and (III) above imply that if we compute the probability of $\ket{\Psi}$ containing ordinary observers (OO) or Boltzmann brains (BB) in vacuum $J$ using the corresponding projection operators ${\cal O}_{{\rm OO},J}$ and ${\cal O}_{{\rm BB},J}$, then we obtain $$\begin{aligned}
&& \bra{\Psi} {\cal O}_{{\rm OO},J} \ket{\Psi} \sim \epsilon_J P^J_0,
\\
&& \bra{\Psi} {\cal O}_{{\rm BB},J} \ket{\Psi}
\sim \Gamma_{{\rm BB},J} \int\! P^J_\tau\, d\tau
= \frac{\Gamma_{{\rm BB},J}}{\Gamma_J} P^J_0.
\label{eq:OO-BB}\end{aligned}$$ Here, the projection operators select observers in a specific range of location and velocity with respect to $p$, although the results do not depend on the chosen location or velocity because of Eq. (\[eq:Psi-cond\]). Under conditions (II) and (III), this gives $$\frac{\bra{\Psi} {\cal O}_{{\rm BB},J} \ket{\Psi}}{\bra{\Psi}
{\cal O}_{{\rm OO},J} \ket{\Psi}}
\sim \frac{\Gamma_{{\rm BB},J}}{\epsilon_J \Gamma_J} \lll 1,
\label{eq:BB-OO-ratio}$$ where we have used the fact that $\Gamma_{{\rm BB},J}$ is double-exponentially suppressed while $\epsilon_J$ is not. (In fact, we only need $\epsilon_J > \Gamma_{{\rm BB},J}/\Gamma_J$ to obtain this result, so $\epsilon_J$ may be double-exponentially suppressed.) We therefore find that the overwhelming majority of observers are indeed ordinary observers, and thus perceive time’s arrow (as discussed in Section \[subsec:max-ig\]).
Perhaps not surprisingly, the conditions described above are similar to the ones obtained in Ref. [@Bousso:2011aa] in the context of the evolving multiverse, despite the fact that the overall physical pictures are rather different. One distinct feature of the present scenario in this respect is that since there is no “initial vacuum,” the absolute nonexistence of Boltzmann brains in such a vacuum ($\Gamma_{{\rm BB},\ast} = 0$ in the notation of Ref. [@Bousso:2011aa]) need not be imposed. In any case, as discussed in Ref. [@Bousso:2011aa], the conditions described above, in particular (I), are likely to be satisfied in the string landscape. It is, therefore, quite promising that the scenario discussed in this paper is indeed consistent with observation in the realistic string theory setup.
Discussions {#sec:discuss}
===========
In this paper we have studied the multiverse in the quantum mechanical framework recently proposed in Refs. [@Nomura:2011dt; @Nomura:2011rb]. By requiring that the laws of quantum mechanics are not violated (Hypothesis I) and that physical predictions do not depend on the reference frame one chooses to described the multiverse (Hypothesis II), we have found that the multiverse state must be static; in particular, the multiverse does not have a beginning or end.
Despite its naive appearance, the scenario does not contradict observation, including the fact that we observe that time flows in a definite direction. [*The arrow of time is simply an emergent phenomenon that is occurring in the branch (terms) corresponding to our universe in the static multiverse state*]{}—the terms that would be obtained by evolving the system from lower entropy states have much larger coefficients than the terms that cannot. The scenario is summarized by the selection conditions in Eq. (\[eq:Psi-cond\]), imposed on the states in ${\cal H}_{\rm QG}$. With these conditions, any multiversal questions can be answered using the Born rule, Eq. (\[eq:prob-final\]) or (\[eq:prob-final-mixed\]), [*without any additional input*]{}, once the explicit form of the operators such as $H$ is known. This scenario, therefore, provides a completion of the framework of the quantum multiverse in Refs. [@Nomura:2011dt; @Nomura:2011rb].
The supplementary condition of the form $H \ket{\Psi} = 0$ has certainly been considered before—indeed, this is nothing but the well-known Wheeler-DeWitt equation [@DeWitt:1967yk]. The scenario presented here, however, differs from standard applications of this equation in several important ways:
- The redundancies associated with gravity are much larger than what are usually imagined. In particular, they reduce the Hilbert space in such a way that it contains only the spacetime region within the reference point’s (stretched) apparent horizon [@Nomura:2011dt; @Nomura:2011rb]. This is important to avoid ambiguities associated with eternally inflating spacetime. The ultimate origin of these large redundancies will, presumably, be string theory.
- We apply the supplementary conditions corresponding to the whole set of time translation and reference frame changes with zero global charges, even if the universe is not closed. Since spacetime is defined only within the apparent horizon, this requires the assumption on the structure of the theory on this surface, which is intrinsically quantum mechanical. Note that it is this assumption that is responsible for the static nature of the multiverse state, which in turn excludes the possibility for the multiverse to have a beginning or end.
- We analyze the consequences of the supplementary conditions at the microscopic level. This selects very special states that are not visible in the analysis at the semi-classical level. In fact, normalizable solutions to the conditions correspond to the states in which the processes of Eq. (\[eq:asympt\]) are balanced with the inverse processes, which put the system back from terminal vacua to unstable vacua.
It is quite satisfying that such simple requirements as Hypotheses I and II lead to a consistent and predictive scheme for the entire multiverse.
Finally, it is instructive to draw a close analogy between the situation in the quantum multiverse described here and that in the standard, hydrogen atom. As is well known, the hydrogen atom cannot be correctly described using classical mechanics. Any orbit of the electron is unstable with respect to the emission of synchrotron radiation. Even if we artificially ignore the emission, the electron can orbit the nucleus at an arbitrary radius, unable to explain the discrete spectral lines. The solution to these problems is intrinsically quantum mechanical, i.e. quantum mechanics is responsible for the very existence of the hydrogen atom, not just providing a correction to the classical picture.
The situation in the quantum multiverse is similar. At the semi-classical level, the multiverse is unstable to the decay to terminal states, such as supersymmetric Minkowski vacua and singularities. Even if we artificially ignore the process of vacuum decays, it would lead to phenomena such as Poincaré recurrence, contradicting observation (the dominance of Boltzmann brains). The picture presented here says that the solution to these problems is [*intrinsically quantum mechanical*]{}—one cannot see it in the usual semi-classical analysis. The multiverse state is very special: a normalizable state satisfying the “quantization conditions” of Eq. (\[eq:Psi-cond\]), as in the case of the hydrogen atom. In the case of the hydrogen atom, these conditions make the dimension of Hilbert space from continuous infinity $\psi(r, \theta, \varphi)$ to countable infinity $(n,l,m)$. In the quantum multiverse, they will presumably make it from countable infinity to finite: ${\rm dim}\,{\cal H}_{\rm QG} \rightarrow
{\rm dim}\,{\cal H}_{\rm Multiverse}$.
After all, quantum mechanics treats the multiverse very similarly to the hydrogen atom. Our job is then to figure out the precise structure of the multiverse, a system which we are a part of. Hopefully, further progress in string theory will serve this purpose.
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to thank Alan Guth and Grant Larsen for useful conversations. This work was supported in part by the Director, Office of Science, Office of High Energy and Nuclear Physics, of the US Department of Energy under Contract DE-AC02-05CH11231, and in part by the National Science Foundation under grant PHY-0855653.
[99]{}
Y. Nomura, JHEP [**11**]{}, 063 (2011) \[arXiv:1104.2324 \[hep-th\]\]. Y. Nomura, arXiv:1110.4630 \[hep-th\]. A. H. Guth and E. J. Weinberg, Nucl. Phys. B [**212**]{}, 321 (1983); A. Vilenkin, Phys. Rev. D [**27**]{}, 2848 (1983); A. D. Linde, Phys. Lett. B [**175**]{}, 395 (1986); Mod. Phys. Lett. A [**1**]{}, 81 (1986). For reviews, see e.g.A. H. Guth, Phys. Rept. [**333**]{}, 555 (2000) \[arXiv:astro-ph/0002156\]; A. Vilenkin, J. Phys. A [**40**]{}, 6777 (2007) \[arXiv:hep-th/0609193\]; S. Winitzki, Lect. Notes Phys. [**738**]{}, 157 (2008) \[arXiv:gr-qc/0612164\]; A. Linde, Lect. Notes Phys. [**738**]{}, 1 (2008) \[arXiv:0705.0164 \[hep-th\]\]. Y. Nomura, AstRv. [**7**]{}, 36 (2012) \[arXiv:1205.2675 \[hep-th\]\]. A. Mithani and A. Vilenkin, arXiv:1204.4658 \[hep-th\]; L. Susskind, arXiv:1204.5385 \[hep-th\]; arXiv:1205.0589 \[hep-th\]. L. Susskind, L. Thorlacius and J. Uglum, Phys. Rev. D [**48**]{}, 3743 (1993) \[arXiv:hep-th/9306069\]; L. Susskind, Phys. Rev. Lett. [**71**]{}, 2367 (1993) \[arXiv:hep-th/9307168\]; C. R. Stephens, G. ’t Hooft and B. F. Whiting, Class. Quant. Grav. [**11**]{}, 621 (1994) \[arXiv:gr-qc/9310006\]. G. ’t Hooft, arXiv:gr-qc/9310026; L. Susskind, J. Math. Phys. [**36**]{}, 6377 (1995) \[arXiv:hep-th/9409089\]; R. Bousso, JHEP [**07**]{}, 004 (1999) \[arXiv:hep-th/9905177\]. R. Bousso and J. Polchinski, JHEP [**06**]{}, 006 (2000) \[arXiv:hep-th/0004134\]; S. Kachru, R. Kallosh, A. Linde and S. P. Trivedi, Phys. Rev. D [**68**]{}, 046005 (2003) \[arXiv:hep-th/0301240\]; L. Susskind, arXiv:hep-th/0302219; M. R. Douglas, JHEP [**05**]{}, 046 (2003) \[arXiv:hep-th/0303194\]. See, e.g., R. Bousso, Rev. Mod. Phys. [**74**]{}, 825 (2002) \[hep-th/0203101\], Section VII A.
S. Weinberg, Phys. Rev. Lett. [**59**]{}, 2607 (1987). G. Larsen, Y. Nomura and H. L. L. Roberts, Phys. Rev. D [**84**]{}, 123512 (2011) \[arXiv:1107.3556 \[hep-th\]\]. L. Dyson, M. Kleban and L. Susskind, JHEP [**10**]{}, 011 (2002) \[arXiv:hep-th/0208013\]. R. Bousso, arXiv:1112.3341 \[hep-th\]. H. Ollivier, D. Poulin and W. H. Zurek, Phys. Rev. Lett. [**93**]{}, 220401 (2004) \[arXiv:quant-ph/0307229\]; R. Blume-Kohout and W. H. Zurek, Phys. Rev. A [**73**]{}, 062310 (2006) \[arXiv:quant-ph/0505031\].
P. A. M. Dirac, [*Lectures on Quantum Mechanics*]{}, (Belfer Graduate School of Science, Yeshiva University, New York, 1964).
B. S. DeWitt, Phys. Rev. [**160**]{}, 1113 (1967). R. Arnowitt, S. Deser and C. W. Misner, in [*Gravitation: An Introduction to Current Research*]{}, ed. L. Witten (John Wiley & Sons, New York, 1962) \[gr-qc/0405109\]. R. Bousso, B. Freivogel and I-S. Yang, Phys. Rev. D [**79**]{}, 063513 (2009) \[arXiv:0808.3770 \[hep-th\]\]; A. De Simone, A. H. Guth, A. Linde, M. Noorbala, M. P. Salem and A. Vilenkin, Phys. Rev. D [**82**]{}, 063520 (2010) \[arXiv:0808.3778 \[hep-th\]\]. S. R. Coleman and F. De Luccia, Phys. Rev. D [**21**]{}, 3305 (1980). S. W. Hawking and I. G. Moss, Phys. Lett. B [**110**]{}, 35 (1982). E. J. Weinberg, Phys. Rev. Lett. [**98**]{}, 251303 (2007) \[hep-th/0612146\].
[^1]: There are several varying, though related, definitions of the measure problem in literature. In this paper we adopt the definition as stated here.
[^2]: For $d=3$, one can add terms $\varDelta [J, K_i] = i\gamma
G_N P_i$ and $\varDelta [K_1,K_2] = -i\gamma G_N H$, where $J \equiv
J_{[12]}$ and $\gamma$ is a real constant, without violating Jacobi identities. The significance of this is not clear.
[^3]: It is, in principle, possible that the predictions are reference frame independent because the multiverse is in an intrinsically mixed state that satisfies $S \rho(t) S^\dagger = \rho(t)$ at all $t$ but each component $\ket{\psi_i(t)}$ in $\rho(t)$ is [*not*]{} a simultaneous eigenstate of all the $S$’s. This is, however, the case only if $\rho(t)$ is the maximally mixed state in ${\cal H}_{\rm QG}$ (because of Schur’s lemma), which is observationally excluded as we saw in Section \[subsec:max-ig\]. We must therefore require that each pure-state component leads to reference-frame independent predictions even if the multiverse is in a mixed state.
[^4]: There is also the possibility that some (or all) of the states given by Eq. (\[eq:sol-cond\]) do not contain any Minkowski or singularity components. This does not affect any of our discussions below.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Positrons are accumulated within a Penning trap designed to make more precise measurements of the positron and electron magnetic moments. The retractable radioactive source used is weak enough to require no license for handling radioactive material and the radiation dosage one meter from the source gives an exposure several times smaller than the average radiation dose on the earth’s surface. The 100 mK trap is mechanically aligned with the 4.2 K superconducting solenoid that produces a 6 tesla magnetic trapping field with a direct mechanical coupling.'
author:
- 'S. Fogwell Hoogerheide'
- 'J. C. Dorr'
- 'E. Novitski'
- 'G. Gabrielse'
title: 'High efficiency positron accumulation for high-precision magnetic moment experiments'
---
Introduction
============
A great triumph of the standard model of particle physics [@StandardModelTriumph] is the remarkable, part-per-trillion agreement between the most precise measurement of a property of an elementary particle and the standard model’s most precise prediction. The electron magnetic moment is measured to 3 parts in $10^{13}$ using one electron in a 100 mK Penning trap [@HarvardMagneticMoment2008] . The standard model’s prediction of the same quantity comes from the Dirac equation, tenth order quantum electrodynamics (QED) [@TenthOrderQED2012], hadronic contributions [@MuonLightByLight], and weak contributions [@Weak1972; @Marciano:1996; @Weak2002; @Weak2003] as well as a measured value of the fine structure constant [@RbAlpha2011].
The positron magnetic moment could be measured just as precisely with a positron suspended in the same trap apparatus. Comparing the positron and electron moments would test the CPT theorem (charge conjugation, parity and time reversal) with leptons about 15 times more precisely than the best previous comparison of a lepton and antilepton [@DehmeltMagneticMoment]. This fundamental theorem of the standard model predicts that the positron moment should be opposite in sign but equal in magnitude to the electron moment.
In some ways, positrons from a radioactive source are compatible with the cryogenic, high vacuum environment needed for the most precise measurements. A Na$^{22}$ source has a 2.6 year half-life that is long enough to make measurements but short enough to avoid long-term disposal issues. The activity of the Na$^{22}$ sources used previously to trap positrons range from about 0.5 mCi to 75 mCi. The 0.5 mCi source, the smallest of these, was fixed to a 4.2 K trap to accumulate trapped positrons at a rate of 0.8 positron/min/mCi [@UwPositrons] – just high enough for measurements that require only a single trapped positron.
One of two drawbacks, however, is that even this source activity is fifty times higher than the 10 $\mu$Ci safety threshold that necessitates a license and a higher level of special handling for the radioactive materials [@Na22LicenseThreshold] – a significant complication for doing the measurement. The second drawback, not well understood, is that a source fixed near the 4.2 K electrodes was observed to result in spontaneous electron loading during electron measurements.
Here we report using a 6.5 $\mu$Ci source (about 77 times weaker) to load positrons at a rate above 150 positrons/min/mCi (about 190 times higher) into a 100 mK trap within a new apparatus (Fig. \[fig:WholeApparatus\]). The loading mechanism used was initially developed for efficiently accumulating positrons directly into high vacuum for antihydrogen production [@PositronsFromPositronium] from very large radioactive sources. The accumulation rate is higher than needed for precise measurements with a trapped positron, so useful positron loading should continue for many Na$^{22}$ half-lives.
In fact, the 6.5 $\mu$Ci source is below the 10 $\mu$Ci threshold that triggers the mentioned licensing requirement[@Na22LicenseThreshold]. No cumbersome heavy metal shielding is required since a point source of this activity at a distance of 1 meter gives a does rate of 0.01 mrem/hour (0.1 $\mu$Sv/hour) – an annual dose of 90 mrem (900 $\mu$Sv). This is much smaller than the average natural background rate of 310 mrem (3.1 mSv) from natural sources and the 310 mrem (3.1 mSv) from manmade sources [@NaturalBackground]. It is also lower than the 100 mrem (1 mSv) that the Nuclear Regulatory Commissions (NRC) has established as the maximum yearly radiation dose to which one of its licensees can expose an individual member of the general public[@NaturalBackground].
Spontaneous electron loading is prevented by making it possible to withdraw the radioactive source away from the trap. In additions, the positron loading is carried out with trap electrodes that are kept at 100 mK by a dilution refrigerator – a substantial additional challenge compared to what is required for a 4.2 K apparatus.
![Apparatus designed for improved comparisons of the magnetic moments and charge-to-mass ratios of the electron and positron.[]{data-label="fig:WholeApparatus"}](WholeApparatus-eps-converted-to.pdf){width="3in"}
Apparatus Overview
==================
A completely new cryogenic apparatus (Fig. \[fig:WholeApparatus\]) was designed for more precise measurements of the magnetic moments and charge-to-mass ratios of the positron and electron. A custom superconducting magnet provides an access diameter of 12.8 cm (5.0 in) into which the dilution refrigerator apparatus and trap apparatus can be inserted. This is significantly larger than the 8.6 cm (3.4 in) available diameter used for the previous measurements[@HarvardMagneticMoment2008; @HarvardMagneticMoment2011] (illustrated in Fig. \[fig:ApparatusComparison\]), enough so to make it possible to admit and retract a radioactive source for positron loading. A 500 liter liquid helium reservoir is within radiation shields cooled to 77 K by a 190 liter liquid nitrogen reservoir. The persistent superconducting solenoid, sitting on the bottom of the helium reservoir and concentric with its vertical symmetry axis, produces up to a 6 tesla vertical field.
A dilution refrigerator with an attached trap is lowered into this fixed dewar from above. The so-called inner vacuum container (IVC) for this fridge is then located within the helium dewar that contains the solenoid, with its lower section located within the solenoid. A separate and isolated trap vacuum enclosure is thermally anchored to the typically 100 mK mixing chamber of the dilution refrigerator within the IVC.
![Comparison of the old (a) and new (b) cryogenic apparatus used for comparing magnetic moments and charge-to-mass ratios of electrons and positrons. The green lines show the much shorter support path between the trap electrodes and the solenoid windings in the new apparatus compared to the old apparatus.[]{data-label="fig:ApparatusComparison"}](FridgeComparison-eps-converted-to.pdf){width="3in"}
![Representation of the trap electrodes for positron loading and for precision measurements, each with an identifying label. []{data-label="fig:TrapDetail"}](TrapCrossSection-eps-converted-to.pdf){width="2.5in"}
The trap vacuum enclosure contains three different types of Penning traps, two of which are represented in Fig. \[fig:TrapDetail\]. An open-endcap cylindrical Penning trap[@OpenTrap] for positron loading and accumulation is at the top of the figure. Below it is shown a closed-endcap cylindrical Penning trap[@CylindricalPenningTrap] for performing precision measurements with a single particle. Below the precision trap and not shown in this figure is a planar Penning trap [@OptimizedPlanarPenningTraps] used for scalable electron qubit studies based upon the methods developed to measure the electron and positron magnetic moments.
Compared to the apparatus we used for previous electron magnetic moment measurements, this positron-electron apparatus design has important advantages:
- The trap enclosure sits mechanically upon the superconducting solenoid so that the trap and magnetic field should change their mechanical position together.
- The center axis of the trap is aligned mechanically to the center of the superconducting solenoid by pins that center the trap within the solenoid.
- To stabilize the mechanical and electrical properties of the trap we have demonstrated the regulation of the height of the liquid helium on the trap apparatus by changing the pressure in the gas above the outer part of the liquid helium dewar [@ThanksDvonLindenfels].
However, supporting the trap from the superconducting solenoid (to keep the trap from changing its location within the superconducting solenoid) also introduce three significant cryogenic challenges.
Cryogenic Challenges
====================
It is important that the 100 mK trap electrodes and the solenoid that produces the magnetic field for the trap be mechanically coupled so that a trapped particle and the magnetic field can only move together. This is accomplished here by supporting the 100 mK trap directly on the 4.2 K superconducting solenoid. The trap hangs below most of the dilution refrigerator on a heat link that flexes as the trap container settles down upon the form on which the solenoid is wound.
The first cryogenic challenge is to avoid quenching the superconducting solenoid when a warm trap apparatus is initially inserted. This is crucial insofar as the field stability required for the precise measurement is only attained if the superconducting solenoid is kept at its 4.2 K operating temperature for months. Inserting a warm trap apparatus into the solenoid must be done so as to cool the trap and dilution refrigerator to 4.2 K before they make contact with the 4.2 K solenoid. Otherwise the solenoid will quench, and only long after the solenoid is re-energized will the needed time stability of the field be restored.
The trap and dilution refrigerator are slowly lowered into the large helium dewar that contains the superconducting solenoid over about 4 hours. During this time the cold helium gas that boils off from the liquid helium cools the trap from room temperature to 4.2 K. It is critical that no air be allowed into the dewar during the insertion. Otherwise, the paramagnetic oxygen ice that builds up within the bore of the solenoid can both keep the trap and IVC from being fully inserted and also reduce the magnetic field homogeneity. The procedure employs the use of a cryogenic o-ring and sliding seal, surrounded by a large plastic glove-bag. A small continuous flow of helium gas prevents air from entering the cold volume of the apparatus.
Once the vacuum enclosure for the trap rests mechanically upon the 4.2 K solenoid, it is cooled from 4.2 K to 100 mK by turning on the dilution refrigerator. The mechanical supports are carbon fiber posts with a very low thermal conductivity. The conduction and radiation losses of the 100 mK vacuum container for the trap are small compared to the 330 $\mu$W that the dilution refrigerator (JDR-500 from Janis Research Company) is rated to sink at 100 mK.
The second cryogenic challenge is that liquid helium consumption has become prohibitively expensive for a dewar of this size given the significant price increases for liquid helium in recent years. When the dilution refrigerator is operating it boils off about $19\pm 1$ liters of liquid helium per day, along with about $24\pm3$ liters of liquid nitrogen per day. The helium consumption drops to about $9\pm 1$ liters per day when the refrigerator and trap are removed to be worked on. The expense is considerable give that the dilution refrigerator must run for many months without stopping to make the precise magnetic moment measurements.
Replacing the nitrogen and helium reservoirs by a pulse tube refrigerator is possible in principle but this replacement could cause significant vibration that could affect our measurements despite the solenoid and trap being mechanically connected. Instead we installed a helium reliquefier (Cryomech PT415 with a remote motor) that turns the cold helium gas that evaporates from the liquid helium back into liquid helium. When the helium reliquefier system is running there is essentially no net liquid helium boiled off from the apparatus. This reliquefier does include a pulse tube refrigerator but there is some mechanical isolation from the trap and it should be possible to turn it off during the most sensitive parts of precise measurements.
The third challenge of the cryogenic operation is preventing radiation from the warm parts of the apparatus from reaching the 100 mK trap enclosure. The radiation must be blocked while allowing an open path for lower the radioactive source down to the 100 mK trap to load positrons, and then to retract it a distance large enough to prevent spontaneous electron loading. To illustrate the challenge, a 0.8 mm (1/32 inch) diameter hole that allows 300 K radiation to reach the 100 mK apparatus would provide around 200 $\mu$W of heating – two-thirds of the total heat load that the mixing chamber of the dilution refrigerator is specified to handle.
The solution is a series of 8 baffles and a special blocking piece that float together on the string that supports the source capsule (see Fig. \[fig:RadiationBaffles\]). The holes in the baffles are offset to block radiation down the center. As the source is lowered from the top of the dilution refrigerator, the blocking piece mates with a conically shape piece that is thermally anchored to the 4.2 K stage. The baffles rest on this blocking piece. This design keeps room temperature radiation from reaching beyond the 4.2 K stage, and the 4.2 K radiation to the 100 mK trap is very small.
![Radiation baffles needed to avoid a thermal load on the dilution refrigerator along the entry and removal path for the radioactive source.[]{data-label="fig:RadiationBaffles"}](RadiationBaffles-eps-converted-to.pdf){width="2.2in"}
Positron Loading
================
Since the positron magnetic moment is best measured with only one trapped positron there is no need to load large numbers of positrons. Also, positrons only need to be loaded very infrequently since once a positron is suspended within the trap for a measurement it can be used for months at a time. When positrons do need to be loaded, it is not a problem to spend an hour to accumulate tens of positrons, since only one needs to be transferred into a precision trap for measurement. The goal of this work was thus to make sure that we could load at least one positron loaded per minute from the weakest source that could accomplish this – preferably from a source so small that even without external shielding it could deliver a dose to an experimenter that is well below the natural background.
We chose to use a Rydberg ionization method of accumulating trapped positrons that our group invented for its early antihydrogen experiments [@PositronsFromPositronium]. Positrons thermalize within a 2 $\mu m$ thick tungsten crystal moderator located at the top of the positron loading trap as shown in Fig. \[fig:sourcedeliv\] and Fig. \[fig:VEconfig\]a. Some of the fast positrons emitted from the radioactive source enter and slow within the moderator. These diffuse around inside the crystal until they get near a surface, whereupon the work function potential that keeps electrons within the material pops the positrons out with a low energy.
The slow positrons that emerge follow magnetic field lines, some of them picking up an electron from the surface as they leave the thin tungsten moderator. The highly magnetized Rydberg positronium atom that is formed, only because the electron and positron are far enough apart to mostly be pinned to field lines, is so weakly bound that it is easily ionized when the electrodes are biased to produce a strong enough electric field within the potential well of a Penning trap. The process is quite stable even though it relies upon an adsorbed layer of gas on the exit surface for reasons that are not yet understood. By reversing the electrode potentials, this method can also be used to load electrons into the same location in the trap.
The first of four reasons for choosing this method is its simplicity; positrons can be accumulated directly into high vacuum without the need for any buffer gas or any additional trapped particles. Second, the method is attractive because the accumulation rate per source activity is a couple of hundred times higher than had been realized for previous precision measurements. Third, the method was shown to be linear in the source activity used (at least at much higher source activities than we use for this report) making it plausible that the loading rate could be extrapolated linearly to sources with less activity. Fourth, the method has proven to be robust when the adsorbed gas is not removed from the surface.
To estimate the source activity required to obtain the desired loading rate of about 1 e$^+$/min, we consider the loading rate achieved in earlier work in our group using a transmission moderator, which was approximately 7 e$^+$/s [@PositronsFromPositronium]. The source used in this earlier work was 2.5 mCi, so the normalized loading rate was 2.8 e$^+$/s/mCi. Assuming the loading rate scales linearly with source size, we would need a 6 $\mathrm{\mu}$Ci source to achieve a 1 e$^+$/min loading rate. However, for such a small source, we would expect the self-absorption observed in the larger source [@HaarsmaThesis] to be minimal and thus would expect the loading rate to be increased by a factor of 2 from the earlier work. Therefore we estimate a 3 $\mathrm{\mu}$Ci source will give a 1 e$^+$/min loading rate, and equivalently, a 6 $\mathrm{\mu}$Ci source will give a 2 e$^+$/min loading rate.
The positron source we use in this work is a $^{22}$Na sealed button source (Isotope Products Laboratories custom diameter POSN source). Its activity was 15.6 $\mathrm{\mu}$Ci when delivered, and it decreased between 6.9 and 6.3 $\mathrm{\mu}$Ci during the studies reported here. The delivered source had the radioactive salt between two 5 $\mu m$ thick Ti foils that were electron-beam welded to completely contain the salt. To facilitate safe handling, we enclosed this source in a two-piece capsule made from a 90% tungsten/10% copper composite, as represented in Fig. \[fig:SourceCapsule\].
A nylon string connects the top source capsule to a rotational vacuum feedthrough at the top of the dilution refrigerator. This allows the source to be raised and lowered inside the inner vacuum chamber of the dilution refrigerator. Pairs of light-emitting diodes and photo-diodes, as well as marks on the string, are used to monitor the location of the source. The positron source can be raised and lowered along the entire length of the dilution refrigerator, although in normal operation it is moved between two locations: an off-axis “storage" position at the mixing chamber of the dilution refrigerator where positrons are prevented from loading into the trap by the distance and the angle; and an on-axis “loading" position where the source capsule is located directly above the trap vacuum enclosure, and separated from it by a thin titanium foil vacuum window. These locations are shown in Fig. \[fig:sourcedeliv\].
![The on-axis loading position allows positrons to be loaded into the trap. The off-axis storage position prevents positrons (or gamma-rays) from entering the trapping region.[]{data-label="fig:sourcedeliv"}](SourceDeliverySystem-eps-converted-to.pdf){width="3.2in"}
Results and Discussion
======================
Particles in the loading trap are detected and counted nondestructively using a cryogenic amplifier connected to a trap electrode. The amplifier and trap electrodes form an RLC circuit and we monitor the Johnson noise spectrum of this circuit. When the trap is empty, the frequency spectrum is a Lorentzian peak centered at the circuit’s resonant frequency. The oscillation of a small number of particles in the trap interacts with the tuned circuit resonance and forms a Lorentzian “dip” in the amplifier resonance whose center is the particles’ axial oscillation frequency and whose FWHM is equal to the number of particles in the trap multiplied by the single-particle damping width, $\gamma_z$.
![An example of the dip in the amplifier noise resonance for a cloud of approximately 160-320 positrons. []{data-label="fig:dip"}](2hourpositrondip-eps-converted-to.pdf){width="3in"}
We have calculated that $\gamma_z = 10$ Hz for this amplifier. The actual value of $\gamma_z$ tends to be somewhat smaller than the calculated value[@Review] and we have some experimental indications that $\gamma_z \approx 5$ Hz. Therefore we use the estimate of $\gamma_z =$ 5-10 Hz in discussing the number of positrons (or electrons) in this paper. Figure \[fig:dip\] shows a cloud of approximately 160-320 positrons. It should eventually be possible to remove the large uncertainty in the number of positrons in the accumulation trap by optimizing the detection amplifier so that it unambiguously resolves the difference between one and two trapped positrons.
Both positrons and electrons have been loaded from our positron source, and the numbers loaded have been used to verify the loading mechanism. Several different electrode potential configurations were used for these loading trials. The main potential configurations are shown in Fig. \[fig:VEconfig\].
![The voltage and electric field profiles are shown for the two main loading configurations used. Solid lines denote the potential and field used for positron loading and dashed lines denote the potential and field used for electron loading. Figures (b) and (c) show the configuration where secondary electrons are unblocked (not prevented from entering the trapping region) while (d) and (e) show the configuration where secondary electrons are blocked (prevented from entering the trapping region). A cross section of the electrode stack is shown (top).[]{data-label="fig:VEconfig"}](PotentialandField_3configurations-eps-converted-to.pdf){width="3in"}
For the field ionization method of accumulating positrons or electrons from positronium, the loading rate for positrons and electrons from the source into the trap should be identical. This is provided that the electrodes are biased to prevent loading the trap with secondary electrons knocked free of the moderator in collisions with positrons passing through. Additionally, the loading rate should be linear in time as the loading mechanism does not depend upon interactions with previously loaded particles. In all cases, the number of particles we observed was indeed linear in time. We found that changing the electrode potentials from the “unblocked” configuration shown in figure \[fig:VEconfig\], which does not prevent secondary electron loading, to the “blocked” configuration, which prevents most secondary electron loading, greatly reduces the large asymmetry seen in electron and positron loading rates as seen in Fig. \[fig:eploadingrates\]. This is as expected for the loading mechanism discussed.
![Electron (dashed lines) and positron (solid lines) loading rates for the two electrode potential configurations shown in figure \[fig:VEconfig\]. Figure (a) shows the configuration where secondary electrons are unblocked and (b) shows the configuration where secondary electrons are blocked. When secondary electrons are prevented from entering the trapping region, (b), the large asymmetry in electron and positron loading rates, (a), is greatly reduced.[]{data-label="fig:eploadingrates"}](eploadingratesv2-eps-converted-to.pdf){width="3.3in"}
Additional factors also affect the loading rate as expected. For example, moving the source 3.8 cm (1.5 in) further from the moderator and loading trap cuts the loading rate of both electrons and positrons in half. Additionally, firing a field emission point (or FEP, located at the bottom of the precision trap and used for quickly loading electrons into either trap) 20 times at approximately 1 nA for approximately 1 minute each time decreased the loading rate by roughly a factor of two. This is because the electron beam from the field emission point strikes the surface of the moderator, and the formation of positronium at the moderator surface depends upon a layer of adsorbed gas on the moderator surface. This phenomenon was observed previously when antiprotons struck such a moderator[@ThesisBowden] and when a laser was used to deliberately heat the moderator[@ThesisEstrada]. Thermal cycling of the apparatus restores the adsorbed gas layer and restores the loading rate.
Our maximum positron loading rate manifests itself as a 687 (10) Hz/hr increase in the measured resonance width, corresponding to about 70-140 e$^+$/hr being accumulated in the trap. For our 6.3 $\mathrm{\mu}$Ci $^{22}$Na source, this gives a loading rate of 3-6 e$^+$/s/mCi, in good agreement with previous work[@PositronsFromPositronium].
Conclusion
==========
Field ionization of slow positronium accumulates the positrons needed to measure the positron magnetic moment at the $3$ parts in $10^{13}$ precision to which we have measured the electron magnetic moment. Compared to the positron accumulation used for previous magnetic moment measurements, an orders of magnitude smaller source accumulates trapped positrons at a rate per source activity that is orders of magnitude larger. The tiny 6.3 $\mu$Ci source that provides positrons to a 100 mK Penning trap is small enough to avoid special licensing and handling methods, provides a dose rate to an unshielded user that is much lower than the average natural background, and can be withdrawn from the trap during precise measurements.
This work was supported by the NSF.
[23]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} “,” (, ) Chap. @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop @noop [****, ()]{} @noop @noop [****, ()]{} @noop [****, ()]{}, @noop [****, ()]{} @noop [****, ()]{} @noop **, @noop [Ph.D. thesis]{}, (), @noop [****, ()]{} **, @noop [Ph.D. thesis]{}, (), **, @noop [Ph.D. thesis]{}, (),
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
L. Markou$^{1}$, E. Stiliaris$^{2}$ and C. N.Papanicolas$^{1,*}$\
\
$^{1}$ The Cyprus Institute, K. Kavafi 20, 2121 Nicosia, Cyprus\
$^{2}$ National and Kapodistrian University of Athens, Physics Department, 15771 Athens, Greece\
\
\
bibliography:
- 'bibliography.bib'
title: 'Multipole sensitivity to phase variation in pion photo-and electroproduction analyses'
---
Introduction {#intro}
============
Compton scattering, pion photoproduction, and pion - nucleon scattering are related by unitarity through a common S matrix [@blanpied2001n] and the Fermi-Watson (FW) [@FW] theorem requires the $( \gamma , \pi )$ and $( \pi , \pi )$ channels to have the same phase below the two-pion threshold. Multipole analyses below this threshold are subject to this theoretical constraint which requires all multipoles with different character but the same quantum numbers $I, l , J$ to have the same phase $\pm n \pi$ which is the same as the corresponding $\pi N$ scattering phase shift. The pion photoproduction multipole phases and the scattering phase shifts are related through [@FW]: $$\label{eq:fw}
A^{I}_{l\pm} = |A^{I}_{l\pm}|e^{i \left ( \delta_{IlJ} + n \pi \right )}$$ where $\delta_{IlJ}$ is the pion - nucleon scattering phase shift, $I$ is the isospin quantum number, $l$ the angular momentum, $J$ the total angular momentum and “$\pm$” is used to distinguish whether $J$ and the spin are parallel or anti-parallel. $A^{I}_{l}=\{E^{I}_{l},M^{I}_{l},L^{I}_{l}\}$ denotes the electric, magnetic or longitudinal nature of the multipole. As $\pi N$ scattering phase shifts are easier to measure and therefore are known with higher precision, this theoretical constraint provides a very powerful tool in photoproduction (and electroproduction) multipole analyses. Multipoles are complex functions of the center mass energy $W$ and by applying the FW theorem the number of unknown parameters is halved since only the moduli of the multipoles $|A_{l\pm}^{I}|$ needs to be determined. It has been widely used in multipole analyses of both pion photoproduction data [@blanpied1997n; @PhysRevLett.78.606; @beck2000determination; @kotulla2007real; @markou1] and pion electroproduction data [@PhysRevLett.82.45; @PhysRevLett.86.2963; @sparveris2005investigation; @stave2006lowest].
The values of the $\pi N$ scattering phase shifts are known from the analyses of $\pi N$ scattering data, *e.g.* ref. [@workman2012parameterization]. The FW applies well beyond the two pion threshold as the $\pi N$ inelasticities are very small [@beck2000determination; @pseudo]. For example, in the pion photoproduction data analysis by Grushin [@grushin] where both the real and imaginary parts of the $\ell \leq 1$ multipoles were determined without using the FW theorem it was found that the mean difference between the $\delta_{33}^{\gamma,\pi N}$ and the pion - nucleon scattering phase shift was only $-(2.3 \pm 0.5)^{\circ}$ over the energy range $E_{\gamma}^{lab} = 250 - 500$ $MeV$.
Fig. \[fig:piN\_phases\_2sigma\] shows the nine $\pi N$ phases reported in ref. [@workman2012parameterization] as part of the WI08 partial wave analysis at $W=1235$ $MeV$ at the photon point. The exact numerical values are available online [@gwuweb2]. Each phase is plotted as a Gaussian, $N[\mu,2\sigma]$, with the same mean value ($\mu$) and double the statistical uncertainty ($\sigma$) derived from the experimental $\pi N$ data. The MAID07 model prediction [@Drechsel:2007if] for the corresponding pion photoproduction multipole phases ($\pm n\pi$), at the same kinematics, is also shown. The $\pi N$ phases are known with limited accuracy, and although multipole analyses use them as if they are known with infinite precision, the effect of this uncertainty on the extracted multipoles has not been explored yet. Using the Athens Model Independent Analysis Scheme (AMIAS) we achieve this by constraining the phases within their experimentally determined uncertainty, making the analysis Bayessian.
The following sections are organized as follows: In Sec. \[sec:fw\_methodology\] we discuss the methodology for multipole extraction with the AMIAS and the inclusion of parameters with known uncertainties. In Sec. \[sec:fw\_pseudo\] we detail the creation of pion photoproduction pseudodata of predetermined statistical precision. The multipole content of those pseudodata is derived in Sec. \[sec:fw\_results\] validating the methodology described in Sec. \[sec:fw\_methodology\]. In Sec. \[sec:fw\_bates\] we apply the same methodology for a re-analysis of the Bates/Mainz electroproduction data [@sparveris2005investigation] measured at $Q^2=0.127$ $GeV^2/c^2$ and $W=1232$ $MeV$. Concluding remarks are given in Sec. \[sec:fw\_conclusions\].
Methodology {#sec:fw_methodology}
===========
The methodology employed is the implementation of the Chew, Goldenberg, Low and Nambu (CGLN) theoretical framework [@chew1957relativistic] for single energy multipole analyses in the Athens Model Independent Analysis Scheme (AMIAS) [@stiliaris2007multipole; @papanicolas2012novel]. The AMIAS method is based on statistical concepts and relies heavily on Monte Carlo and simulation techniques, and it thus requires High Performance Computing as it is computationally intensive. The method identifies and determines with maximal precision parameters that are sensitive to the data by yielding their Probability Distribution Functions (PDF). The AMIAS is computationally robust and numerically stable. It has been successfully applied in the analysis of data from nucleon photo-and electroproduction resonance [@markou1; @stiliaris2007multipole; @RevModPhys.84.1231], lattice QCD simulations [@alexandrou2015novel] and medical imaging [@loizospseudo].
AMIAS requires that the parameters to be extracted from the experimental data are explicitly linked via a theory or a model [@stiliaris2007multipole]. In the case of pion photoproduction this requirement is provided by the CGLN theory as in ref. [@markou1] and in the case of electroproduction as in ref. [@stiliaris2007multipole]. The multipoles are connected to the pion photoproduction observables via the CGLN [@chew1957relativistic] amplitudes $\left ( F_i, i=1,6 \right ) $:[fleqntruemathmargin20pt]{}$$\label{eq:cgln1}
\begin{split}
F_{1} = {} & \sum_{l=0}^{\infty} [ \left ( lM_{l+} + E_{l+} \right ) P'_{l+1}(x) \\
& + \left ( \left (l+1 \right ) M_{l-} + E_{l-} \right ) P'_{l-1}(x) ]
\end{split}$$ $$\label{eq:cgln2}
F_{2} = {} \sum_{l=1}^{\infty} \left [\left (l+1 \right) M_{l+} + lM_{l-} \right ] P'_{l}(x)$$ $$\label{eq:cgln3}
\begin{split}
F_{3} = {} & \sum_{l=1}^{\infty} [ \left ( E_{l+} - M_{l+} \right ) P''_{l+1}(x) \\
& + \left ( E_{l-} + M_{l-} \right ) P''_{l-1}(x) ]
\end{split}$$ $$F_{4} = {} \sum_{l=2}^{\infty} \left [ M_{l+} - E_{l+} - M_{l-} - E_{l-} \right ] P''_{l}(x)
\label{eq:cgln4}$$ $$F_{5} = {} \sum_{l=0}^{\infty} \left [ (l+1)L_{l+} P'_{l+1}(x)-lL_{l-}P'_{l-1}(x) \right ]
\label{eq:cgln5}$$ $$F_{6} = {} \sum_{l=1}^{\infty} \left [ l L_{l-} - (l+1) L_{l+} \right ] P'_{l}(x)
\label{eq:cgln6}$$ where $x = \cos(\theta)$ is the cosine of the scattering angle and $P'_{l}$ are the derivatives of the Legendre polynomials. Multipoles $A_{l\pm}=\{E_{l\pm}, M_{l\pm}, L_{l\pm} \}$ refer to the electric, magnetic or longitudinal nature of the photon respectively. At the real photon point, longitudinal degrees of freedom in the photon’s polarization vanish identically and the $\gamma N \to \pi N$ reaction is described solely by the CGLN amplitudes $F_{1}$ to $F_{4}$.
From isospin conservation in the pion-nucleon system it follows that the multipoles can be expressed in terms of definite isospin [@BERENDS19671; @dreschsel1992threshold], namely, the $A^{1/2}$ and $A^{3/2}$ multipoles. These are obtained from the reaction channel multipoles and the relations [@dreschsel1992threshold]: [fleqnfalse]{}$$\begin{aligned}
A^{1/2} &={} \frac{A_{p \pi^{0}}}{3} + \frac{\sqrt{2} A_{n \pi^{+}}}{3}, & A^{3/2} &={}A_{p \pi^{0}} - \frac{A_{n \pi^{+}}} {\sqrt{2}}
\label{eq:epjaiso}\end{aligned}$$ In contrast to the standard practice adhered up to now where the multipole phases are considered as if known with infinite precision [@beck2000determination; @sparveris2005investigation] and therefore treated as fixed parameters of the problem we allow those phases to vary within their experimentally determined uncertainty obtained from $\pi N$ experiments. This allows the prior knowledge on the multipole phases to be incorporated in the analysis. To ascertain the magnitude of the effect this phase variation induces on the derived multipoles we examine three sets of pseudodata where each set was created with predetermined and increasing statistical precision. For each pseudodata set three multipole analyses were performed differentiated by the manner in which the multipole phases were treated; during the first analysis phases were fixed to the values of the generating model, during the second analysis phases were fixed to the SAID-WI08 [@workman2012parameterization; @gwuweb2] model dependent analysis values and during the third multipole phases were allowed to vary with Gaussian weight, with mean value and twice the standard deviation of that reported by the SAID-WI08 single energy solution [@workman2012parameterization; @gwuweb2]. In implementing the phase variation, and according to eq. \[eq:fw\], we imposed that during the variation procedure all multipoles with the same quantum numbers $I,l,J$ had the exact same phase $\pm \pi$. In contrast, multipole phases with different quantum numbers were varied independently.
Creation of pseudodata {#sec:fw_pseudo}
======================
[fleqntruemathmargin20pt]{}
We have created pseudodata for the four single $(d\sigma_0$, $\hat{\Sigma}$, $\hat{T}$, $\hat{P})$ and four double beam-target $(\hat{E}$, $\hat{F}$, $\hat{G}$, $\hat{H})$ polarization observables for the $\gamma p \to p \pi^0$ and $\gamma p \to n \pi^+$ reactions. The definitions used for the observables are the same as in ref. [@pseudo]. To create the pseudodata the MAID07 multipole solution at the photon point and at center mass energy $W=1234.5$ $MeV$ was inserted in the CGLN multipole series, Eqs. 2-5, which were then used to construct the photoproduction observables defined in Table \[eq:Photo\_kin\_obs\]. A schematic of this “forward procedure” is given in Fig. \[fig:fw\_inverse\]. The observables were subsequently randomized according to the process: [fleqntruemathmargin20pt]{}$$\label{eq:ran1}
\begin{split}
O_{i}^{k} =&{} O_{i}^{k} + N[\mu,\sigma] \cdot O_{i}^{k} \\
\sigma_{O_{i}^{k}} =&{} N[\mu,\sigma] \cdot O_{i}^{k}
\end{split}$$ where $O$ is the MAID07 model prediction, $k$ distinguishes between each of the spin observables, $i$ labels the angle, $N[\mu,\sigma]$ is a normal distribution with known mean $(\mu)$ and standard deviation $\sigma$ and $\sigma_{O_{i}^{k}}$ is the uncertainty attributed to the $i^{th}$ angular measurement of the $k^{th}$ observable.
Using Eq. \[eq:ran1\] we created 9000 sets of pseudodata. Each pseudodata set consisted of $288$ datapoints; $18$ evenly spaced angular measurements in the dynamical region $\theta_{cm} \in[5^{\circ} : 175^{\circ} ]$ for each of the eight polarization observables listed in Table \[eq:Photo\_kin\_obs\] for each proton target reaction. The angle $\theta_{cm}$ is defined as the angle between the incoming photon and the produced pion in the center of mass frame.
The generated values and uncertainties of the pseudodata sets are shown to have the required behavior [@friar1973determination; @friar1975determination]. by examining the resulting $\chi^2$ distribution. The $\chi^2$ distribution resulting by comparing each dataset to the generator, shown in Fig. \[fig:fw\_paper\_distofchi2\], is correctly described by the $\chi^2$ distribution with degrees of freedom equal to the number of datapoints $(=288)$ of the datasets.
The methodology followed in creating the pseudodata through eq. \[eq:ran1\] allows to control the precision of the generated pseudodata. Three distinct classes of pseudodata were created; each class featuring pseudodata of statistical uncertainty a) $2.0$, b) $1.0$ and c) $0.16$ times the statistical uncertainty of the most precise pion photoproduction data [@adlarson2015measurement] available to date. Fig. \[fig:fw\_paper\_ppiz\_pseudo\] show pseudodata for the differential cross section ($d\sigma_0/d\Omega$), the beam asymmetry ($\hat{\Sigma}$), the target asymmetry ($\hat{T}$) and the recoil target asymmetry ($\hat{P}$) for the two proton target reactions. The blue circles are used for pseudodata of relative uncertainty $2.0$, the green diamonds for pseudodata of relative uncertainty $1.0$ and the black triangles for relative uncertainty $0.16$. The continuous magenta curve is the generator. The pseudodata present some qualitative similarities to experimental data; the forward peak in the $\gamma \pi \to n\pi^{+}$ differential cross section, the absence of such peak in the $\gamma \pi \to p\pi^{0}$ differential cross section and larger uncertainties in the very forward and backward angles. The generated pseudodata also provide spin observables which have never been measured before, e.g. the beam-target $\hat{E}$, and full angular coverage.
\
Results {#sec:fw_results}
=======
We applied the methodology presented in Sec. \[sec:fw\_methodology\] to the pseudodata and we extracted values for all multipole amplitudes with relative angular momentum $\ell \leq 2$. Higher multipoles and up to all orders were frozen to the generating model values. The multipole amplitude PDFs derived from the AMIAS analyses were fitted with Gaussians and numerical results were extracted. Table \[tab:fw\_paper\_tab\_moduli\] lists the mean value $\pm 34\%$ uncertainty ($1\sigma$) for the derived multipoles for two distinct analyses: Column “MD07” and “W108” refer to analyses where the multipole phases were fixed to the MAID07 values (which is the generating model) or the WI08 solution respectively. Column “Varied” denotes analyses where the multipole phases were varied in a Gaussian manner with $N[\mu,2\sigma]$ where the mean value is taken from the WI08 $W$-dependent solution and the standard deviation, $\sigma$, the derived uncertainty of the WI08 single energy fit. The derived multipole values and uncertainties are in good statistical agreement with the generator input. The derived multipole uncertainty from each pseudodata set for the analyses with the phases fixed, listed as “MD07” and “W108” in Table \[tab:fw\_paper\_tab\_moduli\], is reduced according to the statistical precision of each set. The pseudodata sets, Set A, Set B and Set C were created with relative uncertainties $2.0$, $1.0$, and $0.16$ respectively. This is reflected in our results as the uncertainty associated with a specific multipole amplitude derived from Set A is reduced by a factor of $2.0$ and $12.5$ when derived from Sets B and C respectively. This behavior indicates that the AMIAS method yields exact uncertainties with a precise statistical meaning [@stiliaris2007multipole; @papanicolas2012novel]. Fig. \[fig:fw\_paper\_amps\_incr\_prec\] shows the PDFs of some selected amplitudes derived from the analysis of each pseudodata set with the multipole phases fixed to the MAID07 (generator) values. As expected the derived uncertainty of each multipole amplitude is seen to decrease according to the statistical precision of the analyzed pseudodata.
\
Regarding the analyses of the pseudodata sets A and B, which are characterized by uncertainties greater or equal to current experimental data, the derived results are statistically equivalent whether the analysis was carried with the multipole phases fixed to the generator values (MAID07), to the WI08 solution, or they were allowed to vary within the allowed experimental uncertainty. This is exhibited in Fig. \[fig:fw\_paper\_mp1rr1\_allsets\] for the case of the $M_{1+}^{1/2}$ amplitude. It demonstrates that data of the currently available precision are not sensitive to such small changes or variations in the multipole phase. The standard practice in multipole analyses, to treat these phases as if known with infinite precision, does not induce additional model bias to the derived multipoles. Regarding the analyses of set C we note significant differences in the derived multipole mean values when phases change from the MAID07 values to the WI08 solution while the derived uncertainty remains unchanged. When the multipole phases are allowed to vary the derived multipole mean values are shifted while their associated uncertainty is increased. This increase is more prominent in the background multipole amplitudes $M_{1+}^{1/2}$ and $E_{0+}^{3/2}$. Fig. \[fig:fw\_paper\_mp1rr1\_allsets\] shows this behavior for the case of the $M_{1+}^{1/2}$ amplitude.
\
For the analyses with “Varied” phases known, Normal distributions were utilized for the phase randomization. The analyses of pseudodata sets A and B yield phases nearly identical to the Normal distributions used for the phase variation. This indicates that data of such precision do not exhibit sensitivity to the magnitude of the phase variation we imposed. The derived phases from set C, the most precise analyzed pseudodata set, emerge significantly narrower than the Normal distributions utilized to vary them. Fig. \[fig:fw\_paper\_phase\_and\_chi2vsep0ph1\] shows the PDF of the $E_{0+}^{1/2}$ phase derived from each pseudodata set. The phase PDFs derived from set A and B exactly match the distribution used to vary the phase and which is marked by a black continuous curve. The $E_{0+}^{1/2}$ phase derived from set C emerges much narrower.
\
Example: Application to the Bates and Mainz data at $Q^2=0.127$ $(GeV/c)^2$ {#sec:fw_bates}
===========================================================================
The methodology of Sec. \[sec:fw\_methodology\] was applied for a re-analysis of the $H(e,e',p)\pi^{0}$ Bates/Mainz measurements performed at $Q^2=0.127$ $GeV^2/c^2$ and $W=1232$ $MeV$. The detailed description and analysis of this data can be found in ref. [@sparveris2005investigation]. The data set consists of cross section results for $\sigma_{TT}$, $\sigma_{LT}$, $\sigma_{0}$, $\sigma_{E2}$ and the polarized beam cross section $\sigma_{LT'}$. The observables are defined as in ref. [@dreschsel1992threshold].
As the data concern $\gamma p \to p \pi^0$ measurements, model input (MAID07) was used to allow the isospin separation of multipoles and only few multipoles were derived. The derived parameters are the $l=0$ charge multipole amplitudes (the $A_{p \pi^{0}}$ multipoles of Eq. \[eq:epjaiso\]) and the $l=1$ resonant multipole amplitudes with isospin $I=3/2$. The $I=1/2$ multipoles were fixed to the MAID07 model values. We performed two new analyses of the data: in the first, the multipole phases were fixed to the $\pi N$ values [@gwuweb2]; in the second the $P_{33}$ phase was varied with Gaussian weight, with mean value the $\pi N$ scattering phase shift value and five times the experimental standard deviation ($\sigma$) of the scattering phase shift. The derived multipoles, which are listed in Table \[tab:fw\_paper\_tab\_ampli\_bates\], were (statistically) identical in both cases; the phase variation did not induce any changes to the derived multipole amplitudes. Our results are in good agreement with earlier analyses of the same data [@sparveris2005investigation; @stiliaris2007multipole]. The Electric-to-Magnetic and Coulomb-to-Magnetic ratios, EMR and CMR respectively, are also given. These are defined as $EMR= E_{1+}^{3/2} / M_{1+}^{3/2}$ and $CMR= E_{1+}^{3/2} / S_{1+}^{3/2}$, where the Coulomb multipole $S_{1+}^{3/2}$ is connected to the longitudinal multipole, the photon’s momentum $q$ and the photon’s energy $\omega$ through the relation $S_{1+}^{3/2} = \frac{\vec{q}_{cm}}{\omega_{cm}} L_{1+}^{3/2}$. EMR and CMR serve as the accepted gauge of the magnitude of the deformation of the proton [@papanicolas2007shapes].
\[ht\]
Summary and Conclusions {#sec:fw_conclusions}
=======================
Using the AMIAS methodology we explored the possible influence of the use of the Fermi-Watson theorem in pion photoproduction analyses. The current practice of using fixed (with no uncertainty) values was examined and compared to analyses where the $\pi N$ phase values and their uncertainty were used as prior knowledge. The AMIAS was used for the first time to allow prior knowledge to be incorporated into experimental analyses. Sets of pseudodata of increasing statistical precision were analyzed and their multipole content was derived. In the case of pseudodata of comparable statistical precision to the most recent pion photoproduction data the derived multipoles emerged nearly identical in mean value and uncertainty. The experimental phase uncertainty induced significant changes in the derived multipole amplitude PDFs when the analyzed pseudodata were created with precision six times the statistical precision of current experimental data.
The same methodology was applied to the $H(e,e',p)\pi^{0}$ Bates/Mainz data measured at $Q^2=0.127$ $GeV^2/c^2$ and $W=1232$ $MeV$ where even a $5\sigma$ phase variation of the experimentally derived phase values did not induce any changes to the derived multipoles. We conclude that for the current precision of pion photo-and electroproduction data the $\pi N$ phases taken from pion-nucleon scattering as perfectly known is justified. However, for the new generation of data aspiring to distinguish among different models of nucleon structure, the type of analysis presented here where the experimentally derived phases are allowed to vary will need to be implemented.
Acknowledgments {#sec:acknowledgements .unnumbered}
===============
This work, part of L. Markou Doctoral Dissertation, was supported by the Graduate School of The Cyprus Institute.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $k$ be an algebraically closed field, $G$ a linear algebraic group over $k$ and $\varphi\in {\textup{Aut}}(G)$, the group of all algebraic group automorphisms of $G$. Two elements $x, y$ of $G$ are said to be $\varphi$-twisted conjugate if $y=gx\varphi(g)^{-1}$ for some $g\in G$. In this paper we prove that for a connected non-solvable linear algebraic group $G$ over $k$, the number of its $\varphi$-twisted conjugacy classes is infinite for every $\varphi\in {\textup{Aut}}(G)$.'
address: 'Indian Institute of Science Education and Research (IISER) Mohali, Knowledge City, Sector 81, S.A.S. Nagar 140306, Punjab, India'
author:
- Sushil Bhunia
- Anirban Bose
title: Twisted Conjugacy in Linear Algebraic Groups
---
[^1]
Introduction
============
Let $G$ be a group and $\varphi$ an endomorphism of $G$. Two elements $x, y\in G$ are said to be $\varphi$-twisted conjugate, denoted by $x\sim_{\varphi} y$, if $y=gx\varphi(g)^{-1}$ for some $g\in G$. Clearly, $\sim_{\varphi}$ is an equivalence relation on $G$. The equivalence classes with respect to this relation are called *$\varphi$-twisted conjugacy classes* or *Reidemeister classes* of $\varphi$. If $\varphi={\textup{Id}}$, then the $\varphi$-twisted conjugacy classes are the usual conjugacy classes. Let $[x]_\varphi$ denote the $\varphi$-twisted conjugacy class containing $x\in G$ and $\mathcal{R}(\varphi):=\{[x]_\varphi\mid x\in G\}$. The cardinality of $\mathcal{R}(\varphi)$, denoted by $R(\varphi)$, is called the *Reidemeister number* of $\varphi$. A group $G$ is said to have the *$R_{\infty}$-property* if $R(\varphi)$ is infinite for every automorphism $\varphi$ of $G$.
The problem of determining groups which have the $R_{\infty}$-property is an active area of research begun by Fel’shtyn and Hill [@fh94] although the study of twisted conjugacy can be traced back to the works of Gantmakher in [@gant]. The reader may refer to [@FN] and the references therein for more literature. The $R_{\infty}$-property of irreducible lattices in a connected semisimple Lie group of real rank at least $2$ has been studied by Mubeena and Sankaran (see [@mstrans Theorem 1]). Nasybullov showed that if $K$ is an integral domain of zero characteristic and ${\textup{Aut}}(K)$ is torsion then $\GL_n(K)$ and $\SL_n(K)$ (for $n> 2$) have the $R_{\infty}$-property (see [@nasy12]). A Chevalley group $G$ (resp. a twisted Chevalley group $G'$) over a field $K$ of characteristic zero possesses the $R_{\infty}$-property if the transcendence degree of $K$ over ${\mathbb Q}$ is finite, see [@FN Theorem 3.2] (resp. [@bdr Theorem 1.2]). It is worth mentioning that a reductive linear algebraic group $G$ over an algebraically closed field $k$ of characteristic zero possesses the $R_{\infty}$-property if the transcendence degree of $k$ over ${\mathbb Q}$ is finite and the radical of $G$ is a proper subgroup of $G$ (see [@FN Theorem 4.1]). The converse of the latter holds if the group $G$ is a Chevalley group of classical type ($A_n, B_n, C_n, D_n$) as shown in [@nas20 Theorem 8], where the author proves that if $k$ has infinite transcendence degree over ${\mathbb Q}$ then there exists an automorphism $\varphi$ of $G$ such that $R(\varphi)=1$. However, it turns out that the automorphism thus obtained is induced by a non-trivial automorphism of $k$ and therefore $\varphi$ is not an algebraic group automorphism of $G$. This motivates the following consideration:
Let $k$ be an algebraically closed field and $G$ a linear algebraic group defined over $k$. Let ${\textup{Aut}}(G)$ denote the group of all algebraic group automorphisms of $G$. We say that $G$ has the *algebraic $R_{\infty}$-property* if $R(\varphi)=\infty$ for all $\varphi\in {\textup{Aut}}(G)$. Since we shall deal with only algebraic automorphisms of algebraic groups in the sequel, we call the algebraic $R_\infty$-property simply as the $R_\infty$-property of $G$. A natural question that arises is the following:
\[question\] Under what conditions does $G$ have the $R_{\infty}$-property ?
Motivation for the present work comes from the results of Springer in [@springer], where he studied the twisted conjugacy classes in simply connected semisimple algebraic groups. A classical result by Steinberg says that for a connected linear algebraic group $G$ over an algebraically closed field $k$ and a surjective endomorphism $\varphi$ of $G$, if $\varphi$ has a finite set of fixed points then $G=\{g\varphi(g)^{-1}\mid g\in G\}=[e]_{\varphi}$, i.e., $R(\varphi)=1$. Therefore, in this scenario, $G$ does not satisfy the $R_{\infty}$-property (see [@St Theorem 10.1]). For an endomorphism $\varphi$ of a simple algebraic group $G$ there exists the following dichotomy: (1) $\varphi$ is an automorphism. (2) $\varphi$ has a finite set of fixed points (see [@St Corollary 10.13]). In an attempt to answer Question \[question\], we prove the following sufficient condition:
Let $G$ be a connected non-solvable linear algebraic group over an algebraically closed field $k$. Then $G$ possesses the $R_{\infty}$-property.
However, the above condition is not necessary as is evident from Example \[eg3\] in [Section \[example\]]{} of this article. Also, see [Proposition \[solvable\]]{}.
Preliminaries
=============
In this section we fix some notations and terminologies and recall some results which will be used throughout this paper.
Linear algebraic groups and Chevalley groups {#agcg}
--------------------------------------------
There are several excellent references for this topic (for example, see [@hum] for linear algebraic groups and [@St2], [@borel] for Chevalley groups). Fix an algebraically closed field $k$. A *linear algebraic group* $G$ (over $k$) is a Zariski-closed subgroup of $\GL_n(k)$ for some $n\geq 1$. Let $G$ be a connected linear algebraic group. The *radical* (denoted by $R(G)$) of $G$ is defined to be the largest closed connected solvable normal subgroup of $G$. The *unipotent radical* (denoted by $R_u(G)$) of $G$ is defined as the largest closed connected unipotent normal subgroup of $G$. The group $G$ is said to be *semisimple* (resp. *reductive*) if $R(G)=1$ (resp. $R_u(G)=1$). A non-commutative algebraic group $G$ is called *simple* if it does not have any non-trivial proper closed connected normal subgroup. It turns out that up to isomorphism, every semisimple algebraic group over $k$ is obtained as a Chevalley group based on $k$ (see [@St2 Chapter 5]), which we now define.
Let $\mathcal{L}$ be a complex semisimple Lie algebra, $\mathcal{H}$ a Cartan subalgebra of $\mathcal{L}$ with the associated root system $\Phi$ and a simple subsystem $\Delta$. Then with respect to the adjoint action ${\mathrm{ad}}:\mathcal{L}\longrightarrow \mathfrak{gl}(\mathcal{L})$, $\mathcal{L}$ has the decomposition $\mathcal{L}=\mathcal{H}\oplus \left(\bigoplus\limits_{\alpha\in \Phi}\mathcal{L}_\alpha\right),$ where $\mathcal{L}_\alpha=\{x\in \mathcal{L}\mid {\mathrm{ad}}(h)(x)=[h,x]=\alpha(h)x\quad \forall\; h\in \mathcal{H}\}$, for each $\alpha\in \Phi$. There exists a Chevalley basis $B=\{h_\delta,e_\alpha:\delta\in \Delta,\alpha\in \Phi\}$ of $\mathcal{L}$, where $h_\alpha\in \mathcal{H}$ is the co-root associated to $\alpha\in \Phi$ and $e_\alpha\in \mathcal{L}_\alpha$ such that $[e_\alpha,e_{-\alpha}]=h_\alpha$ for each $\alpha\in \Phi$. Let $(V,\rho)$ be a finite dimensional complex representation of $\mathcal{L}$. For each $\mu\in \mathcal{H}^\ast$ let $V_\mu:=\{v\in V\mid \rho(h)(v)=\mu(h)v\quad\forall\; h\in \mathcal{H}\}$. The *weights* of the representation $\rho$ are those $\mu\in \mathcal{H}^\ast$ such that $V_\mu\neq 0$ and the additive subgroup of $\mathcal{H}^\ast$ generated by the weights of $\rho$ forms a lattice, denoted by $L_\rho$. Let $L_1$ denote the lattice generated by all the weights of all representations of $\mathcal{L}$ and $L_0$ the sublattice (of $L_1$) generated by the roots $\alpha\in\Phi$. If $\rho$ is faithful then we have $L_0\subset L_\rho\subset L_1$. Moreover, if $L^\prime$ is any lattice such that $L_0\subset L^\prime\subset L_1$, then there exists a faithful representation $(V,\pi) $ of $\mathcal{L}$ such that $L_\pi=L^\prime$.
So let $(V,\pi)$ be a faithful finite dimensional complex representation of $\mathcal{L}$. If $(\mathcal{U},\varphi)$ is the universal enveloping algebra of $\mathcal{L}$ then $\pi$ canonically extends to a representation of $\mathcal{U}$ in $V$ (also denoted by $\pi$). Since $\varphi$ is injective we identify the elements of $\mathcal{L}$ with their images in $\mathcal{U}$ under the homomorphism $\varphi$. Let $\mathcal{U}_\mathbb{Z}$ be the $\mathbb{Z}$-algebra generated by the elements $e_\alpha^n/n!$ ($\alpha\in \Phi, n\geq 0$). A *lattice in $V$* is a finitely generated free abelian subgroup of $V$ with a $\mathbb{Z}$-basis which is also a $\mathbb{C}$-basis of $V$. It turns out that $V$ admits a lattice $M$ (say) which is invariant under $\pi(e_\alpha)^n/n!$ ($\alpha\in \Phi, n\geq 0$) and hence under $\mathcal{U}_\mathbb{Z}$ [@St2 p. 17, Corollary 1(a)]. Now consider the module $M\otimes_\mathbb{Z} \mathbb{Z}[X]$, where $X$ is an indeterminate. Then $X^n\pi(e_\alpha)^n/n!$ acts on $M\otimes_\mathbb{Z} \mathbb{Z}[X]$. Since $\pi(e_\alpha)^n$ acts as zero on $V$ for sufficiently large values of $n$, we note that the expression $\sum\limits_{n=0}^\infty X^n\pi(e_\alpha)^n/n!$ acts on $M\otimes_\mathbb{Z} \mathbb{Z}[X]$ and hence on $M\otimes_\mathbb{Z} \mathbb{Z}[X]\otimes_\mathbb{Z} k$, where $k$ is an algebraically closed field. If $V_k:=M\otimes_\mathbb{Z} k$ then for each $t\in k$ the homomorphism $M\otimes_\mathbb{Z} \mathbb{Z}[X]\otimes_\mathbb{Z} k\longrightarrow V_k$ given by $X\mapsto t$ induces an automorphism $\sum\limits_{n=0}^\infty t^n\pi(e_\alpha)^n/n!$ of the $k$-vector space $V_k$. Set $\exp(t\pi(e_\alpha)):=\sum\limits_{n=0}^\infty t^n\pi(e_\alpha)^n/n!$.
A *Chevalley group of type $\Phi$ over $k$* associated to $\pi$ is defined as the group $G(\pi,k):=\langle \exp (t\pi(e_\alpha))\mid \alpha\in \Phi,t\in k\rangle$. It is a semisimple algebraic group over $k$ and every semisimple algebraic group over $k$ (with an associated root system $\Phi$) is isomorphic to some $G(\pi, k)$. If $L_\pi=L_1$, then the associated Chevalley group (denoted by $G_{sc}$) is said to be of *universal type* and it is simply connected as an algebraic group. If $L_\pi=L_0$ then the associated Chevalley group (denoted by $G_{ad}$) is said to be of *adjoint type* (it is adjoint in the sense of an algebraic group over $k$).
Let $G(\pi ,k)$ and $G(\rho,k)$ be Chevalley groups constructed from the same Lie algebra $\mathcal{L}$ such that $L_{\rho}\subset L_{\pi}$. Then there exists a homomorphism of algebraic groups $f:G(\pi,k)\longrightarrow G(\rho,k)$ such that $f(\exp(t\pi(e_\alpha)))=\exp(t\rho(e_\alpha))$ for all $\alpha\in \Phi,t\in k$. If $L_\pi=L_\rho$ then $f$ is an isomorphism. If $\rho$ is the adjoint representation of $\mathcal{L}$ (i.e., $L_\rho=L_0$) then $\ker(f)=Z(G(\pi,k))$ [@St2 Corollary 5 (p. 44), Corollary 1 (p. 64)].
Automorphisms
-------------
Let $G$ be a connected semisimple algebraic group. Fix a maximal torus $T$ and a Borel subgroup $B$ of $G$ such that $T\subset B$. Let $\Phi$ be the root system of $G$ determined by $T$ and $\Delta$ the simple subsystem of $\Phi$ determined by $B$. Let ${\textup{Aut}}(G)$ denote the group of all algebraic group automorphisms of $G$ and ${\mathrm{Int}}(G)$ the group of all inner automorphisms of $G$ (hence ${\mathrm{Int}}(G)$ is a subgroup of ${\textup{Aut}}(G)$). If $D:=\{\varphi\in{\textup{Aut}}(G)\mid \varphi(T)=T,\varphi(B)=B\}$ and $\Gamma$ is the group of all automorphisms of $\Phi$ which stabilize $\Delta$, then there exists a natural homomorphism $D\longrightarrow \Gamma$. With the above notation, we quote the following result:
[@hum Theorem 27.4]\[staut\] Let $G$ be a semisimple algebraic group. Then the following holds:
1. \[staut1\] ${\textup{Aut}}(G)={\mathrm{Int}}(G)D$.
2. \[staut2\] The natural map $D\longrightarrow \Gamma$ induces a monomorphism ${\textup{Aut}}(G)/{\mathrm{Int}}(G)\longrightarrow \Gamma$.
Furthermore, since we are working over an algebraically closed field, the group $G$ is split and hence we have
[@kmrt Theorem 25.16]\[kernel\] If $G$ is a simply connected or adjoint type semisimple algebraic group, then the sequence $$\xymatrix{1\ar[r]& {\mathrm{Int}}(G)\ar[r]&{\textup{Aut}}(G)\ar[r]&\Gamma\ar[r]&1}$$ is split exact.
\[graphexist\] If $G$ is simple with root system $\Phi$ then the group $\Gamma$ is non-trivial if and only if $\Phi$ is of type $A_l\, (l\geq 2), D_l\, (l\geq 4)$ and $E_6$ (see [@humlie p. 66, Table 1] and Figure \[Figure1\]). Therefore if $\Phi$ is not of type $A_{2l}$ ($l\geq1$) then for every $\gamma\in \Gamma$ there exists a simple root $\alpha\in \Phi$ such that $\gamma(\alpha)=\alpha$.
0.2 cm
at (-0.5,0) [$A_l\, (l\geq 2)$]{}; (a1) at (1,0) [$\bullet$]{}; (a2) at (2,0) [$\bullet$]{}; (al-1) at (3,0) [$\bullet$]{}; (al) at (4,0) [$\bullet$]{}; (1,0) – (2,0); (a2)– (al-1); (3,0) – (4,0); (a1) edge\[bend left=50\] node\[above\] (al); (a2) edge\[bend left=60\] node\[above\] (al-1); at (7,0) [$\Gamma\cong\mathbb{Z}/2\mathbb{Z}$]{};
at (-1,0) [$D_4$]{}; (d1) at (1,0) [$\bullet$]{}; (d2) at (2,0) [$\bullet$]{}; (d3) at (2.5, 0.7) [$\bullet$]{}; (d4) at (2.5,-0.7) [$\bullet$]{}; (1,0) – (2,0); (2,0) – (2.5,0.7); (2,0) – (2.5,-0.7); (d1) edge\[bend left=50\] node\[above\] (d3); (d1) edge\[bend right=50\] node\[above\] (d4); (d3) edge\[bend left=50\] node\[above\] (d4); at (7,0) [$\Gamma\cong S_3$]{};
at (-0.5,0) [$D_l\, (l> 4)$]{}; (d1) at (1,0) [$\bullet$]{}; (d2) at (2,0) [$\bullet$]{}; (dl-3) at (3,0) [$\bullet$]{}; (dl-2) at (4,0) [$\bullet$]{}; (dl-1) at (5, 0.5) [$\bullet$]{}; (dl) at (5,-0.5) [$\bullet$]{}; (1,0) – (2,0); (d2) – (dl-3); (3,0) – (4,0); (4,0) – (5,0.5); (4,0) – (5,-0.5); (dl-1) edge\[bend left=60\] node\[above\] (dl); at (7,0) [$\Gamma\cong\mathbb{Z}/2\mathbb{Z}$]{};
at (-1,0) [$E_6$]{}; (e1) at (1,0) [$\bullet$]{}; (e2) at (2,0) [$\bullet$]{}; (e3) at (3,0) [$\bullet$]{}; (e4) at (4,0) [$\bullet$]{}; (e5) at (5,0) [$\bullet$]{}; (e6) at (3,1) [$\bullet$]{}; (1,0) – (2,0); (2,0) – (3,0); (3,0) – (4,0); (4,0) – (5,0); (3,0) – (3,1); (e1) edge\[bend right=50\] node\[above\] (e5); (e2) edge\[bend right=50\] node\[above\] (e4); at (7,0) [$\Gamma\cong\mathbb{Z}/2\mathbb{Z}$]{};
Some basic results {#general}
------------------
The contents of this section are known for any abstract group. For the sake of completeness, we prove the following results in the context of algebraic groups. The proofs essentially follow the lines of argument used in [@FLT Corollary 3.2] and [@MS Lemma 2.1, Lemma 2.2].
\[inner\] Let $\varphi\in {\textup{Aut}}(G)$ and ${\mathrm{Int}}_g$ the inner automorphism defined by $g\in G$. Then $R(\varphi\circ{\mathrm{Int}}_g)=R(\varphi)$. In particular, $R({\mathrm{Int}}_g)=R({\textup{Id}})$, i.e., the number of inner twisted conjugacy classes in $G$ is equal to the number of conjugacy classes in $G$.
Let $x,y\in G$ such that $[x]_{\varphi\circ{\mathrm{Int}}_g}=[y]_{\varphi\circ{\mathrm{Int}}_g}$. Then there exists a $z\in G$ such that $y=zx(\varphi\circ{\mathrm{Int}}_g)(z^{-1})=zx\varphi(gz^{-1}g^{-1})=zx\varphi(g)\varphi(z^{-1})\varphi(g^{-1})$. This implies that $y\varphi(g)=zx\varphi(g)\varphi(z^{-1})$, i.e, $[x\varphi(g)]_{\varphi}=[y\varphi(g)]_{\varphi}$. Thus we get a well-defined map $$\widehat{\varphi}: \mathcal{R}(\varphi\circ {\mathrm{Int}}_g)\longrightarrow \mathcal{R}(\varphi)$$ given by $\widehat{\varphi}([x]_{\varphi\circ {\mathrm{Int}}_g})=[x\varphi(g)]_{\varphi}$. This is bijective as well. Therefore $R(\varphi\circ{\mathrm{Int}}_g)=R(\varphi)$. Second part of the lemma follows by taking $\varphi={\textup{Id}}$.
\[ses\] Suppose that $1\longrightarrow N \overset{i}{\longrightarrow} G\overset{\pi}{\longrightarrow} Q\longrightarrow 1$ is an exact sequence of algebraic groups. Let $\varphi\in{\textup{Aut}}(G)$ be such that $\varphi(N)=N$. Let $\overline{\varphi}$ denote the automorphism of $Q$ induced by $\varphi$. Then
1. \[qtog\] $R(\varphi)\geq R(\overline{\varphi})$.
2. \[gtoq\] If $N$ is finite and $R(\varphi)=\infty$ then $R(\overline{\varphi})=\infty$.
3. \[ntog\] If $Q$ is finite and $R(\varphi|_N)=\infty$ then $R(\varphi)=\infty$.
\(1) Since $\varphi\in {\textup{Aut}}(G)$ and $\varphi(N)=N$, then the restriction of $\varphi$ to $N$, denoted by $\varphi|_{N}$, is an automorphism of $N$. Clearly, $\varphi$ induces an automorphism of $Q\cong G/N$, denoted by $\overline{\varphi}$ such that the following diagram commutes: $$\xymatrix{
1\ar[r] & N\ar[r]^{i}\ar[d]_{\varphi|_{N}}& G\ar[r]^{\pi} \ar[d]^{\varphi}& Q\ar[r]\ar[d]^{\overline{\varphi}}&1 \\
1\ar[r] & N\ar[r]^{i}& G\ar[r]^{\pi}& Q\ar[r]&1
}.$$ In particular, we have $\overline{\varphi}\circ \pi=\pi\circ \varphi$. Now, observe that $\pi$ induces a surjective map $\widehat{\pi}:\mathcal{R}(\varphi)\longrightarrow\mathcal{R}(\overline{\varphi})$ given by $\widehat{\pi}([x]_{\varphi})=[\pi(x)]_{\overline{\varphi}}$ for all $x\in G$. Therefore $R(\varphi)\geq R(\overline{\varphi})$.
\(2) If possible suppose that not all fibres of $\widehat{\pi}$ are finite. Let $x_r\in G$ ($r\geq 0$) be such that $\widehat{\pi}([x_r]_{\varphi})=\widehat{\pi}([x_0]_{\varphi})$ for all $r\geq 1$ but $[x_r]_{\varphi}\neq [x_s]_{\varphi}$ for all $r\neq s$ (this is possible bacause $R(\varphi)=\infty$). Then for each $r\geq 1$, we have $$\pi(x_0)=\pi(g_r)\pi(x_r)\overline{\varphi}(\pi(g_r))^{-1}=\pi(g_r)\pi(x_r)\pi(\varphi(g_r))^{-1}=\pi(g_rx_r\varphi(g_r)^{-1}),$$ for some $g_r\in G$. Thus, for each $r\geq 1$ there exists $n_r\in N$ such that $x_0n_r=g_rx_r\varphi(g_r)^{-1}$. Hence $[x_0n_r]_{\varphi}=[x_r]_{\varphi}$ for all $r$, which is a contradiction, since $N$ is finite and $[x_r]_{\varphi}\neq [x_s]_{\varphi}$ for $r\neq s$. Hence all the fibres of $\widehat{\pi}$ are finite whenever $N$ is finite. Therefore, by the previous part of this lemma and our assumption that $R({\varphi})=\infty$, we have $R(\overline{\varphi})=\infty$.
\(3) If possible suppose that $R(\varphi)<\infty$. Since $R(\varphi|_{N})=\infty$ then there exist infinitely many elements $n_r\in N$ ($r\geq 0$) such that $[n_r]_{\varphi|_{N}}\neq [n_s]_{\varphi|_{N}}$ ($r\neq s$) but $[n_r]_{\varphi}=[n_0]_{\varphi}$ for all $r\geq 0$. So for all $r\geq 1$ there exists $x_r\in G$ such that $n_r=x_rn_0\varphi(x_r^{-1})$. Again since $Q(\cong G/N)$ is finite then there exist distinct $r,s\in \mathbb{N}$ such that $x_sx_r^{-1}\in N$. Therefore $$n_s=x_sn_0\varphi(x_s^{-1})=x_sx_r^{-1}n_r\varphi(x_r)\varphi(x_s^{-1})=(x_sx_r^{-1})n_r\varphi(x_sx_r^{-1})^{-1}.$$ Hence $[n_r]_{\varphi|_N}=[n_s]_{\varphi|_N}$, which is a contradiction.
This completes the proof.
Proof of the main result {#mainsec}
========================
The following results will be used in the proof of our main theorem.
\[autprod\] Suppose that $G$ is a simple algebraic group. Then $${\textup{Aut}}(G^n)\cong ({\textup{Aut}}(G))^n\rtimes S_n.$$
Let $G^n=G_1\times G_2\times \cdots \times G_n$, where $G_i=G$ and set $H_i=(1\times \cdots \times G_i\times \cdots \times 1)$ for $i\in I:=\{1,...,n\}$. Then $G^n$ is semisimple and $H_1,...,H_n$ are its simple components (i.e., minimal closed connected normal subgroups of $G$ of positive dimension). If $\varphi\in {\textup{Aut}}(G^n)$ then for every $i\in I,$ $\varphi(H_i)$ is also a simple component and hence there exists $k_i\in I$ such that $\varphi(H_i)=H_{k_i}$ (by Theorem 27.5 of [@hum]). Therefore, $\varphi$ gives rise to a bijection $\pi(\varphi):I\longrightarrow I$ defined by $\pi(\varphi)(i)=k_i$ for all $i\in I$. Hence, we have a homomorphism $\pi:{\textup{Aut}}(G^n)\longrightarrow S_n$ such that $\ker(\pi)={\textup{Aut}}(G)^n$, where ${\textup{Aut}}(G)^n$ is identified with a subgroup of $ {\textup{Aut}}(G^n)$ via $$(\varphi_1, \ldots, \varphi_n)\mapsto \varphi_{(1, \ldots, n)}((x_1, \ldots, x_n)\mapsto (\varphi_1(x_1), \ldots, \varphi_n(x_n)).$$
Now, for every $\gamma\in S_n$, define $\overline{\gamma}:G^n\longrightarrow G^n$ by $\overline{\gamma}(x_1, \ldots, x_n)=(x_{\gamma^{-1}(1)}, \ldots, x_{\gamma^{-1}(n)})$ for all $(x_1,...,x_n)\in G^n$ and observe that $\gamma\mapsto \overline{\gamma}$ defines an injective homomorphism $S_n \hookrightarrow {\textup{Aut}}(G^n)$. Observe that $\overline{\gamma}(H_i)=H_{\gamma(i)}$ and hence $\pi(\overline{\gamma})(i)=\gamma(i)$ for all $i\in I, \gamma\in S_n$. This shows that $\pi$ is surjective and the exact sequence $$1\longrightarrow {\textup{Aut}}(G)^n\longrightarrow {\textup{Aut}}(G^n)\overset{\pi}\longrightarrow S_n\longrightarrow 1$$ splits on the right. Hence ${\textup{Aut}}(G^n)\cong {\textup{Aut}}(G)^n\rtimes S_n$ as desired.
We get the following characterization of the $R_{\infty}$-property using the above lemma.
\[productgn\] Suppose that a simple algebraic group $G$ possesses the $R_{\infty}$-property. Then $G^n$ also possesses the $R_{\infty}$-property.
Let $\varphi\in {\textup{Aut}}(G^n)$. Then by [Lemma \[autprod\]]{}, we have $\varphi=(\varphi_1, \varphi_2, \ldots, \varphi_n; \overline{\sigma})$, where $\varphi_i\in {\textup{Aut}}(G)$ for all $i=1, 2, \ldots, n$ and $\sigma\in S_n$. If $\sigma^{-r}(1)=1$ for some smallest positive integer $r$, then choose $a_i$’s from distinct $\varphi_1\varphi_{\sigma^{-1}(1)}\cdots \varphi_{\sigma^{-(r-1)}(1)}$-twisted conjugacy classes in $G$ ($i\in\mathbb{N}$).
**Claim:** $R(\varphi)=\infty$.
If possible let $R(\varphi)<\infty$. Then without loss of generality we may assume that $(a_i, 1, \ldots, 1)\sim_{\varphi} (a_1, 1, \ldots, 1)$ in $G^n$ for $i=2, 3, \ldots$. Then for some fixed $i$ there exists $\widetilde{g}_i\in G^n$ such that $$\begin{aligned}
\label{eq1}
(a_1, 1, \ldots, 1)=\widetilde{g}_i(a_i, 1, \ldots, 1)\varphi(\widetilde{g}_i)^{-1}\end{aligned}$$ where $\widetilde{g}_i=(g_1, g_2, \ldots, g_n)\in G^n$. Now $$\begin{aligned}
\label{eq2}
\varphi(\widetilde{g}_i)=(\varphi_1, \varphi_2, \ldots, \varphi_n; \overline{\sigma})(g_1, g_2, \ldots, g_n)=(\varphi_1(g_{\sigma^{-1}(1)}), \varphi_2(g_{\sigma^{-1}(2)}), \ldots, \varphi_n(g_{\sigma^{-1}(n)})). \end{aligned}$$ Therefore from Equations and , we get $$\begin{aligned}
(a_1, 1, \ldots, 1)=(g_1, g_2, \ldots, g_n)(a_i, 1, \ldots, 1)(\varphi_1(g_{\sigma^{-1}(1)}^{-1}), \varphi_2(g_{\sigma^{-1}(2)}^{-1}), \ldots, \varphi_n(g_{\sigma^{-1}(n)}^{-1})).\end{aligned}$$ This implies $$\begin{aligned}
a_1&=g_1a_i\varphi_1(g_{\sigma^{-1}(1)}^{-1}),\\
g_j^{-1}&=\varphi_j(g_{\sigma^{-1}(j)}^{-1})\end{aligned}$$ for $j=2, 3, \ldots, n$. Therefore, from the above two equations, we get $$\begin{aligned}
a_1&=g_1a_i\varphi_1(g_{\sigma^{-1}(1)}^{-1})=g_1a_i\varphi_1\varphi_{\sigma^{-1}(1)}(g_{\sigma^{-2}(1)}^{-1})\\
&\vdots\\
&=g_1a_i(\varphi_1\varphi_{\sigma^{-1}(1)}\cdots \varphi_{\sigma^{-(r-1)}(1)})(g_{\sigma^{-r}(1)}^{-1})\\
&=g_1a_i(\varphi_1\varphi_{\sigma^{-1}(1)}\cdots \varphi_{\sigma^{-(r-1)}(1)})(g_1^{-1}).\end{aligned}$$ Which is a contradiction to the fact that $a_i$’s are in different $\varphi_1\varphi_{\sigma^{-1}(1)}\cdots \varphi_{\sigma^{-(r-1)}(1)}$-twisted conjugacy classes in $G$. Therefore $R(\varphi)=\infty$.
\[not\] Let $G=G(\pi, k)$ be a Chevalley group of type $\Phi$ over $k$. For each $\alpha\in \Phi$ set $$\begin{aligned}
x_{\alpha}(t)&=\exp(t\pi(e_{\alpha})) \quad t\in k\\
X_{\alpha}&=\langle x_{\alpha}(t)\mid t\in k\rangle\\
n_{\alpha}(t)&=x_{\alpha}(t)x_{-\alpha}(-t^{-1})x_{\alpha}(t) \quad t\in k^{\times} \\
h_{\alpha}(t)&=n_{\alpha}(t)n_{\alpha}(-1) \quad t\in k^{\times}.\end{aligned}$$ Then $G=\langle x_{\alpha}(t)\mid\alpha\in \Phi, t\in k\rangle$. Viewing $G$ as a semisimple linear algebraic group over $k$, we observe that $T=\langle h_{\alpha}(t)\mid\alpha\in \Phi, t\in k^{\times}\rangle$ is a maximal torus of $G$ and $B=T\ltimes U$ is a Borel subgroup of $G$, where $U=\langle x_{\alpha}(t)\mid \alpha\in \Phi^+, t\in k\rangle$ and $\Phi^+$ a positive subsystem. With the above notation we state
[@St2 Corollary 6 (p. 46), Corollary 2 (p. 65)]\[onedim\] For every $\alpha\in \Phi$ there exists a homomorphism of algebraic groups $\varphi_\alpha:\SL_2(k)\longrightarrow \langle X_\alpha,X_{-\alpha}\rangle$ such that $\varphi_\alpha\begin{pmatrix}
1&t\\0&1
\end{pmatrix}=x_\alpha(t)$, $\varphi_\alpha\begin{pmatrix}
1&0\\t&1\end{pmatrix}=x_{-\alpha}(t)$ (for all $t\in k$), $\varphi_\alpha\begin{pmatrix}
0&1\\-1&0
\end{pmatrix}=n_\alpha(1)$, and $\varphi_\alpha\begin{pmatrix}
t&0\\0&t^{-1}
\end{pmatrix}=h_\alpha(t)$ (for all $t\in k^\times$), Moreover, $\ker(\varphi_\alpha)=\{1\}$ or $\{\pm 1\}$, i.e., $\langle X_\alpha,X_{-\alpha}\rangle$ is either $\SL_2(k)$ or $\mathrm{PSL}_2(k)$.
[@St2 p. 44, Corollary (a) to Lemma 28]\[onedim1\] If $G$ is of universal type (i.e., $L_\pi=L_1$) then every element of $h\in T$ can be uniquely written as $h=\prod\limits_{i=1}^nh_{\alpha_i}(t_i),$ $t_i\in k^\times$, where $\Delta=\{\alpha_1,\ldots, \alpha_n\}$ is a simple subsystem of $\Phi$, determined by $B$.
Let $\Gamma$ be as in [Section \[automorphisms\]]{} and $\rho\in \Gamma$. Then by Theorem 29 of [@St2 p. 154], there exists an abstract automorphism $\overline{\rho}:G\longrightarrow G$ such that $\overline{\rho}(x_{\alpha}(t))=x_{\rho(\alpha)}(\epsilon_{\alpha}t)$ for all $\alpha\in \Phi, t\in k$, where $\epsilon_\alpha$ is a sign depending on $\alpha\in \Phi$. Furthermore, $\epsilon_{\alpha}=1$ if $\alpha$ or $-\alpha$ is a simple root.
\[algebraic\] For every $\rho\in \Gamma$, the map $\overline{\rho}$ is an algebraic group automorphism of $G$. Moreover, $\overline{\rho}\in D$ (where $D$ is as in [Section \[automorphisms\]]{}).
We observe that for every $\alpha\in\Phi$, $$\begin{aligned}
\overline{\rho}(h_{\alpha}(t))&=\overline{\rho}\left(n_\alpha(t)n_\alpha(-1)\right)=\overline{\rho}\left(x_\alpha(t)x_{-\alpha}(-t^{-1})x_\alpha(t)x_\alpha(-1)x_{-\alpha}(1)x_\alpha(-1)\right)\\
&=x_{\rho(\alpha)}(\epsilon_{\alpha}t)x_{\rho(-\alpha)}(-\epsilon_{-\alpha}t^{-1})x_{\rho(\alpha)}(\epsilon_{\alpha}t)x_{\rho(\alpha)}(-\epsilon_{\alpha})x_{\rho(-\alpha)}(\epsilon_{-\alpha})x_{\rho(\alpha)}(-\epsilon_{\alpha})\\
&=n_{\rho(\alpha)}(\epsilon_{\alpha}t)n_{\rho(\alpha)}(-\epsilon_{\alpha})\quad (\text{since } \rho(-\alpha)=-\rho(\alpha), \epsilon_{\alpha}\epsilon_{-\alpha}=1)\\
&=h_{\rho(\alpha)}(\epsilon_{\alpha}t). \end{aligned}$$ Therefore $\overline{\rho}(T)=T$. Also, note that $\overline{\rho}(B)=B$, since $\overline{\rho}(U)=U$. It remains to show that $\overline{\rho}$ is an algebraic morphism. In view of [@hum Lemma 32.1], it is sufficient to prove that the restriction of $\overline{\rho}$ on the Borel subgroup $B$ of $G$ is a morphism. Now, observe that $X_{\alpha}\longrightarrow X_{\rho(\alpha)}$ (defined by $x_\alpha(t)\mapsto x_{\rho(\alpha)}(t), t\in k$) and the product map $X_{\alpha_1}\times X_{\alpha_2}\times \cdots \times X_{\alpha_r}\longrightarrow U$, defined by $(x_{\alpha_1}(t_1),\ldots, x_{\alpha_r}(t_r))\mapsto \prod x_{\alpha_i}(t_i)$ ($t_i\in k, i=1,\ldots,r$) are isomorphisms of varieties, where $\Phi^+=\{\alpha_1, \ldots, \alpha_r\}$ is the set of all positive roots (taken in any order). Therefore $\overline{\rho}$ is an isomorphism of varieties when restricted to $U$. Also, we have the following commutative diagram: $$\xymatrix{\mathbb{G}_m^n \ar[r]^{f} \ar[dr]_{\overline{\rho}\circ f}
& T \ar[d]^{\overline{\rho}} \\
&T },$$ where $f(t_1, \ldots, t_n)=\prod_{i=1}^nh_{\alpha_i}(t_i)$ for $t_i\in k^\times$ and $\{\alpha_1,...,\alpha_n\}$ is the set of all simple roots. This in turn induces $$\xymatrix{\frac{\mathbb{G}^n_m}{K} \ar[r]^{\widetilde{f}} \ar[dr]_{\widetilde{\overline{\rho}\circ f}}
& T \ar[d]^{\overline{\rho}} \\
&T }$$ where $K:=\ker(f)=\ker(\overline{\rho}\circ f)$, and $\widetilde{f}$ and $\widetilde{\overline{\rho}\circ f}$ are isomorphism of varieties. Therefore, $\overline{\rho}:T\longrightarrow T$ is a morphism. Thus, $\overline{\rho}$ is a morphism when restricted to $B=TU$. Similarly it can be shown that $\overline{\rho}^{-1}$ is also a morphism of varieties. Hence, the proof.
Before moving on to prove our next set of results we make the following observation: Let $G$ be a simply connected or adjoint type semisimple algebraic group and $\varphi\in {\textup{Aut}}(G)$. Then by [Lemma \[staut\]]{}, $\varphi=f\circ {\mathrm{Int}}_g$ for some $f\in D$ and ${\mathrm{Int}}_g\in {\mathrm{Int}}(G)$. Now let $\rho\in \Gamma$ be the automorphism of $\Phi$ induced by $f\in D$ and consider the element $\overline{\rho}\in D$ induced by $\rho$ as in [Lemma \[algebraic\]]{}. Since $\overline{\rho}(X_\alpha)=X_{\rho(\alpha)}$ for all $\alpha\in \Phi$, we note that $\overline{\rho}$ also induces the automorphism $\rho$ of $\Phi$. Therefore by Lemma \[kernel\], $\overline{\rho}^{-1}\circ f\in {\mathrm{Int}}(G)$, i.e., $f=\overline{\rho}\circ {\mathrm{Int}}_y$ for some $y\in G$ and hence, $\varphi=\overline{\rho}\circ {\mathrm{Int}}_{yg}$. So, by [Lemma \[inner\]]{}, we have $R(\varphi)=R(\overline{\rho})$. We summarise this as:
\[ets\] Let $G$ be a simply connected or adjoint type semisimple algebraic group. Then $G$ has the $R_\infty$-property if and only if $R(\overline{\rho})=\infty$ for all $\rho\in \Gamma$.
Now let $G$ be a reductive algebraic group and $\varphi\in{\textup{Aut}}(G)$. Then $G$ acts on itself via $\varphi$-twisted conjugacy. Let $G//_\varphi G$ be the invariant theoretic quotient. Then there exists a surjective morphism $\pi:G\longrightarrow G//_\varphi G$ which is constant on the orbits in $G$ (cf. [@new78 Theorem 3.5]), i.e., if $x\sim_\varphi y$ for $x,y\in G$, then $\pi(x)=\pi(y)$. Therefore if $[x]_\varphi$ denotes the orbit of $x\in G$, then the assignment $[x]_\varphi\mapsto \pi(x)$ (for all $x\in G$) is well-defined and onto. We record this as
\[git\] Let $G$ be a reductive algebraic group and $\varphi\in {\textup{Aut}}(G)$. Then there exists a surjection $ \mathcal{R}(\varphi)\longrightarrow G{//}_{\varphi}G$, where $\mathcal{R}(\varphi)$ is the set of all $\varphi$-twisted conjugacy classes of $G$.
\[step1\] Every simply connected simple algebraic group $G$ possesses the $R_{\infty}$-property.
Let $\Phi$ be the irreducible root system of $G$ determined by a maximal torus $T$. In view of [Lemma \[ets\]]{}, it suffices to show that $R(\overline{\rho})=\infty$ for every $\rho\in \Gamma$. First, let $\rho=1\in \Gamma$. Then by [Lemma \[kernel\]]{}, $\overline{\rho}={\mathrm{Int}}_y$ for some $y\in G$. Therefore in view of [Lemma \[inner\]]{}, we have $R(\overline{\rho})=R({\mathrm{Int}}_y)=R({\textup{Id}})$ which is the number of conjugacy classes in $G$. Since there are infinitely many semisimple conjugacy classes in $G$ (as $k$ is algebraically closed), we conclude that $R(\overline{\rho})=\infty$.
Now assume that $\rho\neq 1$. Then $\Phi$ is of type $A_l \;(l\geq 2), D_l \;(l\geq 4)$ or $E_6$ (cf. [Remark \[graphexist\]]{}). By [@springer Corollary 2], we have $G{//}_{\overline{\rho}}G\cong \mathbb{A}^d$, where $d=\dim T^{\overline{\rho}}$ and $T^{\overline{\rho}}=\{t\in T\mid \overline{\rho}(t)=t\}$. By [Lemma \[git\]]{} it is enough to prove that $G{//}_{\overline{\rho}}G$ is infinite, i.e., $d\geq1$. We do this by considering the following two cases:
**Case I:** Assume that $\Phi$ is of type $A_{2l+1} (l\geq 1), D_l(l\geq 4)$ or $E_6$. By [Remark \[graphexist\]]{} there exists a simple root (say) $\alpha\in \Phi$, such that $\rho(\alpha)=\alpha$. For each such $\alpha$, we note that $\overline{\rho}(h_{\alpha}(t))=h_{\rho(\alpha)}(t)=h_{\alpha}(t)$ for all $t\in k^{\times}$ (cf. proof of [Lemma \[algebraic\]]{}). Thus $h_{\alpha}(t)\in T^{\overline{\rho}}$ for all $t\in k^{\times}$. Hence $\{h_\alpha(t)\mid t\in k^\times\}$ is a $1$-dimensional torus (by [Lemma \[onedim\]]{}) contained in $T^{\overline{\rho}}$. Therefore $d=\dim(T^{\overline{\rho}})\geq 1$.
**Case II:** Let $\Phi$ be of type $A_{2l}$ ($l\geq 1$). Let $\alpha\in \Phi$ be a simple root. Then $\rho(\alpha)\neq \alpha$. Since $\rho^2=1$ we have $\overline{\rho}(h_{\alpha}(t)h_{\rho(\alpha)}(t))=h_{\rho(\alpha)}(t)h_{\rho^2(\alpha)}(t)=h_{\rho(\alpha)}(t)h_{\alpha}(t)=h_{\alpha}(t)h_{\rho(\alpha)}(t)$ for all $t\in k^{\times}$. Thus $S:=\{h_{\alpha}(t)h_{\rho(\alpha)}(t)\mid t\in k^\times\}\subset T^{\overline{\rho}}$. Following [Lemma \[onedim\]]{}, we have a homomorphism $\psi:D_2(k)\longrightarrow T$ be defined by $\psi\begin{pmatrix}
t&0\\0&t^{-1}
\end{pmatrix}=\varphi_{\alpha}\begin{pmatrix}
t&0\\0&t^{-1}\end{pmatrix}
\varphi_{\rho(\alpha)}\begin{pmatrix}
t&0\\0&t^{-1}\end{pmatrix}$ ($t\in k^\times$), where $D_2(k)$ is the diagonal maximal torus of $\SL_2(k)$. Note that $\psi(D_2(k))=S$, which implies that $S$ is a connected subgroup of $T^{\overline{\rho}}$. Hence $S$ is a torus. We *claim* that $S$ is non-trivial. So if possible let $S$ be trivial, i.e., $h_{\alpha}(t)h_{\rho(\alpha)}(t)=1$ for all $t\in k^\times$. But then by [Lemma \[onedim1\]]{} $h_\alpha(t)=h_{\rho(\alpha)}(t)=1$ for all $t\in k^\times$, a contradiction to the fact that $\{h_\alpha(t)\mid t\in k^\times\}$ is a $1$-dimensional torus. Hence, the claim. Therefore $\dim S=1$ which implies that $d=\dim T^{\overline{\rho}}\geq 1$ as desired.
This completes the proof.
\[step2\] Every simple algebraic group $G$ possesses the $R_{\infty}$-property.
Let $\Phi$ be the irreducible root system associated to $G$. Following [Section \[agcg\]]{} $G$ can be treated as a Chevalley group of type $\Phi$ over $k$. So let $G=G(\pi,k)$ and $G_{sc}=G(\pi_1,k)$ (resp. $G_{ad}=G({\mathrm{ad}},k)$) a universal (resp. adjoint) Chevalley group of type $\Phi$ over $k$ (all constructed from the same complex simple Lie algebra $\mathcal{L}$).
Let $f_1:G_{sc}\longrightarrow G_{ad}$ and $f_2:G\longrightarrow G_{ad}$ be the homomorphisms given by $f_1(\exp(t\pi_1(e_\alpha)))=\exp(t{\mathrm{ad}}(e_\alpha))$ and $f_2(\exp(t\pi(e_\alpha)))=\exp(t{\mathrm{ad}}(e_\alpha))$ respectively (for all $t\in k,\alpha\in \Phi$). Then $\ker(f_1)=Z(G_{sc})$ and $\ker(f_2)=Z(G)$.
For every $\rho\in \Gamma$, let $\overline{\rho}$ (resp. $\widetilde{\rho}$) be the automorphism of $G_{ad}$ (resp. $G_{sc}$) induced by $\rho$ as in [Lemma \[algebraic\]]{}. Then we have the following commutative diagram $$\xymatrix{
G_{sc} \ar[r]^{\widetilde{\rho}} \ar[d]_{f_1} & G_{sc}\ar[d]^{f_1} \\
G_{ad} \ar[r]^{\overline{\rho}} &G_{ad}}$$ and a short exact sequence of groups $$\xymatrix{1\ar[r]&Z(G_{sc})\ar[r]&G_{sc}\ar[r]^{f_1}&G_{ad}\ar[r]&1}.$$ Note that for every $\rho\in \Gamma$, $\widetilde{\rho}(Z(G_{sc}))=Z(G_{sc})$. Since $G_{sc}$ has the $R_{\infty}$-property (by [Proposition \[step1\]]{}) and $Z(G_{sc})$ is finite, by virtue of [Lemma \[ses\]]{}(\[gtoq\]) we have $R(\overline{\rho})=\infty$ for all $\rho\in \Gamma$. Hence by [Lemma \[ets\]]{}, $G_{ad}$ has the $R_{\infty}$-property.
Now consider the short exact sequence $$\xymatrix{1\ar[r]&Z(G)\ar[r]&G\ar[r]^{f_2}&G_{ad}\ar[r]&1}.$$ Observe that for every $\varphi\in {\textup{Aut}}(G)$, we have $\varphi(Z(G))=Z(G)$ and $G_{ad}$ has the $R_\infty$-property (proven above). Therefore in view of [Lemma \[ses\]]{}(\[qtog\]), we conclude that $G$ has the $R_{\infty}$-property. Hence the result.
Now, we are in a position to prove our main theorem of this paper.
\[mainthm\] Let $G$ be a connected non-solvable linear algebraic group. Then $G$ satisfies the $R_{\infty}$-property.
First, suppose that $G$ is a connected semisimple algebraic group with root system $\Phi$. Following [Section \[agcg\]]{}, let $G=G(\pi,k)$ and $G_{ad}=G({\mathrm{ad}}, k)$ an adjoint semisimple group of the same type as $G$. Then $G_{ad}$ is a direct product of connected simple groups, i.e., $G_{ad}= G_1^{n_1}\times \cdots \times G_r^{n_r}$, where $G_i$ is simple (for all $1\leq i\leq r$) and $G_i\ncong G_j$ whenever $i\neq j$. Consider the exact sequence $$1\longrightarrow G_2^{n_2}\times\cdots \times G_r^{n_r}\longrightarrow G_{ad}\longrightarrow G_1^{n_1}\longrightarrow 1 .$$ Note that $G_2^{n_2}\times\cdots \times G_r^{n_r}$ is invariant under ${\textup{Aut}}(G_{ad})$ and $G_1^{n_1}$ possesses the $R_{\infty}$-property (by [Lemma \[productgn\]]{} and [Corollary \[step2\]]{}). Hence by [Lemma \[ses\]]{}(\[qtog\]), $G_{ad}$ has the $R_{\infty}$-property. If we take the exact sequence $$1\longrightarrow Z(G)\longrightarrow G\overset{f}\longrightarrow G_{ad}\longrightarrow 1$$ where $f:G\longrightarrow G_{ad}$ is given by $f(\exp(t\pi(e_\alpha))=\exp(t{\mathrm{ad}}(e_\alpha))$ (for all $t\in k$,$\alpha\in \Phi$), we conclude (by [Lemma \[ses\]]{}(\[qtog\])) that the semisimple group $G$ has the $R_\infty$-property.
Now let $G$ be a connected non-solvable algebraic group. Then we have the following short exact sequence $$\xymatrix{1\ar[r]&R(G)\ar[r]&G\ar[r]&G/R(G)\ar[r]&1},$$ where $R(G)$ is the radical of $G$. Note that $R(G)$ is invariant under ${\textup{Aut}}(G)$ and $G/R(G)$ is connected semisimple. Then by the previous paragraph $G/R(G)$ has the $R_\infty$-property and hence, so does $G$ (again by [Lemma \[ses\]]{}(\[qtog\])).
Let $G$ be a linear algebraic group such that the connected component $G^0$ of $G$ is non-solvable. Then $G$ satisfies the $R_{\infty}$-property.
For any linear algebraic group $G$ we have a short exact sequence $$\xymatrix{1\ar[r]&G^{0}\ar[r]&G\ar[r]&G/G^{0}\ar[r]&1},$$ where $G^{0}$ is the connected component of $G$ containing the identity and $G/G^0$ is finite. Observe that $G^{0}$ is invariant under every element of ${\textup{Aut}}(G)$ and by [Theorem \[mainthm\]]{}, $G^0$ has the $R_{\infty}$-property. Therefore, in view of [Lemma \[ses\]]{}(\[ntog\]), $G$ possesses the $R_{\infty}$-property.
Let $G$ be a non-toral reductive algebraic group. Then $G$ has the $R_{\infty}$-property.
The proof follows from [Theorem \[mainthm\]]{} since $G/R(G)$ is semisimple.
We end this section with the following sufficient condition for a connected solvable group to have the $R_{\infty}$-property.
\[solvable\] Let $G$ be a connected solvable algebraic group and $T$ a maximal torus of $G$. Suppose that $\varphi(T)=T$ implies $R(\varphi|_{T})=\infty$ for all $\varphi\in{\textup{Aut}}(G)$. Then $G$ has the $R_{\infty}$-property.
The connected solvable group $G$ can be decomposed as $G=G_u\rtimes T$, where $G_u$ is the subgroup consisting of all the unipotent elements of $G$ and $T$ is a maximal torus. Let $\varphi\in {\textup{Aut}}(G)$. Then there exists $g\in G$ such that $g\varphi(T)g^{-1}=T$, i.e., ${\mathrm{Int}}_g\circ \varphi(T)=T$. Let $\psi:={\mathrm{Int}}_g\circ \varphi$. We will write $\psi|_{G_u}=\psi_1$ and $\psi|_{T}=\psi_2$. Therefore $$[(1,t)]_{\psi}=\{(ast\cdot(\psi_2(s)^{-1}\cdot\psi_1(a)^{-1}), st\psi_2(s)^{-1})\mid a\in G_u, s\in T\}.$$ Note that $\psi(G_u)=G_u$. Since $R(\psi_2)=\infty$ then $R(\psi)=\infty$. Thus, by [Lemma \[inner\]]{}, $R(\varphi)=\infty$. Hence $G$ has the $R_{\infty}$-property.
Examples {#example}
========
In this section we compute $R(\varphi)$ for certain classes of algebraic groups. Example \[eg3\] shows that the sufficient condition proven in [Theorem \[mainthm\]]{} is not necessary.
1. Let $\mathrm{D}_n(k)=\{{\textup{diag}}(t_1,\ldots, t_n)\mid t_i\in k^{\times}\}$ ($n\geq 1$) be the diagonal subgroup of $\GL_n(k)$.
1. \[diag1\] Let $\varphi$ be an automorphism of $\mathrm{D}_n(k)$ given by $$\varphi({\textup{diag}}(t_1,\ldots, t_n))={\textup{diag}}(t_1^{-1},\ldots, t_n^{-1}).$$ Then $\mathrm{D}_n(k)=[I_n]_{\varphi}$. Thus $\mathrm{D}_n(k)$ does not satisfy the $R_{\infty}$-property.
2. For $r\in \mathbb{N}$, let $\varphi_r$ be an automorphism of $\mathrm{D}_n(k)$ given by $$\varphi_r({\textup{diag}}(t_1, t_2, \ldots, t_n))={\textup{diag}}(t_n, t_1, t_2, \ldots, t_{n-2}, t_{n-1}t_n^{-r}).$$ Then $\mathrm{D}_n(k)=[I_n]_{\varphi_r}$ for all $r\in \mathbb{N}$.
2. Let $\varphi=\varphi_1\times \varphi_2$ be the automorphism of $\mathbb{G}_a\times \mathbb G_m$, where $\varphi_1$ is an automorphism of $\mathbb G_a$ given by $\varphi_1(x)=\alpha x$ for $\alpha\in k^{\times}\setminus \{1\}$ and $\varphi_2$ is an automorphism of $\mathbb{G}_m$ given by $\varphi_2(x)=x^{-1}$. Then $R(\varphi)=R(\varphi_1)R(\varphi_2)=1$. Therefore $\mathbb{G}_a\times \mathbb G_m$ does not satisfy the $R_{\infty}$-property.
3. \[unipotent\] Let $G=\U_n(k)$ ($n\geq 1$) be the group of all upper triangular unipotent matrices of $\GL_n(k)$, $d:={\textup{diag}}(t_1, t_2, \ldots, t_n)\in \mathrm{D}_n(k)$ such that $t_i\neq t_j$ for all $i\neq j$ and define an automorphism $\varphi_d: \U_n(k)\longrightarrow \U_n(k)$ by $\varphi_d(g)=dgd^{-1}$ for all $g\in \U_n(k)$.
**Claim:** $R(\varphi_d)=1$, i.e. $g\sim_{\varphi_d}I_n$ for all $g\in\U_n(k)$.
*Proof of claim:* Let $g=(g_{ij})\in \U_n(k)$. It suffices to find some $y\in \U_n(k)$ such that $g=y^{-1}\varphi_d(y)$ or equivalently, $$\label{solution}
yg=dyd^{-1}$$ For $1\leq i<j\leq n$ let us consider the following system of equations in the variables $x_{ij}$ $$\label{eq}
(t_it_j^{-1}-1)x_{ij}=g_{ij}+\displaystyle\sum_{i< k<j}x_{ik}g_{kj}$$
Now, by the assumption on $d$, we note that $(t_it_j^{-1}-1)\neq 0$ for all $i\neq j$. Therefore, it is clear that Equation admits a unique solution say $y_{ij}\quad (1\leq i<j\leq n)$. If we set $y_{ii}=1$ for all $1\leq i\leq n$ and $y_{ij}=0$ for all $i>j$, then $y=(y_{ij})$ satisfies Equation . Hence, the proof.
(See [@timur19 Theorem 2]).
4. Let $G=\mathrm{D}_l(k)\times \U_n(k)$ ($l, n\geq 1$). The map $\psi:G\longrightarrow G$ defined by $\psi(s,u)=(\varphi(s),\varphi_d(u))$ (for all $(s,u)\in G$) is an automorphism (where $\varphi$ is as in Example \[diag1\] and $d,\varphi_d$ as in Example \[unipotent\]). Observe that $R(\psi)=R(\varphi)R(\varphi_d)=1$.
5. \[eg3\] The standard Borel subgroup $B_2(k)=\Big\{\begin{pmatrix}a&b\\0&a^{-1}\end{pmatrix}\mid a\in k^{\times}, b\in k\Big\}$ of $\SL_2(k)$ satisfies $R_\infty$-property. Note that $B_2(k)=U\rtimes T\cong \mathbb{G}_a\rtimes \mathbb{G}_m$, where $T=\Big\{\begin{pmatrix}a&0\\0&a^{-1}\end{pmatrix}\mid a\in k^{\times}\Big\}\cong\mathbb{G}_m$ and $U=\Big\{\begin{pmatrix}
1&y\\0&1
\end{pmatrix}\mid y\in k\Big\}\cong \mathbb{G}_a$. Let $\varphi\in {\textup{Aut}}(B_2(k))$. Since $T$ and $\varphi(T)$ are two maximal torus then there exists $g\in B_2(k)$ such that $g\varphi(T)g^{-1}=T$. Consider the automorphism $\psi:={\mathrm{Int}}_g\circ \varphi$ of $B_2(k)$. Then the restriction of $\psi$ on $T$ is an automorphism of $T$, with an abuse of notation, call it $\psi$. Then it has two possibilities, either $\psi({\textup{diag}}(t, t^{-1}))={\textup{diag}}(t, t^{-1})$ or $\psi({\textup{diag}}(t, t^{-1}))={\textup{diag}}(t^{-1}, t)$ for all $t\in k^{\times}$.
**Claim:** $\psi({\textup{diag}}(t, t^{-1}))={\textup{diag}}(t^{-1}, t)$ for all $t\in k^{\times}$ is not possible.
On the contrary, if it is possible then $\psi \left(\begin{pmatrix}a&b\\0&a^{-1}\end{pmatrix}\right)=\begin{pmatrix}a^{-1}&\alpha a^{-2}b\\0&a\end{pmatrix}$ on $B_2(k)$, since all the automorphism of $\mathbb{G}_a$ is of the form $x\mapsto \alpha x$ for some $\alpha\in k^{\times}$. But clearly, this is not a homomorphism. Hence the claim.
Therefore $\psi={\textup{Id}}$ on $T$. Thus, $R(\psi|_{T})=R({\textup{Id}})=\infty$. Then, in view of [Proposition \[solvable\]]{}, $R(\psi)=\infty$. By [Lemma \[inner\]]{}, we have $R(\varphi)=R({\mathrm{Int}}_g\circ \varphi)=R(\psi)=\infty$. Hence the group $B_2(k)$ has the $R_{\infty}$-property although it is solvable.
**Acknowledgements:** The authors would like to thank Maneesh Thakur, Shripad Garge and Timur Nasybullov for their valuable suggestions and comments on this work. We also thank the reviewers for some extremely crucial comments and suggestions which improved the exposition considerably.
[Hum75]{}
Sushil Bhunia, Pinka Dey, and Amit Roy. Twisted conjugacy classes in twisted Chevalley groups. , arXiv:2002.01446v1, 2020.
Armand Borel. Properties and linear representations of [C]{}hevalley groups. In [*Seminar on [A]{}lgebraic [G]{}roups and [R]{}elated [F]{}inite [G]{}roups ([T]{}he [I]{}nstitute for [A]{}dvanced [S]{}tudy, [P]{}rinceton, [N]{}.[J]{}., 1968/69)*]{}, Lecture Notes in Mathematics, Vol. 131, pages 1–55. Springer, Berlin, 1970.
Alexander Fel’shtyn and Richard Hill. The [R]{}eidemeister zeta function with applications to [N]{}ielsen theory and a connection with [R]{}eidemeister torsion. , 8(4):367–393, 1994.
Alexander Fel’shtyn, Yuriy Leonov, and Evgenij Troitsky. Twisted conjugacy classes in saturated weakly branch groups. , 134:61–73, 2008.
Alexander Fel’shtyn and Timur Nasybullov. The [$R_\infty$]{} and [$S_\infty$]{} properties for linear algebraic groups. , 19(5):901–921, 2016.
Felix Gantmacher. Canonical representation of automorphisms of a complex semi-simple [L]{}ie group. , 5(47):101–146, 1939.
James E. Humphreys. . Springer-Verlag, New York-Heidelberg, 1975. Graduate Texts in Mathematics, No. 21.
James E. Humphreys. , volume 9 of [*Graduate Texts in Mathematics*]{}. Springer-Verlag, New York-Berlin, 1978. Second printing, revised.
Max-Albert Knus, Alexander Merkurjev, Markus Rost, and Jean-Pierre Tignol. , volume 44 of [*American Mathematical Society Colloquium Publications*]{}. American Mathematical Society, Providence, RI, 1998. With a preface in French by J. Tits.
T. Mubeena and P. Sankaran. Twisted conjugacy classes in abelian extensions of certain linear groups. , 57(1):132–140, 2014.
T. Mubeena and P. Sankaran. Twisted conjugacy classes in lattices in semisimple [L]{}ie groups. , 19(1):159–169, 2014.
Timur Nasybullov. Twisted conjugacy classes in general and special linear groups, , 51, no. 3, 220–231, 2012.
Timur Nasybullov. Twisted conjugacy classes in unitriangular groups. , 22(2):253–266, 2019.
Timur Nasybullov. Chevalley groups of types $B_n, C_n, D_n$ over certain fields do not possess the $R_{\infty}$-property. To appear in , 2020.
P. E. Newstead. , volume 51 of [*Tata Institute of Fundamental Research Lectures on Mathematics and Physics*]{}. Tata Institute of Fundamental Research, Bombay; by the Narosa Publishing House, New Delhi, 1978.
T. A. Springer. Twisted conjugacy in simply connected groups. , 11(3):539–545, 2006.
Robert Steinberg. . Memoirs of the American Mathematical Society, No. 80. American Mathematical Society, Providence, R.I., 1968.
Robert Steinberg. . Yale University, New Haven, Conn., 1968. Notes prepared by John Faulkner and Robert Wilson.
[^1]: Bose is supported by DST-INSPIRE Faculty fellowship(IFA DST/INSPIRE/04/2016/001846)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose differential holography as a method to overcome the long-standing forward-scattering problem in photoelectron holography and related techniques for the three-dimensional imaging of atoms. Atomic images reconstructed from experimental and theoretical Cu 3p holograms from Cu(001) demonstrate that this method suppresses strong forward-scattering effects so as to yield more accurate three-dimensional images of side- and back-scattering atoms.'
author:
- 'S. Omori'
- 'Y. Nihei'
- 'E. Rotenberg'
- 'J. D. Denlinger'
- 'S. Marchesini'
- 'S. D. Kevan'
- 'B. P. Tonner'
- 'M. A. Van Hove'
- 'C. S. Fadley'
title: |
Differential Photoelectron Holography:\
A New Approach for Three-Dimensional Atomic Imaging
---
Holography [@Gabor:1948] is a method of recording both the amplitudes and phases of waves scattered by an object illuminated with coherent radiation, and using this information to directly construct a three-dimensional image of the object. Szöke [@Szoke:1986] first suggested that coherent outgoing waves from atomically-localized sources of photoelectrons, fluorescent x-rays, and $\gamma $-rays could be used to achieve atomic-scale holography. This idea was initially demonstrated theoretically for the case of photoelectrons by Barton [@Barton:1988], and then extended into a multi-energy format by Barton and Terminello and by Tong and co-workers [@Barton:1991]. By now several experimental approaches to such atomic-resolution holography have been demonstrated, including photoelectrons [@Tonner:1991; @Terminello:1993; @Tong:1995; @Len:1999], Auger electrons [@Saldin:1993], Kikuchi electrons [@Wei:1994], diffuse-scattered low-energy electrons [@Saldin:1998], fluorescent x-rays in either a direct mode [@Tegze:1994] or a multi-energy inverse mode [@Gog:1996], $\gamma $-rays [@Korecki:1997], and bremsstrahlung x-rays [@Bompadre:1999].
Among these methods, photoelectron holography (PH) has the advantages of being capable of studying the local atomic structure around each type of emitter without requiring long-range order and of distinguishing emitters through core-level binding-energy shifts [@Len:1999]. Photoelectron holograms also show strong modulations of up to $\pm $50[%]{}, so such effects are easily measurable. However, PH can suffer from serious image aberrations due to the strength of electron scattering. The atomic scattering factor $f$ is a highly anisotropic function of scattering angle, and can depend strongly on electron kinetic energy $E_{k}$. In particular, as $E_{k}$ increases above a few hundred eV, $f$ becomes more and more significant in the forward direction, resulting in a strong forward-scattering (FS) peak [@Fadley:1993] that can induce image aberrations. Beyond this, PH also can suffer from multiple-scattering (MS) effects due to the scattering strength.
Various reconstruction algorithms and measurement methods [@Barton:1991; @Tonner:1991; @Tong:1995; @Greber:1996] have been proposed to correct for the anisotropic $f$ and MS effects, some of which can be summarized via $$\label{eq1}
U\left( {\rm {\bf r}} \right) = \left| {\;\int {\;W\chi \left( {\rm {\bf k}}
\right)\;\exp \left[ { - ikr + i{\rm {\bf k}} \cdot {\rm {\bf r}}}
\right]d^3{\rm {\bf k}}\;} } \right|^2$$ where $U$ is the image intensity at position **r**, *$\chi $* is the normalized 3D hologram, and the function or operator $W$ permits describing the difference between algorithms, with $W$=1 in the original multi-energy formulations [@Barton:1991]. One alternative algorithm [@Tonner:1991] sets $W = f^{ - 1}\left( {k,\theta _{{\rm {\bf r}}}^{{\rm {\bf k}}} }
\right)$ so as to divide out the anisotropic $f$, where $\theta _{{\rm {\bf
r}}}^{{\rm {\bf k}}} $ is the angle between **r** and **k**. In another algorithm [@Tong:1995] based on the more ideal electron back scattering (BS), a window function for $W$ that limits the integral in Eq. (\[eq1\]) to be in a small cone of ${\rm {\bf \hat {k}}}$ around $ - {\rm {\bf
r}}$ is chosen to emphasize the imaging of BS atoms. Although successful in several applications [@Tong:1995; @Luh:1998], it is difficult to apply this small-cone method to many systems where the imaging of FS or even side-scattering (SS) atoms is important, such as epitaxial films and buried interfaces. In fact, imaging of “bulk” atoms surrounded by FS and BS atoms via PH has proven to be especially difficult \[cf. Figs. 7-9 in ref. [@Len:1999]\], with most successful applications being to emitters in the first few layers near a surface.
To overcome such FS effects, we propose in this Letter “differential holography”. By simply replacing $\chi $ in Eq. (\[eq1\]) by its $k$-derivative (i.e. $W = \partial \mathord{\left/ {\vphantom {\partial {\partial k}}}
\right. \kern-\nulldelimiterspace} {\partial k})$ or more conveniently by a numerical difference between two $\chi $’s at different energies ($\delta
\chi = \chi \mbox{(}k + \delta k\mbox{)} - \chi \mbox{(}k\mbox{)})$, FS effects can be greatly suppressed. We have applied this method to multi-energy holograms for Cu 3p emission from Cu(001), and show that this provides images that are improved over prior work in several respects.
To avoid confusion with other methods in PH, we also note that “derivative” PH has been proposed and used successfully by Chiang and co-workers [@Luh:1998]. However, the purpose here is to eliminate uncertainties in $I$ due to the variation of experimental conditions by first taking logarithmic derivatives ${\mbox{[}\partial I /
\partial k\mbox{]}} \mathord{\left/ {\vphantom {{\mbox{[}\partial I /
\partial k\mbox{]}} I}} \right. \kern-\nulldelimiterspace} I$ that are then reintegrated into “self-normalized” intensities; thus, it is still finally $\chi$ that is used in Eq.(\[eq1\]).
The principle of differential photoelectron holography (DPH) is as follows. We consider the single-scattering expression of $\chi $ for an emitter-scatterer pair spaced by a vector ${\rm {\bf r}}$ [@Fadley:1993]: $$\begin{aligned}
\label{eq2}
\chi \left( {\rm {\bf k}} \right) =&\frac{I - I_{0} }{I_{0} }\nonumber \\
\approx&\frac{2\left| {f\left( {k,\theta _{{\rm {\bf r}}}^{{\rm {\bf k}}} }
\right)} \right|}{r}\mbox{cos}\left[ {kr\left( {1 - \mbox{cos}\theta _{{\rm
{\bf r}}}^{{\rm {\bf k}}} } \right) + \varphi \left( {k,\theta _{{\rm {\bf
r}}}^{{\rm {\bf k}}} } \right)} \right],\end{aligned}$$
where $I_{0} $ is the intensity that would be observed without atomic scattering, and $\varphi $ is the scattering phase. If $\delta k$ is sufficiently small so that ${\delta \left| f \right|} \mathord{\left/
{\vphantom {{\delta \left| f \right|} {\left| f \right|}}} \right.
\kern-\nulldelimiterspace} {\left| f \right|} \ll 1$, where $\delta \left| f
\right|$ is the change in $\left| f \right|$, the difference of two holograms at $k_{\pm } = k\pm {\delta k} \mathord{\left/ {\vphantom {{\delta
k} 2}} \right. \kern-\nulldelimiterspace} 2$ can be written in a similar form to Eq. (\[eq2\]) as:
$$\begin{aligned}
\label{eq3}
\delta \chi \left( {\rm {\bf k}} \right) =& \chi \left( {k_{ + } {\rm {\bf
\hat {k}}}} \right) - \chi \left( {k_{ - } {\rm {\bf \hat {k}}}} \right)\nonumber \\
\approx& - \frac{2\left| {f_{\mbox{eff}} } \right|}{r}\mbox{sin}\left[
{kr\left( {1 - \mbox{cos}\theta _{{\rm {\bf r}}}^{{\rm {\bf k}}} } \right) +
\bar {\varphi }\left( {k,\theta _{{\rm {\bf r}}}^{{\rm {\bf k}}} } \right)}
\right], \end{aligned}$$
where direction $\hat {k}$ is defined by angles $\theta $ and $\phi $, the “effective” scattering amplitude is defined as $\left| {f_{\mbox{eff}} }
\right| = 2\left| f \right|\mbox{sin}\left[ {{\delta kr\left( {1 -
\mbox{cos}\theta _{{\rm {\bf r}}}^{{\rm {\bf k}}} } \right)} \mathord{\left/
{\vphantom {{\delta kr\left( {1 - \mbox{cos}\theta _{{\rm {\bf r}}}^{{\rm
{\bf k}}} } \right)} 2}} \right. \kern-\nulldelimiterspace} 2 + {\delta
\varphi } \mathord{\left/ {\vphantom {{\delta \varphi } 2}} \right.
\kern-\nulldelimiterspace} 2} \right]$, and $\bar {\varphi }$ is the average of $\varphi $’s at $k_{\pm } $. In the FS region where $\theta _{{\rm {\bf
r}}}^{{\rm {\bf k}}} \to 0$, $\left| {f_{\mbox{eff}} } \right|$ is thus very small, approaching zero in the limit of $\delta \varphi \to 0$. If *$\delta $k* is also small, $\left| {f_{\mbox{eff}} } \right|$ is proportional to $r$; thus, DPH not only suppresses the FS effects, but also enhances the imaging of distant atoms. In Fig. \[fig1\], $\left| f \right|$ and $\left| {f_{\mbox{eff}} } \right|$ are plotted as a function of $\theta
_{{\rm {\bf r}}}^{{\rm {\bf k}}} $ for Cu-Cu nearest neighbors ($r$=2.56 [Å]{}). For $k$=4.6 [Å]{}$^{ - 1}$ and *$\delta $k*=0.2 [Å]{}$^{ - 1}$, $\left|
{f_{\mbox{eff}} } \right|$ is significant only in the region of $\theta
_{{\rm {\bf r}}}^{{\rm {\bf k}}} > \sim 90^o$. Therefore, the imaging of SS and BS atoms is expected, while it will be difficult for this case to image FS atoms. On the other hand, for $k$=8.8 [Å]{}$^{ - 1}$ and a larger fractional *$\delta $k*=1.0 [Å]{}$^{ - 1}$, $\left| {f_{\mbox{eff}} } \right|$ is significant not only in the BS region but also in the range of $\theta _{{\rm {\bf
r}}}^{{\rm {\bf k}}} \sim $30$^{o}$-90$^{o}$. Since near-neighbor FS diffraction fringes extend out beyond 30$^{o}$ [@Fadley:1993; @Omori:1999], we might expect the latter choice to also permit imaging FS atoms. In this way, the relative sensitivity of DPH to SS and FS atoms can be “tuned” by selecting the range and step width of $k$ scans. Finally, we note that the suppression of MS effects by means of a transform over a volume in **k** space is well known in normal multi-energy PH [@Barton:1991] and this suppression will be equally present in DPH. If anything, the inherent elimination of strong FS effects in DPH should lead to even better MS suppression.
To demonstrate DPH experimentally, photoelectron holograms from Cu(001) were measured at beamline 7.0 of the Advanced Light Source at the Lawrence Berkeley National Laboratory. Photoelectron spectra for Cu 3p emission were collected at 25 energies over $k$=4.5-9.3 [Å]{}$^{ - 1}$ ($E_{k}$=77-330 eV) with a constant step of *$\delta $k*=0.2 [Å]{}$^{ - 1}$ (*$\delta $E*$_{k}$ =7-14 eV), along 65 different directions over a symmetry-reduced 1/8 of the total solid angle above the specimen, and with a polar angle range from *$\theta $*=0$^{o}$ (surface normal) to 70$^{o}$. The photoelectron intensity $I(k$,*$\theta $*,*$\phi $*) was fitted by low-order polynomials to obtain the smooth background intensity [@Len:1999; @Len:1997]:
$$\label{eq4}
I_{0} \left( {k,\theta } \right) = \left( {a_{0} + a_{1} k + a_{2} k^2}
\right)\left( {b_{0} + b_{1} \cos \theta + b_{2} \cos 3\theta } \right).$$
Three kinds of $\chi $ were obtained from this fitting: $\chi _{A} $ by fitting the second factor of Eq. (4) to a scanned-angle pattern $I_{k}
\left( {\theta ,\phi } \right)$ at each fixed $k$ [@Terminello:1993], $\chi _{B}$ by fitting the first factor to a scanned-energy curve $I_{{\rm
{\bf \hat {k}}}} \left( k \right)$ at each fixed direction ${\rm {\bf \hat
{k}}}$ [@Tong:1995] and $\chi _{C} $ by fitting both factors to the full data set of $I\left( {k,\theta ,\phi } \right)$ at one time, with the last expected to be the most accurate from an *a priori* point of view [@Len:1999]. The $k$-differences from $\chi _{C} $ were also used for DPH in what we will term Method D (i.e., $\chi _{D} = \delta \chi _{C}
)$. Since low-frequency fringes due to FS events in $I_{{\rm {\bf \hat
{k}}}} \left( k \right)$ are automatically removed in Method B [@Wei:1994], the resulting $I_{0}$ inherently deviates from the true $I_{0}$ defined as the intensity without scattering, especially in the FS direction. In addition, since $I_{k} \left( {\theta ,\phi } \right)$ and $I_{{\rm {\bf \hat {k}}}} \left( k \right)$ are independently normalized without considering the continuity of $\chi $ in the whole sampled ${\rm
{\bf k}}$ space in Methods A and B, they could degrade holographic fringes in $I_{{\rm {\bf \hat {k}}}} \left( k \right)$ and $I_{k}
\left( {\theta ,\phi } \right)$, respectively. By contrast, Method C takes into account the continuity of $\chi $ over the whole data set, but the FS peaks remain in $\chi _{C} $; however, they should be eliminated in $\chi
_{D}$. The original transform of Eq. (\[eq1\]) was used for all four data sets; but to avoid the abrupt truncation of the integral in Eq. (\[eq1\]), $W$ was taken to be the product of a Gaussian function of $k$ and a Hanning function $\cos ^2\theta $, with an additional multiplication by $r$ to make atoms at larger distances more visible.
![Comparison of the usual scattering amplitude $\left| f \right|$ and the effective scattering amplitude of differential holography $\left|
{f_{\mbox{eff}} } \right|$, calculated for Cu-Cu nearest neighbors (r=2.56 [Å]{}) as a function of scattering angle $\theta _{{\rm {\bf r}}}^{{\rm {\bf
k}}} $ for two different sets of $k$ and $\delta k$ in taking the differential of $\chi $: (a) $k$ = 4.6 [Å]{}$^{ - 1}$ (81 eV), $\delta k$ = 0.2 [Å]{}$^{ - 1}$ (7 eV) and (b) $k$ = 8.8 [Å]{}$^{ - 1}$ (295 eV), $\delta
k$ = 1.0 [Å]{}$^{ - 1}$ (67 eV). The final strong forward-scattering data points of $\left| f \right|$ at the right of panel (b) are truncated.[]{data-label="fig1"}](dph1.ps){width="3.5in"}
![Atomic images in the vertical (100) plane of Cu (001) reconstructed from Cu 3p holograms obtained by Methods A-D, as described in the text. The emitter and scatterer positions are indicated by squares and circles, respectively, and various near-neighbor atoms are numbered. Image intensities above or below $z_{c} ={\rm t}$0.5 [Å]{} have been rescaled by the factor shown in each panel, with this factor being determined so as to make the maximum intensities above and below $z_{c} $ equal. Experimental images (a) Image from Method A: normal holography with $I_{0} $ determined by fitting the angular variation of Eq. (4) at each $k$ value. (b) Image from Method B: normal holography with $I_{0} $ determined by fitting the $k$ variation along each direction. (c) Image from Method C: normal holography with $I_{0}
$ determined by fitting both the angular and k variation. (d) Image from Method D: differential holography, with $I_{0} $ as in (c). Theoretical images: (e) As (c) but theoretical.[]{data-label="fig2"}](dph2.ps){width="3.5in"}
Figure \[fig2\] shows atomic images reconstructed from *$\chi $*$_{{\rm A}}$-*$\chi $*$_{D}$ in the vertical (100) plane of Cu(001). In Methods A and C, only elongated features related to FS effects from atoms of types 6 and 7 are observed above $z_{c}=-0.5$ [Å]{} (the arbitrary location of a change in image multiplication). This is consistent with a previous PH study of W(110) [@Len:1999], in which Method C was used. By contrast, it has been reported [@Terminello:1993] that FS atoms of type 5 have been imaged via Method A from Cu 3p holograms for Cu(001) obtained at 9 energies. Even if possible differences in the two sets of experimental data are taken into account, it is difficult to conclude from our results that the images of these FS atoms can be resolved from strong artifacts via Method A. Below $z_{c}$, several peaks near the BS positions 1-3 are observable for A and C among various strong artifacts, but only with the help of higher image amplifications of 46 and 29, respectively.
In Methods B and D, image intensities are stronger in the BS region, with the relative image amplification factors being reversed in sense and smaller at $\times 5$ compared to A and C. In Method D, a strong, somewhat elongated peak is observed at the FS position 6, with weaker features that appear to be associated with atoms 7 also present in the corners of the image. In both B and D, two strong peaks are observed at the SS positions 4 above $z_{c} $ and five peaks are observed at the BS positions 1-3 below $z_{c} $. However, the most intense features in Method B are the artifacts between the two nearest BS atoms of type 3. In Method D, by contrast, the five strongest peaks below $z_{c} $ are of roughly equal intensity and correspond reasonably well to the near-neighbor BS atoms. Therefore, we find Method D to be the most robust for imaging both SS and BS atoms (as well as to some degree also FS atoms 6), even if there are shifts in position of approximately 0.1 [Å]{} for type 1, 0.6 [Å]{} for 2, and 0.3 [Å]{} for 3. Such peak shifts relative to the true atomic positions, as observed in all methods, can be attributed to the present neglect of corrections for both the scattering phase and the inner potential.
For comparison with experiment, we have also performed MS simulations of $I({\rm {\bf k}})$, using a cluster method fully described elsewhere [@Chen:1998]. The theoretical $I_{0}$ was obtained simply as the square of the zeroth-order wave function without scattering. Images reconstructed from the theoretical $\chi $ and $\delta \chi $ via Methods C and D are shown in Figs. \[fig2\](e) and (f) and can be compared with Figs. \[fig2\](c) and (d), respectively. The main features in Figs. \[fig2\](c) and (d) are well reproduced by our simulations, although the artifacts between the atoms 3 are much stronger in experiment for C, and the relative intensity in the region of FS atom 6 is stronger in experiment for D. Even though the ideal *$\chi $* was used for image reconstruction, no atomically-resolved SS or FS peaks are observable in Fig. \[fig2\](e). Therefore, the corresponding artifacts in Fig. \[fig2\](c) are not purely due to the uncertainties in the experimental data and any errors in the $I_{0}$ subtraction, but must have their origin in the MS effects and basic imaging algorithm. On the other hand, Fig. \[fig2\](f) exhibits well-resolved peaks at the SS and BS positions. Since there are no artifacts below $z_{c}$ in Fig. \[fig2\](f), the artifacts in Fig. \[fig2\](d) are by contrast considered to be purely due to the experimental noise and other non-idealities in the data analysis.
We have also generated full three-dimensional atomic images from the experimental data via $\chi _{D}$, although length limitations prevent showing these here. In these images, we find in addition to the atoms 1-4 and 6 in Fig. \[fig2\], two other types of near-neighbor BS and SS atoms located in the vertical (110) plane (denoted types 2’ and 4’ and situated in the same horizontal layers as 2 and 4, respectively). All of these atoms are reasonably well reconstructed, with only a few, such as 2, being significantly shifted in position, but most within a few tenths of an [Å]{} of the correct positions in all directions. The overall positional errors for all of the atoms compared to the known Cu lattice can be summarized as (radial location shift in $xy)$/(vertical location shift in $z)$, and are: 0.0 [Å]{}/0.1 [Å]{} for atoms 1, 0.6 [Å]{}/0.1 [Å]{} for 2, 0.3 [Å]{}/0.1 [Å]{} for 2’, 0.2 [Å]{}/0.1 [Å]{} for 3, 0.1 [Å]{}/0.0 [Å]{} for 4, 0.3 [Å]{}/0.0 [Å]{} for 4’, and 0.0 [Å]{}/0.4 [Å]{} for 6. As a further indication of the overall image quality obtained by DPH, the reader is referred to an animated comparison of 3D images for the four approaches of Figs. 2(a)-(d), in which DPH is alone in imaging approximately 15 near-neighbor atoms [@Animated:1].
Finally, we compare DPH with a very recently introduced approach for PH termed near-node holography [@Wider:2001], in which FS effects are suppressed by using a special experimental geometry with electron exit nearly perpendicular to light polarization. Although this technique is promising, DPH has the advantages that it does not require a special experimental geometry or s-subshell-like form for the photoelectric cross section, that it seems to yield images of as good or better quality [@Animated:1; @Wider:2001] and that it can be used in other types of holography in which polarization cannot be varied.
In summary, we have demonstrated differential photoelectron holography (DPH) as a powerful method for overcoming the FS problem in PH and enhancing image quality for any kind of system in which FS can arise, as for example, bulk emission and buried interfaces. This method should also be helpful in other types of electron holography in which energy can be stepped in a controlled way (e.g. Kikuchi [@Wei:1994] or LEED [@Saldin:1998] holography). The reconstructed images for Cu 3p/Cu(001) demonstrate that DPH is successful in suppressing the FS effects so as to image SS, BS, and to some degree also FS, atoms with accuracies of 0.1-0.6 [Å]{}.
This work was supported in part by the Director, Office of Energy Research, Basic Energy Science, Materials Sciences Division of the U. S. Department of Energy under Contract No. DE-AC03-76SF00098. S.O. and Y.N. also acknowledge the support of the Japan Society for the Promotion of Science (Grant No. JSPS-RFTF 98R14101).
[4]{} D. Gabor, Nature **161**, 777 (1948). A. Szöke, in *Short Wavelength Coherent Radiation: Generation and Applications*, AIP Conf. Proc. No. 147, edited by D. T. Attwood and J. Boker (AIP, New York, 1986), p. 361. J. J. Barton, Phys. Rev. Lett. **61**, 1356 (1988). J. J. Barton, Phys. Rev. Lett. **67**, 3106 (1991); J. J. Barton and L. J. Terminello, in *Structure of Surfaces III*, edited by S. Y. Tong, M. A. Van Hove, X. Xide and K. Takayanagi (Springer-Verlag, Berlin, 1991), p. 107; S. Y. Tong, H. Li and H. Huang, Phys. Rev. Lett. **67**, 3102 (1991). B. P. Tonner *et al*., Phys. Rev. B **43**, 14423 (1991). L. J. Terminello, J. J. Barton and D. A. Lapiano-Smith, Phys. Rev. Lett. **70**, 599 (1993). S. Y. Tong, H. Li and H. Huang, Phys. Rev. B **51**, 1850 (1995); H. Wu and G. J. Lapeyre, *ibid*. **51**, 14549 (1995). P. M. Len *et al*., Phys. Rev. B **59**, 5857 (1999). D. K. Saldin, G. R. Harp and X. Chen, Phys. Rev. B **48**, 8234 (1993). C. M. Wei, I. H. Hong and Y. C. Chou, Surf. Rev. Lett. **1**, 335 (1994). D. K. Saldin and P. L. DeAndres, Phys. Rev. Lett. **64**, 1270 (1990); K. Reuter *et al*., Phys. Rev. B **58**, 4102 (1998). M. Tegze and G. Faigel, Europhys. Lett. **16**, 41 (1991); P. M. Len *et al*., Phys. Rev. B **50**, 11275 (1994). T. Gog *et al*., Phys. Rev. Lett. **76**, 3132 (1996). P. Korecki, J. Korecki and T. Œlezak, Phys. Rev. Lett. **79**, 3518 (1997). S. G. Bompadre, T. W. Petersen and L. B. Sorensen, Phys. Rev. Lett. **83**, 2741 (1999). C. S. Fadley, Surf. Sci. Rep. **19**, 231 (1993). T. Greber and J. Osterwalder, Chem. Phys. Lett. **256**, 653 (1996). D. –A. Luh, T. Miller and T. –C. Chiang, Phys. Rev. Lett. **81**, 4160 (1998). S. Omori, T. Kozakai and Y. Nihei, Surf. Rev. Lett. **6**, 1085 (1999). P. M. Len, Ph.D. thesis, University of California, Davis (1997). Y. Chen *et al*., Phys. Rev. B **58**, 13121 (1998); multiple scattering photoelectron diffraction program at
http://electron.lbl.gov/mscdpack/mscdpack.html. Animated images in 3D showing the four experimental cases of Figs. 2(a)-2(d) also appear at http://electron.lbl.gov/marchesini/dph/ and http://ftp.aip.org/epaps/phys[\_]{}rev[\_]{}lett/\[epaps[\_]{}code\] J. Wider *et al*., Phys. Rev. Lett. **86**, 2337 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Optical properties of a plasmonic nano-antenna made of two metallic nanospheroids (prolate or oblate) are investigated analytically in quasistatic approximation. It is shown that in clusters of two nanospheroids, three types of plasmonic modes can be present. Two of them can be effectively excited by a plane electromagnetic wave, while the third one can be effectively excited only by a nanolocalized light source (an atom, a molecule, a quantum dot) placed in the gap between the nanoparticles. Analytical expressions for absorption and scattering cross-sections, enhancement of a local field, and radiative decay rate of a dipole source placed near such a nano-antenna are presented and analyzed.'
author:
- 'D.V. Guzatov'
- 'V.V. Klimov'
title: 'Optical properties of a two-nanospheroid cluster: analytical approach'
---
\[intro\]Introduction
=====================
Nowadays, quite a number of works are devoted to the study of optical properties of single nanoparticles and their clusters. Special attention is paid to metal nanoparticles with the help of which it is possible to enhance electric fields at the frequencies of localized plasmon resonances [@ref1; @ref2; @ref3]. On the basis of this effect, a variety of possible applications are considered. The most developed is use of large local fields near a rough surface for increasing of surface-enhanced Raman scattering (SERS) [@ref4]. Modification of fluorescence by means of nanoparticles of different shapes is a basis for creation of nanobiosensors [@ref5; @ref6; @ref7; @ref8; @ref9], nano-antennas [@ref10; @ref11; @ref12; @ref13; @ref14], devices for decoding of DNA structure [@ref15], and etc.
At the present time, optical properties of single metallic nanospheres and their clusters [@ref16; @ref17; @ref18; @ref19; @ref20; @ref21; @ref22; @ref23; @ref24; @ref25; @ref26; @ref27; @ref28; @ref29], single nanospheroids [@ref30; @ref31; @ref32], nanoellipsoids [@ref33; @ref34; @ref35; @ref36], and some other nanobodies [@ref2; @ref3; @ref37] are studied well enough from analytical point of view. Many other nanoparticles’ shapes and also nanoparticles clusters are investigated only numerically. Unfortunately, numerical simulations often do not allow understanding of physical nature of interesting and complicated phenomena in this area. That is why any analytical solutions will have principal importance.
In this work, we present the results of an analytical study of optical properties of clusters of two metallic prolate or oblate spheroidal nanoparticles. Such clusters are investigated both experimentally and numerically and form a basis for various possible applications including nanosensors, nano-antennas, and plasmon waveguides [@ref38; @ref39; @ref40; @ref41; @ref42; @ref43; @ref44]. As far as we know, there is no analytical investigation of optical properties of two-nanospheroid clusters. That is caused by extremely cumbersome mathematical calculations connected with use of general spheroidal wave functions and general rotational-translational addition theorems [@ref45; @ref46; @ref47; @ref48]. However, in the case of nanoparticles which sizes are less than the wavelength, i.e. in a quasistatic case, one can neglect retardation effects. It allows us to find an analytical solution for a cluster of two spheroidal nanoparticles, placed in an arbitrary external field. The geometry of the considered problem is shown in Fig. \[fig1\]. For simplicity, we will consider that the cluster consists of two equal spheroids made of a material with dielectric constant $\varepsilon$ and placed in vacuum. Generalization of our approach to unequal spheroids is straightforward and can be easily done with the help of the general theorem presented in the Appendix.
Main attention will be paid to a case of nearly touching and strongly interacting spheroids since it is the case that seems to be most interesting for applications because a substantial enhancement of electric fields occurs there. The opposite case of weakly interacting spheroids can be easily treated with approximation of the spheroids by point dipoles with corresponding polarizabilities [@ref3].
For illustration of the analytical results obtained, we will consider a case of two identical (prolate or oblate) nanospheroids made of silver [@ref49]. We suppose that the largest size of the nanospheroid is equal to 30 nm, and the aspect ratio of the spheroid is taken to be equal to 0.6.
The rest of the paper is organized as follows. In the Section \[sect2\], free plasmon oscillations of a two-nanospheroid cluster are investigated. The results of this section reveal underlying physics and are necessary for interpretation of results of the other sections. In the Section \[sect3\], we will consider optical properties of the two-nanospheroid cluster placed in the field of a plane electromagnetic wave. Here, we will find absorption and scattering cross-sections and the factor of local field enhancement also. In the Section \[sect4\], an object of examination is optical properties of the two-nanospheroid cluster placed in the field of a radiating atom or a molecule, and their decay rates will be calculated. In the Appendix, derivation of the translational addition theorem for spheroidal functions in a quasistatic case is presented. It is the theorem that allows one to find an analytical presentation of the solution for the two-nanospheroid cluster.
\[sect2\]Plasmon oscillations in a cluster of two nanospheroids
===============================================================
It is well-known that all optical properties of nanoparticles can be derived from their plasmonic spectra, that is, from related plasmon eigenvalues $\varepsilon _{\nu} $ and eigenfunctions ${\mathbf
{e}}_{\nu}$, ${\mathbf {h}}_{\nu} $, that are solution of the sourceless Maxwell equations [@ref50]:
$$\begin{aligned}
{\textrm {rot}} {\mathbf {h}}_{\nu} + i\left( {{\frac{{\omega}} {{v_{c}}} }}
\right)\tilde {\varepsilon} {\mathbf {e}}_{\nu} = 0, \nonumber \\
{\textrm {rot}} {\mathbf {e}}_{\nu} - i\left( {{\frac{{\omega}} {{v_{c}}} }}
\right){\mathbf {h}}_{\nu} = 0, \label{eq1}\end{aligned}$$
where $\tilde {\varepsilon} = \varepsilon _{\nu} $ inside the nanoparticle and $\tilde {\varepsilon} = 1$ outside it, $\omega$ is frequency of electromagnetic oscillations and $v_{c}$ is the speed of light in vacuum. As a result, electric field in the presence of any nanoparticle can be presented in the following form [@ref50]:
$${\mathbf {E}}\left( {{\mathbf {r}}} \right) = {\mathbf {E}}_{0}
\left( {{\mathbf {r}}} \right) + {\sum\limits_{\nu} {{\mathbf
{e}}_{\nu} \left( {{\mathbf {r}}} \right)\left(
{{\frac{{\varepsilon \left( {\omega} \right) - 1}}{{\varepsilon
_{\nu} - \varepsilon \left( {\omega} \right)}}}}
\right){\frac{{{\int\limits_{V} {\left( {{\mathbf {e}}_{\nu} \left(
{{\mathbf {r}}} \right){\mathbf {E}}_{0} \left( {{\mathbf {r}}}
\right)} \right)dV}}} }{{{\int\limits_{V} {\left( {{\mathbf
{e}}_{\nu} \left( {{\mathbf {r}}} \right){\mathbf {e}}_{\nu} \left(
{{\mathbf {r}}} \right)} \right)dV}}} }}}} , \label{eq2}$$
where $\varepsilon \left( {\omega} \right)$ describes the dependence of dielectric permittivity of nanoparticle’s specific material on frequency $\omega$, ${\mathbf {E}}_{0}$ is the excitation field, and $\nu$ is a vector index that defines specific plasmonic mode. From Eq. (\[eq2\]), it is possible to find any optical properties of a nanoparticle or a cluster of nanoparticles. So, to understand very complicated optical properties of a two-nanospheroid cluster, we should investigate plasmonic spectrum of this system at first.
For studying plasmon oscillations and other optical properties of clusters of two nanospheroids, it is enough to solve the quasistatic equations:
$${\textrm {div}} \left( {\tilde {\varepsilon} {\mathbf {e}}_{\nu}} \right) = 0, \quad {\textrm {rot}} {\mathbf {e}}_{\nu} = 0, \label{eq3}$$
that can be reduced to solution of Laplace equations by substituting ${\mathbf {e}}_{\nu} = - \nabla \varphi _{\nu}$:
$$\begin{aligned}
\Delta \varphi _{\nu} ^{in} = 0, \quad && \textrm {inside the nanoparticle}, \nonumber \\
\Delta \varphi _{\nu} ^{out} = 0, \quad && \textrm {outside the
nanoparticle}, \nonumber \\
{\left. {{\begin{array}{*{20}c}
{{\left. {\varphi _{\nu} ^{in}} \right|}_{S} = {\left. {\varphi _{\nu
}^{in}} \right|}_{S} ,} \hfill \\
{{\left. {\varepsilon _{\nu} {\frac{{\partial \varphi _{\nu} ^{in}
}}{{\partial {\mathbf {n}}}}}} \right|}_{S} = {\left.
{{\frac{{\partial \varphi _{\nu} ^{out}}} {{\partial {\mathbf
{n}}}}}} \right|}_{S}} \hfill
\\
\end{array}}} \right\}} \quad &&{\begin{array}{*{20}c}
{\textrm{at the surface}} \hfill \\
{\textrm{of the nanoparticle}.} \hfill \\
\end{array}} \label{eq4}
\end{aligned}$$
In Eq. (\[eq4\]), $\varphi _{\nu} ^{in}$, $\varphi
_{\nu} ^{out}$ are the potentials of plasmonic eigenfunctions inside and outside the nanoparticle correspondingly, and ${\left.
{{\frac{{\partial \varphi _{\nu}} } {{\partial {\mathbf {n}}}}}}
\right|}_{S}$ denotes normal derivative at the nanoparticles’ surface $S$. The last equation in (\[eq4\]) provides continuity of normal components of electrical induction. Note that in this case there is no need to find magnetic fields for description of plasmonic oscillations.
The systems of equations obtained in such a way, have nontrivial solutions only for some negative values of permittivity $\varepsilon
_{\nu}$ defining frequency of plasmon oscillations [@ref2; @ref3]. In the case of Drude theory, $\varepsilon \left( {\omega} \right) =
1 - \omega _{pl}^{2} / \omega ^{2}$, frequency of plasmon oscillations can be found from the expression
$$\omega _{\nu} = {\frac{{\omega _{pl}}} {{\sqrt {1 - \varepsilon
_{\nu}}}}} < \omega _{pl} ,
\label{eq5}$$
where $\omega _{pl} $ is bulk plasmon frequency of a metal from which the nanoparticles are made. Our approach allows us to investigate arbitrary spheroids, but for simplicity in the present section we examine equations for plasmon oscillations in a cluster of two identical metal nanospheroids.
In our case of a two-nanospheroid cluster, we will look for a solution as follows. The total potential outside the spheroids will be the sum of their partial potentials (we will omit the mode index $\nu$ further) [@ref19; @ref51]:
$$\varphi ^{out} = \varphi _{1}^{out} + \varphi _{2}^{out},
\label{eq6}$$
while the potentials inside each nanospheroid shall be denoted as $\varphi _{j}^{in}$ ($j = 1$, 2). To find $\varphi
^{out}$ and $\varphi _{1}^{in}$, $\varphi _{2}^{in}$, it is naturally to use spheroidal coordinates. In the case of a prolate nanospheroid, the relation between Cartesian and spheroidal coordinates ($1 \le \xi < \infty $, $ - 1 \le \eta \le 1$, $0 \le
\phi \le 2\pi )$ is [@ref52]:
$$\begin{aligned}
x &=& f\sqrt {\left( {\xi ^{2} - 1} \right)\left( {1 - \eta ^{2}} \right)}
\cos \phi , \nonumber \\
y &=& f\sqrt {\left( {\xi ^{2} - 1} \right)\left( {1 - \eta ^{2}} \right)}
\sin \phi , \nonumber \\
z &=& f\xi \eta ,
\label{eq7}\end{aligned}$$
where $f = \sqrt {c^{2} - a^{2}}$ is a half of the focal distance in a prolate spheroid ($a < c$) which surface is set by the equation $\left( {x^{2} + y^{2}} \right) / a^{2} + z^{2} / c^{2} =
1$.
In the case of an oblate spheroid ($a > c$), the relation between Cartesian and spheroidal coordinates ($0 \le \xi < \infty $, $ - 1
\le \eta \le 1$, $0 \le \phi \le 2\pi$) has the following form [@ref52]:
$$\begin{aligned}
x &=& f\sqrt {\left( {\xi ^{2} + 1} \right)\left( {1 - \eta ^{2}} \right)}
\cos \phi , \nonumber \\
y &=& f\sqrt {\left( {\xi ^{2} + 1} \right)\left( {1 - \eta ^{2}} \right)}
\sin \phi , \nonumber \\
z &=& f\xi \eta ,
\label{eq8}\end{aligned}$$
where $f = \sqrt {a^{2} - c^{2}}$ is a half of the focal distance in the oblate spheroid. Let us note that this expression can be obtained from Eq. (\[eq7\]) by substitution $\xi \to i\xi $ and $f \to - if$. Further, we will use this formal replacement since it is fundamental and allows us to find a solution for oblate spheroids if the solution for prolate spheroid geometry is known [@ref51; @ref52; @ref53; @ref54].
\[subsect2a\]Plasmon oscillations in a cluster of two identical prolate nanospheroids
-------------------------------------------------------------------------------------
To find plasmonic spectra of a two-nanospheroid cluster, it is naturally to use two local systems of spheroidal coordinates ($\xi
_{j}$, $\eta _{j}$, $\phi _{j}$, $j = 1$, 2) the origins $o_{j}$ of which are placed in the centers of corresponding nanospheroids (see Fig. \[fig1\](a)). Coordinates (and all other values) related to the first or second nanospheroid will be denoted by the index “1” or “2” respectively. The potential inside the $j$-th nanospheroid can be presented in the form [@ref55] ($j = 1$, 2):
$$\varphi _{j}^{in} = {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m =
0}^{n} {P_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta
_{j}} \right)\left( {A_{mn}^{\left( {j} \right)} \cos \left( {m\phi
_{j}} \right) + B_{mn}^{\left( {j} \right)} \sin \left( {m\phi
_{j}} \right)} \right)}}} }, \label{eq9}$$
where $P_{n}^{m} \left( {\eta} \right)$ is an associated Legendre function [@ref56] defined in the region $ - 1 \le \eta
\le 1$, $P_{n}^{m} \left( {\xi} \right)$ is an associated Legendre function [@ref56] defined in a complex plane with the branch cut from $ - \infty $ to +1. The partial potential outside the $j$-th nanospheroid can be presented [@ref55] as ($j = 1$, 2)
$$\varphi _{j}^{out} = {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m
= 0}^{n} {Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta
_{j}} \right)\left( {C_{mn}^{\left( {j} \right)} \cos \left( {m\phi
_{j}} \right) + D_{mn}^{\left( {j} \right)} \sin \left( {m\phi
_{j}} \right)} \right)}}} } , \label{eq10}$$
where $Q_{n}^{m} \left( {\xi} \right)$ is an associated Legendre function of the second kind [@ref56] defined in a complex plane with the branch cut from $ - \infty $ to +1.
By construction, the potentials (\[eq9\]) and (\[eq10\]) are solutions of the Laplace equation [@ref55]. So to find a solution of Eq. (\[eq4\]), one should use only the boundary conditions:
$$\begin{aligned}
{\left. {\varphi _{1}^{in}} \right|}_{\xi _{1} = \xi _{0}} = {\left.
{\varphi ^{out}} \right|}_{\xi _{1} = \xi _{0}} ,\quad {\left.
{\varepsilon {\frac{{\partial \varphi _{1}^{in}}} {{\partial \xi
_{1}}} }} \right|}_{\xi _{1} = \xi _{0}} = {\left.
{{\frac{{\partial \varphi ^{out}}}{{\partial \xi
_{1}}} }} \right|}_{\xi _{1} = \xi _{0}} , \nonumber \\
{\left. {\varphi _{2}^{in}} \right|}_{\xi _{2} = \xi _{0}} = {\left.
{\varphi ^{out}} \right|}_{\xi _{2} = \xi _{0}} ,\quad {\left.
{\varepsilon {\frac{{\partial \varphi _{2}^{in}}} {{\partial \xi
_{2}}} }} \right|}_{\xi _{2} = \xi _{0}} = {\left.
{{\frac{{\partial \varphi ^{out}}}{{\partial \xi _{2}}} }}
\right|}_{\xi _{2} = \xi _{0}}, \label{eq11}
\end{aligned}$$
where $\xi _{0} = c / \sqrt {c^{2} - a^{2}} = c / f$ are local radial coordinates defining surfaces of the nanospheroids, $\varepsilon$ is permittivity of materials from which the nanoparticles are made. To reduce the boundary conditions (\[eq11\]) to a system of linear equations, we apply the translational addition theorem for wave functions of the prolate nanospheroid (see the Appendix). In the case of two identical coaxial nanospheroids, this theorem gives ($j$, $s = 1$, 2, $j \ne
s$, $\phi _{1} = \phi _{2}$)
$$Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right) = {\sum\limits_{q = m}^{\infty} {S_{mqmn}^{\left( {j}
\right)} P_{q}^{m} \left( {\xi _{s}} \right)P_{q}^{m} \left( {\eta
_{s}} \right)}} , \label{eq12}$$
where the functions $S_{mqmn}^{\left( {1} \right)} =
S_{mqmn} \left( {f,f,l,0} \right)$ and $S_{mqmn}^{\left( {2}
\right)} = S_{mqmn} \left( {f,f,l,\pi} \right)$ are defined in the Appendix. Applying the boundary conditions (\[eq11\]) and the theorem (\[eq12\]), one can obtain the following system of equations ($n = 0$, 1, 2, …; $m = 0$, 1, 2, …, $n$):
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {0} \right)}
\left( { -
1} \right)^{m + q}C_{mq}^{\left( {2} \right)}}} &=& 0, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)\left( { - 1} \right)^{m + n}C_{mn}^{\left( {2} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {0} \right)}
C_{mq}^{\left( {1} \right)}}} &=& 0. \label{eq13}
\end{aligned}$$
When deriving Eq. (\[eq13\]), we make use of the fact that for identical nanospheroids $S_{mnmq}^{\left( {2} \right)} =
\left( { - 1} \right)^{n + q} S_{mnmq}^{\left( {1} \right)} $ (see the Appendix) and take $S_{mnmq}^{\left( {0} \right)} = \left( { -
1} \right)^{m + n} S_{mnmq}^{\left( {1} \right)} $. The system of equations for $D_{mn}^{\left( {j} \right)}$ is identical to (\[eq13\]) and gives no additional information for plasmonic spectra of coaxial spheroids. So, we will not consider it further.
As it results from symmetry of the considered cluster and the system (\[eq13\]), there are two independent types of solutions (plasmonic modes) with opposite parity. To select the first type of the modes (symmetric in $z \to - z$ transformation), one should choose $C_{mn}^{\left( {1} \right)} = \left( { - 1} \right)^{m + n}
C_{mn}^{\left( {2} \right)}$ in Eq. (\[eq13\]). As a result, we shall obtain the following system of equations for the symmetric modes:
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {0} \right)}
C_{mq}^{\left( {1} \right)}}} &=& 0. \label{eq14}
\end{aligned}$$
To obtain the second type of the modes (antisymmetric in $z \to - z$ transformation), one should put $C_{mn}^{\left( {1} \right)} = -
\left( { - 1} \right)^{m + n}C_{mn}^{\left( {2} \right)}$ in Eq. (\[eq13\]). As a result, we shall obtain the following system of equations for the antisymmetric modes:
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
- \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {0} \right)}
C_{mq}^{\left( {1} \right)}}} &=& 0. \label{eq15}
\end{aligned}$$
It is important to notice that separation of spectra into symmetric and antisymmetric plasmon modes is possible only in the case when there is a plane of symmetry. When $m$ is even, antisymmetric modes have nonzero dipole moment and they are “bright” modes. On the contrary, symmetric modes have zero dipole moment and are “dark” modes when $m$ is even. In the case of odd $m$, the “bright” and “dark” modes correspond to the symmetric and antisymmetric modes respectively. One can expect that the antisymmetric mode $m = 0$ will have the largest polarizability and thus will be the “brightest” one for excitation of our cluster with a longitudinally (along the z axis) polarized plane wave.
To study plasmon oscillations in clusters of two prolate spheroidal nanoparticles, we have solved the eigenvalue problems (\[eq14\]) and (\[eq15\]) numerically. In Fig. \[fig2\], normalized plasmon frequency $\omega / \omega _{pl}$ of a cluster of two prolate nanospheroids (see Fig. \[fig1\](a) for the geometry), corresponding to the first three plasmon modes, is shown as a function of normalized distances $l/2c$ between the nanoparticles’ centers. Eigenvalues $\varepsilon $ have been obtained as a nontrivial solution of the equations systems (\[eq14\]) and (\[eq15\]) in the case of an axis-symmetric problem ($m = 0$). Then, the found solutions have been substituted into Eq. (\[eq5\]) to obtain plasmon oscillations frequency.
In Fig. \[fig2\], one can observe that plasmon frequencies of a cluster of two prolate nanospheroids tend to plasmon frequencies of a single nanospheroid (see Fig. \[fig2\](c)) if the distances between the nanospheroids are large enough. When width of the gap between the nanospheroids tends to zero, solutions of the equations (\[eq14\]) and (\[eq15\]) behave very differently. For symmetric modes (Fig. \[fig2\](a)), there are two branches: T-modes and M-modes. Modes of “T” type can be obtained by the method of hybridization of plasmon modes of a single prolate nanospheroid, analogously to the case of a cluster of two spherical nanoparticles [@ref25]. When width of the gap between the nanoparticles is decreasing to zero, normalized plasmonic frequencies of T-modes tend to various values in the range from 0 to $1 / \sqrt {2}$, on the analogy with a two-sphere cluster [@ref20; @ref57]. T-modes with higher indices (not shown for clarity) will concentrate near $\omega
/ \omega _{pl} = 1 / \sqrt {2}$. In Fig. \[fig2\](a), one can also see that at very short distances between the nanospheroids ($l / 2c
< 1.1$), a new type of plasmonic modes (M-modes) appears. M-modes are characterized by strong spatial localization in the gap between the nanoparticles. As a result, they can be effectively excited only by a strongly nonuniform electric field of the molecule or the quantum dot. Values of plasmonic frequencies of these modes lie in the range $\omega _{pl} / \sqrt {2} < \omega < \omega _{pl}$. As the gap width decreases to zero, plasmon frequency of M-modes tends to bulk plasmon frequency $\omega _{pl}$.
In Fig. \[fig2\](b), nontrivial solutions of the equations system (\[eq15\]) for the antisymmetric potential in an axial-symmetric case ($m = 0$), are shown. By analogy with a two-sphere cluster, we will call these modes L-modes (longitudinal) because they are “bright” only for longitudinal excitation. These modes can be described by the hybridization method of plasmon oscillations of single nanospheroids forming the considered cluster. As width of the gap between prolate nanospheroids decreases to zero, normalized plasmon frequencies of these modes tend to zero as it also takes place in the case of spherical nanoparticles [@ref20; @ref57]. Plasmonic frequencies of L-modes of higher orders (not shown) tend to $\omega _{pl} / \sqrt {2}$, and concentration of infinite number of L-modes occurs near this value.
In Fig. \[fig3\], distribution of a surface charge of plasmonic modes of the lowest order in clusters of two identical prolate nanospheroids, is shown. It is seen in this figure that the T- and M-modes have symmetric distribution of the surface charge in contrast to the antisymmetric L=1 mode. This behavior, of course, is in agreement with symmetry of the equations (\[eq14\]) and (\[eq15\]). Another interesting feature is that the surface charge of T-modes is distributed over the surface of the all nanoparticles for any distances between them, while for M- and L-modes it is concentrated near the gap between the nanospheroids if the distance between them is sufficiently small. It is interesting to note also that the surface charge of M-modes is more concentrated in comparison with that of L-modes. Indeed, due to an electroneutrality requirement, the total surface charge on each nanospheroid should be equal to zero. Here, both positive and negative charges of M-modes are localized near the gap between the nanoparticles so that on the rest of nanoparticles the charge is almost equal to zero, as it is well seen in Fig. \[fig3\](a). At the same time, in the case of L-modes for each of nanospheroids near the gap a charge of only one sign is concentrated, and a charge of the opposite sign is distributed with small magnitude over the remaining surface of the nanoparticles. Therefore, strictly speaking, the surface charge in an L-mode is distributed over the whole surface of the cluster nanoparticles though it is not clearly seen at small distances between the nanoparticles (see Fig. \[fig3\](a)). As the distance increases, the charge distribution changes in the cluster: it spreads over the nanoparticles’ surface, tending in the limit to a distribution corresponding to single prolate nanospheroids (see Fig. \[fig3\](c)).
\[subsect2b\]Plasmon oscillations in a cluster of two identical oblate nanospheroids
------------------------------------------------------------------------------------
In this geometry, one should also use local systems of coordinates ($\xi _{j}$, $\eta _{j}$, $\phi _{j}$, $j = 1$, 2) which are connected with each nanospheroid and have origins $o_{j}$ in their centers (see Fig. \[fig1\](b)). Now, the electric potential inside the $j$-th nanospheroid can be presented in the form ($j = 1$, 2):
$$\varphi _{j}^{in} = {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m =
0}^{n} {P_{n}^{m} \left( {i\xi _{j}} \right)P_{n}^{m} \left( {\eta
_{j}} \right)\left( {A_{mn}^{\left( {j} \right)} \cos \left( {m\phi
_{j}} \right) + B_{mn}^{\left( {j} \right)} \sin \left( {m\phi
_{j}} \right)} \right)}}} }, \label{eq16}$$
and the partial potential outside the $j$-th oblate nanospheroid will look like ($j = 1$, 2):
$$\varphi _{j}^{out} = {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m
= 0}^{n} {Q_{n}^{m} \left( {i\xi _{j}} \right)P_{n}^{m} \left( {\eta
_{j}} \right)\left( {C_{mn}^{\left( {j} \right)} \cos \left( {m\phi
_{j}} \right) + D_{mn}^{\left( {j} \right)} \sin \left( {m\phi
_{j}} \right)} \right)}}} }. \label{eq17}$$
The total potential outside the nanospheroids will be expressed by Eq. (\[eq6\]). As boundary conditions for the potential, Eq. (\[eq11\]), where $\xi _{0} = c / \sqrt {a^{2} -
c^{2}} = c / f$, is used. In the case of oblate nanospheroids, the addition translation theorem (see the Appendix) has the following form ($j$, $s = 1$, 2, $j \ne s$):
$$Q_{n}^{m} \left( {i\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m\phi _{j}} \right)} \hfill \\
{\sin \left( {m\phi _{j}} \right)} \hfill \\
\end{array}}} \right.}
= {\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q} {P_{q}^{p}
\left( {i\xi _{s}} \right)P_{q}^{p} \left( {\eta _{s}}
\right){\left\{ {{\begin{array}{*{20}c}
{M_{pqmn}^{\left( {j} \right)} \cos \left( {p\phi _{s}} \right)} \hfill \\
{N_{pqmn}^{\left( {j} \right)} \sin \left( {p\phi _{s}} \right)} \hfill \\
\end{array}}} \right.}}}} } , \label{eq18}$$
where the functions $M_{pqmn}^{\left( {1} \right)} =
M_{pqmn} \left( { - if , - if ,l,\pi} \right)$, $N_{pqmn}^{\left(
{1} \right)} = N_{pqmn} \left( { - if , - if ,l,\pi} \right)$ and $M_{pqmn}^{\left( {2} \right)} = M_{pqmn} \left( { - if, - if,l,0}
\right)$, $N_{pqmn}^{\left( {2} \right)} = N_{pqmn} \left( { - if, -
if,l,0} \right)$ are defined in the Appendix. Now, substituting Eqs. (\[eq16\]) and (\[eq17\]) into Eq. (\[eq11\]) and making use of the addition translation theorem (\[eq18\]), we shall obtain the following system of equations ($n = 0$, 1, 2, …; $m
= 0$, 1, 2, …, $n$):
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {0} \right)} \left( { - 1}
\right)^{p}C_{pq}^{\left( {2}
\right)}}} } } &=& 0, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)\left( { - 1} \right)^{m}C_{mn}^{\left( {2} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {0} \right)} C_{pq}^{\left( {1} \right)}}} } }
&=& 0. \label{eq19}
\end{aligned}$$
By deriving Eq. (\[eq19\]), we take into account that $M_{mnpq}^{\left( {2} \right)} = \left( { - 1} \right)^{m + p}
M_{mnpq}^{\left( {1} \right)}$ (see the Appendix) and denote $M_{mnpq}^{\left( {0} \right)} = \left( { - 1} \right)^{m}
M_{mnpq}^{\left( {1} \right)}$. The system of equations for $D_{mn}^{\left( {j} \right)}$ is analogous to Eq. (\[eq19\]), and we will not analyze it here.
Due to symmetry of a cluster of two identical oblate nanospheroids, there are two types of plasmon oscillation: symmetric and antisymmetric relatively the symmetry plane. To select the first (symmetric in $x \to - x$ transformation) type, we take $C_{mn}^{\left( {1} \right)} = \left( { - 1} \right)^{m}
C_{mn}^{\left( {2} \right)} $. As a result, we shall obtain the system of equations:
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {0} \right)} C_{pq}^{\left( {1} \right)}}} } }
&=& 0. \label{eq20}
\end{aligned}$$
To select the second (antisymmetric in $x \to - x$ transformation) type of plasmon modes in Eq. (\[eq19\]), we take $C_{mn}^{\left(
{1} \right)} = - \left( { - 1} \right)^{m}C_{mn}^{\left( {2}
\right)}$. As a result, we shall have
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {1} \right)} \nonumber \\
- \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {0} \right)} C_{pq}^{\left( {1} \right)}}} } }
&=& 0. \label{eq21}
\end{aligned}$$
In Fig. \[fig4\], the dependence of normalized plasmon frequencies $\omega / \omega _{pl} = 1 / \sqrt {1 - \varepsilon}$ of a cluster of two identical oblate nanospheroids on normalized distances $l/2a$ between the nanoparticles’ centers, is shown for the first three plasmon modes. Eigenvalues $\varepsilon$ were obtained as a solution of the equations systems (\[eq20\]) and (\[eq21\]).
One can see in Fig. \[fig4\] that in clusters of two oblate nanospheroids, modes of “T”, “M”, and “L” types, that are analogous to T-, M- and L-modes of a cluster made of two prolate spheroids (see Fig. \[fig2\]), can exist. The T- and M-modes are the solutions of the system (\[eq20\]), while the L-modes are the solution of the system (\[eq21\]). T- and L-modes can be derived by the method of hybridization of plasmonic modes of two oblate nanospheroids and their plasmonic frequencies are lying in the range $0 < \omega < \omega _{pl} / \sqrt {2}$. An infinite number of plasmonic frequencies of higher L-and T-modes lie near $\omega _{pl}
/ \sqrt {2}$. When width of the gap decreases to 0, the ratio $\omega / \omega _{pl} $ for T-modes tends to various values in the range from 0 to $1 / \sqrt {2} $, while plasmonic frequencies of L-modes approach zero by analogy with L-modes in a cluster of two spherical nanoparticles [@ref20; @ref57]. Plasmonic frequencies of strongly localized M-modes (Fig. \[fig4\](a)) lie in the range $\omega _{pl} / \sqrt {2} < \omega < \omega _{pl}$, as it happens in a cluster of two spherical nanoparticles [@ref20; @ref57]. As width of the gap between oblate nanospheroids decreases to zero, plasmon frequencies of M-modes tend to bulk plasmon frequency $\omega _{pl}$ analogously to the case of a two-sphere cluster [@ref20; @ref57]. For large distances between the spheroids, M-modes disappear, and plasmon frequencies of L- and T-modes of a cluster of two oblate nanospheroids tend to plasmonic frequencies of a single spheroid (see. Fig. \[fig4\](c)) and can be found by means of a self-consistent model with approximation of spheroids by anisotropic point dipoles.
In Fig. \[fig5\], the distribution of a surface charge of plasmon modes of lower orders in a cluster of two identical oblate nanospheroids is shown. One can see in this figure that the charge distribution is symmetric in T- and M-modes, while in L-mode it is antisymmetric in agreement with the definition of these modes. For small distances between nanospheroids, charges in M- and L-modes are strongly localized near the gap. On the contrary, when the distance between the spheroids increases, the charge distribution tends to symmetric or antisymmetric combination of a surface charge in a single oblate nanospheroid (see Fig. \[fig5\](c)).
Thus, in a cluster of two oblate or prolate spheroidal nanoparticles, fundamental symmetric and antisymmetric plasmon modes of “T”, “M”, and “L” types can be excited, and it is these modes that define all optical properties of a two-nanospheroid cluster.
\[sect3\]A cluster of two metal nanospheroids in the field of a plane electromagnetic wave
==========================================================================================
In this section, we will consider a two-spheroid cluster in a uniform electric field with the potential
$$\varphi _{0} = - E_{0x} x - E_{0y} y - E_{0z} z, \label{eq22}$$
where the time factor $e^{ - i\omega t}$ is omitted. This case corresponds to a plane wave incidence and is important for transformation of far fields into near fields, for enhancement of electric fields, and for effective excitation of atoms and molecules. Here again, we restrict ourselves to two special cases of nanospheroids’ shape and position which are most interesting for applications. In the first case, two prolate nanospheroids have a general axis of rotation z, and axes x and y are parallel (Fig. \[fig1\](a)). In the second case, two oblate nanospheroids have parallel axes of rotation z and general axis x, i.e., the centers of the nanoparticles are located in the plane that is perpendicular to the axes of rotation (Fig. \[fig1\](b)). The distance between the nanospheroids’ centers is more than a half-sum of the lengths between their foci (the nanoparticles are not overlapping).
\[sect3a\]A cluster of two prolate nanospheroids
------------------------------------------------
Here, we also will use local systems of spheroidal coordinates, the origins of which are placed in the nanospheroids’ centers (see Fig. \[fig1\](a)). The potential inside the $j$-th nanospheroid again can be presented as a series in spheroidal harmonics (\[eq9\]), while the potential outside the nanospheroids now should be presented in the form:
$$\varphi ^{out} = \varphi _{1}^{out} + \varphi _{2}^{out} + \varphi
_{0} , \label{eq23}$$
where $\varphi _{1}^{out}$, $\varphi _{2}^{out}$ are contributions from the first and second nanospheroids (see Eq. (\[eq10\])), and $\varphi _{0}$ is the potential of the external electric field (\[eq22\]).
Electric potential of the incident plane wave (\[eq22\]) in local coordinates of the $j$-th ($j = 1$, 2) prolate nanospheroids, looks like:
$$\begin{aligned}
\varphi _{0}^{\left( {j} \right)} = &&f\left( {E_{0x} \cos \left( {\phi _{j}
} \right) + E_{0y} \sin \left( {\phi _{j}} \right)}
\right)P_{1}^{1} \left(
{\xi _{j}} \right)P_{1}^{1} \left( {\eta _{j}} \right) \nonumber \\
&&- fE_{0z} P_{1} \left( {\xi _{j}} \right)P_{1} \left( {\eta _{j}} \right)
+ \left( { - 1} \right)^{j + 1}E_{0z} {\frac{{l}}{{2}}},
\label{eq24}
\end{aligned}$$
To find electric potentials and electric fields inside and outside the spheroids, one should use the boundary conditions for the potential:
$$\begin{aligned}
{\left. {\varphi _{1}^{in}} \right|}_{\xi _{1} = \xi _{0}} &=& {\left.
{\left( {\varphi _{1}^{out} + \varphi _{2}^{out} + \varphi
_{0}^{\left( {1} \right)}} \right)} \right|}_{\xi _{1} = \xi _{0}}
,\nonumber \\
{\left. {\varepsilon {\frac{{\partial \varphi _{1}^{in}}} {{\partial
\xi _{1}}} }} \right|}_{\xi _{1} = \xi _{0}} &=& {\left.
{{\frac{{\partial \left( {\varphi _{1}^{out} + \varphi _{2}^{out} +
\varphi _{0}^{\left( {1} \right)}}
\right)}}{{\partial \xi _{1}}} }} \right|}_{\xi _{1} = \xi _{0}} , \nonumber \\
{\left. {\varphi _{2}^{in}} \right|}_{\xi _{2} = \xi _{0}} &=& {\left.
{\left( {\varphi _{1}^{out} + \varphi _{2}^{out} + \varphi
_{0}^{\left( {2} \right)}} \right)} \right|}_{\xi _{2} = \xi _{0}}
, \nonumber \\
{\left. {\varepsilon {\frac{{\partial \varphi _{2}^{in}}} {{\partial
\xi _{2}}} }} \right|}_{\xi _{2} = \xi _{0}} &=& {\left.
{{\frac{{\partial \left( {\varphi _{1}^{out} + \varphi _{2}^{out} +
\varphi _{0}^{\left( {2} \right)}} \right)}}{{\partial \xi _{2}}} }}
\right|}_{\xi _{2} = \xi _{0}} , \label{eq25}\end{aligned}$$
where $\xi _{0} = c / \sqrt {c^{2} - a^{2}} = c / f$ is local radial coordinate that defines surfaces of the nanoparticles, $\varepsilon$ is permittivity of the nanoparticle’s material. Making use of the boundary conditions (\[eq25\]) and the addition translation theorem (\[eq12\]), one can obtain the following system of equations for the coefficients $C_{mn}^{\left( {1}
\right)}$ , $C_{mn}^{\left( {2} \right)}$, $D_{mn}^{\left( {1}
\right)}$ and $D_{mn}^{\left( {2} \right)}$ that define the outside field (see Eq. (\[eq10\]) and Eq. (\[eq23\])) ($n = 0$, 1, 2, …; $m = 0$, 1, 2, …, $n$):
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}}
\right)}}{{d\xi _{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) -
P_{n}^{m} \left( {\xi _{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi
_{0}} \right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {1} \right)} && \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {2} \right)}
C_{mq}^{\left( {2} \right)}}} && = a_{mn}^{\left( {1} \right)}
\left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}}} }, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}}
\right)}}{{d\xi _{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) -
P_{n}^{m} \left( {\xi _{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi
_{0}} \right)}}{{d\xi _{02}}} }}
\right)C_{mn}^{\left( {2} \right)}&& \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {1} \right)}
C_{mq}^{\left( {1} \right)}}} &&= a_{mn}^{\left( {2} \right)} \left(
{\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}}}
}, \nonumber \\ \label{eq26}\end{aligned}$$
and ($n = 1$, 2, 3, …; $m = 1$, 2, 3, …, $n$)
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{01}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)D_{mn}^{\left( {1} \right)} && \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {2} \right)}
D_{mq}^{\left( {2} \right)}}} && = b_{mn}^{\left( {1} \right)}
\left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}}} }, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {\xi _{0}} \right) - P_{n}^{m} \left( {\xi
_{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi
_{0}}} }}
\right)D_{mn}^{\left( {2} \right)} && \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = m}^{\infty} {S_{mnmq}^{\left( {1} \right)}
D_{mq}^{\left( {1} \right)}}} && = b_{mn}^{\left( {2} \right)}
\left( {\varepsilon - 1} \right)P_{n}^{m} \left( {\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {\xi _{0}} \right)}}{{d\xi _{0}}}
}, \nonumber \\ \label{eq27}
\end{aligned}$$
where the coefficients
$$\begin{aligned}
a_{mn}^{\left( {j} \right)} &=& - \delta _{n1} f\left( {\delta _{m1} E_{0x} -
\delta _{m0} E_{0z}} \right) + \left( { - 1} \right)^{j}\delta
_{n0} \delta
_{m0} E_{0z} {\frac{{l}}{{2}}}, \nonumber\\
b_{mn}^{\left( {1} \right)} &=& b_{mn}^{\left( {2} \right)} = - \delta _{m1}
\delta _{n1} fE_{0y} , \label{eq28}
\end{aligned}$$
define excitation field. In Eq. (\[eq28\]), $\delta
_{n1}$ is Kronecker delta symbol. It should be noted here that due to axial symmetry of the considered cluster, the systems of equations (\[eq26\]) and (\[eq27\]) allow one to find the coefficients $C_{mn}^{\left( {j} \right)} $ and $D_{mn}^{\left( {j}
\right)}$ for given order $m$, while degree $n$ runs over $n=m$, $m
+ 1$, $m + 2$, …, $m + N$, where $N$ is a large number that defines accuracy of the solution.
Induced dipole moment of a cluster of two prolate nanospheroids, placed in the field of a plane electromagnetic wave, can be calculated by analogy with a single prolate nanospheroid [@ref31], that is, by finding far field asymptotes of the potential (\[eq23\]). As a result, for dipole moment induced in the $j$-th nanospheroid ($j = 1$, 2) we have
$$d_{x}^{\left( {j} \right)} = {\frac{{2f_{}^{2}}}
{{3}}}C_{11}^{\left( {j} \right)} ,\quad d_{y}^{\left( {j} \right)}
= {\frac{{2f_{}^{2} }}{{3}}}D_{11}^{\left( {j} \right)} ,\quad
d_{z}^{\left( {j} \right)} = {\frac{{f_{}^{2}}} {{3}}}C_{01}^{\left(
{j} \right)} , \label{eq29}$$
and the total dipole moment of the cluster will be the sum of the momenta (\[eq29\]). Scattering and absorption cross-sections can be easily found if the dipole momenta (\[eq29\]) are known [@ref58]:
$$\sigma ^{abs} = 4\pi \left( {{\frac{{\omega}} {{v_{c}}} }}
\right){\frac{{ \textrm{Im} \left( {{\mathbf {d}}{\mathbf {E}}_{0}
^{
*}} \right)}}{{{\left| {{\mathbf {E}}_{0}} \right|}^{2}}}},\quad
\sigma ^{scat} = {\frac{{8\pi}} {{3}}}\left( {{\frac{{\omega}}
{{v_{c}}} }} \right)^{4}{\frac{{{\left| {{\mathbf {d}}}
\right|}^{2}}}{{{\left| {{\mathbf {E}}_{0}} \right|}^{2}}}},
\label{eq30}$$
where ${\mathbf {d}} = {\mathbf {d}}^{\left( {1} \right)}
+ {\mathbf {d}}^{\left( {2} \right)}$ denotes dipole momentum of the whole system, and asterix denotes an operation of the complex conjugation.
In Fig. \[fig6\], absorption (a) and scattering (b) cross-sections of a cluster of two identical prolate nanospheroids made from silver, are shown as a function of the wavelength. For longitudinal (z) polarization, both of the cross-sections have two peaks that correspond to longitudinal plasmonic oscillations with L = 1, 2 (see Fig. \[fig2\](b)). It is very important that both of the peaks are split substantially relatively the case of a single spheroid (the “z” dashed curve) due to strong interaction between the nanospheroids.
On the contrary, for transversal (x or y) polarization one can see only one peak due to excitation of the symmetrical T = 1 mode, and this peak is shifted just slightly relatively the single spheroid resonance (the “y” dashed curve). It means that transversal (x or y polarization) excitation of a two-spheroid cluster induces only weak interaction between the nanospheroids (see the dispersion curves for T-modes on Fig. \[fig2\](a)). Due to this weak interaction, absorption and scattering cross-sections are approximately equal to doubled and quadrupled cross-section of a single spheroid respectively.
It should be noticed that in Fig. \[fig6\] the maxima of absorption, corresponding to plasmon oscillations of M-type that should lie in the interval $\omega _{pl} / \sqrt {2} < \omega <
\omega _{pl} $, that corresponds to $326 < \lambda < 337$ nm for silver [@ref49], are not visible. It is related to the fact that M-modes interact with a homogeneous electric field weakly and can be effectively excited only by a source of radiation that is nonuniform in comparison with a size of the gap between the nanoparticles [@ref57] (see Fig. \[fig3\]).
To control correctness and accuracy of our analytical calculations, we have also carried out finite element simulation of this system with Comsol Multiphysics software. The results of this simulation are shown by the circles in Fig. \[fig6\]. One can see that there is a fine agreement between the analytical and pure numerical calculations. This fact confirms correctness and accuracy of the both of approaches.
\[sect3b\]A cluster of two oblate nanospheroids
-----------------------------------------------
The case of two oblate nanospheroids is in many aspects similar to the case of two prolate nanospheroids, considered above. So let us again choose local systems of coordinates that are connected with each of the nanospheroids and have origins in their centers (see Fig. \[fig1\](b)). The potential inside the $j$-th nanospheroid again can be presented as a series in spheroidal harmonics (\[eq16\]), while the potential outside the oblate spheroids can be presented in the form:
$$\varphi ^{out} = \varphi _{1}^{out} + \varphi _{2}^{out} + \varphi
_{0} , \label{eq31}$$
where $\varphi _{1}^{out} ,\varphi _{2}^{out} $ are contributions from the first and second nanospheroids (see Eq. (\[eq17\])) and $\varphi _{0}$ is the potential of the external electric field (\[eq22\]). In the local coordinates of the $j$-th ($j = 1$, 2) oblate nanospheroids, it looks like
$$\begin{aligned}
\varphi _{0}^{\left( {j} \right)} = &&- if\left( {E_{0x} \cos \left( {\phi
_{j}} \right) + E_{0y} \sin \left( {\phi _{j}} \right)}
\right)P_{1}^{1}
\left( {i\xi _{j}} \right)P_{1}^{1} \left( {\eta _{j}} \right) \nonumber \\
&&+ ifE_{0z} P_{1} \left( {i\xi _{j}} \right)P_{1} \left( {\eta _{j}}
\right) + \left( { - 1} \right)^{j}E_{0x} {\frac{{l}}{{2}}} .
\label{eq32}
\end{aligned}$$
Making use of the boundary conditions (\[eq25\]) with $\xi _{0} =
c / \sqrt {a^{2} - c^{2}} = c / f $ and the addition-translation theorem (\[eq18\]), we shall obtain the following systems of equations ($n = 0$, 1, 2, …; $m = 0$, 1, 2, …, $n$):
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {1} \right)}&& \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {2} \right)} C_{pq}^{\left( {2} \right)}}} } }&& =
a_{mn}^{\left( {1} \right)} \left( {\varepsilon - 1}
\right)P_{n}^{m} \left( {i\xi _{0}} \right){\frac{{dP_{n}^{m}
\left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)C_{mn}^{\left( {2} \right)}&& \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p = 0}^{q}
{M_{mnpq}^{\left( {1} \right)} C_{pq}^{\left( {1} \right)}}} } } &&=
a_{mn}^{\left( {2} \right)} \left( {\varepsilon - 1}
\right)P_{n}^{m} \left( {i\xi _{0}} \right){\frac{{dP_{n}^{m}
\left( {i\xi _{0}} \right)}}{{d\xi _{0}}} }, \nonumber \\
\label{eq33}
\end{aligned}$$
and ($n = 1$, 2, 3, …; $m = 1$, 2, 3, …, $n$)
$$\begin{aligned}
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)D_{mn}^{\left( {1} \right)}&& \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 1}^{\infty} {{\sum\limits_{p = 1}^{q}
{N_{mnpq}^{\left( {2} \right)} D_{pq}^{\left( {2} \right)}}} } } &&=
b_{mn}^{\left( {1} \right)} \left( {\varepsilon - 1}
\right)P_{n}^{m} \left( {i\xi _{0}} \right){\frac{{dP_{n}^{m}
\left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }, \nonumber \\
\left( {\varepsilon {\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi
_{0}}} }Q_{n}^{m} \left( {i\xi _{0}} \right) - P_{n}^{m} \left(
{i\xi _{0} } \right){\frac{{dQ_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0}}} }}
\right)D_{mn}^{\left( {2} \right)}&& \nonumber \\
+ \left( {\varepsilon - 1} \right)P_{n}^{m} \left( {i\xi _{0}}
\right){\frac{{dP_{n}^{m} \left( {i\xi _{0}} \right)}}{{d\xi _{0}
}}}{\sum\limits_{q = 1}^{\infty} {{\sum\limits_{p = 1}^{q}
{N_{mnpq}^{\left( {1} \right)} D_{pq}^{\left( {1} \right)}}} } } &&=
b_{mn}^{\left( {2} \right)} \left( {\varepsilon - 1}
\right)P_{n}^{m} \left( {i\xi _{0}} \right){\frac{{dP_{n}^{m}
\left( {i\xi _{0}} \right)}}{{d\xi _{0}}} }, \nonumber
\\ \label{eq34}
\end{aligned}$$
where
$$\begin{aligned}
a_{mn}^{\left( {j} \right)} &=& i\delta _{n1} f \left( {\delta _{m1}
E_{0x} - \delta _{m0} E_{0z}} \right) + \left( { - 1} \right)^{j +
1}\delta
_{m0} \delta _{n0} E_{0x} {\frac{{l}}{{2}}}, \nonumber \\
b_{mn}^{\left( {1} \right)} &=& b_{mn}^{\left( {2} \right)} = i\delta _{m1}
\delta _{n1} f E_{0y} . \label{eq35}
\end{aligned}$$
Apparently, the equations (\[eq33\]) and (\[eq34\]) have more complicated structure than Eqs. (\[eq26\]) and (\[eq27\]) because now, due to lack of axial symmetry, one can not split the system of equations into systems with fixed order $m$ of the Legendre function.
For calculation of absorption and scattering cross-sections of a cluster in the field of a plane electromagnetic wave, one can again use Eq. (\[eq30\]), where dipole momenta of each spheroid can be expressed through the solutions of Eqs. (\[eq33\]) and (\[eq34\]) ($j = 1$, 2)
$$d_{x}^{\left( {j} \right)} = - {\frac{{2f_{}^{2}}}
{{3}}}C_{11}^{\left( {j} \right)} ,\quad d_{y}^{\left( {j} \right)}
= - {\frac{{2f_{}^{2} }}{{3}}}D_{11}^{\left( {j} \right)} ,\quad
d_{z}^{\left( {j} \right)} = - {\frac{{f_{}^{2}}}
{{3}}}C_{01}^{\left( {j} \right)} . \label{eq36}$$
In Fig. \[fig7\], the absorption (a) and scattering (b) cross-sections of a cluster of two identical oblate nanospheroids made from silver are shown as a function of wavelength. For longitudinal (x) polarization, both the cross-sections have two peaks, that correspond to antisymmetric plasmonic oscillations with L = 1, 2 (see Fig. \[fig4\](b)). It is very important that now only one peak (L = 1) is shifted substantially relatively the case of a single spheroid (the “x, y” dashed curves) due to strong interaction between the nanospheroids. The L = 2 mode suffers only a small shift in agreement with Fig. \[fig4\](b).
For transversal (y) polarization, one can see only one peak due to excitation of the symmetrical T =1 mode, and this peak is only slightly shifted relatively the single spheroid resonance (the dashed curve). It means that transversal (y polarization) excitation of a two-spheroid cluster results only in weak interaction between the nanospheroids (see the dispersion curves for the T-modes on Fig. \[fig4\](a)). Due to this weak interaction, absorption and scattering cross-sections for this polarization are approximately equal to doubled and quadrupled cross-sections of a single spheroid respectively. It is also interesting that plasmonic frequency of the L = 2 mode is very close to plasmonic frequency of the T = 1 mode. This fact can be easily understood from analysis of Fig. \[fig4\]. Indeed, when width of the gap tends to zero, plasmonic frequency of L = 2 modes also decreases to zero, while plasmonic frequency of the T = 1 mode increases slightly. So, at some point these modes will intersect and have the same frequencies, and we observe this situation in Fig. \[fig7\].
It should be noticed that in Fig. \[fig7\] the maxima of absorption corresponding to plasmon oscillations of M-type again are not visible. It is related to the fact that M-modes interact with a homogeneous electric field weakly and can be effectively excited only by a source of radiation that is nonuniform in comparison with a size of the gap between nanoparticles [@ref57] (see Fig. \[fig3\]).
To control correctness and accuracy of our analytical calculations for a cluster of two oblate spheroids, we have also carried out finite element simulation of this system with Comsol Multiphysics software. The results of this simulation are shown by the circles in Fig. \[fig7\]. One can see that there is a fine agreement between the analytical and pure numerical calculations. This fact confirms correctness and accuracy of the both of approaches again.
\[sect3c\]Enhancement of local fields
-------------------------------------
The most important characteristic of nanoparticles clusters is the factor of incident field enhancement in the gap between the nanoparticles. It is the characteristic that allows determining the excitation rate of molecules near the nanoparticles or intensity of surface-enhanced Raman scattering [@ref4]. Moreover, achievement of high values of this factor is the main goal of optical nano-antennas development.
Distribution of squared electric field for the L=1 resonance in a cluster of two prolate spheroids is shown in Fig. \[fig8\], that shows that, indeed, the maximal field enhancement takes place in the gap between the nanoparticles on their surfaces. The field maxima are also present on the outer side of the cluster; however, field amplitude is essentially less there. According to general theorems for harmonic functions, the field maximum can be reached only on the region boundaries. In our case, the field maxima are reached in those points of the spheroids’ surface where the distance between the spheroids is minimal.
Using Eqs. (\[eq10\]) and (\[eq17\]), one can find explicit expressions for the field enhancement factor $G$. For clusters of two identitcal prolate spheroids in the considered configuration (Fig. \[fig1\](a)), one can obtain the following expression for the field maximum in the case of an incident field polarized along the z-axis:
$$G = {\frac{{{\left| {{\mathbf {E}}} \right|}^{2}}}{{{\left| {E_{0z}}
\right|}^{2}}}} = {\left| {1 - {\frac{{1}}{{E_{0z}
f}}}{\sum\limits_{n = 1}^{\infty} {C_{0n} \left( {{\frac{{dQ_{n}
\left( {\xi _{1}} \right)}}{{d\xi _{1}}} } + {\frac{{dQ_{n} \left(
{\xi _{2}} \right)}}{{d\xi _{2}}} }} \right)}}} \right|}^{2},
\label{eq37}$$
where $\xi _{1} = c / f$ and $\xi _{2} = \left( {l - c}
\right) / f$, while $C_{0n} = C_{0n}^{\left( {1} \right)} = - \left(
{ - 1} \right)^{n}C_{0n}^{\left( {2} \right)} $. In the most interesting case of a small gap and strongly prolate spheroids, $\xi
_{1}$, $\xi _{2} \approx 1$, and one may use the asymptotic form ${\left. {{\frac{{dQ_{n} \left( {\xi} \right)}}{{d\xi}} }}
\right|}_{\xi \approx 1} \approx - {\frac{{1}}{{2\left( {\xi - 1}
\right)}}}$. As a result, the field enhancement factor takes the form:
$$G = {\frac{{{\left| {{\mathbf {E}}} \right|}^{2}}}{{{\left| {E_{0z}}
\right|}^{2}}}} \approx {\left| {{\frac{{1}}{{2E_{0z} f}}}\left(
{{\frac{{1}}{{\xi _{1} - 1}}} + {\frac{{1}}{{\xi _{2} - 1}}}}
\right){\sum\limits_{n = 1}^{\infty} {C_{0n}}} } \right|}^{2}.
\label{eq38}$$
In the case of clusters of two identical oblate spheroids (Fig. \[fig1\](b)) and incident field polarized along the x-axis, we obtain
$$\begin{aligned}
G = {\frac{{{\left| {{\mathbf {E}}} \right|}^{2}}}{{{\left| {E_{0x}}
\right|}^{2}}}} = {\left| {{\frac{{1}}{{E_{0x} f}}}{\sum\limits_{n =
1}^{\infty} {{\sum\limits_{m = 0}^{n} {C_{mn} {\left.
{{\frac{{dP_{n}^{m} \left( {\eta} \right)}}{{d\eta}} }}
\right|}_{\eta = 0} \left( {{\frac{{1}}{{\xi _{1}}} }Q_{n}^{m}
\left( {i\xi _{1}} \right) - {\frac{{1}}{{\xi _{2}}} }Q_{n}^{m}
\left( {i\xi _{2}} \right)} \right)}}}
}} \right|}^{2} \nonumber \\
+ {\left| {1 - {\frac{{1}}{{E_{0x} f}}}{\sum\limits_{n = 1}^{\infty}
{{\sum\limits_{m = 0}^{n} {C_{mn} P_{n}^{m} \left( {0} \right)\left(
{\sqrt {1 + {\frac{{1}}{{\xi _{1}^{2}}} }} {\frac{{dQ_{n}^{m} \left(
{i\xi _{1}} \right)}}{{d\xi _{1}}} } + \sqrt {1 + {\frac{{1}}{{\xi
_{2}^{2}}} }} {\frac{{dQ_{n}^{m} \left( {i\xi _{2}} \right)}}{{d\xi
_{2}}} }} \right)}}} }} \right|}^{2}, \label{eq39}
\end{aligned}$$
where $\xi _{1} = \sqrt {a^{2} / f^{2} - 1} $ è $\xi _{2}
= \sqrt {\left( {l - a} \right)^{2} / f^{2} - 1} $; $C_{mn} = -
\left( { - 1} \right)^{m}C_{mn}^{\left( {1} \right)} =
C_{mn}^{\left( {2} \right)}$.
In Fig. \[fig9\], the dependence of squared electric field enhancement (\[eq37\]) and (\[eq39\]) for clusters of two identical silver nanospheroids on the wavelength, is shown. Comparing peaks positions to the dispersion curves in Fig. \[fig2\] and Fig. \[fig4\], one can come to a conclusion that only “L” types of plasmon modes are excited in the clusters for the considered configurations of nanospheroids and incident electromagnetic wave polarizations (along the line joining the nanoparticles’ centers). In particular, excitation of the L = 1 and L = 2 modes is noticeable. At that, the position of squared field enhancement peaks agrees with the maxima of the absorption and scattering cross-sections shown in Fig. \[fig6\] and \[fig7\] by the solid lines z and x correspondingly. It should be mentioned that the value of squared field enhancement near a cluster of two nanospheroids can reach up to $10^{6}$. In the case of single nanoparticles, this value is almost two orders less than that of clusters (cf. the solid and dashed curves in Fig. \[fig9\]). This fact determines a greater attractiveness of metal nanoparticles clusters in comparison to single nanoparticles for investigation of Surface Enhance Raman Scattering (SERS) and Surface-Enhanced Fluorescence (SEF). Let us note that the obtained great field enhancement factors can be slightly less in practice since for small particles and for small gaps between them, non-local and other effects not considered in this research become essential.
\[sect4\]A cluster of two nanospheroids in the field of a dipole source
=======================================================================
In the previous section, we have considered the case of a nano-antenna placed in the field of a plane wave. However, highly nonuniform optical fields occur very often in nano-environment. For example, such fields arise when a plasmonic nano-antenna is excited by an atom, or a molecule, or another nanolocalized source of light. So, in this section, we will consider an important case of a two-nanospheroid cluster in the field of electric dipole sources. Excitation of the cluster by magnetic dipole and electric quadrupole sources can be analyzed analogously.
\[sect4a\]A cluster of two prolate nanospheroids
------------------------------------------------
A case of two prolate nanospheroids in the field of a dipole source of radiation can be considered by perfect analogy with a case of the same cluster in a uniform field. One should again look for solution in the form (\[eq9\]), (\[eq10\]) and then apply the boundary conditions (\[eq25\]). The only difference is that now the external potential is the potential $\varphi _{0}$ of the dipole that has the following form in the $j$-th local system of coordinates of a prolate spheroid [@ref55] ($j = 1$, 2):
$$\begin{aligned}
&&\varphi _{0}^{\left( {j} \right)} = \left( {{\mathbf {d}}_{0} {\nabla
}'_{j}} \right){\frac{{1}}{{{\left| {{\mathbf {r}} - {\mathbf
{{r}'}}}
\right|}}}} \nonumber \\
&&= {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m = 0}^{n} {{\left\{
{{\begin{array}{*{20}c}
{P_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left[ {\left( {{\mathbf {d}}_{0} {\nabla} '_{j}}
\right)\alpha _{mn}^{\left( {j} \right)} \cos \left( {m\phi _{j}}
\right) + \left( {{\mathbf {d}}_{0} {\nabla} '_{j}} \right)\beta
_{mn}^{\left( {j} \right)} \sin \left( {m\phi _{j}} \right)}
\right]},} \hfill & {\xi _{j} < {\xi
}'_{j} ,} \hfill \\
{Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left[ {\left( {{\mathbf {d}}_{0} {\nabla} '_{j}}
\right)\gamma _{mn}^{\left( {j} \right)} \cos \left( {m\phi _{j}}
\right) + \left( {{\mathbf {d}}_{0} {\nabla} '_{j}} \right)\sigma
_{mn}^{\left( {j} \right)} \sin \left( {m\phi _{j}} \right)}
\right]},} \hfill & {\xi _{j} > {\xi
}'_{j} .} \hfill \\
\end{array}}} \right.}}}} } \nonumber \\ \label{eq40}\end{aligned}$$
In Eq. (\[eq40\]), ${\mathbf {d}}_{0} $ denotes the dipole momentum of a source placed at ${\mathbf {{r}'}}$, ${\nabla}
'_{j}$ is a gradient over ${\mathbf {{r}'}}$ in local coordinates, and
$$\begin{aligned}
{\left. {{\begin{array}{*{20}c}
{\alpha _{mn}^{\left( {j} \right)}} \hfill \\
{\beta _{mn}^{\left( {j} \right)}} \hfill \\
\end{array}}} \right\}} = {\frac{{1}}{{f_{}}} }\left( {2 - \delta _{m0}}
\right)\left( { - 1} \right)^{m}\left( {2n + 1} \right){\left[
{{\frac{{\left( {n - m} \right)!}}{{\left( {n + m} \right)!}}}}
\right]}^{2}Q_{n}^{m} \left( {{\xi} '_{j}} \right)P_{n}^{m} \left(
{{\eta }'_{j}} \right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m{\phi} '_{j}} \right)} \hfill \\
{\sin \left( {m{\phi} '_{j}} \right)} \hfill \\
\end{array}}} \right.}, \nonumber \\
{\left. {{\begin{array}{*{20}c}
{\gamma _{mn}^{\left( {j} \right)}} \hfill \\
{\sigma _{mn}^{\left( {j} \right)}} \hfill \\
\end{array}}} \right\}} = {\frac{{1}}{{f_{}}} }\left( {2 - \delta _{m0}}
\right)\left( { - 1} \right)^{m}\left( {2n + 1} \right){\left[
{{\frac{{\left( {n - m} \right)!}}{{\left( {n + m} \right)!}}}}
\right]}^{2}P_{n}^{m} \left( {{\xi} '_{j}} \right)P_{n}^{m} \left(
{{\eta }'_{j}} \right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m{\phi} '_{j}} \right)} \hfill \\
{\sin \left( {m{\phi} '_{j}} \right)} \hfill \\
\end{array}}} \right.}, \label{eq41}
\end{aligned}$$
are expansion coefficients of the unit charge potential in local coordinates of a prolate spheroid.
As a result of applying of the boundary conditions, one can obtain a system of equations for the unknown coefficients $C_{mn}^{\left( {j}
\right)} $, $D_{mn}^{\left( {j} \right)}$ in Eq. (\[eq10\]). The new system can be easily derived from Eqs. (\[eq26\]) and (\[eq27\]) if one makes the following replacement for the coefficients $a_{mn}^{\left( {j} \right)} $ and $b_{mn}^{\left( {j}
\right)}$:
$$\begin{aligned}
&&a_{mn}^{\left( {1} \right)} = - \left( {{\mathbf {d}}_{0} {\nabla} '_{1}}
\right)\alpha _{mn}^{\left( {1} \right)} ,\quad a_{mn}^{\left( {2}
\right)} = - \left( {{\mathbf {d}}_{0} {\nabla} '_{2}} \right)\alpha
_{mn}^{\left(
{2} \right)} , \nonumber \\
&&b_{mn}^{\left( {1} \right)} = - \left( {{\mathbf {d}}_{0} {\nabla} '_{1}}
\right)\beta _{mn}^{\left( {1} \right)} ,\quad b_{mn}^{\left( {2}
\right)} = - \left( {{\mathbf {d}}_{0} {\nabla} '_{2}} \right)\beta
_{mn}^{\left( {2} \right)} , \label{eq42}
\end{aligned}$$
where $\alpha _{mn}^{\left( {j} \right)} $ and $\beta
_{mn}^{\left( {j} \right)}$ are defined by Eq. (\[eq41\]).
After Eqs. (\[eq26\]) and (\[eq27\]) have been solved with taking (\[eq42\]) into account, one can find the total induced dipole moment of the prolate nanospheroids
$$\begin{aligned}
d_{x} &=& {\frac{{2f_{}^{2}}} {{3}}}\left( {C_{11}^{\left( {1} \right)} +
C_{11}^{\left( {2} \right)}} \right), \nonumber \\
d_{y} &=& {\frac{{2f_{}^{2}}} {{3}}}\left( {D_{11}^{\left( {1} \right)} +
D_{11}^{\left( {2} \right)}} \right), \nonumber \\
d_{z} &=& {\frac{{f_{}^{2}}} {{3}}}\left( {C_{01}^{\left( {1} \right)} +
C_{01}^{\left( {2} \right)}} \right). \label{eq43}
\end{aligned}$$
Knowing the dipole momenta (\[eq43\]), it is easy to find radiative decay rate of a dipole placed near the prolate nanospheroid [@ref59]:
$$\gamma = {\frac{{P_{}^{rad}}} {{\hbar \omega}} } = {\frac{{\omega
^{3}}}{{3\hbar v_{c}^{3}}} }{\left| {{\mathbf {d}}_{0} + {\mathbf
{d}}} \right|}^{2}. \label{eq44}$$
where $P^{rad}$ is radiation power at frequency $\omega$ and $\hbar \omega $ is emitted photon energy.
Radiative decay rate is a very important characteristics in such applications as surface enhanced Raman scattering, surface-enhanced fluorescence, nanolasers, and so on. To characterize the radiative decay rate, it is naturally to normalize it to radiative decay rate of a dipole in free space, $\gamma _{0} = {\frac{{P_{0}^{rad}}}
{{\hbar \omega}} } = {\frac{{\omega ^{3}}}{{3\hbar v_{c}^{3}}}
}{\left| {{\mathbf {d}}_{0}} \right|}^{2}$.
In Fig. \[fig10\], normalized radiative decay rate of a dipole source placed at the middle point of the gap is shown. As it is well seen in Fig. \[fig10\](a), if the distance between the prolate nanospheroids is small (Fig. \[fig10\](a), the curves $\alpha$ and $\delta $), the dipole source with a moment oriented perpendicular to the cluster’s axis of rotation, can excite both symmetrical T- and M-modes. This fact differs from the case of excitation of the same cluster with a plane wave, when M-modes with peaks located in the region of $\lambda < 337$ nm (see Fig. \[fig2\](a)) are not excited. When the distance between the nanospheroids increases (see Fig. \[fig10\](a)), the peak corresponding to M-modes shifts to $\lambda \approx 337$ nm ($\omega \approx \omega _{pl} / \sqrt {2}
$) and then disappear. After that point, only the peaks corresponding to plasmonic T-modes can be observed. Of course, this picture is in agreement with the behavior of the plasmonic M-modes shown in Fig. \[fig2\](a). We shall also notice that for large enough distances between the nanospheroids (see Fig. \[fig10\](a), the curve $\gamma $) the self-consistent model [@ref3; @ref21], in which nanoparticles are replaced by point dipoles with corresponding polarizabilities [@ref3; @ref34], can be effectively used for calculation of radiative decay rate (the dashed curve).
When dipole moment of a source is oriented along the axis of symmetry (Fig. \[fig10\](b)), only antisymmetric L-modes can be excited due to symmetry reasons. From Fig. \[fig10\](b), one can also see that for small enough distances between the spheroids there are two plasmonic modes (L = 1, 2) that interact with the dipole source. When the distance between the spheroids diminishes, further peaks of radiation power shift towards long wavelengths. At large distances between the nanospheroids, there is only one maximum corresponding to the L = 1 plasmonic mode (see Fig. \[fig10\](b), the curve $\gamma $). In this case, the radiative decay rate of a dipole placed near a cluster of two prolate nanospheroids can be also calculated with making use of the self-consistent analytical model in which the spheroids are approximated by point dipoles (see the dashed curve in Fig. \[fig10\](b)).
\[sect4b\]A cluster of two oblate nanospheroids
-----------------------------------------------
A case of two oblate nanospheroids in the field of a dipole source of radiation can be considered by perfect analogy with a case of the same cluster in a uniform field. One should again look for solution in the forms (\[eq16\]) and (\[eq17\]) and then apply the boundary conditions (\[eq25\]). The only difference is that now the external potential is the potential $\varphi _{0}$ of the dipole that has the following form in the $j$-th local system of coordinates of an oblate spheroid [@ref55] ($j = 1$, 2):
$$\begin{aligned}
&&\varphi _{0}^{\left( {j} \right)} = \left( {{\mathbf {d}}_{0} {\nabla
}'_{j}} \right){\frac{{1}}{{{\left| {{\mathbf {r}} - {\mathbf
{{r}'}}}
\right|}}}} \nonumber \\
&&= {\sum\limits_{n = 0}^{\infty} {{\sum\limits_{m = 0}^{n} {{\left\{
{{\begin{array}{*{20}c}
{P_{n}^{m} \left( {i\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left[ {\left( {{\mathbf {d}}_{0} {\nabla} '_{j}}
\right)\alpha _{mn}^{\left( {j} \right)} \cos \left( {m\phi _{j}}
\right) + \left( {{\mathbf {d}}_{0} {\nabla} '_{j}} \right)\beta
_{mn}^{\left( {j} \right)} \sin \left( {m\phi _{j}} \right)}
\right]},} \hfill & {\xi _{j} < {\xi
}'_{j} ,} \hfill \\
{Q_{n}^{m} \left( {i\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left[ {\left( {{\mathbf {d}}_{0} {\nabla} '_{j}}
\right)\gamma _{mn}^{\left( {j} \right)} \cos \left( {m\phi _{j}}
\right) + \left( {{\mathbf {d}}_{0} {\nabla} '_{j}} \right)\sigma
_{mn}^{\left( {j} \right)} \sin \left( {m\phi _{j}} \right)}
\right]},} \hfill & {\xi _{j} > {\xi
}'_{j} .} \hfill \\
\end{array}}} \right.}}}} } \nonumber \\ \label{eq45}
\end{aligned}$$
In Eq. (\[eq45\]), ${\mathbf {d}}_{0}$ denotes the dipole momentum of a source placed at ${\mathbf {{r}'}}$, ${\nabla}
'_{j}$ is gradient over ${\mathbf {{r}'}}$ in local coordinates, and
$$\begin{aligned}
{\left. {{\begin{array}{*{20}c}
{\alpha _{mn}^{\left( {j} \right)}} \hfill \\
{\beta _{mn}^{\left( {j} \right)}} \hfill \\
\end{array}}} \right\}} = {\frac{{i}}{{f_{}}} }\left( {2 - \delta _{m0}}
\right)\left( { - 1} \right)^{m}\left( {2n + 1} \right){\left[
{{\frac{{\left( {n - m} \right)!}}{{\left( {n + m} \right)!}}}}
\right]}^{2}Q_{n}^{m} \left( {i{\xi} '_{j}} \right)P_{n}^{m} \left(
{{\eta }'_{j}} \right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m{\phi} '_{j}} \right)} \hfill \\
{\sin \left( {m{\phi} '_{j}} \right)} \hfill \\
\end{array}}} \right.}, \nonumber \\
{\left. {{\begin{array}{*{20}c}
{\gamma _{mn}^{\left( {j} \right)}} \hfill \\
{\sigma _{mn}^{\left( {j} \right)}} \hfill \\
\end{array}}} \right\}} = {\frac{{i}}{{f_{}}} }\left( {2 - \delta _{m0}}
\right)\left( { - 1} \right)^{m}\left( {2n + 1} \right){\left[
{{\frac{{\left( {n - m} \right)!}}{{\left( {n + m} \right)!}}}}
\right]}^{2}P_{n}^{m} \left( {i{\xi} '_{j}} \right)P_{n}^{m} \left(
{{\eta }'_{j}} \right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m{\phi} '_{j}} \right)} \hfill \\
{\sin \left( {m{\phi} '_{j}} \right)} \hfill \\
\end{array}}} \right.}, \label{eq46}
\end{aligned}$$
are expansion coefficients of the unit charge potential in local coordinates of an oblate spheroid.
As a result of applying of the boundary conditions, one can obtain the system of equations for the unknown coefficients $C_{mn}^{\left(
{j} \right)} $, $D_{mn}^{\left( {j} \right)}$ in Eq. (\[eq17\]). The system of equations for these coefficients can be easily derived from Eqs. (\[eq33\]) and (\[eq34\]) if one makes the following replacement for the coefficients $a_{mn}^{\left( {j} \right)} $ and $b_{mn}^{\left( {j} \right)} $
$$\begin{aligned}
&&a_{mn}^{\left( {1} \right)} = - i\left( {{\mathbf {d}}_{0} {\nabla} '_{1}}
\right)\alpha _{mn}^{\left( {1} \right)} ,\quad a_{mn}^{\left( {2}
\right)} = - i\left( {{\mathbf {d}}_{0} {\nabla} '_{21}}
\right)\alpha _{mn}^{\left(
{2} \right)} ,\nonumber \\
&&b_{mn}^{\left( {1} \right)} = - i\left( {{\mathbf {d}}_{0} {\nabla} '_{1}}
\right)\beta _{mn}^{\left( {1} \right)} ,\quad b_{mn}^{\left( {2}
\right)} = - i\left( {{\mathbf {d}}_{0} {\nabla} '_{2}} \right)\beta
_{mn}^{\left( {2} \right)} , \label{eq47}
\end{aligned}$$
where $\alpha _{mn}^{\left( {j} \right)} $ and $\beta
_{mn}^{\left( {j} \right)} $ are defined in Eq. (\[eq46\]).
The total induced dipole moment of a cluster of two oblate nanospheroids can be found by using the following expressions (cf. Eq. (\[eq43\])):
$$\begin{aligned}
d_{x} &=& - {\frac{{2f_{}^{2}}} {{3}}}\left( {C_{11}^{\left( {1} \right)} +
C_{11}^{\left( {2} \right)}} \right), \nonumber \\
d_{y} &=& - {\frac{{2f_{}^{2}}} {{3}}}\left( {D_{11}^{\left( {1} \right)} +
D_{11}^{\left( {2} \right)}} \right), \nonumber \\
d_{z} &=& - {\frac{{f_{}^{2}}} {{3}}}\left( {C_{01}^{\left( {1} \right)} +
C_{01}^{\left( {2} \right)}} \right). \label{eq48}
\end{aligned}$$
In Fig. \[fig11\], normalized radiative decay rate of a dipole placed at the middle point between the oblate nanospheroids is shown. As one can see in Fig. \[fig11\](a), if dipole moment of a source is oriented along the line connecting the oblate nanospheroids’ centers, only plasmonic L-modes are excited as it took place for the case of a cluster of two prolate nanospheroids (cf. Fig. \[fig10\](b)). When the distance between the nanospheroids is large enough (the $\gamma$ curve), the properties of radiative decay of a dipole source located near the cluster can be more and more precisely approximated by means of the self-consistent model where spheroids are modeled by point dipoles (see Fig. \[fig11\](a), the dashed curve).
From Fig. \[fig11\](b) and (c) that correspond to the case of dipole moment of a source oriented along the y and z axes, one can conclude that only symmetric T-modes are presented here. Nevertheless, at small enough distances between the nanospheroids one can expect that M-modes are also excited but can not be seen due to large losses (large imaginary part of dielectric permittivity) in silver spheroids. If the imaginary part of permittivity of spheroidal nanoparticles’ material is small enough, for example, in a case of silicized carbon (SiC) [@ref60], all fine features of the spectra and M-modes in particular will be well visible as it was demonstrated in [@ref20; @ref21] by the example of a cluster of two spheres.
\[sect5\]Conclusion
===================
Thus, in the present work optical properties of clusters made of two metal nanospheroids are considered theoretically, and analytical results are obtained. This investigation has become possible due to proving of the addition translation theorem for spheroidal functions in quasistatic regime. Plasmonic eigenoscillations were analyzed in details, and it was found that in a cluster of two prolate or oblate nanospheroids there can be three types of plasmon modes. Two of them (low frequency, $0 < \omega < \omega _{pl} / \sqrt {2}$, L- and T-modes) can be effectively excited by a plane electromagnetic wave, while the third type (high frequency, $\omega _{pl} / \sqrt {2} <
\omega < \omega _{pl} $, M-modes) can be excited only by a strongly nonuniform field of a nanolocalized source of light (a molecule, a quantum dot) located in the gap between two adjacent nanoparticles.
We have also investigated excitation of a nano-antenna made from two silver nanospheroids by fields of a plane wave and an electric dipole. The results of these investigations allow us to find absorption and scattering cross-sections of the nano-antenna as a function of wavelength for various polarization of an incident plane electromagnetic wave and to attribute all observable peaks to excitation of corresponding plasmonic modes. We have also analyzed radiative decay rate (or local density of state) of a dipole placed in the gap between the nanospheroids and have attributed all observable peaks to excitation of corresponding plasmonic modes.
The obtained analytical results can be used in many applications based on plasmonic nano-antennas and on enhancement of local field (SERS, SEF, nanolasers, nano-optical circuits, and so on). Besides, our results are very important for controlling accuracy of different computational software that has no a priori test of accuracy.
The authors thank Russian Foundation of Basic Researches (grant 09-02-13560) for financial support. The authors also express their gratitude to Ulrike Woggon for her helpful comments.
\[appen\]Translational addition theorem for spheroidal wave functions in a quasistatic limit
============================================================================================
A solution of the Helmholtz equation in spheroidal coordinates can be expanded as a series in spheroidal wave functions [@ref52]. In spite of the fact that these functions are thoroughly studied [@ref51; @ref52; @ref53; @ref54; @ref56] , the problem of electromagnetic waves scattering on a spheroid still remains one of the most complex ones, first of all, due to mathematical aspects related to usage of spheroidal wave functions. On the other hand, it is well-known that in a case of spheroids which size is substantially less than the wavelength (quasistatic limit), a solution of the Laplace equation in spheroidal coordinates can be presented as a series in associated Legendre functions [@ref55]. The Legendre functions are widespread and included as standard special functions in majority of mathematical software programs, such as [*Maple*]{}, [*Mathematica*]{}, [*MatLAB*]{}, etc. That is why it is important to obtain the translational addition theorem for spheroidal wave functions in the case when sizes of spheroids are far less than the wavelength in which these functions would be “replaced” by corresponding Legendre functions.
In [@ref45], a rotational-translational addition theorem was obtained for spheroidal wave functions in a case of an arbitrary size of spheroids in comparison to the wavelength. This theorem was derived in the following manner: first, spheroidal wave functions were expanded over spherical wave functions, then the spherical wave functions were expanded over spheroidal ones. Substituting one expansion into the other and using the addition theorem for spherical wave functions [@ref51; @ref61], the required theorem for spheroidal functions was derived as a result. Here, we will do the same in a particular case when spheroids’ sizes are far less than the wavelength.
In a case of a prolate spheroid, we have the following expansions [@ref45]
$$\begin{aligned}
&&R_{mn}^{\left( {3} \right)} \left( {kf,\xi} \right)S_{mn}^{\left(
{1} \right)} \left( {kf,\eta} \right)e^{im\phi} \nonumber \\
&&= {{\sum \limits_{q = {\left| {m} \right|},{\left| {m} \right|} +
1}^{\infty}} {' i^{q - n}d_{q - {\left| {m} \right|}}^{mn} \left(
{kf} \right)h_{q}^{\left( {1} \right)} \left( {kr} \right)P_{q}^{m}
\left( {\cos \theta} \right)e^{im\phi}} } , \label{eqa1}\end{aligned}$$
where $1 \le \xi < \infty $, $ - 1 \le \eta \le 1$ and $0
\le \phi \le 2\pi $ are coordinates of the prolate spheroid [@ref52]; $0 \le r < \infty $, $0 \le \theta \le \pi $ and $0
\le \phi \le 2\pi $ are spherical coordinates; $R_{mn}^{\left( {3}
\right)} \left( {kf,\xi} \right)$ and $S_{mn}^{\left( {1} \right)}
\left( {kf,\eta} \right)$ are radial and angular spheroidal functions of the third and first kinds according to [@ref52]; $h_{q}^{\left( {1} \right)} \left( {kr} \right)$, and $P_{q}^{m}
\left( {\cos \theta} \right)$ are the spherical Hankel function of the first kind [@ref56] and the associated Legendre function [@ref56]; $d_{q - {\left| {m} \right|}}^{mn} \left( {kf}
\right)$ are the expansion coefficients [@ref52]; $k$ is wavenumber, $f$ is a half of the spheroid’s focal distance; the stroke of the sum character implies summation over only even or only odd values of $q$. Another important relation from [@ref45] has the form:
$$\begin{aligned}
&&j_{n} \left( {kr} \right)P_{n}^{m} \left( {\cos \theta} \right)e^{im\phi}
= {\frac{{2}}{{2n + 1}}}{\frac{{\left( {n + m} \right)!}}{{\left( {n
- m}
\right)!}}} \nonumber \\
&&\times {{\sum\limits_{q = {\left| {m} \right|},{\left| {m} \right|} +
1}^{\infty}} {{' \frac{{i^{q - n}}}{{N_{mq} \left( {kf}
\right)}}}d_{n - {\left| {m} \right|}}^{mq} \left( {kf}
\right)R_{mq}^{\left( {1} \right)} \left( {kf,\xi}
\right)S_{mq}^{\left( {1} \right)} \left( {kf,\eta}
\right)e^{im\phi}} } , \label{eqa2}
\end{aligned}$$
where $j_{n} \left( {kr} \right)$ is the spherical Bessel function [@ref56]; $R_{mq}^{\left( {1} \right)} \left( {kf,\xi}
\right)$ is a radial spheroidal function of the first kind [@ref52], and $N_{mq} \left( {kf} \right)$ is normalization factor of angular spheroidal functions of the first kind [@ref52]. In the case of a spheroid which size is far less than the wavelength, that is, in the case $k \to 0$, one can use the following asymptotic expressions [@ref51]:
$$\begin{aligned}
d_{n - m - 2q}^{mn} &\approx& {\frac{{\left( {n + m} \right)!\left( {2n - 4q
+ 1} \right)!!\left( {2n - 2q - 1} \right)!!}}{{\left( {n + m - 2q}
\right)!\left( {2n + 1} \right)!!\left( {2q} \right)!!\left( {2n -
1} \right)!!}}}\left( {kf} \right)^{2q},\quad 0 \le q \le {\frac{{n
-
m}}{{2}}}, \nonumber \\
d_{n - m + 2q}^{mn} &\approx& {\frac{{\left( { - 1} \right)^{q}\left( {n - m
+ 2q} \right)!\left( {2n - 1} \right)!!\left( {2n + 1}
\right)!!}}{{\left( {n - m} \right)!\left( {2n + 4q - 1}
\right)!!\left( {2q} \right)!!\left( {2n + 2q + 1}
\right)!!}}}\left( {kf} \right)^{2q},\quad q \ge 1, \nonumber
\\ \label{eqa3}
\end{aligned}$$
and $N_{mn} \approx {\frac{{2}}{{2n + 1}}}{\frac{{\left(
{n + m} \right)!}}{{\left( {n - m} \right)!}}}$. Radial and angular spheroidal functions $R_{mn}^{\left( {1} \right)} \left( {kf,\xi}
\right)$, $R_{mn}^{\left( {3} \right)} \left( {kf,\xi} \right)$ and $S_{mn}^{\left( {1} \right)} \left( {kf,\eta} \right)$ can be also presented in the asymptotic form [@ref51; @ref53]
$$\begin{aligned}
R_{mn}^{\left( {1} \right)} \left( {kf,\xi} \right) &\approx& {\frac{{\left(
{n - m} \right)!}}{{\left( {2n - 1} \right)!!\left( {2n + 1}
\right)!!}}}\left( {kf} \right)^{n}P_{n}^{m} \left( {\xi} \right), \nonumber \\
R_{mn}^{\left( {3} \right)} \left( {kf,\xi} \right) &\approx& -
i{\frac{{\left( { - 1} \right)^{m}\left( {2n - 1} \right)!!\left(
{2n + 1} \right)!!}}{{\left( {n + m} \right)!}}}\left( {kf}
\right)^{ - n - 1}Q_{n}^{m} \left( {\xi} \right), \nonumber
\\ S_{mn}^{\left( {1} \right)} \left( {kf,\eta} \right) &\approx&
P_{n}^{m} \left( {\eta} \right), \label{eqa4}
\end{aligned}$$
where $P_{n}^{m} \left( {\eta} \right)$ is the associated Legendre function [@ref56] defined on the segment $ - 1 \le \eta
\le 1$; $P_{n}^{m} \left( {\xi} \right)$, $Q_{n}^{m} \left( {\xi}
\right)$ are associated Legendre functions of the first and second kinds [@ref56], correspondingly, defined in a complex plane with branch cut from $ - \infty $ to +1. Finally, asymptotic forms for spherical Bessel and Henkel functions have the form [@ref56]:
$$j_{n} \left( {kr} \right) \approx {\frac{{2^{n}n!}}{{\left( {2n + 1}
\right)!}}}\left( {kr} \right)^{n},\quad h_{n}^{\left( {1} \right)}
\left( {kr} \right) \approx - i{\frac{{\left( {2n}
\right)!}}{{2^{n}n!}}}\left( {kr} \right)^{ - n - 1}. \label{eqa5}$$
Making use of Eqs. (\[eqa3\])-(\[eqa5\]), one can obtain the following relations from Eqs. (\[eqa1\]) and (\[eqa2\]):
$$\begin{aligned}
&&Q_{n}^{m} \left( {\xi} \right)P_{n}^{m} \left( {\eta} \right) =
{\sum\limits_{q = 0}^{\infty} {{\frac{{\left( { - 1}
\right)^{m}\left( {n - m + 2q} \right)!\left( {n + m}
\right)!}}{{\left( {n - m} \right)!\left( {2n + 2q + 1}
\right)!!\left( {2q} \right)!!}}}\left( {{\frac{{f}}{{r}}}}
\right)^{n + 2q + 1}P_{n + 2q}^{m} \left( {\cos \theta} \right)}} , \nonumber \\
&&\left( {{\frac{{r}}{{f}}}} \right)^{n}P_{n}^{m} \left( {\cos \theta}
\right) = {\frac{{\left( {n - m} \right)!}}{{\left( {2n - 1}
\right)!!}}}P_{n}^{m} \left( {\xi} \right)P_{n}^{m} \left( {\eta}
\right)
\nonumber \\
&&+ {{\sum\limits_{q = 0,1}^{n - m - 2} } { ' {\frac{{\left( {n + m}
\right)!\left( {2m + 2q + 1} \right)q!}}{{\left( {n - m - q}
\right)!!\left( {n + m + q + 1} \right)!!\left( {2m + q}
\right)!}}}P_{m + q}^{m} \left( {\xi} \right)P_{m + q}^{m} \left(
{\eta} \right)}} . \label{eqa6}
\end{aligned}$$
To obtain the required addition theorem for wave functions of prolate spheroids which size is far less than the wavelength from Eq. (\[eqa6\]), it is necessary to apply the corresponding theorem for spherical wave functions [@ref19]:
$$\begin{aligned}
&&\left( {{\frac{{l}}{{r_{j}}} }} \right)^{n + 1}P_{n}^{m} \left( {\cos
\theta _{j}} \right)e^{im\phi _{j}} = {\sum\limits_{q =
0}^{\infty} {{\sum\limits_{p = - q}^{q} {R_{pqmn} \left( {\theta
_{js} ,\phi _{js}} \right)\left( {{\frac{{r_{s}}} {{l}}}}
\right)^{q}P_{q}^{p} \left( {\cos
\theta _{s}} \right)e^{ip\phi _{s}}} }} } , \nonumber \\
&&R_{pqmn} \left( {\theta _{js} ,\phi _{js}} \right) = {\frac{{\left( { - 1}
\right)^{q + m}\left( {q + n - p + m} \right)!}}{{\left( {q + p}
\right)!\left( {n - m} \right)!}}}P_{q + n}^{p - m} \left( {\cos
\theta _{js}} \right)e^{i\left( {m - p} \right)\phi _{js}} .
\label{eqa7}
\end{aligned}$$
In the expression (\[eqa7\]), the indices $j$, $s = 1$, 2 ($j \ne s$) denote the first “1” and the second “2” local systems of spherical coordinates, that are based on local Cartesian coordinates with parallel axes and attached to the spheres’ centers; ($r_{js}$, $\theta _{js}$, $\phi _{js}$) are spherical coordinates of the origin of the $s$-th coordinate system in the $j$-th local system of coordinates; $l = r_{js} = r_{sj} $ is the distance between the origins of local coordinates systems (see Fig. \[fig12\]). Substituting Eq. (\[eqa7\]) into Eq. (\[eqa6\]), we obtain after series of transformations the required translational addition theorem for wave functions of prolate spheroids in the quasistatic limit:
$$\begin{aligned}
&&Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right)e^{im\phi _{j}} = {\sum\limits_{q = 0}^{\infty}
{{\sum\limits_{p = - q}^{q} {O_{pqmn} \left( {f_{j} ,f_{s} ,l,\theta
_{js} ,\phi _{js}} \right)P_{q}^{p} \left( {\xi _{s}}
\right)P_{q}^{p} \left( {\eta _{s}}
\right)e^{ip\phi _{s}}} }} } , \nonumber \\
&&O_{pqmn} \left( {f_{j} ,f_{s} ,l,\theta _{js} ,\phi _{js}} \right) =
{\frac{{\left( { - 1} \right)^{q}\left( {2q + 1} \right)\left( {q -
p} \right)!\left( {n + m} \right)!}}{{\left( {q + p} \right)!\left(
{n - m}
\right)!}}} \nonumber \\
&&\times {\sum\limits_{r = 0}^{\infty} {{\sum\limits_{k = 0}^{\infty}
{{\frac{{\left( {q + n + 2r + 2k - p + m} \right)!}}{{\left( {2q +
2r + 1} \right)!!\left( {2n + 2k + 1} \right)!!\left( {2r}
\right)!!\left( {2k}
\right)!!}}}}}} } \nonumber \\
&&\times \left( {{\frac{{f_{s}}} {{l}}}} \right)^{2r + q}\left(
{{\frac{{f_{j}}} {{l}}}} \right)^{2k + n + 1}P_{q + n + 2r + 2k}^{p
- m} \left( {\cos \theta _{js}} \right)e^{i\left( {m - p}
\right)\phi _{js}} . \label{eqa8}
\end{aligned}$$
In Eq. (\[eqa8\]), the indices $j$, $s = 1$, 2 ($j \ne
s$) denote the first and second local prolate spheroid coordinate systems with the parallel Cartesian axes introduced above; the spherical angles $\theta _{js}$ and $\phi _{js}$ are the same as in Eq. (\[eqa7\]); $l$ is the distance between the origins of local coordinate systems (see Fig. \[fig12\] for details). Let us add that an obligatory condition for usage of the theorem (\[eqa8\]) is parallelism of Cartesian axes in the local coordinate systems. In the case when the second local coordinate system is turned around the first one, the theorem (\[eqa7\]) is to be changed [@ref19], resulting in corresponding modification of the theorem (\[eqa8\]). Let us also note that mathematical applicability condition for (\[eqa8\]) is convergence of the series in this expression. In particular, one may state that Eq. (\[eqa8\]) is applicable in the case of non-overlapping spheroids [@ref51].
The translational addition theorem for functions of oblate spheroids which size is much less than the wavelength, can be obtained from Eq. (\[eqa8\]) if the formal substitution is used [@ref52]: $\xi \to i\xi$ and $f \to - if$. In expressions derived in this manner, $\xi$ is a coordinate in the system of the oblate spheroid ($0 \le \xi < \infty )$; $f$ is a half of the focal distance of the oblate spheroid. Let us note that (\[eqa8\]) is a very general theorem, and many other theorems relating prolate, oblate, and spherical functions can be derived from it.
The obtained theorem (\[eqa8\]) is not very convenient for numerical calculations since it contains summation both over positive and negative $p$. To simplify it, let us use properties of the associated Legendre functions [@ref51; @ref62], basing on which we shall find that
$$\begin{aligned}
&&O_{ - p,q, - m,n} \left( {f_{j} ,f_{s} ,l,\theta _{js} ,\phi _{js}}
\right) \nonumber \\
&&= \left( { - 1} \right)^{p - m}{\left[ {{\frac{{\left( {q + p}
\right)!\left( {n - m} \right)!}}{{\left( {q - p} \right)!\left( {n
+ m} \right)!}}}} \right]}^{2}O_{pqmn} \left( {f_{j} ,f_{s}
,l,\theta _{js} ,\phi _{js}} \right)e^{i2\left( {p - m} \right)\phi
_{js}} . \label{eqa9}
\end{aligned}$$
Combining Eqs. (\[eqa8\]) and (\[eqa9\]), one can derive the following relation ($m$, $p = 0$, 1, 2, …):
$$\begin{aligned}
&&Q_{n}^{m}\left( {\xi _{j}}\right) P_{n}^{m}\left( {\eta _{j}}\right)
\left\{ {{\begin{array}{*{20}c} {\cos \left( {m\phi _{j}} \right)} \hfill \\
{\sin \left( {m\phi _{j}} \right)} \hfill \\ \end{array}}}\right. ={
\sum\limits_{q=0}^{\infty }{{\sum\limits_{p=0}^{q}{P_{q}^{p}\left(
{\xi _{s}}
\right) P_{q}^{p}\left( {\eta _{s}}\right) }}}} \nonumber \\
&&\times {\left\{ {{\begin{array}{*{20}c} {L_{pqmn} \left( {f_{j}
,f_{s} ,l,\theta _{js}} \right)\cos \left( {p\phi _{s} - \left( {p -
m} \right)\phi _{js}} \right) } \hfill \\ {L_{pqmn} \left( {f_{j}
,f_{s} ,l,\theta _{js}} \right)\sin \left( {p\phi _{s} - \left( {p -
m} \right)\phi _{js}} \right) }
\hfill \\ \end{array}}}\right. } \nonumber \\
&&\left. {{\begin{array}{*{20}c} { + T_{pqmn} \left( {f_{j} ,f_{s}
,l,\theta _{js}} \right)\cos \left( { - p\phi _{s} + \left( {p + m}
\right)\phi _{js}} \right)} \hfill \\ { + T_{pqmn} \left( {f_{j}
,f_{s} ,l,\theta _{js}} \right)\sin \left( { - p\phi _{s} + \left(
{p + m} \right)\phi _{js}} \right)} \hfill \\ \end{array}}}\right\}
, \label{eqa10}
\end{aligned}$$
where
$$\begin{aligned}
&&L_{pqmn}\left( {f_{j},f_{s},l,\theta _{js}}\right) = {\frac{{\left( {-1}
\right) ^{q}\left( {2-\delta _{0p}}\right) \left( {2q+1}\right)
\left( {q-p} \right) !\left( {n+m}\right) !}}{{2\left( {q+p}\right)
!\left( {n-m}\right) !
}}} \nonumber \\
&&\times {\sum\limits_{r=0}^{\infty }{{\sum\limits_{k=0}^{\infty
}{{\frac{{ \left( {q+n+2r+2k-p+m}\right) !}}{{\left(
{2q+2r+1}\right) !!\left( {2n+2k+1}
\right) !!\left( {2r}\right) !!\left( {2k}\right) !!}}}}}}} \nonumber \\
&&\times {{{{\left( {{\frac{{f_{s}}}{{l}}}}\right) ^{2r+q}\left(
{{\frac{{ f_{j}}}{{l}}}}\right) ^{2k+n+1}P_{q+n+2r+2k}^{p-m}\left(
{\cos \theta _{js}}
\right) }}}}, \nonumber \\
&&T_{pqmn}\left( {f_{j},f_{s},l,\theta _{js}}\right) ={\frac{{\left(
{-1} \right) ^{q+m}\left( {2-\delta _{0p}}\right) \left(
{2q+1}\right) \left( {q-p }\right) !\left( {n+m}\right) !}}{{2\left(
{q+p}\right) !\left( {n-m}\right)
!}}} \nonumber \\
&&\times {\sum\limits_{r=0}^{\infty }{{\sum\limits_{k=0}^{\infty
}{{\frac{{ \left( {q+n+2r+2k-p-m}\right) !}}{{\left(
{2q+2r+1}\right) !!\left( {2n+2k+1}
\right) !!\left( {2r}\right) !!\left( {2k}\right) !!}}}}}}} \nonumber \\
&&\times {{{{\left( {{\frac{{f_{s}}}{{l}}}}\right) ^{2r+q}\left(
{{\frac{{ f_{j}}}{{l}}}}\right) ^{2k+n+1}P_{q+n+2r+2k}^{p+m}\left(
{\cos \theta _{js}} \right) }}}}, \label{eqa11}
\end{aligned}$$
and $\delta _{0p} $ is Kronecker delta symbol.
In the case of coaxial prolate spheroids with a common axis $z,$ one should take $\phi _{j} = \phi _{s} $ and $\theta _{js} = 0$ or $\theta _{js} = \pi$ in Eq. (\[eqa10\]). This results in the following relation ($m = 0$, 1, 2, …):
$$\begin{aligned}
&&Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}} \right) =
{\sum\limits_{q = m}^{\infty} {S_{mqmn} \left( {f_{j} ,f_{s}
,l,\theta _{js}} \right)P_{q}^{m} \left( {\xi _{s}}
\right)P_{q}^{m} \left( {\eta
_{s}} \right)}} , \nonumber \\
&&S_{mqmn} \left( {f_{j} ,f_{s} ,l,\theta _{js}} \right) = {\frac{{\left( {
- 1} \right)^{q}\left( {2q + 1} \right)\left( {q - m} \right)!\left(
{n + m} \right)!}}{{\left( {q + m} \right)!\left( {n - m}
\right)!}}}P_{q + n}
\left( {\cos \theta _{js}} \right) \nonumber \\
&&\times {\sum\limits_{r = 0}^{\infty} {{\sum\limits_{k = 0}^{\infty}
{{\frac{{\left( {q + n + 2r + 2k} \right)!}}{{\left( {2q + 2r + 1}
\right)!!\left( {2n + 2k + 1} \right)!!\left( {2r} \right)!!\left(
{2k} \right)!!}}}\left( {{\frac{{f_{s}}} {{l}}}} \right)^{2r +
q}\left( {{\frac{{f_{j}}} {{l}}}} \right)^{2k + n + 1}}}} } .
\nonumber \\ \label{eqa12}
\end{aligned}$$
In the case when local coordinate systems of the spheroids have a common axis $x$, one should take $\theta _{js} = \pi / 2$ and $\phi
_{js} = 0$ or $\phi _{js} = \pi $ in Eq. (\[eqa10\]). As a result, we have in this case
$$\begin{aligned}
&&Q_{n}^{m} \left( {\xi _{j}} \right)P_{n}^{m} \left( {\eta _{j}}
\right){\left\{ {{\begin{array}{*{20}c}
{\cos \left( {m\phi _{j}} \right)} \hfill \\
{\sin \left( {m\phi _{j}} \right)} \hfill \\
\end{array}}} \right.} \nonumber \\
&&= {\sum\limits_{q = 0}^{\infty} {{\sum\limits_{p =
0}^{q} {P_{q}^{p} \left( {\xi _{s}} \right)P_{q}^{p} \left( {\eta
_{s}} \right){\left\{ {{\begin{array}{*{20}c}
{M_{pqmn} \left( {f_{j} ,f_{s} ,l,\phi _{js}} \right)\cos \left( {p\phi
_{s}} \right)} \hfill \\
{N_{pqmn} \left( {f_{j} ,f_{s} ,l,\phi _{js}} \right)\sin \left( {p\phi
_{s}} \right)} \hfill \\
\end{array}}} \right.}}}} } , \label{eqa13}\end{aligned}$$
where
$$\begin{aligned}
&&M_{pqmn} \left( {f_{j} ,f_{s} ,l,\phi _{js}} \right) \nonumber \\
&&= \left( {L_{pqmn}
\left( {f_{j} ,f_{s} ,l,\pi / 2} \right) + T_{pqmn} \left( {f_{j}
,f_{s} ,l,\pi / 2} \right)} \right)\cos \left( {\left( {p - m}
\right)\phi _{js}}
\right), \nonumber \\
&&N_{pqmn} \left( {f_{j} ,f_{s} ,l,\phi _{js}} \right) \nonumber \\
&&= \left( {L_{pqmn}
\left( {f_{j} ,f_{s} ,l,\pi / 2} \right) - T_{pqmn} \left( {f_{j}
,f_{s} ,l,\pi / 2} \right)} \right)\cos \left( {\left( {p - m}
\right)\phi _{js}} \right). \label{eqa14}
\end{aligned}$$
To simplify Eq. (\[eqa14\]), it is convenient to use the relation [@ref56]:
$$P_{n}^{m} \left( {0} \right) = {\frac{{2^{m}}}{{\sqrt {\pi}} } }\cos
\left( {\left( {n + m} \right){\frac{{\pi}} {{2}}}}
\right){\frac{{\Gamma \left( {\left( {n + m + 1} \right) / 2}
\right)}}{{\Gamma \left( {\left( {n - m + 2} \right) / 2}
\right)}}}, \label{eqa15}$$
where $\Gamma \left( {x} \right)$ is Euler’s gamma function [@ref56].
[99]{} H. Raether, *Surface Plasmons on Rough and Smooth Surfaces and on Gratings* (Springer-Verlag, Berlin, 1998).
S.A. Maier, *Plasmonics: Fundamentals and Applications* (Springer, New York, 2007).
V.V. Klimov, *Nanoplasmonics* \[in Russian\] (Fizmatlit, Moscow, 2009).
*Surface-Enhanced Raman Scattering*, K. Kneipp, M. Moskovits, H. Kneipp (eds.) (Springer-Verlag, Berlin, 2006).
Y. Liu, J. Bishop, L. Williams, S. Blair, and J. Herron, Nanotechnology **15**, 1368 (2004).
A.J. Haes, S.L. Zou, G.C. Schatz, and R.P. van Duyne, J. Phys. Chem. B **108**, 6961 (2004).
A.J. Haes, S.L. Zou, G.C. Schatz, and R.P. van Duyne, J. Phys. Chem. B **108**, 109 (2004).
S. Chah, M.R. Hammond, R.N. Zare, Chemistry and Biology **12**, 323 (2005).
C. Sonnichsen, B.M. Reinhard, J. Liphardt, and A.P. Alivisatos, Nat. Biotechnol. **23**, 741 (2005).
J.N. Farahani, D.W. Pohl, H.-J. Eisler, and B. Hecht, Phys. Rev. Lett. **95**, 017402 (2005).
P. Muhlschlegel, H.-J. Eisler, O.J.F. Martin, B. Hecht, and D.W. Pohl, Science **308**, 1607 (2005).
L. Novotny, Phys. Rev. Lett. **98**, 266802 (2007).
T.H. Taminiau, F.D. Stefani, and N.F. van Hulst, Opt. Exp. **16**, 10858 (2008).
H. Fischer and O.J.F. Martin, Opt. Exp. **16**, 9144 (2008).
J.R. Lakowicz, J. Malicka, I. Gryczynski, Z. Gryczynski and C.D. Geddes, J. Phys. D: Appl. Phys. **36**, R240 (2003).
R. Ruppin, J. Chem. Phys. **76**, 1681 (1982).
R. Ruppin, Phys. Rev. B **26**, 3440 (1982).
F. Claro, Phys. Rev. B **25**, 7875 (1982).
J.M. Gerardy and M. Ausloos, Phys. Rev. B **22**, 4950 (1980).
V.V. Klimov and D.V. Guzatov, Phys. Rev. B **75**, 024303 (2007).
V.V. Klimov, D.V. Guzatov, Quantum Electron. **37**, 209 (2007).
P. Chu and D.L. Mills, Phys. Rev. B **77**, 045316 (2008).
K. Li, M.I. Stockman, and D.J. Bergman, Phys. Rev. Lett. **91**, 227402 (2003).
D.A. Genov, A.K. Sarychev, V.M. Shalaev, and A. Wei, Nano Lett. **4**, 153 (2004).
P. Nordlander and C. Oubre, E. Prodan, K. Li and M.I. Stockman, Nano Lett. **4**, 899 (2004).
I. Romero, J. Aizpurua, G.W. Bryant, F.J. Garcia de Abajo, Opt. Exp. **14**, 9988 (2006).
M. Danckwerts and L. Novotny, Phys. Rev. Lett. **98**, 026104 (2007).
M. Quinten, A. Leitner, J.R. Krenn, and F.R. Aussenegg, Opt. Lett. **23**, 1331 (1998).
S.A. Maier, P.G. Kik, and H.A. Atwater, Appl. Phys. Lett. **81**, 1714 (2002).
D.-S. Wang and M. Kerker, Phys. Rev. B **24**, 1777 (1981).
V.V. Klimov, M. Ducloy, and V.S. Letokhov, Eur. Phys. J. D **20**, 133 (2002).
A. Trugler and U. Hohenester, Phys. Rev. B **77**, 115403 (2008).
A.F. Stevenson, J. Appl. Phys. **24**, 1143 (1953).
D.V. Guzatov, V.V. Klimov, Chem. Phys. Lett. **412**, 341 (2005).
J. Grand, P.-M. Adam, A.-S. Grimault, A. Vial, M. Lamy de la Chapelle, J.-L. Bijeon, S. Kostcheev and P. Royer, Plasmonics **1**, 135 (2006).
D.V. Guzatov, V.V. Klimov, and M.Yu. Pikhota, Laser Phys. **10**, 85 (2010).
*Surface Plasmon Nanophotonics*, M.L. Brongersma, P.G. Kik (eds.) (Springer, Dordrecht, 2007).
S.A. Maier, M.L. Brongersma, P.G. Kik, S. Meltzer, A.A.G. Requicha, and H.A. Atwater, Adv. Mater. **13**, 1501 (2001).
O.L. Muskens, V. Giannini, J.A. Sanchez-Gil, J. Gomez Rivas, Opt. Exp. **15**, 17736 (2007).
W. Rechberger, A. Hohenau, A. Leitner, J.R. Krenn, B. Lamprecht, F.R. Aussenegg, Optics Comm. **220**, 137 (2003).
E. Cubukcu, F. Degirmenci, C. Kocabas, M.A. Zimmler, J.A. Rogers, and F. Capasso, Proc. Nat. Acad. Sc. **106**, 2495 (2009).
V. Poponin, A. Ignatov, J. Kor. Phys. Soc. **47**, 222 (2005).
L. Rogobete, F. Kaminski, M. Agio, and V. Sandoghdar, Opt. Lett. **32**, 1623 (2007).
A. Ghoshal and P.G. Kik, J. Appl. Phys. **103**, 113111 (2008).
R.H. MacPhie, J. Dalmas, and R. Deleuil, Quart. Appl. Math. **XLIV**, 737 (1987).
J. Dalmas, R. Deleuil and R.H. MacPhie, Quart. Appl. Math. **XLVII**, 351 (1989).
M.F.R. Cooray and I.R. Ciric, IEEE Trans. Antennas Propag. **37**, 608 (1989).
M.F.R. Cooray and I.R. Ciric, Computer Phys. Comm. **68**, 279 (1991).
P.B. Johnson, R.W. Christy, Phys. Rev. B **6**, 4370 (1972).
N.N. Voitovich, B.Z. Katsenelenbaum, A.N. Sivov, *Generalized Method of Eigen-Oscillations in the Theory of Diffraction* \[in Russian\] (Nauka, Moscow, 1977).
E.A. Ivanov, *Diffraction of Electromagnetic Waves on Two Bodies* \[in Russian\] (Nauka i Tekhnika, Minsk, 1969).
C. Flammer, *Spheroidal Wave Functions* (Stanford University Press, Stanford, 1957).
I.V. Komarov, L.I. Ponomaryev, S.Yu. Slavyanov, *Spheroidal and Coulomb’s Spheroidal Functions* \[in Russian\] (Nauka, Moscow, 1976).
L.-W. Li, X.-K. Kang, M.-S. Leong, *Spheroidal Wave Functions in Electromagnetic Theory* (John Wiley and Sons, New York, 2002).
W.R. Smythe, *Static and Dynamic Electricity* (McGraw-Hill, New York, 1950).
M. Abramowitz and I.A. Stegun, *Handbook of Mathematical Functions* (Dover, New York, 1965).
V.V. Klimov, D.V. Guzatov, Appl. Phys. A **89**, 305 (2007).
L.D. Landau and E.M. Lifshits, *Electrodynamics of Continuous Media* \[in Russian\] (Nauka, Moscow, 1982).
V.V. Klimov, M. Ducloy, Phys. Rev. A **69**, 013812 (2004).
F. Engelbrecht and R. Helbig, Phys. Rev. B **48**, 698 (1993).
B. Friedman and J. Russek, Quart. Appl. Math. **XII**, 13 (1954).
I.S. Gradshteyn and I.M. Ryzhik, *Tables of Integrals, Sums, Series and Products* \[in Russian\] (Fizmatgiz, Moscow, 1963).
![\[fig1\] The geometry of a two-nanospheroid cluster. (a) The case of prolate nanospheroids, (b) the case of oblate nanospheroids.](fig1.eps){width="9.0cm"}
![\[fig2\] (Color online) Normalized frequencies of the first three plasmonic oscillations in a cluster of two identical prolate spheroidal nanoparticles as a function of the distance between the nanospheroids’ centers. The axis-symmetrical case ($m =
0$) is considered, and the aspect ratio of a single spheroid is $a/c
= 0.6$ ($\xi _{0} = 1 / \sqrt {1 - \left( {a / c} \right)^{2}} $). (a) Symmetrical modes (eigenvalues of Eq. (\[eq14\])), (b) antisymmetric modes (eigenvalues of Eq. (\[eq15\])). Dashed lines show plasmon frequencies of a single prolate nanospheroid ${\frac{{\omega}} {{\omega _{pl}}} } = {\left\{ {{\frac{{dP_{n}^{m}
\left( {\xi _{0}} \right)}}{{d\xi _{0}}} }Q_{n}^{m} \left( {\xi
_{0}} \right) / \left( {{\frac{{dP_{n}^{m} \left( {\xi _{0}}
\right)}}{{d\xi _{0} }}}Q_{n}^{m} \left( {\xi _{0}} \right) -
P_{n}^{m} \left( {\xi _{0}} \right){\frac{{dQ_{n}^{m} \left( {\xi
_{0}} \right)}}{{d\xi _{0}}} }} \right)} \right\}}^{1 / 2}$. The figure (c) shows plasmon frequencies of the single prolate nanospheroid as a function of the inverse aspect ratio $c/a$. The vertical line corresponds to $a/c = 0.6$ and allows to select asymptotic values for the figures (a) and (b).](fig2.eps){width="6.6cm"}
![\[fig3\] (Color online) Distribution of a surface charge (a.u.) of the lowest plasmon mode in a cluster of two identical prolate spheroidal nanoparticles according to solution of the equations (\[eq14\]) and (\[eq15\]). The axial-symmetrical case ($m = 0$) is considered, and the aspect ratio of a single spheroid is $a/c = 0.6$. The distance between the nanospheroids’ centers is $l/2c = 1.03$ (a), $l/2c = 1.2$ (b). Distribution of a surface charge in a single spheroid is shown in panel (c). The red color corresponds to the positive charge, and blue - to the negative one.](fig3.eps){width="11.0cm"}
![\[fig4\] (Color online) Normalized plasmon frequencies of the first three modes in a cluster of two identical oblate nanospheroids as a function of the distance between their centers. (a) Symmetrical modes (the eigenvalues of Eq. (\[eq20\])), (b) antisymmetric modes (the eigenvalues of Eq. (\[eq21\])). By dashed lines, plasmon frequencies of a single oblate nanospheroid ${\frac{{\omega}} {{\omega _{pl}}} } = {\left\{ {{\frac{{dP_{n}^{m}
\left( {i\xi _{0}} \right)}}{{d\xi _{0}}} }Q_{n}^{m} \left( {i\xi
_{0}} \right) / \left( {{\frac{{dP_{n}^{m} \left( {i\xi _{0}}
\right)}}{{d\xi _{0} }}}Q_{n}^{m} \left( {i\xi _{0}} \right) -
P_{n}^{m} \left( {i\xi _{0}} \right){\frac{{dQ_{n}^{m} \left( {i\xi
_{0}} \right)}}{{d\xi _{0}}} }} \right)} \right\}}^{1 / 2}$ are shown. The spect ratio of a single oblate nanospheroids is $c/a =
0.6$ ($\xi _{0} = 1 / \sqrt {\left( {a / c} \right)^{2} - 1} $). The figure (c) shows plasmon frequencies of a single oblate nanospheroid as a function of the inverse aspect ratio $a/c$. The vertical line corresponds to $c/a = 0.6$ and allows to select asymptotic values for the figures (a) and (b).](fig4.eps){width="6.6cm"}
![\[fig5\] (Color online) Distribution of a surface charge (a.u.) of the lowest plasmon modes in a cluster of two identical oblate spheroidal nanoparticles according to solution of the equations (\[eq20\]) and (\[eq21\]). The aspect ratio is $c/a = 0.6$. The distance between the nanospheroids’ centers $l/2a =
1.05$ (a), $l/2a = 1.4$ (b). Distribution of a surface charge in a single spheroid is shown in panel (c). The red color corresponds to the positive charge, and blue - to the negative one.](fig5.eps){width="11.0cm"}
![\[fig6\] (Color online) Absorption (a) and scattering (b) cross-sections of a cluster of two identical prolate nanospheroids made from silver as a function of the wavelength. The large semi-axes of the nanospheroids are $c=15$ nm, the aspect ratios are $a/c=0.6$, and the distance between the nanospheroids’ centers is $l/2c = 1.05$. The labels x, y, z correspond to polarization of an incident wave along the x, y, z axis. The circles correspond to finite element simulations with Comsol Multiphysics software. Doubled absorption cross-section and quadrupled scattering cross-section of a single nanospheroid are shown by dashed curves.](fig6.eps){width="8.0cm"}
![\[fig7\] (Color online) Absorption (a) and scattering (b) cross-sections of a cluster of two identical oblate nanospheroids made from silver as a function of the wavelength. The large semi-axes of the nanospheroids are $a=15$ nm, the aspect ratios are $c/a=0.6$, the distance between the nanospheroids’ centers is $l/2a = 1.05$. The labels x, y, z correspond to polarization of an incident wave along the x, y, z axis. The circles correspond to finite element simulations with Comsol Multiphysics software. Doubled absorption cross-section and quadrupled scattering cross-section of a single nanospheroid are shown by dashed curves.](fig7.eps){width="8.0cm"}
![\[fig8\] (Color online) Spatial distribution of ${\left| {{\mathbf {E}}} \right|}^{2} / {\left| {{\mathbf {E}}_{0}}
\right|}^{2}$ in a cluster of two prolate nanospheroids for $\lambda
= 425$ nm at longitudinal polarization of the excitation field (L = 1 plasmonic resonance, common logarithmic scale).](fig8.eps){width="8.0cm"}
![\[fig9\] (Color online) Enhancement of ${\left|
{{\mathbf {E}}} \right|}^{2}$ in the gap between two identical silver nanospherodis as a function of the wavelength. P and O labels correspond to prolate and oblate spheroids correspondigly. The large semi-axes of the nanospheroids are 15 nm, the aspect ratios are 0.6, $l/2c = 1.05$ for a cluster of prolate spheroids and $l/2a = 1.05$ for a cluster of oblate spheroids. Enhancements for single nanospheroids are shown by the dashed curve.](fig9.eps){width="8.0cm"}
![\[fig10\] (Color online) Normalized radiative decay rate of a dipole placed at the middle point between two identical prolate nanospheroids made from silver as a function of the wavelength. The dipole source moment is oriented along the x or y axes (a) and along the z axis (b). The large semi-axes of the nanospheroids are $c=15$ nm, the aspect ratios are $a/c=0.6$. The curves $\alpha $, $\beta $, $\gamma $, $\delta $ correspond to $l/2c
= 1.05$, 1.1, 1.3 and 1.03, respectively. The asymptotic expression obtained by approximation of the spheroids by point dipoles ($l/2c =
1.3$) is shown by the dashed curve.](fig10.eps){width="8.0cm"}
![\[fig11\] (Color online) Normalized radiative decay rate of a dipole placed at the middle point between two identical oblate nanospheroids made from silver as a function of the wavelength. The dipole source moment is oriented along the x (a), y (b), and z (c) axis. The large semi-axes of the nanospheroid are $a=15$ nm, the aspect ratio is $c/a=0.6$. The curves $\alpha $, $\beta $, $\gamma $ correspond to $l/2a = 1.05$, 1.1, and 1.3, respectively. The asymptotic expression obtained by approximation of the spheroids by point dipoles ($l/2a = 1.3$) is shown by the dashed curve.](fig11.eps){width="7.5cm"}
![\[fig12\] Geometry of derivation of the translational addition theorem.](fig12.eps){width="8.0cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study strange stars in the framework of $f\left(R,\mathcal{T}\right)$ theory of gravity. To provide exact solutions of the field equations it is considered that the gravitational Lagrangian can be expressed as the linear function of the Ricci scalar $R$ and the trace of the stress-energy tensor $\mathcal{T}$, i.e. $f(R,\mathcal{T})=R+2\chi\mathcal{T}$, where $\chi$ is a constant. We also consider that the strange quark matter (SQM) distribution inside the stellar system is governed by the phenomenological MIT Bag model equation of state (EOS), given as $p_r=\frac{1}{3}\left(\rho-4\,B \right)$, where $B$ is the Bag constant. Further, for a specific value of $B$ and observed values of mass of the strange star candidates we obtain the exact solution of the modified Tolman-Oppenheimer-Volkoff (TOV) equation in the framework of $f\left(R,\mathcal{T}\right)$ gravity and have studied in detail the dependence of the different physical parameters, like the metric potentials, energy density, radial and tangential pressures and anisotropy etc., due to the chosen different values of $\chi$. Likewise in GR, as have been shown in our previous work \[Deb et al., Ann. Phys. (Amsterdam) 387, 239 (2017)\] in the present work also we find maximum anisotropy at the surface which seems an inherent property of the strange stars in modified $f\left(R, \mathcal{T}\right)$ theory of gravity. To check the physical acceptability and stability of the stellar system based on the obtained solutions we have performed different physical tests, viz., the energy conditions, Herrera cracking concept, adiabatic index etc. In this work, we also have explained the effects, those are arising due to the interaction between the matter and the curvature terms in $f\left(R,\mathcal{T}\right)$ gravity, on the anisotropic compact stellar system. It is interesting to note that as the values of $\chi$ increase the strange stars become more massive and their radius increase gradually so that eventually they gradually turn into less dense compact objects. The present study reveals that the modified $f\left(R,\mathcal{T}\right)$ gravity is a suitable theory to explain massive stellar systems like recent magnetars, massive pulsars and super-Chandrasekhar stars, which can not be explained in the framework of GR. However, for $\chi=0$ the standard results of Einsteinian gravity are retrieved.'
author:
- Debabrata Deb
- Farook Rahaman
- Saibal Ray
- 'B.K. Guha'
title: 'Anisotropic strange stars under simplest minimal matter-geometry coupling in the $f\left(R,\mathcal{T}\right)$ gravity'
---
Introduction
============
We are living in the age of the accelerated expansion of the universe which is well supported by the evidences of recent observations, like CMB, LSS, supernovae-Ia and BAO [@Riess1998; @Perlmutter1999; @Bernardis2000; @Perlmutter2003]. Thus the modern cosmology is mainly dependent on the recent observational evidences of the accelerated expansion of the universe. However, Einstein’s general theory of relativity failed to answer the satisfactory reason behind this accelerated expansion of the universe. Although in this connection many researchers [@Caldwell2002; @Nojiri2003; @Odinstov2003; @Padmanabhan2002; @Kamenshchik2001; @Bento2002] predicted that the sole reason behind this phenomenon is the presence of an unknown form of exotic energy dominated by the negative pressure which is widely known as the dark energy. The gravitational interaction is the most fundamental but least understood force of the nature. According to the strings/M-theory (also known as theory of everything) general relativity is an approximation and consistent to the small curvature. Though in the early days some unknown gravitational theory described the evolution of the universe but now it is well accepted that the modified gravity which is a classical generalization of the general relativity, can explain the early-time inflation and the late-time acceleration without introducing any form of the dark component. Also, some of the modified gravity theories with the gravitational term are well valid in the high energy realm which produced inflationary epoch. The curvature decreases during the evolution of the universe and in the intermediate universe general relativity provides a sufficient approximation. Interestingly, the early-time as well as the late-time acceleration happen due to the fact that some sub-dominant terms of gravitational action may become essential to the large or small curvatures. Though the complete gravitational action should be described by some fundamental theory which is yet to be achieved, but such approach of the alternative theory of gravity can be considered as a dynamical solution of the cosmological constant problem. The modified gravity approach in the absence of the fundamental quantum gravity showed a promising way out as it is well consistent with the observational data and data from local tests [@Nojiri2011]. Few well-known relevant alternative gravity theories are $f \left(R\right)$ gravity [@Nojiri2003a; @Carroll2004; @Allemandi2005; @Nojiri2007; @Bertolami2007], Brans-Dicke (BD), $f\left(G\right)$ [@Bamba2010a; @Bamba2010b; @Rodrigues2014] gravity, $f\left(\mathbb{T}\right)$ gravity [@Bengocheu2009; @Linder2010], scalar tensor theories of gravity and $f\left(R,G\right)$ gravity, etc., where $R$, $G$ and $\mathbb{T}$ are the scalar curvature, the Gauss-Bonnet scalar and the torsion scalar, respectively.
In his pioneering work Capozziello [@Capozziello2002] proposed a new modified theory of gravity to tackle the issue of dark energy. Later, Allemandi et al. [@Allemandi2005] have introduced the nonlinear scalar-gravity theories in the Palatini formulation. In their important review article Nojiri and Odintsov [@Nojiri2011] have presented a detailed study on the various extended gravity models, viz., traditional $f\left(R\right)$ and Ho[ř]{}ava-Lifshitz $f\left(R\right)$ gravity, scalar-tensor theory, string-inspired and Gauss–Bonnet theory, non-local gravity, non-minimally coupled models, and power-counting renormalizable covariant gravity. In this large volume of works they have investigated relation between the discussed modified gravity theories and their different representations. Further, the authors also have demonstrated how these extended gravity theories are showing well agreement with the local tests and featuring well justified description of the inflation with the dark energy epoch. Again, Capozziello and Laurentis [@Capozziello2011] presented an extended study on the different modified theories of gravity, viz., $f\left(R\right)$ gravity, scalar-tensor gravity, Brans-Dicke gravity and $f\left(R,\phi\right)$ gravity, etc., to address the shortcomings of GR at the scale of ultraviolet and infrared. Astashenok et al. [@Astashenok2013] and Capozziello [@Capozziello2016] presented models for neutron stars under different form of the $f\left(R\right)$ gravity. In another work Astashenok et al. [@Astashenok2015] have studied non-perturbative models for strange quark stars in $f\left(R\right)$ gravity.
Recently Harko et al. [@Harko2011] presented a more generalized form of $f \left(R\right)$ gravity theory by choosing the matter Lagrangian consists of an arbitrary function of the Ricci scalar ($R$) and the trace of the energy momentum tensor ($\mathcal{T}$) given as $f\left(R,\mathcal{T}\right)$. This is known as $f\left(R,\mathcal{T}\right)$ theory of gravity. Immediately it has drawn attention of many researchers and in the framework of many cosmological models [@Myrzakulov2012; @Jamil2012; @Shabani2013; @Shabani2014; @Moraes2015d; @Momeni2015; @Zaregonbadi2016; @Shabani2017a; @Shabani2017b] have been studied. Besides cosmology this gravity has successfully been studied in the realm of astrophysics too. Under astrophysics it is observed that Sharif et al. [@sharif2014] explored the factors that affect the stability of a locally isotropic spherically symmetric self gravitating system. By employing the perturbation scheme Noureen et al. [@noureen2015; @noureen2015b; @noureen2015c] have presented a series of works on the dynamical instability of spherically symmetric anisotropic collapsing stars under different conditions. Further, Zubair and Noureen [@zubair2015a] studied the dynamical stability of axially symmetric anisotropic sources whereas Zubair et al. [@zubair2015b] investigated the possible formation of compact stars by employing Krori and Barua metric. Alhamzawi and Alhamzawi [@Ahmed2015] have shown the effect of $f\left(R,\mathcal{T}\right)$ gravity on the gravitational lensing and also compared their result with the standard results of general relativity (GR).
Furthermore, general relativity and its possible extension [@Psaltis2008] can be distinguished due to the strong gravitational field regimes of the relativistic stars. Various developments of the new stellar structures constitute the signature of the extended gravity model [@Capozziello2011a; @Capozziello2012] as they have important observational consequences. Also, in particular some simplest extension of the general relativity, for example $f\left(R\right)$ gravity do not support existence of the stable stellar system [@Briscese2007; @Abdalla2005; @Bamba2008; @Kobayashi2008; @Babichev2010; @Nojiri2009; @Bamba2011]. On the other hand, the stability of the stellar system in modified gravity in the certain cases can be achieved using the so-called Chameleon Mechanism [@Khoury2004a; @Khoury2004b; @Upadhye2009].
Although all the above-mentioned literature are studied on the basis of the analytical solution Moraes et al. [@Moraes2015] first presented the exact solution of the Tolman-Oppenheimer-Volkoff (TOV) equation in $f(R,\mathcal{T})$ gravity, using Runge-Kutta 4th-order method and studied hydrostatic equilibrium configurations for neutron stars and strange stars. Here, we would like to mention that unfortunately in the TOV equation \[Eq. $\left(3.9\right)$\] a minus sign has been missed in their paper [@Moraes2015]. Using the results of Moraes et al. [@Moraes2015] later on Das et al. [@Amit2016] presented an analytical model of compact stars in $f\left(R,\mathcal{T}\right)$ gravity by employing the Lie algebra with the conformal Killing vectors. However, in another work on gravastars Das et al. [@Amit2017] have corrected the form of the TOV equation and provided an analytical model in $f\left(R,\mathcal{T}\right)$ gravity.
Harko et al. [@Harko2011] in their pioneering work mentioned that the motivation behind considering $T$-dependence in the $f\left(R,\mathcal{T}\right)$ theory of gravity is the possible existence of exotic imperfect fluids or quantum effects, such as the particle production [@Harko2014]. The authors in their study [@Harko2011] showed that the covariant derivative of the energy-momentum tensor is not zero and an extra acceleration will always be present in $f\left(R,\mathcal{T}\right)$ gravity due to the coupling between the matter and the curvature terms. Hence particles will follow non-geodesic path in $f\left(R,\mathcal{T}\right)$ gravity. Later, Chakraborty [@SC2013] addressed this issue and showed that for a specific form of the function $f\left(R,\mathcal{T}\right)$, as $f\left(R,\mathcal{T}\right)=R+h\left(\mathcal{T}\right)$, the test particles follow the geodesic path. Hence the author [@SC2013] demonstrated that the whole system would act like non-interacting two fluid system where the second type of fluid is originated due to the interaction between the geometry and the matter.
Here we would like to highlight the fact that in the framework of GR one can find a vast number of works [@Ivanov2002; @SM2003; @MH2003; @Usov2004; @Varela2010; @Rahaman2010; @Rahaman2011; @Rahaman2012; @Kalam2012; @Deb2016; @Shee2016; @Maurya2016; @Maurya2017], where influence of the anisotropy on the static spherically symmetric compact objects have been studied. It is to note that when the radial component of the pressure, ${p_r}(r)$, differs from the angular component, ${p_{\theta}}(r) = {p_{\phi}}(r) \equiv {p_{t}}(r)$ the system can be said anisotropic in nature. Clearly, the condition ${p_{\theta}}(r) = {p_{\phi}}(r)$ is rising due to the effect of the spherical symmetry. In a physical system, the pressures are anisotropic when the associated scalar field has a non-zero spatial gradient. The anisotropic stress in the present case may be arising due to the presence of the anisotropic nature of the two-fluid system.
We have arranged the present article as follows: Basic mathematical formulation of $f\left(R,\mathcal{T}\right)$ gravity is presented in Sec. \[sec1\]. In Sec. \[sec2\] we formulate basic stellar equations and present the solution of the Einstein field equations in Sec. \[sec3\]. We examine physical acceptability and stability of the stellar system in Sec. \[sec4\] by studying energy conditions \[subsec4.1\], mass-radius relation \[subsec4.2\], stability of the stellar model \[subsec4.3\] and compactification factor as well as reddshift \[subsec4.4\]. Finally, we conclude our study with a discussion in Sec. \[sec5\].
Basic formulation of $f\left(R,\mathcal{T}\right)$ theory of gravity {#sec1}
====================================================================
Following Harko et al. [@Harko2011], the modified form of Einstein-Hilbert (EH) action in $f\left(R,\mathcal{T}\right)$ gravity reads $$\label{1.1}
S=\frac{1}{16\pi}\int d^{4}xf(R,\mathcal{T})\sqrt{-g}+\int
d^{4}x\mathcal{L}_m\sqrt{-g},$$ where $g$ and $\mathcal{L}_m$ are the determinant of the metric $g_{\mu\nu}$ and the matter Lagrangian density, respectively. We adopt throughout the article $G=1=c$.
Variation of the modified EH action (\[1.1\]) in $f\left(R,\mathcal{T}\right)$ gravity with respect to $g_{\mu\nu}$ yields the modified field equation as follows $$\begin{aligned}
\label{1.2}
&\qquad\hspace{-3cm} G_{\mu\nu}=\frac{1}{{f_R}\left(R,\mathcal{T}\right)}\Big[\lbrace{8\pi+f_\mathcal{T}}\left(R,\mathcal{T}\right)\rbrace{T_{\mu\nu}}\nonumber \\
&\qquad -\rho{g_{\mu\nu}}{f_\mathcal{T}}\left(R,\mathcal{T}\right)+\frac{1}{2}\lbrace{f_\mathcal{T}}\left(R,\mathcal{T}\right)-R{f_R}\left(R,\mathcal{T}\right)\rbrace{g_{\mu\nu}}\nonumber\\
&\qquad\hspace{1cm} +\left(\nabla_{\mu} \nabla_{\nu}-g_{\mu\nu}{\mathop{}\!\mathbin\Box}\right) {{f_R}\left(R,\mathcal{T}\right)}\Big],\end{aligned}$$ where $f_R (R,\mathcal{T})=\partial f(R,\mathcal{T})/\partial R$ and $f_\mathcal{T}(R,\mathcal{T})=\partial f(R,\mathcal{T})/\partial\mathcal{T}$ whereas ${\mathop{}\!\mathbin\Box}\equiv\partial_{\mu}(\sqrt{-g} g^{\mu\nu} \partial_{\nu})/\sqrt{-g}$ is the D’Alambert operator and $R_{\mu\nu}$ is the Ricci tensor. We assume that $\mathcal{L}_m=\rho$, where $\rho$ is the energy density of SQM distribution.
Now we define $T_{\mu\nu}$, which represents the stress-energy tensor for the anisotropic fluid distribution, in the following form $$\begin{aligned}
\label{1.3}
T_{\mu\nu}=\left(\rho+{p_t}\right){u_{\mu}}{u_{\nu}}-{p_t}{g_{\mu\nu}}+\left({p_r}-{p_t}\right){v_{\mu}}{v_{\nu}},\end{aligned}$$ where $p_r$ and $p_t$ represent the radial and tangential pressures of the SQM distribution, respectively whereas $u_{\mu}$ and $v_{\mu}$ represent four-velocity and radial four-vector, respectively.
The covariant divergence of the stress-energy tensor (\[1.3\]) is given by $$\begin{aligned}
\label{1.4}
&\qquad\hspace{-0.5cm}\nabla^{\mu}T_{\mu\nu}=\frac{f_\mathcal{T}(R,\mathcal{T})}{8\pi -f_\mathcal{T}(R,\mathcal{T})}\big[\left(-T_{\mu\nu}+\rho{g_{\mu\nu}}\right)\nabla^{\mu}\ln f_\mathcal{T}(R,\mathcal{T}) \nonumber\\
&\qquad\hspace{-0.5cm}-2\nabla^{\mu}T_{\mu\nu}+\frac{1}{2}{g_{\mu\nu}}{\nabla^{\mu}}\left(2\rho-\mathcal{T}\right)\big].\end{aligned}$$
Following Harko et al. [@Harko2011], in the present article we consider simple linear form of the function $f\left(R,\mathcal{T}\right)$ as $f\left(R,\mathcal{T}\right)=f\left(R\right)+2f\left(\mathcal{T}\right)$, where $f\left(R\right)=R$ and $f\left(\mathcal{T}\right)=2\chi \mathcal{T}$. This form of the function $f\left(R,\mathcal{T}\right)$ has been broadly used by several authors [@singh2015; @moraes2014b; @moraes2015a; @moraes2015b; @moraes2017; @singh2014; @baffou2015; @shabani2013; @shabani2014; @sharif2014b; @reddy2013b; @kumar2015; @shamir2015; @Fayaz2016].
Now substituting the assumed form of the function $f\left(R,\mathcal{T}\right)$ into Eq. (\[1.2\]) we find $$\begin{aligned}
\label{1.5}
G_{\mu\nu}=8\pi T_{\mu\nu}+\chi \mathcal{T}g_{\mu\nu}+2\chi(T_{\mu\nu}-\rho g_{\mu\nu})=8\pi T_{\mu\nu,eff},\nonumber\\\end{aligned}$$ where $G_{\mu\nu}$ is the standard Einstein tensor and $T_{\mu\nu,eff}=T_{\mu\nu}+\frac{\chi}{8\pi} \mathcal{T}g_{\mu\nu}+\frac{\chi}{4\pi}(T_{\mu\nu}-\rho g_{\mu\nu})$. The usual general relativistic results can be achieved by substituting $\chi=0$ into Eq. (\[1.5\]).
Now, substituting $f(R,\mathcal{T})=R+2\chi\mathcal{T}$ in Eq. (\[1.4\]) we have $$\begin{aligned}
\label{1.6}
\left(4\pi+\chi\right)\nabla^{\mu}T_{\mu\nu}=-\frac{1}{2}\chi\left[g_{\mu\nu}\nabla^{\mu}\mathcal{T}-2\,\nabla^{\mu}(\rho g_{\mu\nu})\right].\end{aligned}$$
We can write Eq. (\[1.6\]) as follows $$\label{1.7}
\nabla^{\mu}T_{\mu\nu,eff}=0.$$
Here also one may achieve the standard form of the conservation of stress-energy tensor as GR by substituting $\chi=0$ into Eq. (\[1.6\]).
Basic stellar equations in $\bf{\textit{f}\left(R,\mathcal{T}\right)}$ theory of gravity {#sec2}
========================================================================================
We consider the spherically symmetric metric in its usual form $$\label{2.1}
ds^2=e^{\nu(r)}dt^2-e^{\lambda(r)}dr^2-r^2(d\theta^2+\sin^2\theta
d\phi^2),$$ where $\nu$ and $\lambda$ are metric potentials and function of the radial coordinate only.
Hence, using Eqs. (\[1.3\]), (\[1.5\]) and (\[2.1\]) we find the Einstein field equations for the spherically symmetric anisotropic stellar system given as $$\begin{aligned}
\label{2.2}
&\qquad\hspace{-2cm} {{\rm e}^{-\lambda}} \left( {\frac {\lambda^{{\prime}}}{r}}-\frac{1}{{r}^{2}} \right) +\frac{1}{{r}^{2}}=\left( 8\,\pi +\chi \right) \rho-\chi\,p_{{r}}-2\,\chi\,p_{{t}}\nonumber\\
&\qquad\hspace{-1.7cm} =8\pi{{\rho}_{eff}}, \\ \label{2.3}
&\qquad\hspace{-2cm} {{\rm e}^{-\lambda}} \left( {\frac {\nu^{{\prime}}}{r}}+\frac{1}{{r}^{2}} \right) -\frac{1}{{r}^{2}}=\chi\,\rho+ \left( 8\,\pi +3\,\chi \right) p_{{r}}+2\,\chi\,p_{{t}}\nonumber\\
&\qquad\hspace{-1.8cm}=8\pi{{p}_{effr}}, \\ \label{2.4}
&\qquad\hspace{-1.5cm} \frac{{\rm e}^{-\lambda}}{2}\left( \nu^{{\prime\prime}}+\frac{{{\nu^{{\prime}}}}^{2}}{2}+{\frac {\nu^{{\prime}}-\lambda^{{\prime}}}{r}}-\frac{\nu^{{\prime}}\lambda^{{\prime}}}{2} \right) =\chi\,\rho+\chi\,p_{{r}}+ \left( 8\,\pi +4\,\chi \right) p_{{t}}\nonumber\\
&\qquad\hspace{0.5cm}=8\pi{{p}_{efft}},\end{aligned}$$ where a $\prime$ denotes differentiation with respect to the radial coordinate $r$. Here ${\rho}_{eff}$, ${p}_{effr}$ and $p_{efft}$ represents the effective energy density, radial pressure and tangential pressure for our system and given as $$\begin{aligned}
\label{2.5}
{{\rho}_{eff}}=\rho+{\frac {\chi }{8\pi }}\left( \rho-p_{{r}}-2\,p_{{t}} \right),\\ \label{2.6}
{{p}_{effr}}=p_{{r}}+{\frac {\chi }{8\pi }}\left( \rho+3\,p_{{r}}+2\,p_{{t}} \right),\\ \label{2.7}
{{p}_{efft}}=p_{{t}}+{\frac {\chi}{8\pi }\left( \rho+p_{{r}}+4\,p_{{t}} \right)}.\end{aligned}$$
We assume that the SQM distribution inside the strange stars is governed by the simple phenomenological MIT Bag model EOS [@Chodos1974]. In bag model, by introducing [*ad hoc*]{} bag function all the corrections of energy and pressure functions of SQM have been maintained. We also consider that the quarks are non-interacting and massless in a simplified bag model. The quark pressure therefore can be defined as $$\label{2.8}
{p_r}={\sum_{f=u,d,s}}{p^f}-{B},$$ where $p^f$ is the individual pressure of the up $\left(u\right)$, down $\left(d\right)$ and strange $\left(s\right)$ quark flavors and $B$ is the vacuum energy density (also well known as Bag constant) which is a constant quantity within a numerical range. In the present article we consider the value of Bag constant as $B=83~MeV/{{fm}^3}$ [@Rahaman2014].
Now the individual quark pressure ($p^f$) can be defined as $p^f=\frac{1}{3}{{\rho}^f}$, where ${{\rho}^f}$ is the energy density of the individual quark flavor. Hence, the energy density, $\rho$ of the de-confined quarks inside the bag is given by $${{\rho}}={\sum_{f=u,d,s}}{{\rho}^f}+B. \label{2.9}$$
Using Eqs. (\[2.8\]) and (\[2.9\]) we have the EOS for SQM given as $${p_r}=\frac{1}{3}({{\rho}}-4B).\label{2.10}$$
It is observed that ignoring critical aspects of the quantum particle physics in the framework of GR several authors [@1; @2; @3; @4; @5; @6; @7; @8] successfully have been introduced this simplified form of the MIT Bag EOS to study stellar systems made of SQM.
To have non-singular monotonically decreasing matter density inside the spherically symmetric stellar system, following Mak and Harko [@Harko2002], we assume simplified form of $\rho$ given as $$\label{2.11}
\rho(r)=\rho_c\left[1-\left(1-\frac{\rho_0}{\rho_c}\right)\frac{r^{2}}{R^{2}}\right],$$ where $\rho_c$ and $\rho_0$ are constants and denote the maximum and minimum values of $\rho$ at the center and on the surface, respectively.
Now following [@Moraes2017] we consider $p_t$ is related to $\rho$ by a relation given as $$\begin{aligned}
p_{t}=c_{1}\rho+c_{2},\label{2.11a}\end{aligned}$$ where $c_1$ and $c_2$ are constants.
We define the mass function of the spherically symmetric stellar system as $$\label{2.12}
m \left( r \right) =4\,\pi\int_{0}^{r}\!{{\rho}_{eff}} \left( r \right) {r}^{2}{dr}.$$
At this juncture we consider the Schwarzschild metric to represent the exterior spacetime of our system given as $$\begin{aligned}
\label{2.13}
{ds}^2=\left(1-\frac{2M}{r}\right)dt^2- \frac{{dr}^2}{\left(1-\frac{2M}{r}\right)}-r^2(d\theta^2+\sin^2\theta
d\phi^2),\nonumber\\
\end{aligned}$$ where $M$ is the total mass of the stellar system.
Now, substituting Eq. (\[2.12\]) into Eq. (\[2.2\]) we find $$\begin{aligned}
\label{2.14}
{{\rm e}^{-\lambda \left( r \right) }}=1-{\frac {2m}{r}}.
\end{aligned}$$
Solution of the Einstein Field equations {#sec3}
========================================
Solving Einstein field equations (\[2.2\])-(\[2.4\]) and using the Eqs. (\[2.5\])-(\[2.7\]),(\[2.10\])-(\[2.12\]) we obtain expressions for the different physical parameters, which are given as $$\begin{aligned}
\label{3.1}
&\qquad\hspace{-2cm} \lambda \left(r \right) =-\ln \left[{\frac {\lambda_{{3}}{r}^
{4}-80 \lambda_{{1}} \left(B\chi+2\pi \rho_{{c}} \right) {r}^{2}
{R}^{2}+60\lambda_{{1}}{R}^{2}}{ 15\left(3\chi c_{{1}}+4\pi +3
\chi \right) {R}^{2}}} \right],\\\label{3.2}
&\qquad\hspace{-1cm} \nu\left(r\right)=\frac {1}{36864 \nu_{{4}} \left(-3 \chi c_{{1}}+12 \pi +\chi \right) } \Bigg[\nu_{{3}}{\rm arctanh} \Big\lbrace\Big(\big[16\nu_{{2}}B{R}^{5}\nonumber\\
&\qquad\hspace{-1cm}-32 B{R}^{3}{r}^{2}\nu_{{2}}+6 M{r}^{2} \left(\pi +\chi \right) \big] \lambda_{{2}}-5 \lambda_{{1}}M\pi {R}^{2}\Big)/{16\nu_{{4}}}\Big\rbrace\nonumber\\
&\qquad\hspace{-1cm} -294912\big\lbrace\left( \frac{3}{8}c_{{1}}+\frac{1}{2} \right) \chi+\pi \big\rbrace \nu_{{4}}\ln\Big[384 {R}^{5}{r}^{2}\lambda_{{2}}\nu_{{2}}B\nonumber\\
&\qquad\hspace{-1cm}-384 \lambda_{{2}}{r}^{4}\nu_{{2}}B{R}^{3}-120{R}^{2}{r}^{2}\lambda_{{1}}M\pi +72 M{r}^{4} \left( \pi +\chi \right) \lambda_{{2}}\nonumber\\
&\qquad\hspace{-1cm}+\nu_{{5}}\Big] -\nu_{{3}}{\rm arctanh}\Big[\big\lbrace -\lambda_{{2}}\nu_{{2}}B{R}^{5}-{\frac {5 \lambda_{{1}}M\pi {R}^{2}}{16}}\nonumber\\
&\qquad\hspace{-1cm}+\frac{3}{8}\lambda_{{2}}{R}^{2}M \left(\pi +\chi \right)\big\rbrace/{\nu_{{4}}}\Big] +442368 \nu_{{4}} \Big\lbrace\Big[ \frac{2}{3}\pi + \left(\frac{1}{4}c_{{1}}+\frac{1}{3}\right) \chi\Big]\nonumber\\
&\qquad\hspace{-1cm} \ln\Big\lbrace 24 {R}^{4}\nu_{{1}} \left(R-2M \right) \Big\rbrace +\ln \left(1-{\frac {2M}{R}} \right) \lambda_{{2}} \Big\rbrace\Bigg], \\\label{3.3}
&\qquad\hspace{-3.8cm} \rho_{{{\it eff}}}={\frac {12 {{\lambda}_2} \left( \rho_{{0}}-\rho_{{c}} \right) {r}^{2}+6{R}^{2} \left( B\chi+2\pi\rho_{{c}} \right) }{12\pi{R}^{2}}},\\ \label{3.4}
&\qquad\hspace{-3.5cm} p_{{{\it effr}}}=-{\frac { \left(3\chi c_{{1}}+4 \pi +3\chi \right) \left( \rho_{{c}}-\rho_{{0}} \right) {r}^{2}+p_{{1}}{R}^{2}}{12\pi{R}^{2}}},\\ \label{3.5}
&\qquad\hspace{-1cm} p_{{{\it efft}}}= \Big[\Big\lbrace 3\left( c_{{1}}+\frac{1}{3}\right) \left(\rho_{{0}}-\rho_{{c}} \right) {r}^{2}-5\left( B-\frac{2}{5}\rho_{{c}} \right) {R}^{2}\Big\rbrace \chi\nonumber\\
&\qquad\hspace{-1cm} -8\pi\Big\lbrace -\frac{3}{4}c_{{1}} \left( \rho_{{0}}-\rho_{{c}} \right) {r}^{2}+{R}^{2} \left(B-\frac{1}{4}\rho_{{c}} \right)\Big\rbrace\Big]\Bigg/6\pi{R}^{2},\nonumber\\\end{aligned}$$ where $\lambda_{{1}}$, $\lambda_{{2}}$, $\lambda_{{3}}$, $\nu_{{1}}$, $\nu_{{2}}$, $\nu_{{3}}$, $\nu_{{4}}$, $\nu_{{5}}$, ${\rho}_c$ and ${\rho}_0$ are constants and their expressions are shown in Appendix.
![Variation of (i) ${e}^{\nu(r)}$ (upper panel) and (ii) ${e}^{\lambda(r)}$ (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$. Here $B=83~ MeV/{{fm}^3}$ and $c_1=0.2$.[]{data-label="Fig1"}](nu "fig:"){width="6cm"} ![Variation of (i) ${e}^{\nu(r)}$ (upper panel) and (ii) ${e}^{\lambda(r)}$ (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$. Here $B=83~ MeV/{{fm}^3}$ and $c_1=0.2$.[]{data-label="Fig1"}](lambda "fig:"){width="6cm"}
![Variation of i) ${\rho}_{\it eff}$ (upper panel), (ii) $p_{{{\it effr}}}$ (middle panel) and (iii) $p_{{{\it efft}}}$ (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig2"}](density "fig:"){width="6cm"} ![Variation of i) ${\rho}_{\it eff}$ (upper panel), (ii) $p_{{{\it effr}}}$ (middle panel) and (iii) $p_{{{\it efft}}}$ (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig2"}](rad_press "fig:"){width="6cm"} ![Variation of i) ${\rho}_{\it eff}$ (upper panel), (ii) $p_{{{\it effr}}}$ (middle panel) and (iii) $p_{{{\it efft}}}$ (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig2"}](tan_press "fig:"){width="6cm"}
The variation of the physical parameters, viz., ${e}^{\lambda}$, ${e}^{\nu}$, ${\rho}_{{{\it eff}}}$, ${p}_{{{\it effr}}}$ and ${p}_{{{\it efft}}}$ with respect to the radial coordinate ($r/R$) in the framework of $f\left( R,\mathcal{T}\right)$ gravity theory are shown in Figs. \[Fig1\] and \[Fig2\].
The anisotropy $\left(\Delta\right)$ for our system reads as $$\label{3.6}
\Delta=\frac { \left( c_{{1}}-\frac{1}{3}\right) \left( \pi +\frac{\chi}{4} \right) \left( \rho_{{0}}-\rho_{{c}} \right) {r}^{2}}{\pi \,{R}^{2}}.$$
![Variation of anisotropy as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig3"}](anisotropy){width="6cm"}
The variation of anisotropy with respect to radial coordinate $r/R$ is featured in Fig. \[Fig3\]. We find here in $f(R,\mathcal{T})$ gravity model that the anisotropy for the strange star is minimum at the center and maximum on the surface as prediction by Deb et al. [@Deb2017] in the case of GR.
The modified form of the energy conservation equation for the stress-energy tensor in the framework of $f\left(R,\mathcal{T}\right)$ can be written explicitly from Eq. (\[1.6\]) as $$\begin{aligned}
\label{3.7}
&\qquad\hspace{-4cm} -p_r^{{\prime}}-\frac{1}{2}\nu^{{\prime}} \left( \rho+p_r \right)+\frac{2}{r}\left({p_t}-{p_r}\right)\nonumber\\
&\qquad\hspace{1cm} -{\frac {\chi}{8\,\pi +2\,\chi}}\left(\rho^{{\prime}}+p_r^{{\prime}} +2p_t^{{\prime}} \right)=0.\end{aligned}$$
Now using Eqs. (\[2.3\]), (\[2.11a\]), (\[2.14\]) and (\[3.7\]) we find the hydrostatic equation for an anisotropic spherically symmetric compact stars in the framework of $f\left(R,\mathcal{T}\right)$ theory of gravity as follows $$\begin{aligned}
\label{3.8}
&\qquad p_r^{\prime}=-\Big[\big\lbrace 4 \pi {r}^{2}p_{{r}}+{\frac {m}{r}}+\frac{1}{2}\chi \left( \rho+3 p_{{r}}+2 p_{{t}} \right) {r}^{2} \big\rbrace \nonumber\\
&\qquad \left( \rho+p_{{r}} \right) -2 \left( p_{{t}}-p_{{r}} \right) \left( 1-{\frac {2m}{r}} \right)\Big]\Big/\Big[ r \left( 1-{\frac {2m}{r}} \right)\nonumber\\
&\qquad \big\lbrace 1+\frac {\chi}{8 \pi +2 \chi}\big[ 1+ {\frac {{\rm d}\rho}{{\rm d}p_{{r}}}}\left( 1+2 c_{{1}} \right) \big] \big\rbrace\Big],\end{aligned}$$ where we assume that SQM density parameter $\rho$ depends on it’s radial pressure ${p}_{r}$ as ${\rho}={\rho} \left({{p}_{r}}\right)$. For ${\chi}=0$ Eq. (\[3.8\]) reduces to the standard form of the TOV equation as found in GR. Now using Eqs. (\[2.10\]) and (\[2.11\]) and also considering Bag constant $B=83~MeV/{{fm}^{3}}$ [@Rahaman2014] with ${c_1}={0.2}$ we obtain exact solution of the Eq. (\[3.8\]). Here, using the observed values of the mass of different strange stars as presented in Table \[Table 2\] we can predict radii of the strange stars.
Physical properties of the anisotropic stellar system in $f\left(R,\mathcal{T}\right)$ theory of gravity {#sec4}
========================================================================================================
In this section we shall test physical validity of the obtained solutions in the framework of $f \left(R,\mathcal{T}\right)$ theory of gravity. To this end, we study the energy conditions, Herrera cracking concept, adiabtic index, etc., in the following subsections.
Energy conditions in $f\left(R,\mathcal{T}\right)$ gravity {#subsec4.1}
----------------------------------------------------------
Our system will be consistent with the energy conditions, viz., the null energy condition (NEC), weak energy condition (WEC), strong energy condition (SEC) and dominant energy condition (DEC) only if it satisfy the following inequalities simultaneously, given as [@SC2013] $$\begin{aligned}
\label{4.1.1}
&\qquad\hspace{-1cm} NEC:{\rho}_{eff}+p_{effr}\geq 0,~{{\rho}_{eff}}+{p_{efft}}\geq 0, \\ \label{4.1.2}
&\qquad\hspace{-1.2cm} WEC: {{\rho}_{eff}}+p_{effr}\geq 0,~{{\rho}_{eff}}\geq 0,~{{\rho}_{eff}}+{p_{efft}}\geq 0, \nonumber \\ \\ \label{4.1.3}
&\qquad\hspace{-1.4cm} SEC: {{\rho}_{eff}}+p_{effr}\geq 0,~{{\rho}_{eff}}+p_{efft}\geq 0,\nonumber \\
&\qquad\hspace{2cm} ~{{\rho}_{eff}}+{p_{effr}}+2 {p_{efft}}\geq 0, \\ \label{4.1.4}
&\qquad\hspace{-1.4cm} DEC: {{\rho}_{eff}}\geq 0,~{{{\rho}_{eff}}-{p_{effr}}}\geq 0,~{{{\rho}_{eff}}-{p_{efft}}}\geq 0. \end{aligned}$$
![Variation of energy conditions with the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig4"}](EC0 "fig:"){width="4.5cm"} ![Variation of energy conditions with the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig4"}](EC04 "fig:"){width="4.5cm"} ![Variation of energy conditions with the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig4"}](EC08 "fig:"){width="4.5cm"} ![Variation of energy conditions with the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig4"}](EC12 "fig:"){width="4.5cm"} ![Variation of energy conditions with the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig4"}](EC16 "fig:"){width="4.5cm"}
The variation of the energy conditions with respect to the radial coordinate $r/R$ for the parametric values of $\chi$ are presented in Fig. \[Fig4\], which clearly suggests that our system is consistent with all the energy conditions.
Mass-radius relation {#subsec4.2}
--------------------
Substituting Eqs. (\[2.6\]), (\[2.10\]) and (\[2.11a\]) into Eq. (\[2.12\]) the mass function for the present system is given as $$\begin{aligned}
\label{4.2.1}
\hspace{-0.5cm} m=\tilde{m} \Big\lbrace 1+{\frac {\chi}{12\pi }}\left( 1-3 c_{{1}} \right) \Big\rbrace +\frac{2}{9}\chi {r}^{3} \left( B-\frac{3}{2}c_{{2}} \right),\end{aligned}$$ where $\tilde{m}=4 \pi \int_{0}^{r}\!{{\rho}} \left( r \right) {r}^{2}{dr}$ is the mass function of the SQM fluid distribution. For $\chi=0$ Eq. (\[4.2.1\]) reduce to the mass function $m(r)=\tilde{m}(r)$ as achieved in GR. Hence, clearly for $\chi \neq 0$ the coupling between matter and curvature terms produces a new kind of matter distribution having mass given as $m_{new}=\chi \lbrace {\frac {\tilde{m}}{12\pi }}\left( 1-3 c_{{1}} \right) +\frac{2}{9}{r}^{3} \left( B-\frac{3}{2}c_{{2}} \right) \rbrace$. We have shown variation of the total mass ($M$), normalized in solar masses ($M_{\odot}$), with respect to the total radius ($R$) in Fig. \[Fig5\] for different values of $\chi$ and for a specific value of the bag constant as $B=83~ MeV/{{fm}^3}$ [@Rahaman2014]. Fig. \[Fig5\] shows that the mass-radius relation for the strange stars in $f\left(R,\mathcal{T}\right)$ gravity has achieved typical behaviour as in GR. Also, we find that for the chosen increasing values of $\chi$, i.e., $\chi=0$, $0.4$, $0.8$, $1.2$ and $1.6$ the values of the maximum masses are increasing gradually.
![Mass $(M/{M_{\odot}})$ vs Radius ($R$ in km) curve for the strange stars due to the different values of $\chi$. The solid circles are representing the maximum mass points for the strange stars.[]{data-label="Fig5"}](MR_curve){width="8cm"}
![Variation of (i) $M/{M_{\odot}}$ (left panel) and (ii) $R$ in $km$ (right panel) as a function of the central density $({{\rho}_{effc}}~in~MeV/{fm}^3)$ due to different values of $\chi$. The solid circles are representing the maximum mass points for the strange stars.[]{data-label="Fig6"}](M_max "fig:"){width="6cm"} ![Variation of (i) $M/{M_{\odot}}$ (left panel) and (ii) $R$ in $km$ (right panel) as a function of the central density $({{\rho}_{effc}}~in~MeV/{fm}^3)$ due to different values of $\chi$. The solid circles are representing the maximum mass points for the strange stars.[]{data-label="Fig6"}](R_max "fig:"){width="6cm"}
In Fig. \[Fig6\] the variation of $M$, normalized in $M_{\odot}$ and the variation of $R$ with respect to the central density, ${\rho}_{effc}$ are shown in the upper and lower panel, respectively. The upper panel of Fig. \[Fig6\] features that as the values of $\chi$ increases the maximum mass points are achieved for the lower values of $({\rho}_{effc})$. We find, for $\chi=0$ the maximum mass $M_{max}=2.951~{{M}_{\odot}}$ is obtained for ${\rho}_{effc}=2.14\times {{10}^{15}}~gm/{{cm}^3}$. On the other hand, for $\chi=1.6$ the maximum mass increases to the value $M_{max}=~3.464 {{M}_{\odot}}$ and the corresponding value of the central density decreases to ${\rho}_{effc}=1.78\times {{10}^{15}}~gm/{{cm}^3}$. The lower panel of Fig. \[Fig6\] presents that as the value of $\chi$ increases the value of the radius increases gradually. We find, for $\chi=0$ the radius corresponding to the maximum mass point is $R_{Mmax}=10.498~km$ and as $\chi$ increases to the value $\chi=1.6$ the radius corresponding to the maximum mass point also increases to the value $R_{Mmax}=11.779~km$. Hence, as the value of $\chi$ increases both the mass and the radius of the strange stars increase and the stars become less compact.
Stability of the stellar model {#subsec4.3}
------------------------------
To discuss stability of the stellar model we shall study (i) Modified form of the TOV equation in $f\left(R,\mathcal{T}\right)$ gravity, (ii) Herrera cracking concept and (iii) Adiabatic index in the following sub-subsections.
### Modified form of the TOV equation in $f\left(R,\mathcal{T}\right)$ gravity {#subsubsec4.3.1}
We have presented the modified form of the energy conservation equation for the stress-energy tensor in the framework of $f\left(R,\mathcal{T}\right)$ theory of gravity in Eq. (\[1.6\]) and later we have shown it in a more concise form in Eq. (\[1.7\]). Hence the modified form of the Tolman-Oppenheimer-Volkoff (TOV) equation as already presented in Eq. (\[3.7\]) is given as $$\begin{aligned}
&\qquad\hspace{-4cm} -p_r^{{\prime}}-\frac{1}{2}\nu^{{\prime}} \left( \rho+p_r \right)+\frac{2}{r}\left({p_t}-{p_r}\right)\nonumber\\
&\qquad\hspace{1cm} -{\frac {\chi}{8\,\pi +2\,\chi}}\left(\rho^{{\prime}}+p_r^{{\prime}} +2p_t^{{\prime}} \right)=0,\end{aligned}$$ where the first term represents the hydrodynamic force $(F_h)$, the second term denotes gravitational force $F_g$ and the third term indicates anisotropic force $(F_a)$. Here, the last term is the resultant of the coupling between the matter and the geometry and we are introducing it as the ’modified force’ $F_m$. Hence, the modified TOV equation predicts that in $f\left(R,\mathcal{T}\right)$ gravity also, sum of all the forces are zero, i.e., ${F_h}+{F_g}+{F_a}+{F_m}=0$. So, in terms of equilibrium of the forces our system is completely stable. Clearly, for $\chi=0$ the extra force term $F_m$ will be zero and the usual form of the TOV equation as in GR will be retrieved.
![Variation of the different forces with respect to the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig7"}](F0 "fig:"){width="4.5cm"} ![Variation of the different forces with respect to the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig7"}](F04 "fig:"){width="4.5cm"} ![Variation of the different forces with respect to the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig7"}](F08 "fig:"){width="4.5cm"} ![Variation of the different forces with respect to the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig7"}](F12 "fig:"){width="4.5cm"} ![Variation of the different forces with respect to the radial coordinate $r/R$ for $LMC\,X-4$ due to different chosen values of $\chi$.[]{data-label="Fig7"}](F16 "fig:"){width="4.5cm"}
In Fig. \[Fig7\] we have shown variation of the different forces against the radial coordinate $r/R$ due to different chosen values of $\chi$. We find that the equilibrium of the forces is achieved due to all the values of $\chi$ and confirms stability of the system. Fig. \[Fig7\] features that the inward pull of $F_g$ is counter balanced by the combined effect of $F_h$, $F_a$ and $F_m$ which acts along the outward direction. Hence, we find that the nature of the modified force, $F_m$ is repulsive and acts along the outward directions.
### Herrera cracking concept {#subsubsec4.3.2}
To establish stability of the stellar system now we shall study the concept of Herrera’s cracking. For a physically acceptable stellar system the causality condition must be satisfied, which demands that square of the radial $(v^2_{sr})$ and tangential $(v^2_{st})$ sound speeds should lie within the limit $\left[0,1\right]$, i.e., explicitly $0 \leq v^2_{sr} \leq 1$ and $0 \leq v^2_{st} \leq 1$. According to Herrera [@Herrera1992] and Abreu [@Abreu2007] for a physically stable stellar system made of anisotropic fluid distribution the difference of square of the sound speeds should maintain it’s sign inside the stellar system and specially for a potentially stable region square of the radial sound speed should be greater than the square of the tangential sound speeds. Hence, according to Herrera’s cracking concept the required condition is $|{v^2_{st}}- {v^2_{sr}}|\leq 1$. For our system $(v^2_{sr})$ and $(v^2_{st})$ are given as $$\begin{aligned}
\label{4.3.2.1}
&\qquad {v^2_{{{\it sr}}}}={\frac {3 \chi c_{{1}}+4 \pi +3 \chi}{-3 \chi c_{{1}}+12 \pi +\chi}}, \\ \label{4.3.2.2}
&\qquad {v^2_{{{\it st}}}}={\frac {2 \left(6 \pi c_{{1}}+3 \chi c_{{1}}+\chi\right)}{-3 \chi c_{{1}}+12 \pi +\chi}}.\end{aligned}$$
![Variation of $|{v^2_{st}}- {v^2_{sr}}|$ with the radial coordinate $r/R$ for $LMC\,X-4$.[]{data-label="Fig8"}](diff_soundvel){width="6cm"}
We have featured variation of $|{v^2_{st}}- {v^2_{sr}}|$ with respect to the radial coordinate in Fig. \[Fig8\] and as $|{v^2_{st}}- {v^2_{sr}}|\leq 1$, so our system is consistent with the concept of Herrera’s cracking, which again confirms the stability of our stellar system.
### Adiabatic Index {#subsubsec4.3.3}
The stability of both the relativistic and non-relativistic stars can be examined by studying adiabatic index $(\Gamma)$ of the system. For a given density it also can characterize the stiffness of the EOS. Following the pioneering work by Chandrasekhar [@Chandrasekhar1964] several authors [@Hillebrandt1976; @Chan1994; @Herrera1997; @Horvat2010; @Doneva2012; @Silva2015] studied the dynamical stability of the stellar system against an infinitesimal radial perturbation. For a dynamically stable stellar system Heintzmann and Hillebrandt [@Heintzmann1975] have shown that adiabatic indices should exceed $4/3$ inside the stellar system. Now the radial $({\Gamma}_r)$ and tangential $({\Gamma}_t)$ adiabatic indices can be defined as $$\begin{aligned}
\label{4.3.3.1}
& \qquad {\Gamma}_r=\frac{{p_{effr}}+{{\rho}_{eff}}}{{p_{effr}}}\,\frac{d{p_{effr}}}{d{{\rho}_{eff}}}=\frac{{p_{effr}}+{{\rho}_{eff}}}{{p_{effr}}}\,{v^2_{sr}},\\\label{4.3.3.2}
& \qquad {\Gamma}_t=\frac{{p_{efft}}+{{\rho}_{eff}}}{{p_{efft}}}\,\frac{d{p_{efft}}}{d{{\rho}_{eff}}}=\frac{{p_{efft}}+{{\rho}_{eff}}}{{p_{efft}}}\,{v^2_{st}}. \end{aligned}$$
![Variation of i) ${\Gamma}_r$ (upper panel) and ii) ${\Gamma}_t$ (lower panel) with the radial coordinate $r/R$ for $LMC\,X-4$.[]{data-label="Fig9"}](rad_adiaindex "fig:"){width="6cm"} ![Variation of i) ${\Gamma}_r$ (upper panel) and ii) ${\Gamma}_t$ (lower panel) with the radial coordinate $r/R$ for $LMC\,X-4$.[]{data-label="Fig9"}](tan_adiaindex "fig:"){width="6cm"}
In Fig. \[Fig8\] we have shown the variation of ${\Gamma}_r$ (upper panel) and ${\Gamma}_t$ (lower panel) against the radial coordinate $r/R$ which demonstrate that in both the cases the values of the adiabatic indices are greater than $4/3$ through out the system. Hence, our system is completely stable against the radial pulsations.
Compactification factor and redshift {#subsec4.4}
------------------------------------
The compactification factor $(u)$ for our system is expressed by $$\begin{aligned}
\label{4.4.1}
&\qquad\hspace{-2cm} u(r)=\frac{m(r)}{r}=-\frac {1}{ \lbrace \left( \frac{3}{8}c_{{1}}-\frac{1}{8}\right) {\chi}^{2}+\frac{9}{4}\pi\left( c_{{1}}+\frac{1}{9}\right) \chi+{\pi }^{2} \rbrace {R}^{5}}\nonumber \\
&\qquad\hspace{-2cm}\Big[ 8r^2 \Big\lbrace -\frac{1}{32}B \left( R^2-r^2 \right) \left( c_{{1}}-\frac{1}{3}\right){R}^{3}{\chi}^{3}\nonumber \\
&\qquad\hspace{-2cm}+ \big[ -\frac{3}{16}B{R}^{3} \left( R^2-r^2 \right)\left( c_{{1}}-1 \right) \pi -{\frac { \left( 3
c_{{1}}-1 \right) M{r}^{2}}{64}} \big] {\chi}^{2}\nonumber \\
&\qquad\hspace{-2cm}-\frac{1}{4}\pi \big\lbrace B \left( R^2-r^2 \right) \left( c_{{1}}-\frac{10}{3}
\right) {R}^{3}\pi +\frac{1}{16}\big[15 M \big\lbrace \left( c_{{1}}+1 \right) {R}^{2}\nonumber \\
&\qquad\hspace{-2cm}+\frac{1}{5} \left( c_{{1}}-\frac{13}{3} \right) {r}^{2} \big\rbrace \big] \big\rbrace \chi+ \big[ \left( B{R}^{5}-B{R}^{3}{r}^{2} \right) \pi\nonumber \\
&\qquad\hspace{-2cm} -\frac {1}{16}\left( 5{R}^{2}-3{r}^{2} \right) M \big] {\pi }^{2}\Big\rbrace \Big].\end{aligned}$$
Again, expression for the redshift function in the present model is given as $$\begin{aligned}
&\qquad\hspace{-5cm} Z={e^{-{\nu(r)}/2}}-1\nonumber \\
&\qquad\hspace{-1cm} =exp\Bigg\lbrace -\frac {1}{73728\,\nu_{{4}} \left( -3\,\chi\,c_{{1}}+12\,\pi +\chi \right) } \Big[ \nu_{{3}}{\rm arctanh} \Big\lbrace\Big(\big[ 16\,\nu_{{2}}B{R}^{5}\nonumber\\
&\qquad\hspace{-1cm}-32\,B{R}^{3}{r}^{2}\nu_{{2}}+6\,M{r}^{2} \left(\pi +\chi \right) \big] \lambda_{{2}}-5\,\lambda_{{1}}M\pi \,{R}^{2}\Big)/{16\nu_{{4}}}\Big\rbrace\nonumber\\
&\qquad\hspace{-1cm} -294912\big\lbrace\left( \frac{3}{8}c_{{1}}+\frac{1}{2} \right) \chi+\pi \big\rbrace \nu_{{4}}\ln\Big[ 384\,{R}^{5}{r}^{2}\lambda_{{2}}\nu_{{2}}B\nonumber\\
&\qquad\hspace{-1cm}-384\,\lambda_{{2}}{r}^{4}\nu_{{2}}B{R}^{3}-120{R}^{2}{r}^{2}\lambda_{{1}}M\pi +72\,M{r}^{4} \left( \pi +\chi \right) \lambda_{{2}}\nonumber\\
&\qquad\hspace{-1cm}+\nu_{{5}}\Big] -\nu_{{3}}{\rm arctanh}\Big[\big\lbrace -\lambda_{{2}}\nu_{{2}}B{R}^{5}-{\frac {5\,\lambda_{{1}}M\pi \,{R}^{2}}{16}}\nonumber\\
&\qquad\hspace{-1cm}+\frac{3}{8}\lambda_{{2}}{R}^{2}M \left( \pi +\chi \right)\big\rbrace/{\nu_{{4}}}\Big] +442368\,\nu_{{4}} \Big\lbrace\Big[ \frac{2}{3}\pi + \left( \frac{1}{4}c_{{1}}+\frac{1}{3}\right) \chi\Big]\nonumber\\
&\qquad\hspace{-1cm} \ln\Big\lbrace 24\,{R}^{4}\nu_{{1}} \left(R-2M \right) \Big\rbrace +\ln \left( 1-{\frac {2M}{R}} \right) \lambda_{{2}} \Big\rbrace\Big] \Bigg\rbrace-1.\nonumber\\\end{aligned}$$
![Variation of the (i) compactification factor (upper panel) and (ii) redshift (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig10"}](compactness "fig:") ![Variation of the (i) compactification factor (upper panel) and (ii) redshift (lower panel) as a function of the radial coordinate $r/R$ for the strange star $LMC~X-4$.[]{data-label="Fig10"}](redshift "fig:")
We have featured the variation of the compactification factor and the redshift function with respect to the radial coordinate $r/R$ in Fig. \[Fig10\].
Discussion and Conclusion {#sec5}
=========================
Present article serves our motivation to explore the possibility of existence of the anisotropic ultra dense strange quark stars in the framework of the $f\left(R,\mathcal{T}\right)$ theory of gravity. To this end, following Harko et al. [@Harko2011] we have considered simplified linear form of the arbitrary function $f\left(R,\mathcal{T}\right)$ given as $f(R,\mathcal{T})=R+2\chi\mathcal{T}$. In Eq. (\[1.5\]) we present the field equation due to the modified EH action in $f\left(R,\mathcal{T}\right)$ gravity. Eq. (\[1.5\]) clearly indicates that our system is not made of only the SQM but also a second kind of unknown matter is produced as a coupling effect of the matter and geometry. In this context, one may consult the article by Chakraborty [@SC2013], where he studied the nature and origin of this second kind of matter distribution which has been produced due to the effect of $f\left(R,\mathcal{T}\right)$ gravity.
Now considering the stress-energy tensor due to the effective matter distribution as $T_{\mu\nu,eff}$ we find the standard form of the energy conservation equation in Eq. (\[1.7\]). To solve the Einstein field equations \[(\[2.2\])-(\[2.4\])\] we have considered that the SQM matter distribution is governed by the simplified MIT Bag EOS (\[2.10\]) and assumed a relation between $p_t$ and $\rho$, as given in Eq. (\[2.11a\]). Throughout the study we have been considering $B=83~MeV/{{fm}^3}$ [@Rahaman2014], $c_1=0.2$ and the values of $\chi$ as $0$, $0.4$, $0.8$, $1.2$ and $1.6$. We have illustrated the results graphically for $LMC~X-4$ as the representative of the strange quark stars.
We have presented features of the metric potentials, viz., $e^{\nu}$ (in the upper panel) and $e^{\lambda}$ (in the lower panel) in Fig. \[Fig1\], which indicate that our stellar model is free from the geometrical singularity. In Fig. \[Fig2\] we have shown the variation of the energy density $({\rho}_{eff})$, radial pressure $(p_{effr})$ and tangential pressure $(p_{efft})$ in the upper, middle and lower panel, respectively. We find that ${\rho}_{eff}$, $p_{effr}$ and $p_{efft}$ are maximum at the center and decrease monotonically inside the spherical system to achieve the minimum value on the surface. The anisotropy of the system is featured in Fig. \[Fig3\], which shows that anisotropy is minimum, i.e., zero at the center and maximum on the surface in the present $f\left(R,\mathcal{T}\right)$ gravity model as the prediction made by Deb et al. [@Deb2017] in the case of GR that maximum anisotropy on the surface is the inherent property of the anisotropic strange stars.
Values of $\chi$ $\chi=0$ $\chi=0.4$ $\chi=0.8$ $\chi=1.2$ $\chi=1.6$
--------------------------------- --------------------------- --------------------------- --------------------------- --------------------------- --------------------------- -- -- -- -- -- -- -- -- -- -- -- --
Predicted Radius *(Km)* 9.678 9.821 9.926 10.001 10.054
${\rho}_{effc}$ ($gm/{{cm}^3}$) $7.985\times {{10}^{14}}$ $7.535\times {{10}^{14}}$ $7.213\times {{10}^{14}}$ $6.986\times {{10}^{14}}$ $6.820\times {{10}^{14}}$
${\rho}_{eff0}$ ($gm/{{cm}^3}$) $5.927\times {{10}^{14}}$ $5.743\times {{10}^{14}}$ $5.619\times {{10}^{14}}$ $5.538\times {{10}^{14}}$ $5.488\times {{10}^{14}}$
${p}_{effc}$ ($dyne/{{cm}^2}$) $6.166\times {{10}^{34}}$ $5.961\times {{10}^{34}}$ $5.822\times {{10}^{34}}$ $5.755\times {{10}^{34}}$ $5.724\times {{10}^{34}}$
${2M}/{R}$ 0.393 0.387 0.383 0.381 0.379
$Z_s$ 0.284 0.277 0.273 0.271 0.269
---------------- ------------------------------- -------------------- ---------------------------- ---------------------------- ---------------------------- ---------- ---------------- -- -- -- -- -- -- -- --
Strange Observed Predicted ${{\rho}_{effc}}$ ${{\rho}_{eff0}}$ ${{p}_{effc}}$ Surface $\frac{2M}{R}$
Stars Mass ($M_{\odot}$) Radius $(Km)$ $(gm/{cm}^3)$ $(gm/{cm}^3)$ $(dyne/{cm}^2)$ Redshift
$Vela~X-1$ $1.77 \pm 0.08$ [@dey2013] $10.866 \pm 0.133$ $7.965 \times {{10}^{14}}$ $5.597 \times {{10}^{14}}$ $8.646 \times {{10}^{34}}$ 0.388 0.481
$4U~1820-30$ $1.58 \pm 0.06$ [@guver2010b] $10.529 \pm 0.113$ $7.642\times {{10}^{14}}$ $5.606\times {{10}^{14}}$ $7.433\times {{10}^{34}}$ 0.340 0.443
$Cen~X-3$ $1.49 \pm 0.08$ [@dey2013] $10.354 \pm 0.160$ $7.502\times {{10}^{14}}$ $5.611\times {{10}^{14}}$ $6.907\times {{10}^{34}}$ 0.319 0.425
$LMC~X - 4$ $1.29 \pm 0.05$ [@dey2013] $9.926 \pm 0.115$ $7.213\times {{10}^{14}}$ $5.619\times {{10}^{14}}$ $5.822\times {{10}^{34}}$ 0.273 0.383
$SMC~X - 1$ $1.04 \pm 0.09$ [@dey2013] $9.299 \pm 0.248$ $6.895\times {{10}^{14}}$ $5.628 \times {{10}^{14}}$ $4.624 \times {{10}^{34}}$ 0.222 0.330
---------------- ------------------------------- -------------------- ---------------------------- ---------------------------- ---------------------------- ---------- ---------------- -- -- -- -- -- -- -- --
To examine the physical acceptability of the proposed anisotropic stellar model in $f\left(R,\mathcal{T}\right)$ gravity, we have studied the energy conditions, mass-radius relation, stability of the stellar system, etc. Fig. \[Fig4\] features that our system is consistent with all the energy conditions. Further, in Fig. \[Fig5\] we have presented the total mass $M$ (normalized in $M_{\odot}$) versus the total radius $R$ relations for the chosen values of $\chi$ and $B=83~MeV/{{fm}^3}$. The solid circles in Fig. \[Fig5\] are representing the maximum mass points for the strange stars. The figure shows that as the value of $\chi$ increases the value of $M$ and $R$ also gradually increases and thus provides a proportionality relation. We find for $\chi=1.6$ the value of $M_{max}$ increases to $17.36\%$ and $R_{Mmax}$ increases to $12.2\%$ than its corresponding value in GR and becomes $M_{max}=3.464~{M_{\odot}}$ and $R_{Mmax}=11.774~km$, respectively. In Fig. \[Fig6\] we have presented variation of $M$ (in the upper panel) and $R$ (in the lower panel) with respect to the central density of the effective matter distribution, ${\rho}_{effc}$. Fig. \[Fig6\] shows that with increase of $\chi$ the value of density decreases. For example, due to $\chi=1.6$ the maximum mass point $M_{max}=3.464~{M_{\odot}}$ is achieved for ${{\rho}_{effc}}=7.739\,{{\rho}_{nuclear}}$, which is $16.823\%$ lower than its value in GR. So, as the value of $\chi$ increases the strange stars become massive and bigger and thus show a gradual derease in its density.
To show stability of the system in terms of the equilibrium of forces we have studied modified TOV equation in the framework of $f\left(R,\mathcal{T}\right)$ theory of gravity. The variation of all the forces are featured in Fig. \[Fig7\], which confirms that our system is stable in terms of the equilibrium of forces. Fig. \[Fig7\] also features an interesting fact that an extra force $F_m$ is produced due to the coupling effect between the matter and geometry. We introduced this force, $F_m$, as the modified force. We find that $F_m$ is repulsive in the nature and acts along the outward direction in the stellar system. To examine stability we also studied the Herrera cracking concept [@Herrera1992; @Abreu2007] and presented variation of the difference in square of the sound speeds, $|{v^2_{st}}- {v^2_{sr}}|$ against the radial coordinate $r/R$ in Fig. \[Fig8\]. We found our system is consistent with the causality condition and the Herrera cracking concept. Further, in Fig. \[Fig9\] we presented variation of both the adiabatic indices ${\Gamma}_r$ and ${\Gamma}_t$ with respect to the radial coordinate $r/R$ and have concluded that as both the adiabatic indices ${\Gamma}_r$ and ${\Gamma}_t$ are greater than $4/3$ so our system is stable against the radial pulsation. We also presented the variation of the compactification factor and the redshift in the upper and the lower panel, respectively in Fig. \[Fig10\].
In TABLE \[Table 1\] we have predicted different physical parameters for the observed values of the mass of $LMC\,X-4$ for $B=83~MeV/{{fm}^3}$ [@Rahaman2014], $c_1=0.2$ and the chosen values of $\chi$ as $0$, $0.4$, $0.8$, $1.2$ and $1.6$. We find as the coupling parameter, $\chi$ increases the mass $(M)$ and the radius $(R)$ of the star also increases gradually. However, TABLE \[Table 1\] shows that with the increasing value of $\chi$ the central $({\rho}_{effc})$ and surface density ${\rho}_{eff0}$, central pressure ${p}_{effc}$, surface redshift $(Z_s)$ and the value of $2M/R$ decreases gradually. Again, in TABLE \[Table 2\] we have presented above-mentioned physical parameters for different strange star candidates due to $\chi=0.8$. The high surface redshift values $(0.388-0.222)$ and surface density values $(7.965 \times {{10}^{14}}-6.895 \times {{10}^{14}}~gm/{cm}^3 )$ as presented in TABLE \[Table 2\] clearly indicate that the stellar candidates are ultra-dense strange stars [@Ruderman1972; @Glendenning1997; @Herjog2011]. It is also clear from both TABLES \[Table 1\] and \[Table 2\] that for all the values of $\chi$ our system is consistent with the Buchdahl condition [@Buchdahl1959], which demands a stringent condition $2M/R<8/9$. Now, as the compact stellar systems become gradually massive with the increment of $\chi$, hence our study reveals that the modified $f\left(R,\mathcal{T}\right)$ theory of gravity is a suitable theory to explain massive stellar systems like recent magnetars, massive pulsars and super-Chandrasekhar stars, which can not be explained in the framework of GR.
Again, by introducing $f\left(R,\mathcal{T}\right)=R+2\chi \mathcal{T}$ to consider the simplest minimal matter-geometry coupling, we have presented a similar and interesting result as presented by Astashenok et al. [@Astashenok2015]. The authors [@Astashenok2015] in their study presented a nonperturbative model of strange stars in $f\left(R\right)=R+\alpha {R}^2$ theory of gravity, where $\alpha$ is a constant. They showed that as the value of the constant parameter $\alpha$ increases mass of the strange star candidates increases gradually. In our present study we have also obtained the similar result for the increasing values of $\chi$, i.e., as the value of $\chi$ increases the stellar system becomes more massive gradually. It is interesting to note that the extra gravitational mass was arising in the case of $f\left(R\right)$ [@Astashenok2015] theory of gravity due to the extra geometrical term $\alpha {R}^2$, whereas in our study the same is obtained due to the extra material term $2\chi \mathcal{T}$. Hence, it is difficult to distinguish the effect of both the extra geometrical term and the material term on the ultra dense stellar configuration.
However, in the case of $f\left(R\right)$ gravity Astashenok et al. [@Astashenok2015] have obtained the maximum mass points due to the different values of $\alpha~(>0)$ for the higher values of central densities as compared to GR. On the contrary, in our study we find that for the different values of $\chi~(>0)$ the maximum mass points are achieved for the lower values of central densities as compared to GR. Also, Astashenok et al. [@Astashenok2015] showed that with the increasing values of $\alpha$ from $\alpha=0$ in $f\left(R\right)$ gravity values of the surface redshift increases gradually, whereas we find with the increasing values of $\chi$ from $\chi=0$ in $f\left(R, \mathcal{T}\right)$ theory of gravity the surface redshift decreases gradually. Thus by studying the central density and the surface redshift one may easily distinguish the effects and predictions of $f\left(R\right)$ and $f\left(R, \mathcal{T}\right)$ theory of gravities.
We can easily discriminate modified $f\left(R, \mathcal{T}\right)$ gravity from GR by noting at the surface redshift which has an inverse relationship with the parameter $\chi$. We also find that a stellar system becomes more massive in modified $f\left(R, \mathcal{T}\right)$ gravity compared to GR. It is worth mentioning that likewise GR in the present extended gravity theory too MIT bag model takes a suitable role to discuss strange star candidates.
As a final comment, in this paper we have successfully presented a stable and physically acceptable anisotropic stellar model, which is suitable to study ultra-dense strange stars in the framework of $f\left(R,\mathcal{T}\right)$ theory of gravity.
ACKNOWLEDGMENTS {#acknowledgments .unnumbered}
===============
SR and FR are thankful to the Inter-University Centre for Astronomy and Astrophysics (IUCAA), Pune, India for providing Visiting Associateship under which a part of this work was carried out. SR is also thankful to the authority of The Institute of Mathematical Sciences, Chennai, India for providing all types of working facility and hospitality under the Associateship scheme. FR is also grateful to DST-SERB (EMR/2016/000193), Govt. of India for providing financial support. A part of this work was completed while DD was visiting IUCAA and the author gratefully acknowledges the warm hospitality and facilities at the library there. We all are thankful to the anonymous referee for the pertinent comments which has helped us to upgrade the manuscript substantially.
[99]{}
A.G. Riess et al., Astron. J. **116**, 1009 (1998).
S. Perlmutter et al., Astrophys. J. **517**, 565 (1999).
P. de Bernardis et al, Nature **404**, 955 (2000).
R.A. Knop et al., Astrophys. J. **598**, 102 (2003).
R.R. Caldwell, Phys. Lett. B **545**, 23 (2002).
S. Nojiri and S.D. Odinstov, Phys. Lett. B **562**, 147 (2003).
S. Nojiri and S.D. Odinstov, Phys. Lett. B **565**, 1 (2003).
T. Padmanabhan and T.R. Choudhury, Phys. Rev. D **66**, 081301 (2002).
A. Kamenshchik, U. Moschella, and V. Pasquier, Phys. Lett. B **511**, 265 (2001).
M.C. Bento, O. Bertolami, and A.A. Sen, Phys. Rev. D **66**, 043507 (2002).
S. Nojiri and S.D. Odintsov, Phys. Rep. **505**, 59 (2011).
S. Nojiri and S.D. Odintsov, Phys. Rev. D **68**, 123512 (2003).
S.M. Carroll, V. Duvvuri, M. Trodden, and M.S. Turner, Phys. Rev. D **70**, 043528 (2004).
G. Allemandi, A. Borowiec, M. Francaviglia, and S.D. Odintsov, Phys. Rev. D **72**, 063505 (2005).
S. Nojiri and S.D. Odintsov, Int. J. Geom. Meth. Mod. Phys. **04**, 115 (2007).
O. Bertolami, C.G. Bohmer, T. Harko, and F.S.N. Lobo, Phys. Rev. D **75**, 104016 (2007).
K. Bamba, C.Q. Geng, S. Nojiri, and S.D. Odintsov, EPL **89**, 50003 (2010).
K. Bamba, S.D. Odintsov, L. Sebastiani, and S. Zerbini, Eur. Phys. J. C **67**, 295 (2010).
M.E. Rodrigues, M.J.S. Houndjo, D. Mommeni, and R. Myrzakulov, Can. J. Phys. **92**, 173 (2014).
G.R. Bengochea and R. Ferraro, Phys. Rev. D **79**, 124019 (2009).
E.V. Linder, Phys. Rev. D **81**, 127301 (2010).
S. Capozziello, Int. J. Mod. Phys. D, **11**, 483 (2002).
S. Capozziello and M. DeLaurentis, Phys. Rep. **509**, 167 (2011).
A.V. Astashenok, S. Capozziello, and S.D. Odintsov, J. Cosmol. Astropart. Phys. 12 (2013) 040.
S. Capozziello, M. DeLaurentis, R. Farinelli, and S.D. Odintsov, Phys. Rev. D **93**, 023501 (2016).
A.V. Astashenoka, S. Capozziello, and S.D. Odintsov, Phys. Lett. B **742**, 160 (2015).
T. Harko, F.S.N. Lobo, S. Nojiri, and S.D. Odintsov, Phys. Rev. D **84**, 024020 (2011).
R. Myrzakulov, Eur. Phys. J. C **72**, 2203 (2012).
M. Jamil, D. Momeni, and R. Myrzakulov, Chin. Phys. Lett. **29**, 109801 (2012).
H. Shabani and M. Farhoudi, Phys. Rev. D **88**, 044048 (2013).
H. Shabani and M. Farhoudi, Phys. Rev. D **90**, 044031 (2014).
P.H.R.S. Moraes, Eur. Phys. J. C **75**, 168 (2015).
D. Momeni, R. Myrzakulov, and E. G[ü]{}dekli, Int. J. Geom. Meth. Mod. Phys. **12**, 1550101 (2015).
R. Zaregonbadi and M. Farhoudi, Gen. Rel. Grav. **48**, 142 (2016).
H. Shabani, Int. J. Mod. Phys. D **26**, 1750120 (2017).
H. Shabani and A.H. Ziaie, Eur. Phys. J. C **77**, 31 (2017).
M. Sharif and Z. Yousaf, Astrophys. Space Sci. [**354**]{}, 471 (2014).
I. Noureen and M. Zubair, Astrophys. Space Sci. [**356**]{}, 103 (2015).
I. Noureen and M. Zubair, Eur. Phys. J. C [**75**]{}, 62 (2015).
I. Noureen, M. Zubair, A.A. Bhatti, and G. Abbas, Eur. Phys. J. C [**75**]{}, 323 (2015).
M. Zubair and I. Noureen, Eur. Phys. J. C **75**, 265 (2015).
M. Zubair, G. Abbas, and I. Noureen, Astrophys. Space Sci. **361**, 8 (2016).
A. Alhamzawi and R. Alhamzawi, Int. J. Mod. Phys. D [**25**]{}, 1650020 (2016).
D. Psaltis, Living Rev. Relativ. **11**, 9 (2008).
S. Capozziello, M. DeLaurentis, S.D. Odintsov, and A. Stabile, Phys. Rev. D **83**, 064004 (2011).
S. Capozziello, M. DeLaurentis, I. DeMartino, M. Formisano, and S.D. Odintsov, Phys. Rev. D **85**, 044022 (2012).
F. Briscese, E. Elizalde, S. Nojiri, and S.D. Odintsov, Phys. Lett. B **646**, 105 (2007).
M.C.B. Abdalla, S. Nojiri, and S.D. Odintsov, Class. Quantum Gravity **22**, L35 (2005).
K. Bamba, S. Nojiri, and S.D. Odintsov, J. Cosmol. Astropart. Phys. 10 (2008) 045.
T. Kobayashi and K.I. Maeda, Phys. Rev. D **78**, 064019 (2008).
E. Babichev and D. Langlois, Phys. Rev. D **81**, 124051 (2010).
S. Nojiri and S.D. Odintsov, Phys. Lett. B **676**, 94 (2009).
K. Bamba, S. Nojiri, and S.D. Odintsov, Phys. Lett. B **698**, 451 (2011).
J. Khoury and A. Weltman, Phys. Rev. D **69**, 044026 (2004).
J. Khoury and A. Weltman, Phys. Rev. Lett. **93**, 171104 (2004).
A. Upadhye and W. Hu, Phys. Rev. D **80**, 064002 (2009).
P.H.R.S. Moraes, J.D.V. Arba[ñ]{}il, and M. Malheiro, J. Cosmol. Astropart. Phys. 06 (2016) 005.
A. Das, F. Rahaman, B.K. Guha, and S. Ray, Eur. Phys. J. C **76**, 654 (2016).
A. Das, S. Ghosh, B.K. Guha, S. Das, F. Rahaman, and S. Ray, Phys. Rev. D **95**, 124011 (2017).
T. Harko, Phys. Rev. D **90**, 044067 (2014).
S. Chakraborty, Gen. Relativ. Gravit. **45**, 2039 (2013).
B.V. Ivanov, Phys. Rev. D **65**, 104011 (2002).
F.E. Schunck and E.W. Mielke, Class. Quantum Gravit. **20**, R301 (2003).
M.K. Mak and T. Harko, Proc. R. Soc. A **459**, 393 (2003).
V.V. Usov, Phys. Rev. D **70**, 067301 (2004).
V. Varela, F. Rahaman, S. Ray, K. Chakraborty, and M. Kalam, Phys. Rev. D **82**, 044052 (2010).
F. Rahaman, M. Jamil, A. Ghosh, and K. Chakraborty, Mod. Phys. Let. A **25**, 835 (2010).
F. Rahaman, P.K.F. Kuhfittig, M. Kalam, A.A. Usmani, and S. Ray, Class. Quantum Gravit. **28**, 155021 (2011).
F. Rahaman, R. Maulick, A.K. Yadav, S. Ray, and R. Sharma, Gen. Relativ. Gravit. **44**, 107 (2012).
M. Kalam, F. Rahaman, S. Ray, Sk.M. Hossein, I. Karar, and J. Naskar, Eur. Phys. J. C **72**, 2248 (2012).
D. Deb, S. R. Chowdhury, B.K. Guha, and S. Ray, arXiv:1611.02253 \[gr-qc\].
D. Shee, F. Rahaman, B.K. Guha, and S. Ray, Astrophys. Space Sci. **361**, 167 (2016).
S.K. Maurya, Y.K. Gupta, S. Ray, and D. Deb, Eur. Phys. J. C **76**, 693 (2016).
S.K. Maurya, D. Deb, S. Ray, and P.K.F. Kuhfittig, eprint arXiv:1703.08436.
V. Singh and C.P. Singh, Astrophys. Space Sci. [**356**]{}, 153 (2015).
P.H.R.S. Moraes, Astrophys. Space Sci. **352**, 273 (2014).
P.H.R.S. Moraes, Eur. Phys. J. C **75**, 168 (2015).
P.H.R.S. Moraes, Int. J. Theor. Phys. **55**, 1307 (2016).
P.H.R.S. Moraes, R.A.C. Correa, and R.V. Lobato, J. Cosmol. Astropart. Phys. 07 (2017) 029.
C.P. Singh and P. Kumar, Eur. Phys. J. C **74**, 3070 (2014).
E.H. Baffou, A.V. Kpadonou, M.E. Rodrigues, M.J.S. Houndjo, and J. Tossa, Astrophys. Space Sci. [**356**]{}, 173 (2015).
H. Shabani and M. Farhoudi, Phys. Rev. D **88**, 044048 (2013).
H. Shabani and M. Farhoudi, Phys. Rev. D **90**, 044031 (2014).
M. Sharif and M. Zubair, Astrophys. Space Sci. [**349**]{}, 457 (2014).
D.R.K. Reddy and R. Shantikumar, Astrophys. Space Sci. [**344**]{}, 253 (2013).
P. Kumar and C.P. Singh, Astrophys. Space Sci. [**357**]{}, 120 (2015).
M.F. Shamir, Eur. Phys. J. C [**75**]{}, 354 (2015).
V. Fayaz, H. Hossienkhani, Z. Zarei, and N. Azimi, Eur. Phys. J. Plus [**131**]{}, 22 (2016).
A. Chodos, R.L. Jaffe, K. Johnson, C.B. Thorn, and V.F. Weisskopf, Phys. Rev. D **9**, 3471 (1974).
F. Rahaman, K. Chakraborty, P.K.F. Kuhfittig, G.C. Shit, and M. Rahman, Eur. Phys. J. C **74**, 3126 (2014).
M. Brilenkov, M. Eingorn, L. Jenkovszky, and A. Zhuk, J. Cosmol. Astropart. Phys. 08 (2013) 002.
N.R. Panda, K.K. Mohanta, and P.K. Sahu, J. Physics: Conference Series **599**, 012036 (2015).
A.A. Isayev, Phys. Rev. C **91**, 015208 (2015).
S.D. Maharaj, J.M. Sunzu, and S. Ray, Eur. Phys. J. Plus. **129**, 3 (2014).
L. Paulucci and J.E. Horvath, Phys. Lett. B **733**, 164 (2014).
G. Abbas, S. Qaisar, and A. Jawad, Astrophys. Space Sci. **359**, 57 (2015).
J.D.V. Arba[ñ]{}il and M. Malheiro, J. Cosmol. Astropart. Phys. 11 (2016) 012.
G. Lugones and J.D.V. Arba[ñ]{}il, Phys. Rev. D **95**, 064022 (2017).
M.K. Mak and T. Harko, Chin. J. Astron. Astrophys. **2**, 248 (2002).
P.H.R.S. Moraes, R.A.C. Correa, and R.V. Lobato, J. Cosmol. Astropart. Phys. 07 (2017) 029.
D. Deb, S. Roy Chowdhury, S. Ray, F. Rahaman, and B.K. Guha, Ann. Phys. **387**, 239 (2017).
T. Gangopadhyay, S. Ray, X.D. Li, J. Dey, and M. Dey, Mon. Not. R. Astron. Soc. [**431**]{}, 3216 (2013).
T. G[ü]{}ver, F. [Ö]{}zel, A. Cabrera-Lavers, and P. Wroblewski, Astrophys J. [**712**]{}, 964 (2010).
L. Herrera, Phys. Lett. A **165**, 206 (1992).
H. Abreu, H. Her[ń]{}andez, and L.A. N[[ú]{}]{}${\tilde{n}}$ez, Class. Quantum Gravit. **24**, 4631 (2007).
S. Chandrasekhar, Astrophys J. **140**, 417 (1964).
W. Hillebrandt and K.O. Steinmetz, Astron. Astrophys. **53**, 283 (1976).
R. Chan, L. Herrera, and N.O. Santos, Mon. Not. R. Astron. Soc. **267**, 637 (1994).
L. Herrera and N.O. Santos, Phys. Rep. **286**, 53 (1997).
D. Horvat, S. Ilijić, and A. Marunović, Class. Quantum Gravit. **28**, 025009 (2011).
D.D. Doneva and S.S. Yazadjiev, Phys. Rev. D **85**, 124023 (2012).
H.O. Silva, C.F.B. Macedo, E. Berti, and L.C.B. Crispino, Class. Quantum Gravit. **32**, 145008 (2015).
H. Heintzmann and W. Hillebrandt, Astron. Astrophys. **38**, 51 (1975).
R. Ruderman, Rev. Astron. Astrophys. **10**, 427 (1972).
N.K. Glendenning, Compact Stars: Nuclear Physics, Particle Physics and General Relativity (Springer, New York, 1997).
M. Herzog and F.K.R[ö]{}pke, Phys. Rev. D **84**, 083002 (2011).
H.A. Buchdahl, Phys. Rev. D **116**, 1027 (1959).
APPENDIX: Expressions of constants {#appn .unnumbered}
==================================
The expressions of the constants used in Eqs. (\[3.1\])-(\[3.5\]) are given as $$\begin{aligned}
\label{appn1}
&\qquad\hspace{-2cm} \lambda_{{1}}= \left( \frac{3}{4}c_{{1}}+\frac{3}{4} \right) \chi+\pi, \\ \label{appn2}
&\qquad\hspace{-2cm} \lambda_{{2}}= \left( -\frac{1}{4}c_{{1}}+\frac{1}{12}\right) \chi+\pi, \\ \label{appn3}
&\qquad\hspace{-1cm} \lambda_{{3}}=-384\, \left[ \left( \frac{5}{8}B-\frac{1}{4}\rho_{{c}} \right)
\chi+\pi \, \left( B-\frac{1}{4}\rho_{{c}} \right) \right] \lambda_{{2}}, \\ \label{appn4}
&\qquad\hspace{-1cm} \nu_{{1}}= \left( \frac{3}{8}c_{{1}}-\frac{1}{8} \right) {\chi}^{2}+\frac{9}{4}\left( c_{{1}}+\frac{1}{9}\right) \pi \,\chi+{\pi }^{2}, \\ \label{appn5}
&\qquad\hspace{-1cm} \nu_{{2}}= \left( \pi +\frac{\chi}{2}\right) \left( \pi +\frac{\chi}{4}\right), \\ \label{appn5a}
&\qquad\hspace{-1cm} \nu_{{3}}=589824\, \left( \pi +\frac{\chi}{4} \right) {R}^{2} \Bigg[ B{\pi }^{3}{R}^{3}+ \Big\lbrace \frac{1}{8}B{R}^{3} \left( c_{{1}}+{\frac {26}{3}} \right) \chi \nonumber\\
&\qquad\hspace{-1cm} -{\frac {5M}{32}} \Big\rbrace {\pi }^{2}-\frac {3\chi\pi }{32} \Big\lbrace {R}^{3} \left( {c_{{1}}}^{2}+\frac{1}{3}c_{{1}}-{\frac {32}{9}} \right) B\chi-\frac{5}{4}\left( c_{{1}}
-\frac{5}{3}\right) M \Big\rbrace\nonumber\\
&\qquad\hspace{-1cm}-\frac {3\,{\chi}^{2} \left( c_{{1}}-\frac{1}{3}\right) }{64} \Big\lbrace B{R}^{3} \left( c_{{1}}+\frac{4}{3}\right) \chi-{\frac {15\,M \left( c_{{1}}+1 \right) }{4}} \Big\rbrace \Bigg], \\ \label{appn6}
&\qquad\hspace{-1cm} \nu_{{4}}={R}^{2}\Big[{\nu_{{2}}}^{2}{\lambda_{{2}}}^{2}{B}^{2}{R}^{6}+\frac{1}{4}\nu_{{1}}\nu_{{2
}}\lambda_{{2}}B{R}^{4}-\frac{5}{8}\nu_{{2}}\lambda_{{1}}\pi \,M\lambda_{{2}
}B{R}^{3}\nonumber\\
&\qquad\hspace{-1cm}-{\frac {3\,\nu_{{1}}M \left( \pi +\chi \right) \lambda_{{2}}
R}{64}}+{\frac {25\,{\lambda_{{1}}}^{2}{\pi }^{2}{M}^{2}}{256}}
\Big]^{\frac{1}{2}}, \\ \label{appn7}
&\qquad\hspace{-1cm} \nu_{{5}}=24{R}^{5} \Big[\frac{9}{4}\chi\left(\pi +\frac{\chi}{6}\right) c_{{1}}
+ \left( \pi +\frac{\chi}{2}\right)\left( \pi -\frac{\chi}{4}\right)\Big], \\ \label{appn8}
&\qquad\hspace{-1cm} p_{{1}}=16\,B\pi +10\,B\chi-4\,\pi \,\rho_{{c}}-4\,\chi\,\rho_{{c}}, \\ \label{appn9}
&\qquad\hspace{-1cm} {{\rho}_c}=-\Big[\Big\lbrace-48\,B\pi \,{R}^{3}\chi\,c_{{1}}+192\,B{\pi }^{2}{R}^{3}+176\,B\pi \,{R}^{3}\chi\nonumber\\
&\qquad\hspace{-1cm}+40\,B{R}^{3}{\chi}^{2}-45\,M\chi\,c_{{1}}-60\,M\pi -45\,M\chi\Big\rbrace\Big/\nonumber\\
&\qquad\hspace{-1cm}\Big\lbrace 4\,{R}^{3} \left( 18\,\pi \,\chi\,c_{{1}}+3\,{\chi}^{2}c_{{1}}+8\,{\pi }^{2}+2\,\pi \,\chi-{\chi}^{2} \right)\Big\rbrace\Big],\\ \label{appn10}
&\qquad\hspace{-1cm} \rho_{{0}}={\frac {3\,\chi\,c_{{1}}\rho_{{c}}+16\,B\pi +10\,B\chi-\chi
\,\rho_{{c}}}{3\,\chi\,c_{{1}}+4\,\pi +3\,\chi}}.\end{aligned}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The normalized maximum likelihood (NML) is a recent penalized likelihood that has properties that justify defining the amount of discrimination information (DI) in the data supporting an alternative hypothesis over a null hypothesis as the logarithm of an NML ratio, namely, the alternative hypothesis NML divided by the null hypothesis NML. The resulting DI, like the Bayes factor but unlike the p-value, measures the strength of evidence for an alternative hypothesis over a null hypothesis such that the probability of misleading evidence vanishes asymptotically under weak regularity conditions and such that evidence can support a simple null hypothesis. Unlike the Bayes factor, the DI does not require a prior distribution and is minimax optimal in a sense that does not involve averaging over outcomes that did not occur. Replacing a (possibly pseudo-) likelihood function with its weighted counterpart extends the scope of the DI to models for which the unweighted NML is undefined. The likelihood weights leverage side information, either in data associated with comparisons other than the comparison at hand or in the parameter value of a simple null hypothesis. Two case studies, one involving multiple populations and the other involving multiple biological features, indicate that the DI is robust to the type of side information used when that information is assigned the weight of a single observation. Such robustness suggests that very little adjustment for multiple comparisons is warranted if the sample size is at least moderate.'
bibliography:
- 'refman.bib'
title: Statistical inference optimized with respect to the observed sample for single or multiple comparisons
---
**Keywords:** indirect information; information criteria; information for discrimination; minimum description length; model selection; multiple comparison procedure; multiple testing; normalized maximum likelihood; penalized likelihood; reduced likelihood; weighted likelihood
\[sec:Introduction\]Introduction
================================
\[sub:Quantifying-statistical-evidence\]Quantifying statistical evidence
------------------------------------------------------------------------
Many areas of science involve investigations of whether some effect is present and thus call for statistical methods that assess the evidence pertaining to whether a null hypothesis or an alternative hypothesis is closer to the system studied. For example, many experimental biologists are more interested in whether gene expression levels differ between control and treatment groups than in the effect size itself.
Because not all samples are representative of their populations, the amount of evidence against the null hypothesis is misleadingly high for some samples. Although the probability of observing such an unrepresentative sample should decrease as the size of the sample increases, that is not the case if proximity of a p-value to 0 is interpreted as the strength of evidence against the null hypothesis. Indeed, the distribution of the p-value associated with a simple (point) null hypothesis remains the same at all sample sizes if the null hypothesis holds, making the p-value impossible to interpret as a level of evidence apart from considering the sample size, as @RefWorks:122, @Blume2003439, and others have argued; cf. @RefWorks:293 on the sample-size incoherence of significance testing. @RefWorks:435 defined the lacking property by calling a measure of evidence *interpretable* if its probability of misleading evidence vanishes asymptotically. That is, a measure of evidence satisfies the interpretability condition only if the frequentist probability of observing a sample that has misleading evidence exceeding some fixed threshold converges to 0 as the sample size diverges.
Another adverse consequence of treating the p-value as a measure of evidence is its inability to indicate evidence in favor of a simple null hypothesis. In general, the amount of information in the data that favors a simple null hypothesis cannot be quantified by the p-value since it can only indicate whether there is evidence against it.
The Bayes factor in principle overcomes the above limitations of the p-value but poses the notorious problem of specifying the prior distribution of a nuisance parameter that is not random in the frequentist sense. Any solution to the problem has practical implications since the Bayes factor is sensitive to prior specification [@RefWorks:183].
Subjective prior distributions have the advantage of coherence and yet are rarely used in data analysis since they depend on arbitrary choices in prior specification. On the other hand, the improper prior distributions generated by conventional algorithms cannot be directly applied to model selection since they would leave the Bayes factor undefined. That has been overcome to some extent by dividing the data into training and test samples, with the training samples generating proper priors for use with test samples, but at the expense of requiring the specification of training samples and, in the presence of multiple training samples, a method of averaging [@RefWorks:1023]. Further, the interpretation of the resulting posterior probability is not clear except perhaps as an approximation to an agent’s level of belief [@RefWorks:17].
\[sub:Optimality-mean\]Repeated-sampling optimality
---------------------------------------------------
Since there are many potential measures of evidence, most notably the Bayes factors defined by different priors, that satisfy the criteria that a measure of evidence be interpretable and that it can support a simple null hypothesis, an optimality criterion will be applied to determine a unique method of hypothesis testing and more general model selection. Before doing so, that criterion will be distinguished from standards of optimality in the received framework of statistics, that of Neyman and Pearson as generalized by Wald.
The goal of minimizing *risk*, the expected loss with respect to a sampling distribution [@Wald1950], has provided a unified framework of estimation and testing and, as briefly reviewed in @conditional2009 [§3.1], has led to recent multiple comparison procedures. However, @RefWorks:1302, @FraserAncillaries2004a, @RefWorks:1334, and other frequentist statisticians have criticized the framework for promoting opportunistic trade-offs between hypothetical samples, thereby potentially misleading scientists and yielding unacceptably pathological procedures. The main non-Bayesian alternative involves replacing the marginal sampling distribution with a conditional sampling distribution given an exact or approximate ancillary statistic [e.g., @RefWorks:SprottBook2000 §3.3].
While conditioning on an ancillary statistic makes the reference distribution more relevant to inference on the basis of the observed sample [@RefWorks:985], it still does not permit statements about the actual loss incurred. For example, the confidence level remains the proportion of confidence intervals corresponding to repeated samples that cover the parameter of interest. Although the use of exact confidence intervals minimizes a risk ([@RefWorks:1187]; [@frequentistReasoning]), it is silent regarding the loss associated with the observed sample.
\[sub:Optimality-observed\]Observed-sample optimality
-----------------------------------------------------
### Information-theoretic inference
In order to address the issues outlined above, this paper continues the development of a new information-theoretic alternative to previous approaches to statistical inference. The concept of a predictive distribution will enable defining minimax optimality without repeated-sampling or posterior-distribution averages. This approach is presented here largely without the terminology of its origin in universal source coding [@RefWorks:404].
Consider the observed data vector $x\in\mathcal{X}^{n}$. Let $\mathcal{E}\left(\Omega\right)$ denote the set of all probability density functions on any sample space $\Omega$, and let $\mathcal{F}=\left\{ f_{\phi}:\phi\in\Phi\right\} \subset\mathcal{E}\left(\mathcal{X}^{n}\right)$ denote a parametric family of density functions on $\mathcal{X}^{n}$ for parameter space $\Phi$. (Herein, the probability densities are Radon-Nikodym derivatives, reducing to probability masses if $\mathcal{X}$ is countable.) The maximum likelihood estimate of $\phi$, denoted by $\hat{\phi}\left(x\right)$, is assumed to be unique.
The *regret* of a predictive density $\bar{f}\in\mathcal{E}\left(\mathcal{X}^{n}\right)$ is the logarithmic loss$$\begin{aligned}
\operatorname{reg}\left(\bar{f},x;\Phi\right) & = & -\log\bar{f}\left(x\right)-\inf_{\phi\in\Phi}\left(-\log f_{\phi}\left(x\right)\right)=\log\frac{f_{\hat{\phi}\left(x\right)}\left(x\right)}{\bar{f}\left(x\right)}\label{eq:untargeted-regret}\end{aligned}$$ for any $x\in\mathcal{X}^{n}$. The *$\mathcal{E}\left(\mathcal{X}^{n}\right)$-optimal predictive density function* *relative to* $\mathcal{F}$, $$\bar{f}_{0}=\arg\inf_{\bar{f}\in\mathcal{E}\left(\mathcal{X}^{n}\right)}\sup_{u\in\mathcal{X}^{n}}\operatorname{reg}\left(\bar{f},u;\Phi\right),\label{eq:optimality}$$ while by definition in $\mathcal{E}\left(\mathcal{X}^{n}\right)$, is not necessarily in $\mathcal{F}$. Rather, $\bar{f}_{0}$ is a probability density function that represents the entire family $\mathcal{F}$ with a single distribution, much as does a prior predictive density function. Instead of averaging the members of $\mathcal{F}$ with respect to a prior distribution, the present definition employs $\mathcal{F}$ in equation for each $u\in\mathcal{X}^{n}$ through the maximization of the likelihood over $\phi\in\Phi$, as seen by substituting $u$ for $x$ in equation .
Originally motivated in the information theory literature by a need to minimize codelength [@RefWorks:404], equation defines the type of minimax optimality employed as opposed to the optimality of Section \[sub:Optimality-mean\]. (According to the minimum description length principle, each family of distributions corresponds to an algorithm of most efficiently encoding the information in $x$ [@RefWorks:374; @RefWorks:375; @Rissanen2009b].) The predictive density function $\bar{f}_{0}$ is optimal in that it solves the minimax problem involving all $u\in\mathcal{X}^{n}$, and thus for the observed sample $x\in\mathcal{X}^{n}$, rather than the more usual minimax problem involving an expectation value over all samples, as in the standard decision theory of frequentism. The following result [@RefWorks:404; @RefWorks:374; @RefWorks:375], to be proved in Section \[sub:NMWL\] for a more general optimization problem, sheds light on the nature of the optimality considered.
\[thm:NML\]If $\int_{\mathcal{X}^{n}}f_{\hat{\phi}\left(u\right)}\left(u\right)du<\infty$, then the $\mathcal{E}\left(\mathcal{X}^{n}\right)$-optimal predictive density function relative to $\mathcal{F}$ is$$\bar{f}_{0}\left(\bullet\right)=\bar{f}_{0}\left(\bullet;\Phi\right)=\frac{f_{\hat{\phi}\left(\bullet\right)}\left(\bullet\right)}{\int_{\mathcal{X}^{n}}f_{\hat{\phi}\left(u\right)}\left(u\right)du}.\label{eq:NML}$$
This proof by contradiction is based on the direct proof given by @RefWorks:375 [§6.2.1]. Assume, contrary to the claim, that the density function $\bar{f}_{0}$ that satisfies equation is not the optimal predictive density function. Since, for any $v\in\mathcal{X}^{n}$, the ratio $\bar{f}_{0}\left(v\right)/f_{\hat{\phi}\left(v\right)}\left(v\right)$ does not depend on $v$, it follows that, for any $\breve{f}\in\mathcal{E}\left(\mathcal{X}^{n}\right)\backslash\left\{ \bar{f}_{0}\right\} $, there is a $v\in\mathcal{X}^{n}$ such that $\breve{f}\left(v\right)/f_{\hat{\phi}\left(v\right)}\left(v\right)<\bar{f}_{0}\left(v\right)/f_{\hat{\phi}\left(v\right)}\left(v\right)$. Therefore, given any $\breve{f}\in\mathcal{E}\left(\mathcal{X}^{n}\right)\backslash\left\{ \bar{f}_{0}\right\} $, there is a $v\in\mathcal{X}^{n}$ such that $\operatorname{reg}\left(\bar{f}_{0},v;\Phi\right)<\operatorname{reg}\left(\breve{f},v;\Phi\right)$, which contradicts the assumption.
Note that $u$, the dummy variable of integration over $\mathcal{X}^{n}$, appears twice in the integrand. For the observed $x\in\mathcal{X}^{n}$, the quantity $\bar{f}_{0}\left(x\right)=\bar{f}_{0}\left(x;\Phi\right)$ is called the *normalized maximum likelihood* (NML) with respect to $\Phi$.
According to Theorem \[thm:NML\], the minimax optimality of $\bar{f}_{0}$ guarantees that $\operatorname{reg}\left(\bar{f}_{0},x;\Phi\right)$, the regret due to the observed sample, cannot exceed $\sup_{u\in\mathcal{X}^{n}}\operatorname{reg}\left(\bar{f}_{0},u;\Phi\right)$, the regret due to the worst-case sample. In that sense, $\bar{f}_{0}$ is optimal for the observed sample. By contrast, standard frequentist optimality, concerned only with loss averaged over all possible samples, guarantees no bound on the loss inflicted by any individual sample.
Such observed-sample optimality justifies selecting the model or hypothesis corresponding to the family $\mathcal{F}_{i}$ of distributions that minimizes $-\log\bar{f}_{0}\left(x;\Phi_{i}\right)$, the observed prediction error of the $i$th among a finite number of distribution families under consideration. Following the terminology of @RefWorks:292 and @mediumScale, $-\log\bar{f}_{0}\left(x;\left\{ \phi_{0}\right\} \right)-\left(-\log\bar{f}_{0}\left(x;\Phi\right)\right)$ would be the *information in $x$ for discrimination* in favor of the alternative hypothesis that $\phi\ne\phi_{0}$ over the null hypothesis that $\phi=\phi_{0}$. Such information is an interpretable measure of evidence under general conditions and can quantify the strength of any evidence in favor of the null hypothesis as well as that of any evidence against it. More importantly, the information for discrimination optimally quantifies the difference in how well each model or hypothesis predicts relative to ideal predictors of individual samples rather than relative to unknown true distributions, the ideal predictors in the sense of averages over samples. The Kullback-Leibler risk, for example, only measures mean discrimination information relative to unknown ideal predictors in the average sense.
Since the base of the logarithm is inconsequential, it may be chosen for convenience of interpretation. The binary logarithm $\left(\log_{2}\right)$, yielding the number of *bits* of information, enables not only immediate exponentiation back to the ratio domain but also the use of grades of evidence that are both broad enough and refined enough for applications across scientific disciplines (Table \[tab:Grades-of-hypothesis-evidence\]). Except for the distinction between negligible and weak evidence, the grades closely mirror those @RefWorks:182 originally proposed for the Bayes factor; cf. @RefWorks:435. Accordingly, the $\left[3,5\right)$ grade of Table \[tab:Grades-of-hypothesis-evidence\] is what @RefWorks:122 [§1.12] considers “fairly strong evidence” for one simple hypothesis over another, and the $\left[5,7\right)$ and $\left[7,\infty\right)$ grades together constitute his “quite strong evidence.”
**Information (bits)** $\left(0,1\right)$ $\left[1,2\right)$ $\left[2,3\right)$ $\left[3,5\right)$ $\left[5,7\right)$ $\left[7,\infty\right)$
------------------------ -------------------- -------------------- -------------------- -------------------- -------------------- -------------------------
**Evidence grade** Negligible Weak Moderate Strong Very strong Overwhelming
: \[tab:Grades-of-hypothesis-evidence\]Heuristic grades of evidence for an alternative hypothesis over a null hypothesis corresponding to intervals of the information for discrimination. The absolute value of a negative amount of information gives the grade of evidence favoring the null hypothesis.
### Extension of information-theoretic inference
Despite the unique observed-sample optimality of the NML for quantifying discrimination information, three shortcomings make it impractical for use in many biostatistics applications. First, since such applications typically partition $\phi$ into an interest parameter $\theta$ and a nuisance parameter $\lambda$, the regret is relative to an ideal distribution determined by maximizing the likelihood not only over $\theta$ but also over $\lambda$. As a result, the ideal member of the family of distributions would be considered a better predictor than another member that has the same value of $\theta$ on the basis of having a different value of $\lambda$, which should be irrelevant. Thus, the NML is inadequate for testing hypotheses about $\theta$ in the presence of $\lambda$.
Second, the NML only uses information that is in $x$, but considering such information about the parameter in isolation from other available information can be misleading unless the sample size is sufficiently large. Additional information may be available in data from other populations, from other biological features such as genes or SNPs, or from other feature-feature comparisons. Even in the absence of such incidental information, there would be some information in the fact that the null hypothesis that $\phi=\phi_{0}$ is seriously considered.
Third, the normalizing denominator of equation , the logarithm of which is called the *parametric complexity* of $\mathcal{F}$, is infinite for typical families of distributions, including the normal family. Each of the variant NMLs proposed to address the problem introduces its own conceptual difficulties [@Lanterman2005b; @RefWorks:375]. For example, @RefWorks:374 [§5.2.4], @RissanenRoos2007b, and @RefWorks:375 [§11.4.2] proposed conditional versions of the NML. Cf. related work by @TakimotoWarmuth2000b.
To overcome the first of the three identified problems with NML, it is generalized in Section \[sec:NMWL\] by replacing the original data with a statistic that is a function of the data and that has a distribution depending on $\theta$ but not on $\lambda$. Since the information in the data relevant to the interest parameter is largely confined to the statistic, that information can be better quantified in terms of the distribution of the statistic than in terms of the distribution of the original data, the latter depending on the value of the nuisance parameter. In terms of the minimum description length (MDL) metaphor [@RefWorks:374; @RefWorks:375], the data are first compressed with little information loss by reduction to a smaller-dimensional statistic and then further compressed by the family of distributions.
The use of a weighted likelihood addresses the second problem in the same section, which also includes some results relevant to the probability of observing misleading information. (The weighted likelihood was originally proposed for bias-variance trade-offs given relatively small $n_{i}$ but potentially large $N$ [@Feifang2002347]. More formally, @RefWorks:509 derived the weighted likelihood from the minimization of Kullback-Leibler loss.)
As a by-product for commonly used distribution families, that solution to the second problem automatically solves the third problem, as illustrated in Section \[sec:Case-study\] with a multiple-population data set and a multiple-feature data set. Finally, Section \[sec:Discussion\] concludes by highlighting desirable properties of the new NML-based measure of information for discrimination.
\[sec:NMWL\]Optimal inference
=============================
\[sub:Preliminaries\]Preliminaries
----------------------------------
### Weighted likelihood
The framework of Section \[sub:Optimality-observed\] is generalized by the use of data reduction to eliminate a nuisance parameter in $\phi$. Consider a measurable map $\tau_{n}:\mathcal{X}^{n}\rightarrow\mathcal{T}\left(n\right)$. Let $\theta:\Phi\rightarrow\Theta$ denote a subparameter function such that the probability density of $\tau_{n}\left(X\right)$ is $g_{\theta\left(\phi\right)}\left(\tau_{n}\left(X\right)\right)$, abbreviated as $g_{\theta}\left(\tau\left(X\right)\right)$; the dependence of the density function $g_{\theta}$ on $n$ is suppressed. Thus, the reduction of the data $X$ to a statistic $\tau\left(X\right)$ has the effect of replacing the full parameter $\phi$ with the interest parameter $\theta$. Important special cases of $L\left(\theta;\tau\left(x\right)\right)=g_{\theta}\left(\tau\left(x\right)\right)$ as a function of $\theta$ are conditional likelihood functions and marginal likelihood functions ([@RefWorks:122; @Severini2000]; [@mediumScale]).
The framework is now extended to $N$ hypotheses or comparisons. Let $\mathcal{G}_{n,i}=\left\{ g_{n,i,\theta}:\theta\in\Theta\right\} \subset\mathcal{E}\left(\mathcal{T}\left(n\right)\right)$ denote the parametric family of density functions on $\mathcal{T}\left(n\right)$ for parameter space $\Theta$. The “$n$” and “$i$” subscripts will be dropped when their values are clear. For the $i$th of $N$ null hypotheses or comparisons, suppose $x_{i}\in\mathcal{X}^{n_{i}}$ is a realization of the random vector $X_{i}$ of $n_{i}$ independent components. Then each $T_{i}=\tau\left(X_{i}\right)$ is distributed with density $g_{\theta_{i}}=g_{n_{i},i,\theta_{i}}$, and each outcome $t_{i}=\tau\left(x_{i}\right)$ is an element of $\mathcal{T}_{i}=\mathcal{T}\left(n_{i}\right)$. Let $L_{i}\left(\theta;\tau\left(x_{i}\right)\right)=g_{n,i,\theta}\left(\tau\left(x_{i}\right)\right)=g_{\theta}\left(\tau\left(x_{i}\right)\right)$, giving each comparison its own likelihood function.
Mapping $\mathcal{X}^{n_{i}}$ to $\mathcal{T}_{i}=\mathbb{R}^{D}$ is common in data reduction applications in which $\Theta=\mathbb{R}^{D}$. Assigning a common parametric family to all comparisons ($\mathcal{G}_{n,i}=\mathcal{G}_{n,1}$ for all $i$) is usually appropriate when each comparison corresponds to a biological feature, as in Section \[sub:Biological-features\].
The observation $\mathbf{x}=\left\langle x_{1},\dots,x_{N}\right\rangle $ generates the test statistic vector $\mathbf{t}=\left\langle t_{1},\dots,t_{N}\right\rangle =\left\langle \tau\left(x_{1}\right),\dots,\tau\left(x_{N}\right)\right\rangle $, an outcome of $\mathbf{T}=\left\langle T_{1},\dots,T_{N}\right\rangle =\left\langle \tau\left(X_{1}\right),\dots,\tau\left(X_{N}\right)\right\rangle $. For inference about $\theta_{i}$ on the basis of $\mathbf{t}$, the *weighted likelihood function* $\bar{L}_{i}\left(\bullet;\mathbf{t}\right):\Theta\rightarrow\left[0,\infty\right)$ is defined by $$\log\bar{L}_{i}\left(\theta_{i};\mathbf{t}\right)=\sum_{j=1}^{N}w_{ij}\log L_{j}\left(\theta_{i};t_{j}\right),\label{eq:weighted-likelihood}$$ where the weights $w_{i}=\left\langle w_{i1},\dots,w_{iN}\right\rangle $ are real numbers that may depend on $\left\langle n_{1},\dots,n_{N}\right\rangle $ and that satisfy $w_{ii}\ge w_{ij}$ [@Feifang2002347]. The weights normally also conform to $\sum_{j=1}^{N}w_{ij}=1$, a requirement that will be temporarily relaxed in Section \[sub:NMWL\].
\[exa:microarray\]In most microarray studies, the expression levels of $N$ genes are measured with the goal of determining which genes are differentially expressed between a treatment/perturbation group of $m$ replicates and a control group of $n$ replicates; each of these biological replicates represents one or more organisms. (Single-channel arrays do not require the pairing of replicates between groups as did the dual-channel arrays.) Following the typical assumption that intensity values are lognormally distributed, let $x_{i}=\left\langle x_{i1},\dots,x_{im}\right\rangle $ and $y_{i}=\left\langle y_{i1},\dots,y_{in}\right\rangle $ denote the logarithms of the $m$ and $n$ intensities of the $i$th gene in the perturbation and control group, respectively. For small numbers of replicates, the assumption of a common variance within each group is useful: $X_{ij}\sim\operatorname{N}\left(\xi_{i},\sigma_{i}^{2}\right)$ and $Y_{ij}\sim\operatorname{N}\left(\eta_{i},\sigma_{i}^{2}\right)$ with realized values $X_{ij}=x_{ij}$ for $j=1,\dots,m$ and $Y_{ij}=y_{ij}$ for $j=1,\dots,n$. If $\theta_{i}$ is the absolute value of the *inverse coefficient of variation* $\left(\xi_{i}-\eta_{i}\right)/\sigma_{i}$, then $t_{i}$ is conveniently taken as the absolute value of the two-sample, equal-variance $t$-statistic, which has a noncentral $t$ distribution with noncentrality parameter $\left(m^{-1}+n^{-1}\right)^{-1/2}\theta_{i}$ and $m+n-2$ degrees of freedom.
The sampling distribution of $\mathbf{T}$ is denoted by $P$ to specify properties of the weights while accommodating model misspecification, the case that there is not a $\theta_{i}\in\Theta$ such that $g_{\theta_{i}}$ is a density admitted by the marginal distribution $P\left(T_{i}\in\bullet\right)$ for all $i\in\left\{ 1,\dots,N\right\} $. With suitable weights and the assumption that $\hat{\theta}_{i}\left(\mathbf{T}\right)=\arg\sup_{\theta\in\Theta}\bar{L}_{i}\left(\theta;\mathbf{T}\right)$ is almost surely unique for all $i\in\left\{ 1,\dots,N\right\} $, the difference between $\hat{\theta}_{i}\left(\mathbf{T}\right)$ and the conventional maximum likelihood estimator of $\theta_{i}$ almost surely converges to 0 as $n_{i}$ diverges with $N$ held fixed. Specifically, $w_{ii}=1+o_{P}\left(1\right)$ and $i\ne j\implies w_{ij}=o_{P}\left(1\right)$ ensure that $\hat{\theta}_{i}\left(\mathbf{T}\right)=\arg\sup_{\theta\in\Theta}L\left(\theta;x_{i}\right)+o_{P}\left(1\right)$, where the term $o_{P}\left(1\right)$ converges to 0 with $P$-probability 1 as $n_{i}\rightarrow\infty$ with any ratio $n_{j}/n_{k}$ bounded by constants: $n_{j}=O\left(n_{k}\right)$ for all $j,k\in\left\{ 1,\dots,N\right\} $.
### Predictive loss
For some $\bar{g}\in\mathcal{E}\left(\mathcal{T}_{i}\right)$, the *generalized regret$$\operatorname{reg}_{i}\left(\bar{g},\mathbf{t};\Theta\right)=-\log\bar{g}\left(\tau\left(x_{i}\right)\right)-\inf_{\theta\in\Theta}\left(-\log\bar{L}_{i}\left(\theta;\mathbf{t}\right)\right)=\log\frac{\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}\right);\mathbf{t}\right)}{\bar{g}\left(\tau\left(x_{i}\right)\right)}\label{eq:generalized-regret}$$* measures loss incurred by the likelihood associated with $\bar{g}$, the predictive distribution, relative to $\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}\right);\mathbf{t}\right)$, the maximum weighted likelihood of $\theta_{i}$. In other words, $\operatorname{reg}_{i}\left(\bar{g},\mathbf{t};\Theta\right)$ is the discrepancy between error in predicting the value of $\tau\left(x\right)$ on the basis of $\bar{g}$ and the prediction error minimized over the interest parameter. The latter error is more relevant to hypotheses about the value of $\theta$ than a prediction error minimized over the full parameter $\phi$, including the nuisance parameter $\lambda$ (§\[sub:Optimality-observed\]). Thus, $\operatorname{reg}_{i}\left(\bar{g},\mathbf{t};\Theta\right)$ replaces $\operatorname{reg}\left(\bar{f},x;\Phi\right)$ as the regret in the presence of the nuisance parameter or a nonzero weight other than $w_{ii}$.
\[sub:Optimal-predictive-distribution\]Optimal predictive distribution
----------------------------------------------------------------------
### \[sub:NMWL\]Exact predictive distribution
For each $t\in\mathcal{T}_{i}$, let $\mathbf{t}_{i}\left(t\right)$ denote the $N$-tuple of statistics that is equal to $\mathbf{t}$ in all components except the $i$th, which has $t$ in place of $t_{i}$. For example, $\mathbf{t}_{1}\left(t\right)=\left\langle t,t_{2},\dots,t_{N}\right\rangle $, but $\mathbf{t}_{i}\left(t\right)=\left\langle t_{1},\dots,t_{i-1},t,t_{i+1},\dots,t_{N}\right\rangle $ if $3\le i\le N-2$.
The optimal predictive density function of Section \[sub:Optimality-observed\] is a special case of $$\bar{g}_{i}=\arg\inf_{\bar{g}\in\mathcal{E}\left(\mathcal{T}_{i}\right)}\sup_{t\in\mathcal{T}_{i}}\operatorname{reg}_{i}\left(\bar{g},\mathbf{t}_{i}\left(t\right);\Theta\right),$$ the $\mathcal{E}\left(\mathcal{T}_{i}\right)$-*optimal predictive density function relative to* $\left\langle \mathcal{G},w_{i}\right\rangle $.
\[thm:Shtarkov\]Given some $i\in\left\{ 1,\dots,N\right\} $ and $\mathbf{t}\in\mathcal{T}_{1}\times\cdots\times\mathcal{T}_{N}$, if $\int\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t\right)\right);\mathbf{t}_{i}\left(t\right)\right)dt<\infty$, then, for all $t_{i}\in\mathcal{T}_{i}$, the -optimal predictive density function relative to satisfies$$\bar{g}_{i}\left(t_{i}\right)=\frac{\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}\right);\mathbf{t}\right)}{\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t\right)\right);\mathbf{t}_{i}\left(t\right)\right)dt}.\label{eq:NMWL}$$
The present argument follows that used to prove Theorem \[thm:NML\]. Assume, contrary to the claim, that the density function $\bar{g}_{i}$ that satisfies equation for all $t_{i}\in\mathcal{T}_{i}$ is not the optimal predictive density function relative to $\left\langle \mathcal{G},w_{i}\right\rangle $. The substitution $\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}\right);\mathbf{t}\right)=\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t_{i}\right)\right);\mathbf{t}_{i}\left(t_{i}\right)\right)$ demonstrates that the ratio $\bar{g}_{i}\left(t_{i}\right)/\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t_{i}\right)\right);\mathbf{t}_{i}\left(t_{i}\right)\right)$ does not depend on $t_{i}$. It follows that, for any $\breve{g}_{i}\in\mathcal{E}\left(\mathcal{T}_{i}\right)\backslash\left\{ \bar{g}_{i}\right\} $, there is a $t_{i}\in\mathcal{T}_{i}$ such that $\breve{g}_{i}\left(t_{i}\right)/\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t_{i}\right)\right);\mathbf{t}_{i}\left(t_{i}\right)\right)<\bar{g}_{i}\left(t_{i}\right)/\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t_{i}\right)\right);\mathbf{t}_{i}\left(t_{i}\right)\right)$. Therefore, given any $\breve{g}_{i}\in\mathcal{E}\left(\mathcal{T}_{i}\right)\backslash\left\{ \bar{g}_{i}\right\} $, there is a $t_{i}\in\mathcal{T}_{i}$ such that $\operatorname{reg}_{i}\left(\bar{g}_{i},\mathbf{t}_{i}\left(t_{i}\right);\Theta\right)<\operatorname{reg}_{i}\left(\breve{g}_{i},\mathbf{t}_{i}\left(t_{i}\right);\Theta\right)$, which contradicts the assumption.
For any $x_{i}\in\mathcal{X}^{n}$, the quantity $\bar{g}_{i}\left(\tau\left(x_{i}\right)\right)=\bar{g}_{i}\left(\tau\left(x_{i}\right);\Theta\right)$ is the *normalized maximum weighted likelihood (NMWL)* with respect to $\Theta$ or, more precisely, with respect to $\left\langle \mathcal{G},w_{i}\right\rangle $.
\[exa:conditional-NML\]When the constraint that $\sum_{j=1}^{N}w_{ij}=1$ is relaxed, NMWL generalizes various previous NMLs as follows. If $\tau\left(x_{1}\right)=x_{1}$ for some $x_{1}\in\mathcal{X}^{n_{1}}$ and if $N=1$, the NMWL reduces to the probability density $\bar{g}_{1}\left(x_{1}\right)$ with $w_{1,1}=1$ and thus to $\bar{f}_{0}\left(x\right)$, the NML of equation . For an observed vector $\left\langle y_{1},\dots,y_{n_{1}}\right\rangle \in\mathcal{X}^{n_{1}}$, assigning $N=2$, $\theta_{1}=\theta_{2}$, $t_{1}=\left\langle y_{1},\dots,y_{n_{1}-1}\right\rangle $, and $t_{2}=y_{n_{1}}$ demonstrates that the prominent conditional NMLs are NMWLs in the case of IID data. In particular, @RefWorks:375 [§11.4.2] considered $\bar{g}_{1}\left(\left\langle t_{1},t_{2}\right\rangle \right)$ with $w_{1,1}=w_{1,2}=1$. Conversely, @RefWorks:374 [§5.2.4] and @RissanenRoos2007b studied $\bar{g}_{2}\left(\left\langle t_{1},t_{2}\right\rangle \right)$ with $w_{2,1}=w_{2,2}=1$, thereby facilitating computation of the normalizing constant in equation since the integration is only over a scalar. The main drawback of applying conditional NMLs to the IID setting is the arbitrary nature of choosing an observation $x_{2}$ to leave out since the observations are not ordered in time [@RefWorks:375 §11.4.3]. The same issue arises in Bayesian model selection when an improper prior is conditioned on a minimal training sample before computing the Bayes factor. A popular solution is to take geometric or arithmetic averages over all possible minimal training samples [@Berger2004841]. Analogous approaches to IID applications of conditional NMLs would likewise depend on arbitrary choices of averages and of training sample sizes (§\[sub:Quantifying-statistical-evidence\]).
### Approximate predictive distribution
A computationally efficient approximation to the NMWL is available if:
1. The weight of any comparison in focus is equal to that of any other comparison when it is in focus, i.e., $w_{ii}=w_{1,1}$ for all $i$.
2. The weight of each comparison not in focus is equal to that of any other comparison not in focus, i.e., $w_{ij}=w_{1,2}$ for all $i\ne j$.
3. The sample sizes and sample spaces are equal, i.e., $n_{i}=n_{1}$ and $\mathcal{T}_{i}=\mathcal{T}_{1}$ for all $i$.
4. All comparisons share a single family, i.e., $\mathcal{G}_{n_{1},i}=\mathcal{G}_{n_{1},1}$ and $L_{1}=L_{i}$ for all $i$.
Under those *equal weight conditions*, there is an approximate weight $\tilde{w}_{N+1}$ such that $\tilde{w}_{N+1}=w_{ii}$ for all $i$ and an approximate weight $\tilde{w}_{j}$ such that $\tilde{w}_{j}=\left(N-1\right)N^{-1}w_{ij}$ for all $i$ and $j$ except $i=j$. Then $\sum_{j=1}^{N+1}\tilde{w}_{j}=N\tilde{w}_{1}+\tilde{w}_{N+1}=\left(N-1\right)w_{1,2}+w_{1,1}=1$.
For any $t\in\mathcal{T}_{1}$, let $\tilde{\mathbf{t}}\left(t\right)$ denote $\left\langle t_{1},\dots,t_{N},t\right\rangle \in\mathcal{T}_{1}^{N+1}$. For inference about $\theta_{i}$ on the basis of $\mathbf{t}$, the *approximate weighted likelihood function* $\tilde{L}\left(\bullet;\tilde{\mathbf{t}}\left(t\right)\right):\Theta\rightarrow\left[0,\infty\right)$ is defined by $$\begin{aligned}
\log\tilde{L}\left(\theta_{i};\tilde{\mathbf{t}}\left(t\right)\right) & = & \sum_{j=1}^{N}\tilde{w}_{j}\log L_{j}\left(\theta_{i};t_{j}\right)+\tilde{w}_{N+1}\log L_{j}\left(\theta_{i};t\right)\\
& = & \frac{1-w_{1,1}}{N}\sum_{j=1}^{N}\log L_{1}\left(\theta_{i};t_{j}\right)+w_{1,1}\log L_{1}\left(\theta_{i};t\right),\end{aligned}$$ the second equality implied by the equal weight conditions. Let $\hat{\theta}\left(\tilde{\mathbf{t}}\left(t\right)\right)=\arg\sup_{\theta\in\Theta}\tilde{L}\left(\theta;\tilde{\mathbf{t}}\left(t\right)\right)$.
The following theorem indicates that the exact NMWL is approximated by$$\tilde{g}_{i}\left(t_{i}\right)=\frac{\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}\right);\mathbf{t}\right)}{\int_{\mathcal{T}_{1}}\tilde{L}\left(\hat{\theta}\left(\tilde{\mathbf{t}}\left(t\right)\right);\tilde{\mathbf{t}}\left(t\right)\right)dt},$$ which may be quickly calculated even for large $N$ since the denominator, not depending on $i$, need only be computed once. In the theorem and its supporting lemmas, $\tilde{\mathbf{T}}\left(t\right)=\left\langle T_{1},\dots,T_{N},t\right\rangle $, and $\overset{\text{a.s.}}{\rightarrow}$ denotes almost sure convergence as $N$ increases with $n_{1}$ fixed.
\[lem:approximation\]If the equal weight conditions hold and if $T_{1},\dots,T_{N}$ are drawn independently from a mixture distribution, then, for all $\theta\in\Theta$ and $t\in\mathcal{T}_{i}$,$$\log\tilde{L}\left(\theta;\tilde{\mathbf{T}}\left(t\right)\right)-\log\bar{L}_{i}\left(\theta;\mathbf{T}_{i}\left(t\right)\right)\overset{\text{a.s.}}{\rightarrow}0.\label{eq:approximation}$$
According to the equal weight conditions,$$\begin{aligned}
\log\tilde{L}\left(\theta;\tilde{\mathbf{T}}\left(t\right)\right)-\log\bar{L}_{i}\left(\theta;\mathbf{T}_{i}\left(t\right)\right) & =\end{aligned}$$ $$\left(\frac{1-\tilde{w}_{N+1}}{N}\sum_{j=1}^{N}\log L_{1}\left(\theta;T_{j}\right)+\tilde{w}_{N+1}\log L_{1}\left(\theta;T_{i}\right)\right)-\left(\frac{1-w_{ii}}{N-1}\sum_{j\ne i;i=1}^{N}\log L_{1}\left(\theta;T_{j}\right)+w_{ii}\log L_{1}\left(\theta;T_{i}\right)\right)$$ $$=\left(1-\tilde{w}_{N+1}\right)\left(\frac{1}{N}\sum_{j=1}^{N}\log L_{1}\left(\theta;T_{j}\right)-\frac{1}{N-1}\sum_{j\ne i;i=1}^{N}\log L_{1}\left(\theta;T_{j}\right)\right).$$ The second factor almost surely vanishes by the law of large numbers.
\[lem:MLEs\]Under the assumptions of Lemma \[lem:approximation\], the stipulations that $\hat{\theta}\left(\tilde{\mathbf{T}}\left(t\right)\right)$ and $\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right)\right)$ are almost always unique for all $t\in\mathcal{T}_{i}$ and that $L_{1}\left(\bullet;T_{i}\right)$ is almost surely continuous on $\Theta$ for all $i=1,\dots,N$ imply that, for all $i=1,\dots,N$ and $t\in\mathcal{T}_{i}$, $\hat{\theta}\left(\tilde{\mathbf{T}}\left(t\right)\right)-\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right)\right)\overset{\text{a.s.}}{\rightarrow}0.$
By Lemma \[lem:approximation\], equation holds for all $\theta\in\Theta$. Thus, since almost sure convergence is preserved under almost surely continuous transformations [@Serfling:1254212 §1.7],\
$\arg\sup_{\theta\in\Theta}\tilde{L}\left(\theta;\tilde{\mathbf{T}}\left(t\right)\right)-\arg\sup_{\theta\in\Theta}\bar{L}_{i}\left(\theta;\mathbf{T}\right)\overset{\text{a.s.}}{\rightarrow}0.$
\[thm:approximate-complexity\]Under the assumptions of Lemma \[lem:MLEs\], the difference between the approximate and exact parametric complexities almost surely vanishes: $$\int_{\mathcal{T}_{1}}\tilde{L}\left(\hat{\theta}\left(\tilde{\mathbf{T}}\left(t\right)\right);\tilde{\mathbf{T}}\left(t\right)\right)dt-\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right)\right);\mathbf{T}_{i}\left(t\right)\right)dt\overset{\text{a.s.}}{\rightarrow}0.\label{eq:approximate-complexity}$$
Combining the results of Lemmas \[lem:approximation\] and \[lem:MLEs\] gives$$\tilde{L}\left(\hat{\theta}\left(\tilde{\mathbf{T}}\left(t\right)\right);\tilde{\mathbf{T}}\left(t\right)\right)-\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right)\right);\mathbf{T}_{i}\left(t\right)\right)\overset{\text{a.s.}}{\rightarrow}0$$ for all $t\in\mathcal{T}_{i}$ since $\mathcal{T}_{i}=\mathcal{T}_{1}$ and the functions are almost surely continuous by assumption.
\[sub:Optimal-discrimination-information\]Optimal discrimination information
----------------------------------------------------------------------------
For any $\Theta^{\prime}\subseteq\Theta$, let $\bar{g}_{i}\left(t_{i};\Theta^{\prime}\right)$ denote the optimal predictive density function relative to $\left\langle \left\{ g_{\theta}:\theta\in\Theta^{\prime}\right\} ,w_{i}\right\rangle $ as defined in Section \[sub:NMWL\]. For any $\Theta_{0},\Theta_{1}\subseteq\Theta$, the *optimal information in $x$ for discrimination* in favor of the hypothesis that $\theta_{i}\in\Theta_{1}$ over the hypothesis that $\theta_{i}\in\Theta_{0}$ is$$\bar{I}_{i}\left(\Theta_{1},\Theta_{0}\right)=-\log\bar{g}_{i}\left(t_{i};\Theta_{0}\right)-\left(-\log\bar{g}_{i}\left(t_{i};\Theta_{1}\right)\right),$$ generalizing quantities in @RefWorks:292, @RefWorks:342, @RefWorks:435, and @mediumScale. The approximate optimal information $\tilde{I}_{i}\left(\Theta_{1},\Theta_{0}\right)$ is defined identically except with $\tilde{g}_{i}$ in place of $\bar{g}_{i}$. $\bar{I}_{i}\left(\Theta_{1},\Theta_{0}\right)$ is not restricted to the case of smoothness conditions on $\left\{ g_{\theta}:\theta\in\Theta\right\} $, but applies to any problem of selecting one of two models.
Since $\bar{g}_{i}\left(t_{i};\Theta_{1}\right)/\bar{g}_{i}\left(t_{i};\left\{ \theta_{0}\right\} \right)$ for $\theta_{0}\in\Theta$is a likelihood ratio, the discrimination information has the universal bound on the probability of misleading evidence under $\theta=\theta_{0}$ ([@RefWorks:123]; [@mediumScale]). The next lemma and theorem bear on whether the optimal information for discrimination is an interpretable measure of evidence in that the probability of observing misleading information converges to 0 as $n_{i}\rightarrow\infty$. Let $\hat{\theta}_{i}\left(\mathbf{T};\Theta^{\prime}\right)=\arg\sup_{\theta\in\Theta^{\prime}}\bar{L}_{i}\left(\theta;\mathbf{T}\right)$ for $i=1,\dots,N$ and $\hat{\theta}_{0}\left(t;\Theta^{\prime}\right)=\arg\sup_{\theta\in\Theta^{\prime}}L_{i}\left(\theta;t\right)$ given any $\Theta^{\prime}\subseteq\Theta$.
\[lem:complexity\]Suppose $\Theta=\mathbb{R}^{D}$, $\theta=\left\langle \theta_{1},\dots,\theta_{D}\right\rangle ^{\operatorname{T}}$, $n_{j}=O\left(n_{k}\right)$ for all $j,k\in\left\{ 1,\dots,N\right\} $, $\mathcal{G}_{n,i}=\mathcal{G}_{n,1}$ and $L_{i}=L_{1}$ for all $i\in\left\{ 1,\dots,N\right\} $ and sufficiently large $n$, and $\tau\left(x\right)=x$ for all $x\in\mathcal{X}^{n}$, which implies that $\mathcal{T}_{i}=\mathcal{X}^{n}$, $t_{i}=x_{i}$, and $T_{i}=X_{i}$. Assume also that for some $i\in\left\{ 1,\dots,N\right\} $, there exists an open, bounded set $\Theta^{\prime}\subseteq\Theta$ on which $L\left(\bullet;X_{i}\right)$ is almost surely continuous and such that $$\log\int_{\mathcal{X}^{n}}\bar{L}_{i}\left(\hat{\theta}_{0}\left(t;\Theta^{\prime}\right);t\right)dt=\frac{D}{2}\log\frac{n_{i}}{2\pi}+\log\int_{\Theta^{\prime}}\sqrt{\frac{1}{n_{i}}\left|E\frac{\partial^{2}\ln g_{\theta}\left(X_{i}\right)}{\partial\theta\partial\theta^{\operatorname{T}}}\right|}d\theta+o\left(1\right).\label{eq:complexity}$$ $$\therefore\log\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)dt=\frac{D}{2}\log\frac{n_{i}}{2\pi}+\log\int_{\Theta^{\prime}}\sqrt{\frac{1}{n_{i}}\left|E\frac{\partial^{2}\ln g_{\theta}\left(X_{i}\right)}{\partial\theta\partial\theta^{\operatorname{T}}}\right|}d\theta+o\left(1\right)$$ almost surely holds for any weights that satisfy $P\left(\lim_{n_{i}\rightarrow\infty}w_{ii}=1\right)=1$ and $\sum_{j=1}^{N}w_{ij}=1$.
The continuity condition and the constraints on the weights and sample sizes ensure that $\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)\overset{\text{a.s.}}{\rightarrow}\bar{L}_{i}\left(\hat{\theta}_{0}\left(t;\Theta^{\prime}\right);t\right)$ as $n_{i}\rightarrow\infty$ for all $t\in\mathcal{T}_{i}$.
The assumptions of Lemma \[lem:complexity\] are broadly applicable since equation holds under general regularity conditions [@RefWorks:396]. The result will now be extended to non-bounded parameter spaces.
\[thm:unbounded-complexity\]Suppose that $\Theta_{1}\subseteq\Theta$, that $n_{j}=O\left(n_{k}\right)$ for all $j,k\in\left\{ 1,\dots,N\right\} $, and that $P$ is the sampling distribution of $\mathbf{T}$. Assume also that for any $i\in\left\{ 1,\dots,N\right\} $, there exists an open, bounded set $\Theta^{\prime}\subseteq\Theta_{1}$ such that $$P\left(\lim_{n_{i}\rightarrow\infty}\log\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)dt=\infty\right)=1.\label{eq:unbounded-complexity}$$ $$\therefore P\left(\lim_{n_{i}\rightarrow\infty}\log\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)dt=\infty\right)=1.$$
Let $\mathfrak{T}=\left\{ t\in\mathcal{T}_{i}:\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right)\in\Theta^{\prime}\right\} $ to expand $\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)dt$ as $$\int_{\mathfrak{T}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)dt+\int_{\mathcal{T}_{i}\backslash\mathfrak{T}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)dt.$$ Thus, since $\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)=\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)$ for all $t\in\mathfrak{T}$ and $$\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)>\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)$$ for all $t$ in non-empty $\mathcal{T}_{i}\backslash\mathfrak{T}$ given any sufficiently large $n_{i}$,$$\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta_{1}\right);\mathbf{T}_{i}\left(t\right)\right)dt\ge\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right);\Theta^{\prime}\right);\mathbf{T}_{i}\left(t\right)\right)dt$$ follows, where the equality and both inequalities hold with $P$-probability 1.
Since the claim of Lemma \[lem:complexity\] implies equation , Theorem \[thm:unbounded-complexity\] applies to the wide class of models satisfying the regularity conditions of @RefWorks:396. The largely overlapping regularity conditions of @RefWorks:387 then ensure that $\lim_{n_{i}\rightarrow\infty}P\left(\bar{I}_{i}\left(\Theta_{1},\left\{ \theta_{0}\right\} \right)>0\right)=0$ when there is no $\theta\in\Theta$ such that $g_{\theta}$ is closer in Kullback-Leibler divergence than $g_{\theta_{0}}$ to the marginal distribution $P\left(T_{i}\in\bullet\right)$. In the special case of correct model specification considered in @RefWorks:123 and @mediumScale, the equation holds for all $P$ admitting $g_{\theta_{0}}$ as the marginal density of $T_{i}$.
\[sub:Single-observation-weights\]Single-observation weights
------------------------------------------------------------
This section defines *single-observation weights* as the components of $w_{i}$ such that for every $i\in1,\dots,N$ that all incidental data (all $x_{j}$ with $j\ne i$) together have the weight of one observation in the focus vector $x_{i}$ $\left(\sum_{j\ne i}w_{ij}=w_{ii}/n_{i}\right)$ and that each comparison other than the $i$th has equal weight $\left(\forall j,k\ne i\, w_{ij}=w_{ik}\right)$. Solving those equations and $\sum_{j=1}^{N}w_{ij}=1$ uniquely gives $w_{ii}=1-\left(n_{i}+1\right)^{-1}$ and $i\ne j\implies w_{ij}=\left(n_{i}+1\right)^{-1}\left(N-1\right)^{-1}$.
If there is only a single comparison, then its observed statistic $t_{1}=\tau\left(x_{1}\right)$ is supplemented by a pseudo-statistic $t_{0}$, a scientifically meaningful value in $\mathcal{T}_{1}$ that does not depend on $x_{1}$. For example, $t_{0}$ might be $\int tg_{\theta_{0}}\left(t\right)dt$, the expectation value of $T_{1}$ under $\theta=\theta_{0}$. (Similarly, @RefWorks:1147 considered the use of a prior with the Fisher information of a single pseudo-observation.) The use of $\mathbf{t}=\left\langle t_{0},t_{1}\right\rangle $ and $N=2$ with single-observation weights then entails that $$\log\bar{L}_{1}\left(\theta_{1};\mathbf{t}\right)=\left(n_{1}+1\right)^{-1}\log L_{1}\left(\theta_{1};t_{0}\right)+\left(1-\left(n_{1}+1\right)^{-1}\right)\log L_{1}\left(\theta_{1};t_{1}\right).\label{eq:null-weights}$$ For a smoother transition from a single comparison to multiple comparisons, the pseudo-statistic may be assigned the same weight as each of the $N-1$ incidental statistics among $t_{1},\dots,t_{N}$, i.e., $w_{ij}=\left(n_{i}+1\right)^{-1}N^{-1}$ for all $j\in\left\{ 0,1,\dots,N\right\} \backslash\left\{ i\right\} $.
The following result applies whether there is a single comparison or multiple comparisons.
Assume the components of $w_{i}$ are single-observation weights, that $n_{i}=n_{1}$ for all $i=1,\dots,N$, and that $\mathcal{G}_{n_{1},i}=\mathcal{G}_{n_{1},1}$ for all $i=1,\dots,N$. If $T_{1},\dots,T_{N}$ are independent and drawn from a mixture distribution, if $\hat{\theta}\left(\tilde{\mathbf{T}}\left(t\right)\right)$ and $\hat{\theta}_{i}\left(\mathbf{T}_{i}\left(t\right)\right)$ are almost always unique for all $t\in\mathcal{T}_{i}$, and if $L_{i}\left(\bullet;T_{i}\right)$ is almost surely continuous on $\Theta$ for all $i=1,\dots,N$, then equation holds.
All the conditions of Theorem \[thm:approximate-complexity\] are given except for the equal weights condition, which follows from the single-observation weights assumption, the equality of the sample sizes, and the commonality of the family of distributions.
\[sec:Case-study\]Case studies
==============================
In the following models, $\int f_{\hat{\phi}\left(x\right)}\left(x\right)dx=\infty$, rendering the unweighted NML useless. The NMWL can be used instead since $\int_{\mathcal{T}_{i}}\bar{L}_{i}\left(\hat{\theta}_{i}\left(\mathbf{t}_{i}\left(t\right)\right);\mathbf{t}_{i}\left(t\right)\right)dt<\infty$.
Results of two separate NMWL analyses are presented for each application. The first uses multiple comparisons for inference relevant to each comparison . The second uses $\int tg_{\theta_{0}}\left(t\right)dt=0$ in place of data associated with other comparisons, as if there were only a single comparison . All plots use the binary logarithm to express information in bits and display a different value for each comparison.
\[sub:Populations\]Single and multiple populations
--------------------------------------------------
Before addressing a problem in contemporary biology, the proposed methodology will be illustrated using a simple data set that has motivated both Bayesian [@Rubin1981b] and weighted likelihood [@Wang2006279] approaches. The reduced data consist of the estimated average effect of a training program on SAT scores and an estimated standard error of the effect estimate for each of eight test sites. Following the tradition continued by @Wang2006279, the standard errors $\sigma_{1},\dots,\sigma_{8}$ are considered known, and the effect estimates are modeled as normal observations with unknown means $\theta_{1},\dots,\theta_{8}$. Thus, $N=8$ and $\left\{ g_{i,\theta_{i}}:\theta\in\Theta\right\} $ is the family of distributions, where $g_{i,\theta_{i}}$ is the normal density of mean $\theta_{i}$ and standard deviation $\sigma_{i}$. For the $i$th site, $\theta_{i}\ne0$ is the alternative hypothesis and $\theta_{i}=0$ is the null hypothesis.
Fig. \[fig:infoSAT\] displays $\tilde{I}_{i}\left(\mathbb{R}\backslash\left\{ 0\right\} ,\left\{ 0\right\} \right)$, the resulting approximate discrimination information, with $\bar{I}_{i}\left(\mathbb{R}\backslash\left\{ 0\right\} ,\left\{ 0\right\} \right)$, the exact discrimination information. As in Section \[sub:Single-observation-weights\], the weight of a single observation is assigned either to 0, the null hypothesis value (“information from null”), or to the incidental testing sites (“information from sites”). The resulting information values are barely distinguishable.
![\[fig:infoSAT\]Information (bits) favoring the hypothesis that the test score at a site was affected by the treatment.](infoSAT.png)
\[sub:Biological-features\]Single and multiple biological features
------------------------------------------------------------------
In typical experiments measuring gene expression or the abundance of proteins or metabolites, the primary question is whether the expectation value of a logarithm of the expression or abundance of each feature is affected by a treatment, disease, or other perturbation. Since that question is equivalent to that of whether $\text{CV}_{i}^{-1}$, the inverse coefficient of variation ** for the $i$th feature, is 0, the data reduction strategy of Example \[exa:microarray\] often proves effective even if the magnitude of $\text{CV}_{i}^{-1}$ is not of direct interest. $\text{CV}_{i}^{-1}$ has a one-to-one correspondence to the proportion of the feature-feature pairs with abundance ratios greater than 1 [@RefWorks:21; @RefWorks:26]. In addition, $\text{CV}_{i}^{-1}$ is often of more scientific interest than the mean since small changes in numbers of biomolecules can have a strong influence on downstream processes.
The method of Example \[exa:microarray\] is applied to the proteomics data set of Alex Mirons lab at the Dana-Farber Cancer Institute [@ProData2009b], with $x_{ij}$ and $y_{ij}$ as the logarithms of the abundance levels of the $i$th of $N=20$ proteins in the $j$th woman with and without breast cancer, respectively, after the preprocessing of @mediumScale. Likewise, $\xi_{i}$ and $\eta_{i}$ are the expectation values of the random variables $X_{ij}$ and $Y_{ij}$. Each of two breast cancer groups (one of 55 HER2-positive women and the other of 35 women mostly-ER/PR-positive) were compared to a control group of 64 women. Since $\theta_{i}=\left|\text{CV}_{i}^{-1}\right|$ and thus $\Theta=\left[0,\infty\right)$, the competing hypotheses for the $i$th protein are $\theta_{i}>0$ and $\theta_{i}=0$.
The left panel of Fig. \[fig:infoAB\] displays the approximate information for discrimination in favor of the alternative hypothesis that $\theta_{i}\ne0$ over the null hypothesis that $\theta_{i}=0$ by weighing the incidental proteins as a single observation (§\[sub:Single-observation-weights\]). $\tilde{I}_{i}\left(\left(0,\infty\right),\left\{ 0\right\} \right)$, the approximate optimal information, is compared to $\log\left(g_{i}\left(\hat{\theta}_{\text{MLE}};t_{i}\right)/g_{i}\left(0;t_{i}\right)\right)$. Here, $\hat{\theta}_{\text{MLE}}$ is common to all proteins, denoting the maximum likelihood estimate (MLE) defined under the assumptions that $\theta_{i}\in\left\{ 0,\theta_{\text{alt.}}\right\} $ for some $\theta_{\text{alt.}}>0$ for all $i$ and that the test statistics are independent [@mediumScale]. The right panel of Fig. \[fig:infoAB\] contrasts the widely varying regret of the MLE information with the constant regret of the optimal information.
![\[fig:infoAB\]*Left panel:* Discrimination information (bits) favoring the hypothesis that the abundance level of a protein differs by disease status versus $\widehat{\text{CV}}_{i}^{-1}$. *Right panel:* The corresponding regret versus $\widehat{\text{CV}}_{i}^{-1}$. ($\widehat{\text{CV}}_{i}^{-1}$ denotes the difference in sample means divided by the sample standard deviation for the $i$th protein.)](infoAB.png)
Giving the null hypothesis the weight of a single observation , as if the abundance level of only one protein were measured, results in information values that are visually indistinguishable from those of Fig. \[fig:infoAB\]. Nonetheless, some effect of the weighting method is perceptible for much smaller sample sizes. For example, Fig. \[fig:nullABCB\] displays the effect of using the null hypothesis weights instead of the protein weights on $\tilde{I}_{i}\left(\left(0,\infty\right),\left\{ 0\right\} \right)$ for two randomly selected patients from each breast cancer group and from the healthy group. Even in this extreme case, only one protein out of 20 in the right-side panel has a different evidence grade (Table \[tab:Grades-of-hypothesis-evidence\]) depending on how the weights are computed.
![\[fig:nullABCB\]Discrimination information (bits) favoring the hypothesis that the abundance level of a protein differs by disease status ($n=2$ women per group) using weights from the null hypothesis versus that using weights from the incidental proteins. ](nullABCB.png)
\[sec:Discussion\]Discussion
============================
In both of the case studies of Section \[sec:Case-study\], the use of data associated with comparisons other than the comparison currently in focus in place of an artificial data point determined by the null hypothesis has little effect on the information for discrimination. In the second application, little information was lost for inference about a single protein were the other 19 absent except when the sample size was reduced to $n_{i}=4$. Thus, the use of single-observation weights robustly addresses the infinite-complexity issue with NML raised in Section \[sub:Optimality-observed\].
The insensitivity to the use of incidental information also suggests that the NMWL solution to the incidental-information issue raised in the same section is a measure of evidence that has the same interpretation for any number of comparisons. By contrast, p-values adjusted to control error rates and, to a lesser extent, posterior probabilities from hierarchical Bayesian models, tend to vary so greatly between a single comparison and a large number of comparisons that they require researchers to separately build the intuition needed to interpret statistical reports for small numbers of comparisons, medium numbers of comparisons, large numbers of comparisons, etc. This shortcoming of traditional approaches to the multiple comparisons problem is especially glaring when an article reports various degrees of adjusting p-values for data types involving very different numbers of features.
As seen in Section \[sub:Populations\], the optimal information for discrimination can indicate strong evidence for a simple null hypothesis. While in principle the Bayes factor can also favor the null hypothesis, prior distributions commonly used in practice often can provide only weak Bayes-factor support for a simple null hypothesis that corresponds to the data-generating distribution [@Johnson2010143]. The ability of the information for discrimination to indicate whether the evidence in the data is strongly in favor of the alternative hypothesis, strongly in favor of the null hypothesis, or insufficient to strongly favor either hypothesis (Table \[tab:Grades-of-hypothesis-evidence\]) guards against the prevalent misinterpretation of a high p-value as evidence for a null hypothesis. More important, the discrimination information provides scientists a reliable tool designed to objectively answer the questions they ask of their data.
Acknowledgments {#acknowledgments .unnumbered}
===============
The author thanks Corey Yanofsky for comments on the manuscript. `Biobase` [@RefWorks:161] facilitated data management. This research was partially supported by the Canada Foundation for Innovation, by the Ministry of Research and Innovation of Ontario, and by the Faculty of Medicine of the University of Ottawa.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The empirical copula has proved to be useful in the construction and understanding of many statistical procedures related to dependence within random vectors. The empirical beta copula is a smoothed version of the empirical copula that enjoys better finite-sample properties. At the core lie fundamental results on the weak convergence of the empirical copula and empirical beta copula processes. Their scope of application can be increased by considering weighted versions of these processes. In this paper we show weak convergence for the weighted empirical beta copula process. The weak convergence result for the weighted empirical beta copula process is stronger than the one for the empirical copula and its use is more straightforward. The simplicity of its application is illustrated for weighted Cramér–von Mises tests for independence and for the estimation of the Pickands dependence function of an extreme-value copula.'
address:
- 'Ruhr-Universität Bochum, Fakultät für Mathematik, Universitätsstr. 150, 44780 Bochum, Germany'
- 'Université catholique de Louvain, ISBA, Voie du Roman Pays 20, 1348 Louvain-la-Neuve, Belgium'
author:
- Betina Berghaus
- Johan Segers
bibliography:
- 'biblioN.bib'
title: Weak convergence of the weighted empirical beta copula process
---
Copula ,empirical beta copula ,empirical copula ,weighted weak convergence ,Pickands dependence function.
Introduction
============
In many statistical questions related to multivariate dependence, a crucial role is played by the copula function. A basic nonparametric copula estimator is the empirical copula, dating back to [@Rus76; @Deh79] and defined as the empirical distribution function of the vectors of component-wise ranks. The asymptotic behavior of the empirical copula has been established under various assumptions on the true copula and the serial dependence of the observed random vectors [see, e.g., @GanStu87; @FerRadWeg04; @Seg12; @BucVol13]. The upshot is that the empirical copula process converges weakly to a centered Gaussian field with covariance function depending on the true copula and the serial dependence of the observations.
Recently, @BerBucVol17 investigated the weak convergence of the weighted empirical copula process. They showed that the empirical copula process divided by a weight function, that can be zero on parts of the boundary of the unit cube, still converges weakly to a Gaussian field. As illustrated in the latter reference, this stronger result allows for additional applications of the continuous mapping theorem or the functional delta method. However, this result is only valid for a clipped version of the process. Since the empirical copula itself is not a copula, weak convergence fails on the upper boundaries of the unit cube [@BerBucVol17 Remark 2.3].
The empirical beta copula [@SegSibTsu17] arises as a particular case of the empirical Bernstein copula [see, e.g., @SanSat04; @JanSwaVer12] if the degrees of the Bernstein polynomials are set to the sample size. In the numerical experiments in [@SegSibTsu17], the empirical beta copula exhibited a better performance than the empirical copula, both in terms of bias and variance.
In contrast to the empirical copula, the empirical beta copula is a genuine copula, a property that it shares with the checkerboard copula, whose limit is derived in [@GenNesRem2017] and which is very close to the empirical copula if the margins are continuous. Since the empirical beta copula is itself a copula, it is possible to prove weighted weak convergence for the empirical beta copula process on the whole unit cube. This is the main result of the paper. Weak convergence on the whole unit cube rather than on a subset thereof is quite handy since it allows for a direct application of, e.g., the continuous mapping theorem. In particular, there is no longer any need to treat the boundary regions separately.
We consider two applications. First, we modify the Cramér–von Mises test statistic for independence in [@genest+r:2004] by using the empirical beta copula and, more importantly, adding a weight function in the integral, emphasizing the tails. The asymptotic distribution of the statistic under the null hypothesis is an easy corollary of our main result. More interestingly, the inclusion of a weight function leads to a markedly better power against difficult alternatives such as the *t* copula with zero correlation parameter, with favorable comparisons even to the novel statistics introduced recently by @belalia+b+l+t:2017. As a second application we consider the Capéràa–Fougères–Genest estimator [@CapFouGen97] of the Pickands dependence function of a multivariate extreme-value copula. Under weak dependence, replacing the empirical copula by the empirical beta copula yields a more accurate estimator. Its asymptotic distribution is again an immediate consequence of our main result.
The paper is organized as follows. In Section \[sec:main\] we introduce the various empirical copula processes and we state the main result of the paper, the weighted convergence of the empirical beta copula process on the whole unit cube. We illustrate the ease of application of the main result to the analysis of weighted Cramér–von Mises tests of independence (Section \[sec:indep\]) and nonparametric estimation of multivariate extreme-value copulas (Section \[sec:pick\]). The proofs are deferred to Section \[sec:proof\], whereas a number of technical arguments are worked out in detail in Section \[sec:aux\].
Notation and main result {#sec:main}
========================
Let $({\bm }X_n)_n$ be a strictly stationary time series whose $d$-variate stationary distribution function $F$ has continuous marginal distribution functions $F_1,\dots, F_d$ and copula $C$. Writing ${\bm }X_i = (X_{i,1}, \ldots, X_{i,d})$, we have, for ${\bm }x \in {\mathbb{R}}^d$, $$\begin{aligned}
{\mathbb{P}}( X_{i,j} {\leqslant}x_j ) &= F_j(x_j), &
{\mathbb{P}}( {\bm }X_i {\leqslant}{\bm }x ) &= F({\bm }x) = C\{ F_1(x_1), \ldots, F_d(x_d) \}.\end{aligned}$$ For vectors ${\bm }x, {\bm }y \in {\mathbb{R}}^d$, the inequality $ {\bm }x {\leqslant}{\bm }y$ means that $x_j {\leqslant}y_j$ for $j = 1, \dots, d$. Similar conventions apply for other inequalities and for minima and maxima, denoted by the operators $\wedge$ and $\vee$, respectively. Given the sample ${\bm }X_1, \ldots, {\bm }X_n$, the aim is to estimate $C$ and functionals thereof.
Although the copula $C$ captures the instantaneous (cross-sectional) dependence, the setting is still general enough to include questions about serial dependence. For instance, if $(Y_n)_{n}$ is a univariate, strictly stationary time series, then the $d$-variate time series of lagged values ${\bm }X_n = (Y_n, Y_{n-1}, \ldots, Y_{n-d+1})$ is strictly stationary too and the instantaneous dependence within the series $({\bm }X_n)_n$ corresponds to serial dependence within the original series $(Y_n)_n$ up to lag $d-1$.
For $i = 1, \ldots, n$ and $j = 1, \ldots, d$, let $R_{i,j}$ denote the rank of $X_{i,j}$ among $X_{1,j}, \ldots, X_{n,j}$. For convenience, we omit the sample size $n$ in the notation for ranks. The random vectors $\hat{{\bm }U}_i = (\hat{U}_{i,1}, \ldots, \hat{U}_{i,d})$, with $\hat{U}_{i,j} = n^{-1} R_{i,j}$ and $i = 1, \ldots, n$, are called pseudo-observations from $C$. Letting ${\operatorname{\mathds{1}}}_A$ denote the indicator of the event $A$, the empirical copula is $$\hat C_n({\bm }u) = \frac{1}{n} \sum_{i=1}^n {\operatorname{\mathds{1}}}_{ \{ \hat{{\bm }U}_i {\leqslant}{\bm }u\} }, \qquad {\bm }u \in [0,1]^d.$$
Under mixing conditions on the sequence $( {\bm }X_n )_n$ and smoothness conditions on $C$, @BucVol13 showed that $$\label{eq:hCn:weak}
\hat{{\mathbb{C}}}_n = \sqrt n(\hat C_n - C) {\rightsquigarrow}{\mathbb{C}}_C, \qquad n \to \infty
$$ in the metric space $\ell^\infty([0,1]^d) = \{ f : [0, 1]^d \to {\mathbb{R}}\mid \sup_{{\bm }u \in [0, 1]^d} \lvert f({\bm }u) \rvert < \infty \}$ equipped with the supremum distance. The arrow ${\rightsquigarrow}$ in denotes weak convergence in metric spaces as exposed in [@VanWel96]. The limit process in is $${\mathbb{C}}_C({\bm }u)
= \alpha_C({\bm }u) - \sum_{j=1}^d \dot C_j({\bm }u) \, \alpha_C(1, \dots, 1, u_j, 1, \dots, 1),
\qquad {\bm }u \in [0, 1]^d,$$ where $\dot{C}_j({\bm }u) = \partial C({\bm }u) / \partial u_j$ and where $\alpha_C$ is a tight, centered Gaussian process on $[0, 1]^d$ with covariance function $$\label{eq:cov}
{\operatorname{Cov}}\bigl( \alpha_C({\bm }u), \alpha_C({\bm }v) \bigr)
= \sum_{i \in {\mathbb{Z}}} {\operatorname{Cov}}\bigl( {\operatorname{\mathds{1}}}_{ \{ {\bm }U_0 {\leqslant}{\bm }u \} }, {\operatorname{\mathds{1}}}_{ \{ {\bm }U_i {\leqslant}{\bm }v \} } \bigr),
\qquad {\bm }u, {\bm }v \in [0, 1]^d,$$ where $\bm{U}_i = (U_{i,1}, \ldots, U_{i,d})$ and $U_{i,j} = F_j(X_{i,j})$. Since $F_j$ is continuous, the random variables $U_{i,j}$ are uniformly distributed on $[0, 1]$. The joint distribution function of $\bm{U}_i$ is $C$. The margins $F_1, \ldots, F_d$ being unknown, we cannot observe the ${\bm }U_i$, and this is why we use the $\hat{{\bm }U}_i$ instead. In the case of serial independence, weak convergence of $\hat{{\mathbb{C}}}_n$ has been investigated by many authors, see the survey by @BucVol13; the series in simplifies to ${\operatorname{Cov}}( {\operatorname{\mathds{1}}}_{ \{{\bm }U_0 {\leqslant}{\bm }u\} }, {\operatorname{\mathds{1}}}_{ \{ {\bm }U_0 {\leqslant}{\bm }v \} } ) = C( {\bm }u \wedge {\bm }v ) - C( {\bm }u) C( {\bm }v)$ so that $\alpha_C$ is a $C$-Brownian bridge. In the stationary case, convergence of the series in is a consequence of the mixing conditions imposed on $({\bm }X_n)_n$.
Weak convergence in is helpful for deriving asymptotic properties of estimators and test statistics based upon the empirical copula, such as estimators of Kendall’s tau or Spearman’s rho or such as Kolmogorov–Smirnov and Cramér–von Mises statistics for testing independence. However, as argued by @BerBucVol17, sometimes weak convergence with respect to a stronger metric is required, i.e., a weighted supremum norm. Examples mentioned in the cited article include nonparametric estimators of the Pickands dependence function of an extreme-value copula and bivariate rank statistics with unbounded score functions such as the van der Waerden rank (auto-)correlation. This motivates the study of the weighted empirical copula process $\hat {{\mathbb{C}}}_n/g^\omega$, with $\omega \in(0,1/2)$ and a suitable weight function $g$ on $[0, 1]^d$. The limit of the empirical copula process is zero almost surely as soon as one of its arguments is zero or if all arguments but at most one are equal to one. We can thus hope to obtain weak convergence with respect to a weight function that vanishes at such points. A possible function with this property is $$\label{eq:g}
g({\bm }u ) =
\bigwedge_{j=1}^d \biggl\{ u_j \wedge \bigvee_{k \ne j} (1 - u_k) \biggr\},
\qquad {\bm }u \in [0, 1]^d.$$ Note that $g({\bm }u)$ is small as soon as there exists $j$ such that either $u_j$ is small or else all other $u_k$ are close to $1$. The trajectories of the processes $\hat{{\mathbb{C}}}_n/g^\omega$ are not bounded on the unit cube, hence the processes cannot converge weakly in $\ell^\infty([0,1]^d)$. A solution is to restrict the domain from $[0, 1]^d$ to sets of the form $[c/n, 1 - c/n]^d$ for $c \in (0, 1)$, or, more generally, to $\{ {\bm }v \in [0, 1]^d : g({\bm }v) {\geqslant}c/n \}$. Relying on such a workaround, Theorem 2.2 in [@BerBucVol17] states weak convergence of the weighted empirical copula process $\hat{{\mathbb{C}}}_n/g^\omega$ to ${\mathbb{C}}_C/g^\omega$. Note that $g({\bm }v) = 0$ if and only if $v_j = 0$ for some $j$ or if there exists $j$ such that $v_k = 1$ for all $k \ne j$, and that ${\mathbb{C}}_C({\bm }v) = 0$ almost surely for such ${\bm }v$ too. The empirical copula is a piecewise constant function whereas the estimation target is continuous. It is natural to consider smoothed versions of the empirical copula. @SegSibTsu17 defined the empirical beta copula as $$\label{eq:empBetaCop}
{C_n^{\beta}}({\bm }u )= \frac{1}{n} \sum_{i=1}^n\prod_{j=1}^d F_{n,R_{i,j}}(u_j),
\qquad {\bm }u =(u_1,\dots , u_d) \in [0,1]^d,$$ where $F_{n,r}$ is the distribution function of the beta distribution $\mathcal{B}(r,n+1-r)$, i.e., $F_{n,r}(u) = \sum_{s=r}^n \binom{n}{s} u^s (1-u)^{n-s}$, for $u \in [0, 1]$ and $r \in \{1, \ldots, n\}$. Note that $$\label{eq:empCop2empBetaCop}
{C_n^{\beta}}({\bm }u ) = \int_{[0,1]^d} \hat C_n ({\bm }w) {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w),$$ where $\mu_{n,{\bm }u}$ is the law of the random vector $(S_1/n, \dots , S_d/n)$, with $S_1, \dots, S_d$ being independent binomial random variables, $S_j \sim {\operatorname{Bin}}(n, u_j)$. In the absence of ties, the rank vector $(R_{1,j}, \ldots, R_{n,j})$ of the $j$-th coordinate sample is a permutation of $(1, \ldots, n)$. As a consequence, the empirical beta copula can be shown to be a genuine copula, unlike the empirical copula.
Under a smoothness condition on $C$, it follows from Theorem 3.6(ii) in [@SegSibTsu17] that weak convergence in $\ell^\infty([0, 1]^d)$ of the empirical copula process $\hat{{\mathbb{C}}}_n$ in to a limit process ${\mathbb{C}}$ with continuous trajectories is sufficient to conclude the weak convergence of the empirical beta copula process: in the space $\ell^\infty([0, 1]^d)$, we have $$\label{eq:betacop:weak}
{\mathbb{C}}_n^{\beta} = \sqrt n ({C_n^{\beta}}- C) = \hat{{\mathbb{C}}}_n + {\mathrm{o}}_{{\mathbb{P}}}(1) {\rightsquigarrow}{\mathbb{C}}, \qquad n \to \infty.$$ The asymptotic distribution of the empirical beta copula is thus the same as the one of the empirical copula. Still, for finite samples, numerical experiments in [@SegSibTsu17] revealed the empirical beta copula to be more accurate.
Our aim is to extend the convergence statement in for weighted versions ${\mathbb{C}}_n^{\beta} / g^\omega$, with $g$ as in and for suitable exponents $\omega > 0$. As the empirical beta copula is a genuine copula, the zero-set of ${\mathbb{C}}_n^{\beta}$ includes the zero-set of $g$, and on this set we implicitly define ${\mathbb{C}}_n^{\beta} / g^\omega$ to be zero. With this convention, the sample paths of ${\mathbb{C}}_n^{\beta} / g^\omega$ are bounded on $[0, 1]^d$; see Lemma \[lem:boundary\] below. We can therefore hope to prove weak convergence of ${\mathbb{C}}_n^{\beta} / g^\omega {\rightsquigarrow}{\mathbb{C}}_C/g^\omega$ in $\ell^\infty([0, 1]^d)$ without having to exclude those border regions of $[0, 1]^d$ where $g$ is small, as was necessary in [@BerBucVol17].
The analysis of ${\mathbb{C}}_n^{\beta}/g^\omega$ will be based on the one of $\hat{{\mathbb{C}}}_n/g^\omega$ via . We will therefore need the same smoothness condition on $C$ as imposed in @BerBucVol17 [Condition 2.1], combining Conditions 2.1 and 4.1 in [@Seg12]. Condition \[cond:second\] below is satisfied by many copula families: in [@Seg12 Section 5], part (i) of the condition is verified for strict Archimedean copulas with continuously differentiable generators, whereas both parts of the condition are verified for the non-singular bivariate Gaussian copula and for bivariate extreme-value copulas with twice continuously differentiable Pickands dependence function and a growth condition on the latter’s second derivative near the boundary points of its domain.
\[cond:second\] (i) For every $j \in \{ 1, \dots, d \}$, the first-order partial derivative $\dot C_j({\bm }u) := \partial C({\bm }u)/\partial u_j$ exists and is continuous on $V_j=\{ {\bm }u \in [0,1]^d: u_j \in (0,1) \}$.
\(ii) For every $j_2, j_2 \in \{1, \dots, d\}$, the second-order partial derivative $\ddot C_{j_1 j_2}({\bm }u) := \partial^2 C({\bm }u)/\partial u_{j_1}\partial u_{j_2}$ exists and is continuous on $V_{j_1} \cap V_{j_2}$. Moreover, there exists a constant $K>0$ such that, for all $j_1, j_2 \in \{1, \ldots, d\}$, we have $$\label{eq:second}
\bigl\lvert \ddot C_{j_1j_2}({\bm }u) \bigr\rvert
{\leqslant}K \min \left\{ \frac{1}{u_{j_1}(1-u_{j_1})}, \frac{1}{u_{j_2}(1-u_{j_2})} \right\}, \qquad \forall\, {\bm }u \in V_{j_1} \cap V_{j_2}.$$
The alpha-mixing coefficients of the sequence $({\bm }X_n)_n$ are defined as $$\alpha(k) =
\sup \left\{
\lvert {\mathbb{P}}(A \cap B) - {\mathbb{P}}(A) \, {\mathbb{P}}(B) \rvert :
A \in \sigma({\bm }X_j, j {\leqslant}i), B \in \sigma({\bm }X_{j+k}, j {\geqslant}i), i \in {\mathbb{Z}}\right\},$$ for $k = 1, 2, \ldots$. The sequence $({\bm }X_n)_n$ is said to be strongly mixing or alpha-mixing if $\alpha(k) \to 0$ as $k \to \infty$. Now we can state the main result.
\[thm:main\] Suppose that ${\bm }X_1, {\bm }X_2, \dots$ is a strictly stationary, alpha-mixing sequence with $\alpha(k) = {\mathrm{O}}(a^k)$, as $k \to \infty$, for some $a \in (0,1)$. Assume that within each variable, ties do not occur with probability one. If the copula $C$ satisfies Condition \[cond:second\], then, for any $\omega\in[0,1/2)$, we have, in $\ell^\infty([0, 1]^d)$, $${\mathbb{C}}_n^\beta/g^\omega {\rightsquigarrow}{\mathbb{C}}_C/g^\omega, \qquad n \to \infty.$$
The tie-excluding assumption is needed to ensure that the empirical beta copula is a genuine copula almost surely. The assumption implies that the $d$ stationary marginal distributions are continuous. For iid sequences, continuity of the margins is also sufficient. In the strictly stationary case, ties may occur with positive probability even if the margins are continuous; for instance, take a Markov chain where the current state is repeated with positive probability.
The result also holds under weaker assumptions on the serial dependence. In [@BerBucVol17] it is shown that weak convergence of the weighted empirical copula process is still valid under more general assumptions on the marginal empirical processes and quantile processes and an assumption on the multivariate empirical process. In this case, however, the range of $\omega$ is smaller [@BerBucVol17 Theorem 4.5].
Application: weighted Cramér–von Mises tests for independence {#sec:indep}
=============================================================
Testing for independence is a classical subject which still attracts interest today. One approach consists of comparing the multivariate empirical cumulative distribution function to the product of empirical cumulative distribution functions. Integrating out the difference with respect to the sample distribution yields a Cramér–von Mises style test statistic going back to @hoeffding:1948 and @blum+k+r:1961. To achieve better power, one may, in the spirit of the Anderson–Darling goodness-of-fit test statistic, introduce a weight function in the integral that tends to infinity near (parts of) the boundary of the domain; see @dewet:1980.
@deheuvels:1981 [@deheuvels:1981:jmva] was perhaps the first to reformulate the question in the copula framework: for continuous variables, the problem consists in testing whether the true copula, $C$, is equal to the independence copula, $\Pi({\bm }u) = \prod_{j=1}^d u_j$. The empirical copula process $\sqrt{n} ({\hat{C}_n}- \Pi)$, for which he proposed an ingenious combinatorial transformation, can thus be taken as a basis for the construction of test statistics. @genest+r:2004 relied on his ideas to test the white noise hypothesis and considered Cramér–von Mises statistics based on the empirical copula process. @genest+q+r:2006 studied the power of such statistics against local alternatives, while @kojadinovic+h:2009 developed an extension to the case of testing for independence between random vectors. For the latter problem, @fan:2017 proposed an alternative approach based on empirical characteristic functions.
Recently, @belalia+b+l+t:2017 proposed to use the Bernstein empirical copula [@SanSat04; @JanSwaVer12] rather than the empirical copula in the Cramér–von Mises test statistic. Moreover, they constructed new test statistics based on the Bernstein copula density estimator by @bouezmarni+r+t:2010. Recall that the empirical beta copula arises from the Bernstein empirical copula by a specific choice of the degree of the Bernstein polynomials.
A situation of particular interest is when the true copula differs from the independence copula mainly in the tails. For instance, the bivariate *t* copula with zero correlation parameter has both Spearman’s rho and Kendall’s tau equal to zero. Still, the common value of its coefficients of upper and lower tail dependence is positive and depends on the degrees-of-freedom parameter. In their numerical experiments, @belalia+b+l+t:2017 found that for such alternatives, the power of the Cramér–von Mises test based on both the empirical copula and the Bernstein empirical copula is particularly weak. Their test statistics based on the Bernstein copula density estimator performed much better.
To increase the power of the Cramér–von Mises statistic against such difficult alternatives, a natural approach is to follow @dewet:1980 and introduce a weight function emphasizing the tails. For $\gamma \in [0, 2)$, we propose the weighted Cramér–von Mises statistic $$\label{eq:CvM}
T_{n,\gamma}
=
n \int_{[0,1]^d} \frac{\{ C_n^\beta({\bm }u) - C({\bm }u) \}^2}{\{g({\bm }u)\}^\gamma} \, \mathrm{d} {\bm }u.$$ We are mostly interested in the case where $C({\bm }u) = \Pi({\bm }u) = \prod_{j=1}^d u_j$, the independence copula. If $\gamma = 0$, the weight function disappears and we are back to the original Cramér–von Mises statistic, but with the empirical beta copula replacing the empirical copula.
Under the assumptions of Theorem \[thm:main\], we have, for every $\gamma \in [0, 2)$, the weak convergence $$T_{n,\gamma} {\rightsquigarrow}T_\gamma = \int_{[0, 1]^d} \frac{\{{\mathbb{C}}_C({\bm }u)\}^2}{\{g({\bm }u)\}^\gamma} \, \mathrm{d} {\bm }{u},
\qquad n \to \infty.$$ This is particularly true in case of independent random sampling from a $d$-variate distribution with continuous margins and independent components ($C = \Pi$).
We have $$T_{n, \gamma}
=
\int_{[0,1]^d}
\left( \frac{{\mathbb{C}}_n^\beta({\bm }u)}{\{g({\bm }u)\}^{\gamma/4}}\right)^2 \,
\frac{1}{\{g({\bm }u)\}^{\gamma/2}} \,
\mathrm{d} {\bm }u.$$ By Theorem \[thm:main\] applied to $\omega = \gamma / 4 \in [0, 1/2)$, the first part of the integrand converges weakly, in $\ell^\infty([0, 1]^d)$, to the stochastic process $({\mathbb{C}}_C/g^{\gamma/4})^2$. Further, since $\gamma/2 \in [0, 1)$, the integral $\int_{[0, 1]^d} \{ g({\bm }u) \}^{-\gamma/2} \, \mathrm{d} {\bm }u$ is finite. The linear functional that sends a measurable function $f \in \ell^\infty([0, 1]^d)$ to the scalar $\int_{[0, 1]^d} f( {\bm }u ) \, \{ g({\bm }u) \}^{-\gamma/2} \, \mathrm{d} {\bm }u$ is therefore bounded. The conclusion follows from the continuous mapping theorem.
A comprehensive simulation study comparing the performance of the weighted Cramér–von Mises statistic against all competitors and for a wide range of tuning parameters and data-generating processes is out of this paper’s scope. We limit ourselves to the case identified as the most difficult one in @belalia+b+l+t:2017, the bivariate *t* copula with zero correlation parameter. We copy the settings in their Section 5: the degrees-of-freedom parameter is $\nu = 2$ and we consider independent random samples of size $n \in \{100, 200, 400, 500\}$. We compare the power of our statistic $T_{n,\gamma}$ with the powers of their statistics $T_n,\delta_n,I_n$ at the $\alpha = 5\%$ significance level based on $1\,000$ replications.
We implemented our estimator in the statistical software environment [@Rlanguage] using the package [@KojYan10R]. The critical values were computed by a Monte Carlo approximation based on $10\,000$ random samples from the uniform distribution on the unit square. For the statistics in [@belalia+b+l+t:2017], we copied the relevant values from their Tables 4, 5, and 6. Their statistics depend on the degree, $k$, of the Bernstein polynomials, which they selected in $\{5, 10, \ldots, 30\}$. Note that for $\gamma = 0$ and $k = n$, our statistic $T_{n,\gamma}$ coincides with their statistic $T_n$. Their statistics $\delta_n$ and $I_n$ are based on the Bernstein copula density estimator in [@bouezmarni+r+t:2010].
The results are presented in Table \[tab:power\]. The unweighted Cramér–von Mises statistic $T_n$ does a poor job in detecting the alternative. The novel statistics $\delta_n$ and $I_n$ in [@belalia+b+l+t:2017] are more powerful, especially the statistic $I_n$, which is a Cramér–von Mises statistic based on the Bernstein copula density estimator. For the weighted Cramér–von Mises statistic $T_{n,\gamma}$, the power increases with $\gamma$. For the largest considered value, $\gamma = 1.75$, the power is higher than the one of $T_n$, $\delta_n$ and $I_n$ for any value of $k$ considered.
$k$ $T_n$ $\delta_n$ $I_n$ $\gamma$ $T_{n,\gamma}$
--------- ---------------- ------- ------------ ------- ---------- ----------------
$n=100$ $\phantom{0}5$ 0.056 0.114 0.094 $0.25$ 0.102
$10$ 0.064 0.130 0.168 $0.50$ 0.091
$15$ 0.066 0.166 0.254 $0.75$ 0.138
$20$ 0.070 0.132 0.270 $1.00$ 0.179
$25$ 0.070 0.102 0.284 $1.25$ 0.216
$30$ 0.068 0.114 0.294 $1.50$ 0.292
$1.75$ 0.401
$n=200$ $\phantom{0}5$ 0.076 0.176 0.094 $0.25$ 0.123
$10$ 0.080 0.222 0.308 $0.50$ 0.161
$15$ 0.088 0.226 0.442 $0.75$ 0.233
$20$ 0.094 0.210 0.466 $1.00$ 0.335
$25$ 0.096 0.176 0.472 $1.25$ 0.428
$30$ 0.086 0.148 0.458 $1.50$ 0.605
$1.75$ 0.705
$n=400$ $\phantom{0}5$ 0.044 0.366 0.230 $0.25$ 0.278
$10$ 0.038 0.492 0.588 $0.50$ 0.427
$15$ 0.048 0.472 0.702 $0.75$ 0.555
$20$ 0.044 0.432 0.762 $1.00$ 0.777
$25$ 0.048 0.382 0.772 $1.25$ 0.864
$30$ 0.050 0.354 0.780 $1.50$ 0.930
$1.75$ 0.964
$n=500$ $\phantom{0}5$ 0.072 0.398 0.192 $0.25$ 0.406
$10$ 0.096 0.542 0.688 $0.50$ 0.588
$15$ 0.100 0.552 0.746 $0.75$ 0.773
$20$ 0.110 0.506 0.806 $1.00$ 0.883
$25$ 0.106 0.476 0.824 $1.25$ 0.966
$30$ 0.096 0.458 0.824 $1.50$ 0.986
$1.75$ 0.992
: \[tab:power\]Testing the independence hypothesis when the true copula is equal to the *t* copula with zero correlation parameter and degrees-of-freedom parameter $\nu = 2$. Powers based on $1\,000$ random samples of sizes $n \in \{100, 200, 400, 500\}$ at significance level $\alpha = 5\%$. Comparison between, on the one hand, the statistics $T_n,\delta_n,I_n$ in @belalia+b+l+t:2017 with degree $k$ of the Bernstein polynomials and, on the other hand, the weighted Cramér–von Mises statistic $T_{n,\gamma}$ in Eq. with weight parameter $\gamma$. The values in the columns headed $T_n$, $\delta_n$ and $I_n$ have been copied from Tables 4–6 in [@belalia+b+l+t:2017].
Application: nonparametric estimation of a Pickands dependence function {#sec:pick}
=======================================================================
A $d$-variate copula $C$ is a multivariate extreme-value copula if and only if it can be written as $$C({\bm }u)
=
\exp \left\{
\left( \sum_{j=1}^d {\ln}u_j \right) \,
A \left( \frac{{\ln}u_1}{\sum_{j=1}^d {\ln}u_j}, \dots, \frac{{\ln}u_{d-1}}{\sum_{j=1}^d {\ln}u_j} \right)
\right\},$$ for ${\bm }u \in (0,1]^d \setminus \{ (1, \ldots, 1) \}$. The function $A:\Delta_{d-1} \to [1/d,1]$ is called the Pickands dependence function [after @Pic81], its domain being the unit simplex $\Delta_{d-1} = \{ {\bm }t=(t_1, \dots , t_{d-1}) \in [0,1]^{d-1}: \sum_{j=1}^{d-1}t_j{\leqslant}1\}$. Writing $t_d = t_d({\bm }t) = 1 - t_1 - \cdots - t_{d-1}$ for ${\bm }t \in \Delta_{d-1}$, we have $C(u^{t_1}, \ldots, u^{t_d}) = u^{A({\bm }t)}$ for $0 < u < 1$, and thus $${\ln}\{ A( {\bm }t ) \}
=
- \gamma + \int_0^1 \left\{ C(u^{t_1}, \ldots, u^{t_d}) - {\operatorname{\mathds{1}}}_{[e^{-1}, 1]}(u) \right\} \, \frac{{{\,\mathrm{d}}}u}{u {\ln}u},$$ where $\gamma = 0.5772156649\ldots$ is the Euler–Mascheroni constant. The rank-based Capéràa–Fougères–Genest (CFG) estimator, ${\hat{A}_n^{\mathrm{CFG}}}({\bm }t)$, arises by replacing $C$ in the above formula by the empirical copula, ${\hat{C}_n}$; see [@CapFouGen97] for the original estimator and see [@GenSeg09; @gudendorf+s:2012] for the rank-based versions in dimensions two and higher, respectively. We now propose to replace $C$ by the empirical beta copula instead, which gives the estimator $$\label{eq:CFG:b}
{\ln}\{ {\hat{A}_{n,\beta}^{\mathrm{CFG}}}({\bm }t) \}
=
- \gamma + \int_0^1 \left\{ {C_n^{\beta}}(u^{t_1}, \ldots, u^{t_d}) - {\operatorname{\mathds{1}}}_{[e^{-1}, 1]}(u) \right\} \, \frac{{{\,\mathrm{d}}}u}{u {\ln}u}.$$ The technique could also be used for other estimators based upon the empirical copula [@BucDetVol11; @BerBucDet13].
For the CFG-estimator on usually employs the endpoint-corrected version $$\label{eq:CFG:c}
{\ln}\{ {\hat{A}_{n,\mathrm{c}}^{\mathrm{CFG}}}({\bm }t) \}
=
{\ln}\{ {\hat{A}_n^{\mathrm{CFG}}}({\bm }t) \}
-
\sum_{j=1}^d t_j {\ln}\{ {\hat{A}_n^{\mathrm{CFG}}}( {\bm }e_j ) \},$$ where ${\bm }e_j = (0, \ldots, 0, 1, 0, \ldots, 0)$ is the $j$-th canonical unit vector in $\mathbb{R}^d$. For the estimator based on the empirical beta copula the endpoint correction is immaterial, since ${C_n^{\beta}}$ is a copula itself and thus ${\ln}{\hat{A}_{n,\beta}^{\mathrm{CFG}}}( {\bm }e_j ) = 0$ for all $j = 1, \ldots, d$.
Thanks to Theorem \[thm:main\], the limit of the beta CFG estimator can be derived from Theorem \[thm:main\] by a straightforward application of the continuous mapping theorem. The result does not require serial independence and can be extended to higher dimensions.
Let $C$ be a $d$-variate extreme-value copula with Pickands dependence function $A : \Delta_{d-1} \to [1/d, 1]$. Under the assumptions of Theorem \[thm:main\] we have, as $n \to \infty$, $$\sqrt n \left\{ {\hat{A}_{n,\beta}^{\mathrm{CFG}}}({\,\cdot\,}) - A({\,\cdot\,}) \right\}
{\rightsquigarrow}\mathbb{A}({\,\cdot\,}) \; \text{ in } \ell^\infty(\Delta_{d-1}),$$ where, for ${\bm }t \in \Delta_{d-1}$, we define $\mathbb{A}({\bm }t) = A( {\bm }t ) \int_0^1 \mathbb{C}_C(u^{t_1}, \ldots, u^{t_d}) \, (u {\ln}u)^{-1} \, \mathrm{d}u$.
Let $0 < \omega < 1/2$. We have $$\begin{aligned}
\sqrt n \left[ {\ln}\{ {\hat{A}_{n,\beta}^{\mathrm{CFG}}}({\bm }t) \} - {\ln}\{ A({\bm }t) \} \right]
=
\int_0^1
{\mathbb{C}}_n^{\beta}(u^{t_1}, \ldots, u^{t_d})
\frac{\mathrm{d}u}{u {\ln}u}
=
\int_0^1
\frac{{\mathbb{C}}_n^{\beta}(u^{t_1}, \ldots, u^{t_d})}{\{g(u^{t_1}, \ldots, u^{t_d})\}^\omega} \,
\{g(u^{t_1}, \ldots, u^{t_d})\}^\omega \,
\frac{\mathrm{d}u}{u {\ln}u}.\end{aligned}$$ The integral $
\int_0^1 \{g(u^{t_1}, \ldots, u^{t_d})\}^\omega \, (u {\ln}u)^{-1} \, \mathrm{d}u
$ is bounded, uniformly in ${\bm }t \in \Delta_{d-1}$. Therefore, the linear map that sends a measurable function $f \in \ell^\infty([0, 1]^d)$ to the bounded function ${\bm }t \mapsto \int_0^1 f(u^{t_1}, \ldots, u^{t_d}) \, \{g(u^{t_1}, \ldots, u^{t_d})\}^\omega \, (u {\ln}u)^{-1} \, \mathrm{d} u$ is continuous. By Theorem \[thm:main\] and the continuous mapping theorem, we find, as $n \to \infty$, $$\sqrt n \left[{\ln}\{{\hat{A}_{n,\beta}^{\mathrm{CFG}}}({\,\cdot\,})\} - {\ln}\{A({\,\cdot\,})\}\right]
{\rightsquigarrow}\left(\int_0^1 {\mathbb{C}}_C(u^{t_1}, \ldots, u^{t_d}) \, \frac{\mathrm{d}u}{u {\ln}u}\right)_{{\bm }t \in \Delta_{d-1}}
\; \text{ in } \ell^\infty(\Delta_{d-1}).$$ Finally, the result follows by an application of the functional delta method.
We compare the finite-sample performance of the endpoint-corrected CFG estimator with the variant based on the empirical beta copula. As performance criterion for an estimator $\hat{A}$, we use the integrated mean squared error, $$\int_{\Delta_{d-1}} {\operatorname{E}}\left[ \left\{ \hat{A}({\bm }t) - A({\bm }t) \right\}^2 \right] \, {{\,\mathrm{d}}}{\bm }t
=
{\operatorname{E}}\left[ \left\{ \hat{A}({\bm }T) - A({\bm }T) \right\}^2 \right],$$ where the random variable ${\bm }T$ is uniformly distributed on $\Delta_{d-1}$ and is independent of the sample from which $\hat{A}$ was computed. We approximate the integrated mean squared error through a Monte Carlo procedure: for a large integer $M$, we generate $M$ random samples of size $n$ from a given copula and we calculate $$\frac{1}{M} \sum_{m=1}^M \left\{ \hat{A}_n^{(m)}({\bm }T^{(m)}) - A({\bm }T^{(m)}) \right\}^2$$ where $\hat{A}_n^{(m)}$ denotes the estimator based upon sample number $m$, and where the random variables ${\bm }T^{(1)}, \ldots, {\bm }T^{(m)}$ are uniformly distributed on $\Delta_{d-1}$ and are independent of each other and of the copula samples. The approximation error is ${\mathrm{O}}_{{\mathbb{P}}}(1/\sqrt{M})$, aggregating both the sampling error and the integration error. A similar trick was used in [@SegSibTsu17] and is more efficient then first estimating the pointwise mean squared error through a Monte Carlo procedure and then integrating this out via numerical integration.
We considered the following data-generating processes:
- independent random sampling from the bivariate Gumbel copula [@gumbel:1961], which has Pickands dependence function $A(t) = \{t^{1/\alpha} + (1-t)^{1/\alpha}\}^{\alpha}$ for $t \in [0, 1]$ and with parameter $\alpha \in [0, 1]$, for which Kendall’s tau is $\tau = 1-\alpha$. We also considered independent random samples from the bivariate Galambos, Hüsler–Reiss and t-EV copula families, yielding similar results as for the bivariate Gumbel copula, not shown to save space. See, e.g., [@GudSeg2010] for the definitions of these copulas;
- independent random sampling from a special case of the trivariate asymmetric logistic extreme-value copula [@tawn:1990], with Pickands dependence function $A(t_1, t_2) = \sum_{(i,j) \in \{(1, 2), (2, 3), (3, 1)\}} \{(\theta t_i)^{1/\alpha} + (\phi t_j)^{1/\alpha} \}^\alpha + 1 - \theta - \phi$ for $(t_1, t_2) \in \Delta_2$ and $t_3 = 1-t_1-t_2$. As in [@GudSeg12 Section 5], we set $\phi = 0.3$ and $\theta = 0.6$, and $\alpha$ varies between $0$ and $1$;
- sampling from the strictly stationary bivariate moving maximum process $(U_{t1}, U_{t2})_{t \in \mathbb{Z}}$ given by $$U_{t1} = \max \left\{ W_{t-1,1}^{1/a}, W_{t1}^{1/(1-a)} \right\}
\qquad
\text{and}
\qquad
U_{t2} = \max \left\{ W_{t-1,2}^{1/b}, W_{t2}^{1/(1-b)} \right\},$$ where $a, b \in [0, 1]$ are two parameters and where $(W_{t1}, W_{t2})_{t \in \mathbb{Z}}$ is an iid sequence of bivariate random vectors whose common distribution is an extreme value-copula with some Pickands dependence function $B$. By Eq. (8.1) in [@bucher+s:2014], the stationary distribution of $(U_{t1}, U_{t2})$ is an extreme-value copula too, and its Pickands dependence function can be easily calculated to be $$A(t)
=
\{ a(1-t)+bt \} \, B \left( \frac{bt}{a(1-t)+bt} \right)
+
\{ (1-a)(1-t) + (1-b)t \} \, B \left( \frac{(1-b)t}{(1-a)(1-t) + (1-b)t} \right),$$ for $t \in [0, 1]$. We let $B(t) = \{t^{1/\alpha} + (1-t)^{1/\alpha}\}^{\alpha}$ for $t \in [0, 1]$ and $\alpha \in [0, 1]$ (the bivariate Gumbel copula as above, with Kendall’s tau $\tau = 1-\alpha$) and we set $a = 0.1$ and $b = 0.7$, so that $A$ is asymmetric.
The results are shown in Figure \[fig:Pickands\]. Each plot is based on $10\,000$ samples of size $n \in \{20, 50, 100\}$. For weak dependence (small $\tau$, large $\alpha$), the beta variant is the more efficient one, whereas for strong dependence (large $\tau$, small $\alpha$), it is the usual CFG estimator which is more accurate.
In order to gain a better understanding, we have also traced some trajectories of estimated Pickands dependence functions for independent random samples of the bivariate Gumbel copula at $\tau \in \{0.3, 0.9\}$ and $n \in \{20, 50, 100\}$; see Figure \[fig:Pickands:A\]. For each trajectory of the CFG estimator, there is a corresponding trajectory of the new estimator that is based on the same sample. For large $\tau$, the true extreme-value copula $C$ is close to the Fréchet–Hoeffding upper bound, $M(u_1, u_2) = \max(u_1, u_2)$. As a result, $C$ is strongly curved around the main diagonal $u_1 = u_2$, and this implies a strong curvature of the Pickands dependence function $A$ around $t = 1/2$. The empirical beta copula can be seen as a smoothed version of the empirical copula with an implicit bandwidth of the order $1/\sqrt{n}$ [@SegSibTsu17 p. 47]. For smaller $n$, oversmoothing occurs, producing a negative bias for the empirical beta copula around the diagonal $u_1 = u_2$ and thus a positive bias for the beta variant of the CFG estimator around $t = 1/2$.
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=20 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=50 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=100 "fig:"){width="32.00000%"}
![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/asymmetric_logistic_n=20 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/asymmetric_logistic_n=50 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/asymmetric_logistic_n=100 "fig:"){width="32.00000%"}
![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/mm_gumbel_n=20 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/mm_gumbel_n=50 "fig:"){width="32.00000%"} ![\[fig:Pickands\]Integrated mean squared error (vertical axis) of the endpoint-corrected CFG-estimator (dashed, blue) and the empirical beta variant (solid, red) based on samples of the data-generating process (M1), (M2) and (M3) (top to bottom) for various choices of the parameter $\alpha$ or $\tau = 1-\alpha$. Each point is based on $10\,000$ random samples of size $n \in \{20, 50, 100\}$ (left to right).](art/mm_gumbel_n=100 "fig:"){width="32.00000%"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=20_tau=03_A "fig:"){width="32.00000%"} ![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=50_tau=03_A "fig:"){width="32.00000%"} ![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=100_tau=03_A "fig:"){width="32.00000%"}
![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=20_tau=09_A "fig:"){width="32.00000%"} ![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=50_tau=09_A "fig:"){width="32.00000%"} ![\[fig:Pickands:A\]Plots of trajectories of the CFG-estimator (dashed blue) and the empirical beta variant (solid red) of the Pickands dependence function (solid black) based on samples from the Gumbel copula with Kendall’s $\tau \in \{0.3, 0.9\}$ (top to bottom) and $n \in \{20, 50, 100\}$ (left to right).](art/gumbel_n=100_tau=09_A "fig:"){width="32.00000%"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Proof of Theorem \[thm:main\] {#sec:proof}
=============================
Recall the empirical copula process $\hat{{\mathbb{C}}}_n = \sqrt{n} (\hat{C}_n - C)$ and the empirical beta copula process $\hat{{\mathbb{C}}}_n^\beta = \sqrt{n} ({C_n^{\beta}}- C)$. The link between the empirical copula $\hat{C}_n$ and the empirical beta copula ${C_n^{\beta}}$ is given in . In the derivation of the limit of the weighted empirical beta copula process the following decomposition plays a central role: $$\begin{gathered}
\label{eq:decomp}
\frac {{\mathbb{C}}_n^\beta({\bm }u )}{g({\bm }u)^\omega}
~=~
\frac{\hat{{\mathbb{C}}}_n({\bm }u)}{g({\bm }u )^\omega}\int_{[0,1]^d} \frac{g({\bm }w)^\omega}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w) \\
~+ ~
\int_{[0,1]^d} \left\{ \frac{\hat {{\mathbb{C}}}_n({\bm }w)}{g({\bm }w )^\omega} - \frac{\hat{{\mathbb{C}}}_n({\bm }u)}{g({\bm }u)^\omega} \right\} \frac{g({\bm }w)^\omega}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
~ +~
\int_{[0,1]^d} \sqrt n \frac{C({\bm }w) - C({\bm }u)}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w).\end{gathered}$$ It is reasonable to assume that the last two terms on the right-hand side vanish as $n\to \infty$. Indeed, the measure $\mu_{n,{\bm }u}$ concentrates around its mean ${\bm }u$, if the sample size grows, and both integrands are small if ${\bm }w$ is close to ${\bm }u$. By the same reason, the integral in the first term should be close to one. The decomposition can thus be used to obtain weak convergence of ${\mathbb{C}}_n^\beta / g^\omega$ on the interior of the unit cube. The boundary of the unit cube has to be treated separately.
The case $\omega = 0$ corresponds to the unweighted case, so we assume henceforth that $0 < \omega < 1/2$. Fix a scalar $\gamma$ such that $1 / \{2(1-\omega)\} < \gamma < 1$. Consider the abbreviations $\{ g {\geqslant}n^{-\gamma} \} = \{ {\bm }v \in [0, 1]^d \mid g({\bm }v) {\geqslant}n^{-\gamma} \}$ and similarly $\{ g < n^{-\gamma} \}$. By Lemma \[lem:boundary\], we have $$\begin{aligned}
{\mathbb{C}}_n^\beta/g^\omega
&=
{{\mathbb{C}}_n^\beta/g^\omega {\operatorname{\mathds{1}}}_{\{ g {\geqslant}n^{-\gamma} \}}} + {\mathbb{C}}_n^\beta/g^\omega {\operatorname{\mathds{1}}}_{ \{ g < n^{-\gamma} \} } \\
&=
{{\mathbb{C}}_n^\beta/g^\omega {\operatorname{\mathds{1}}}_{\{ g {\geqslant}n^{-\gamma} \}}} + {\mathrm{o}}(1),
\qquad n \to \infty, \quad \text{a.s.}\end{aligned}$$ The three terms on the right-hand side of are treated in Lemmas \[lem:bias\], \[lem:int\] and \[lem:bias2\]. We find $$\label{eq:CbnhatCbn}
{\mathbb{C}}_n^\beta/g^\omega
=
\hat{{\mathbb{C}}}_n/g^\omega {\operatorname{\mathds{1}}}_{ \{ g {\geqslant}n^{-\gamma} \} } (1 + {\mathrm{o}}(1)) + {\mathrm{o}}_{{\mathbb{P}}}(1),
\qquad n \to \infty.$$
Recall ${\bm }U_i = (U_{i,1}, \ldots, U_{i,d})$ with $U_{i,j} = F_j(X_{i,j})$. The empirical distribution function and the empirical process associated to the unobservable sample ${\bm }U_1, \ldots, {\bm }U_n$ are $$\begin{aligned}
C_n({\bm }u)
&= \frac{1}{n} \sum_{i=1}^n {\operatorname{\mathds{1}}}_{ \{{\bm }U_i {\leqslant}{\bm }u\} }, &
\alpha_n({\bm }u)
&= \sqrt n \{ C_n({\bm }u) - C({\bm }u) \},\end{aligned}$$ respectively, for ${\bm }u \in [0, 1]^d$. Consider the process $$\bar {\mathbb{C}}_n({\bm }u)
=
\alpha_n({\bm }u)
- \sum_{j=1}^d \dot C_j({\bm }u) \, \alpha_n(1,\dots,1,u_j,1,\dots,1),
\qquad {\bm }u \in [0, 1]^d,$$ with $u_j$ appearing at the $j$-th coordinate. Note the slight but convenient abuse of notation in the definition of $\bar{{\mathbb{C}}}_n$: if ${\bm }u$ is such that $u_j \in \{0, 1\}$, then $\alpha_n(1,\dots,1,u_j,1,\dots,1) = 0$ almost surely, so that the fact that for such ${\bm }u$, the partial derivative $\dot{C}_j({\bm }u)$ has been left undefined in Condition \[cond:second\] plays no role.
By above and by Theorem 2.2 in [@BerBucVol17] (see also Remark \[rem:proof\] below), $$\begin{aligned}
{\mathbb{C}}_n^\beta/g^\omega
&=
\{ \bar{{\mathbb{C}}}_n/g^\omega + {\mathrm{o}}_{{\mathbb{P}}}(1) \} {\operatorname{\mathds{1}}}_{ \{ g {\geqslant}n^{-\gamma} \} } (1 + {\mathrm{o}}(1)) + {\mathrm{o}}_{{\mathbb{P}}}(1) \\
&=
{\bar{{\mathbb{C}}}_n/g^\omega {\operatorname{\mathds{1}}}_{ \{ g {\geqslant}n^{-\gamma} \} }} + {\mathrm{o}}_{{\mathbb{P}}}(1),
\qquad n \to \infty.\end{aligned}$$ In view of Lemma 4.9 in [@BerBucVol17], the indicator function can be omitted, and, applying Theorem 2.2 in the same reference again, we obtain $${\mathbb{C}}_n^\beta/g^\omega
=
\bar{{\mathbb{C}}}_n/g^\omega
+ {\mathrm{o}}_{{\mathbb{P}}}(1)
{\rightsquigarrow}{\mathbb{C}}_C/g^\omega,
\qquad n \to \infty,$$ as required. This finishes the proof of Theorem \[thm:main\].
The theorem follows immediately from the decomposition in , Lemmas \[lem:boundary\]-\[lem:bias2\] and Theorem 2.2 in [@BerBucVol17]. These results immediately imply that for $\gamma \in (1/(2(1-\omega)),1] $ $$\begin{aligned}
{\mathbb{C}}_n^\beta/g^\omega
~=~&
{\mathbb{C}}_n^\beta/g^\omega {\operatorname{\mathds{1}}}_{ \{ {\bm }v \vert g({\bm }v){\geqslant}n^{-\gamma} \} } +{\mathbb{C}}_n^\beta/g^\omega {\operatorname{\mathds{1}}}_{ \{\ {\bm }v \vert g({\bm }v)<n^{-\gamma} \} } \\
~=~&
\hat {\mathbb{C}}_n/g^\omega {\operatorname{\mathds{1}}}_{ \{ {\bm }v \vert g({\bm }v){\geqslant}n^{-\gamma} \} } + o_{\mathbb{P}}(1)\\
~=~&
\bar{{\mathbb{C}}}_n/g^\omega +o_{\mathbb{P}}(1),\end{aligned}$$ which converges weakly to ${\mathbb{C}}_C/g^\omega $ in $\ell^\infty([0,1]^d)$ and where, for any ${\bm }u \in [0,1]^d$, $$\bar {\mathbb{C}}_n({\bm }u) := \alpha_n({\bm }u) - \sum_{j=1}^d \dot C_j({\bm }u) \alpha_n(1,\dots ,1,u_j,1, \dots,1 ),$$ and where $$\alpha_n({\bm }u) = \sqrt n \{ C_n({\bm }u) - C({\bm }u) \}, \qquad C_n({\bm }u) = n^{-1} {\textstyle \sum_{i=1}^n} {\operatorname{\mathds{1}}}( {\bm }U_i {\leqslant}{\bm }u),$$ denotes the empirical process based on $U_{i,j}=F_j(X_{i,j})$, $j=1,\dots,d$ and $i=1,\dots,n$.
\[rem:proof\] Some of the results in [@BerBucVol17] have to be adapted to the present situation.
- In the latter reference, the pseudo-observations are defined as $\hat U_{i,j} = (n+1)^{-1} R_{i,j}$ rather than $n^{-1} R_{i,j}$. However, this does not affect the asymptotics, since the difference of the two empirical copulas is at most $d/n$, almost surely. For ${\bm }u \in \{ g {\geqslant}n^{-\gamma} \}$, this modification makes a difference of the order ${\mathrm{O}}_{\mathbb{P}}(n^{\gamma\omega+1/2-1}) = {\mathrm{o}}_{\mathbb{P}}(1)$, as $n \to \infty$.
- In Theorem 2.2 in [@BerBucVol17], the approximation of $\hat {\mathbb{C}}_n$ by $\bar {\mathbb{C}}_n$ is stated on the interior of the set $ [c/n,1-c/n]^d$ for any $c \in (0,1)$. But it can be seen in the proof of the latter statement that the result can be easily extended to the set $\{ g {\geqslant}c/n \}$. See Section \[sec:BerBucVol17\] below for details.
Auxiliary results {#sec:aux}
=================
Throughout and unless otherwise stated, we assume the conditions of Theorem \[thm:main\].
Negligibility of the boundary regions
-------------------------------------
\[lem:boundary\] For $\gamma > 1 / \{2(1-\omega)\}$, we have $$\sup_{{\bm }u \in \{g {\leqslant}n^{-\gamma}\}} \lvert {\mathbb{C}}_n^\beta({\bm }u)/g({\bm }u)^\omega \rvert = {\mathrm{o}}(1),
\qquad n \to \infty,
\quad \text{a.s.}$$
Let $\gamma > 1 / \{2(1-\omega)\}$ and ${\bm }u \in \{ g {\leqslant}n^{-\gamma}\}$. Without loss of generality, we only need to consider the cases $g ({\bm }u) = u_1$ and $g({\bm }u)= 1-u_1$. The remaining cases can be treated analogously.
Let us start with the case $g({\bm }u)=u_1 {\leqslant}n^{-\gamma}$. Since ${C_n^{\beta}}$ is a copula almost surely, we have ${C_n^{\beta}}({\bm }u) {\leqslant}u_1$. This in turn gives us $$\lvert {\mathbb{C}}_n^\beta({\bm }u)/g({\bm }u)^\omega \rvert
{\leqslant}\sqrt n \, u_1^{-\omega} \lvert {C_n^{\beta}}({\bm }u) + C({\bm }u) \rvert
{\leqslant}2 \sqrt n \, u_1^{1-\omega}
{\leqslant}2 n^{1/2 +\gamma\omega -\gamma},
\qquad \text{a.s.},$$ an upper bound which vanishes as $n \to \infty$ by the choice of $\gamma$.
Now suppose that $g({\bm }u)=1-u_1 {\leqslant}n^{-\gamma}$. By the definition of $g({\bm }u)$, we can assume without loss of generality that $1-u_j {\leqslant}1-u_1$ for $j=3,\dots ,d$. Again, we will use the fact that ${C_n^{\beta}}$ is a copula almost surely. Note that ${C_n^{\beta}}(1,u_2,1,\dots ,1)= u_2$. Hence, by the Lipschitz continuity of copulas we obtain, almost surely, $$\begin{aligned}
\lvert {\mathbb{C}}_n^\beta({\bm }u)/g({\bm }u)^\omega \rvert
&{\leqslant}\sqrt n (1-u_1)^{-\omega}
\{ \lvert {C_n^{\beta}}({\bm }u) -u_2 \rvert + \lvert C({\bm }u)-u_2 \rvert \} \\
&{\leqslant}\textstyle{2 \sqrt n (1-u_1)^{-\omega} \sum_{j\ne 2} (1-u_j)} \\
&{\leqslant}\textstyle{2 \sqrt n \sum_{j \ne 2} (1-u_j)^{1-\omega}} \\
&{\leqslant}2(d-1) n^{1/2+\gamma\omega-\gamma}
= {\mathrm{o}}(1),
\qquad n \to \infty.\end{aligned}$$
The upper bounds do not depend on ${\bm }u \in \{ g {\leqslant}n^{-\gamma} \}$, whence the uniformity in ${\bm }u$.
The three terms in the decomposition (\[eq:decomp\])
----------------------------------------------------
The following lemma is to be compared with Proposition 3.5 in [@SegSibTsu17]. There, a pointwise approximation rate of ${\mathrm{O}}(n^{-1})$ was established. Here, we state a rate which is slightly slower, ${\mathrm{O}}(n^{-1} {\ln}n)$, but uniformly in ${\bm }u$.
\[lem:bias\] If $C$ satisfies Condition \[cond:second\], then $$\sup_{{\bm }u \in [0, 1]^d}
\left\lvert
\int_{[0, 1]^d} \{ C({\bm }w) - C({\bm }u) \} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right\rvert
=
{\mathrm{O}}( n^{-1} {\ln}n ),
\qquad n \to \infty.$$
Put ${\varepsilon}_n = n^{-1} {\ln}n$. First, we show that we can ignore those ${\bm }u$ for which $u_j {\leqslant}{\varepsilon}_n$ for some $j \in \{1, \ldots, d\}$. Indeed, for such ${\bm }u$, the absolute value in the statement is bounded by $$\int_{[0, 1]^d} w_j {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) + u_j = 2u_j {\leqslant}2{\varepsilon}_n.$$
Let ${\bm }u \in [{\varepsilon}_n, 1]^d$. We show how to reduce the analysis to the case where ${\bm }u \in [{\varepsilon}_n, 1-{\varepsilon}_n]^d$. Let $J = J({\bm }u)$ denote the set of indices $j = 1, \ldots, d$ such that $u_j > 1-{\varepsilon}_n$ and suppose that $J$ is not empty. Consider the vector ${\bm }e \in \{0, 1\}^d$ which has components $e_j = 1$ for $j \in J$ and $e_j = 0$ otherwise. For ${\bm }v \in [0, 1]^d$, the vector ${\bm }v \vee {\bm }e$ has components $({\bm }v \vee {\bm }e)_j$ equal to $v_j$ if $j \not\in J$ and to $1$ if $j \in J$. Recall that copulas are Lipschitz continuous with respect to the $L^1$ norm with Lipschitz constant $1$. It follows that $$\begin{gathered}
\left\lvert
\int_{[0, 1]^d} \{ C({\bm }w) - C({\bm }u) \} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right\rvert
{\leqslant}\left\lvert
\int_{[0, 1]^d} \{ C({\bm }w \vee {\bm }e) - C({\bm }u \vee {\bm }e) \} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right\rvert \\
+ \int_{[0, 1]^d} \lvert C({\bm }w) - C({\bm }w \vee {\bm }e) \rvert {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
+ \lvert C({\bm }u \vee {\bm }e) - C({\bm }u) \rvert.\end{gathered}$$
- The first integral on the right-hand side does not depend on the variables $w_j$ for $j \in J$. It can therefore be reduced to an integral as in the statement of the lemma with respect to the variables in the set $\{1, \ldots, d\} \setminus J$. The copula of those variables is a multivariate margin of the original copula and Condition \[cond:second\] applies to it as well. By construction, all remaining $u_j$ are in the interval $[{\varepsilon}_n, 1-{\varepsilon}_n]$, as required.
- We have $\lvert C({\bm }w) - C({\bm }w \vee {\bm }e) \rvert {\leqslant}\sum_{j \in J} \lvert w_j - 1 \rvert {\leqslant}\sum_{j \in J} (\lvert w_j - u_j \rvert + {\varepsilon}_n)$. By the Cauchy–Schwarz inequality, $\int_{[0, 1]^d} \lvert w_j - u_j \rvert {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) {\leqslant}\{n^{-1} u_j(1-u_j)\}^{1/2} {\leqslant}n^{-1/2} {\varepsilon}_n^{1/2} {\leqslant}{\varepsilon}_n$. Hence $\int_{[0,1]^d}\lvert C({\bm }w) - C({\bm }w \vee {\bm }e) \rvert {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w){\leqslant}2d{\varepsilon}_n$.
- Finally, $\lvert C({\bm }u \vee {\bm }e) - C({\bm }u) \rvert {\leqslant}\sum_{j \in J} (1 - u_j) {\leqslant}d {\varepsilon}_n$.
It remains to consider the case ${\bm }u \in [{\varepsilon}_n, 1-{\varepsilon}_n]^d$. As in the proof of Proposition 3.4 in [@SegSibTsu17], we have $$\int_{[0, 1]^d} \{ C({\bm }w) - C({\bm }u) \} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
=
\sum_{j=1}^d
\int_0^1
\left[
\int_{[0, 1]^d}
(w_j - u_j)
\bigl\{ \dot{C}_j({\bm }u + t({\bm }w - {\bm }u)) - \dot{C}_j({\bm }u) \bigr\}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right]
{{\,\mathrm{d}}}t.$$ It is sufficient to show that the absolute value of the integral in square brackets is ${\mathrm{O}}({\varepsilon}_n)$, uniformly in $j \in \{1, \ldots, d\}$ and $t \in (0, 1)$ and ${\bm }u \in [{\varepsilon}_n, 1-{\varepsilon}_n]^d$.
The integral over $[0, 1]^d$ can be reduced to an integral over $(0, 1)^d$: indeed, the integrand is bounded in absolute value by $1$ (recall $0 {\leqslant}\dot{C}_j {\leqslant}1$), and the mass on the boundary is $\mu_{n, {\bm }u}([0, 1]^d \setminus (0, 1)^d) = {\mathbb{P}}[\exists j : S_j \in \{0, n\}] {\leqslant}2d (1 - {\varepsilon}_n)^n {\leqslant}2d \exp(-n{\varepsilon}_n) = 2d n^{-1} = {\mathrm{o}}({\varepsilon}_n)$ as $n \to \infty$.
In view of the second part of Condition \[cond:second\], we have $$\begin{gathered}
\int_{(0, 1)^d}
(w_j - u_j)
\bigl\{ \dot{C}_j({\bm }u + t({\bm }w - {\bm }u)) - \dot{C}_j({\bm }u) \bigr\}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
=
t
\sum_{k=1}^d
\int_0^1
\left[
\int_{(0, 1)^d}
(w_j - u_j)(w_k - u_k)
\ddot{C}_{jk} ({\bm }u + st({\bm }w - {\bm }u))
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right]
{{\,\mathrm{d}}}s.\end{gathered}$$ It is sufficient to show that the absolute value of the integral in square brackets is ${\mathrm{O}}({\varepsilon}_n)$, uniformly in $j,k \in \{1, \ldots, d\}$ and $s, t \in (0, 1)$ and ${\bm }u \in [{\varepsilon}_n, 1-{\varepsilon}_n]^d$.
We apply the bound in to $\ddot{C}_{jk}( {\bm }u + st ({\bm }w - {\bm }u) )$. We have $\min(a^{-1}, b^{-1}) {\leqslant}(ab)^{-1/2}$, and the latter is a convex function of $(a, b) \in (0, \infty)^2$. The point ${\bm }u + st ({\bm }w - {\bm }u)$ is located on the line segment connecting ${\bm }u$ and ${\bm }w$. Therefore, $$\begin{aligned}
\bigl\lvert \ddot{C}_{jk}( {\bm }u + st ({\bm }w - {\bm }u) ) \bigr\rvert
{\leqslant}K \left[ \frac{1}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}} + \frac{1}{\{w_j(1-w_j)w_k(1-w_k)\}^{1/2}} \right].\end{aligned}$$ We obtain $$\begin{gathered}
\left\lvert
\int_{(0, 1)^d}
(w_j - u_j)(w_k - u_k)
\ddot{C}_{jk} ({\bm }u + st({\bm }w - {\bm }u))
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
\right\rvert \\
{\leqslant}K
\int_{(0, 1)^d}
\biggl[
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}}
+
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{w_j(1-w_j)w_k(1-w_k)\}^{1/2}}
\biggr]
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w).\end{gathered}$$ First, by the Cauchy–Schwarz inequality and the fact that ${\operatorname{E}}[ (S_i/n - u_i)^2 ] = n^{-1} u_i(1-u_i)$ for all $i \in \{1, \ldots, d\}$, we have $$\int_{(0, 1)^d}
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
{\leqslant}\prod_{i \in \{j,k\}} \left\{ \int_{(0, 1)^d} \frac{(w_i - u_i)^2}{u_i(1-u_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \right\}^{1/2}
{\leqslant}n^{-1} {\leqslant}{\varepsilon}_n.$$ Second, again by Cauchy–Schwarz inequality, $$\int_{(0, 1)^d}
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{w_j(1-w_j)w_k(1-w_k)\}^{1/2}}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
{\leqslant}\prod_{i \in \{j,k\}} \left\{ \int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i(1-w_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \right\}^{1/2}.$$ Each of the two integrals ($i = j$ and $i = k$), and therefore their geometric mean, will be bounded by the same quantity. Note that $\tfrac{1}{w_i(1-w_i)} = \tfrac{1}{w_i} + \tfrac{1}{1-w_i}$ and that the integral involving $\tfrac{1}{1-w_i}$ is equal to the one involving $\tfrac{1}{w_i}$ when $u_i$ is replaced by $1-u_i$, which we are allowed to do since ${\bm }u \in [{\varepsilon}_n, 1-{\varepsilon}_n]^d$ anyway. Therefore, we can replace $w_i(1-w_i)$ by $w_i$ in the denominator at the cost of a factor two. Further, $$\begin{aligned}
\int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
&{\leqslant}\int_{[0, 1]^d} {\operatorname{\mathds{1}}}_{(0, 1]}(w_i) \frac{(w_i - u_i)^2}{w_i} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
&=
\int_{[0, 1]^d} {\operatorname{\mathds{1}}}_{(0, 1]}(w_i) \bigl( w_i - 2u_i + \tfrac{u_i^2}{w_i} \bigr) {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
&=
u_i - 2u_i {\mathbb{P}}[S_i/n > 0] + u_i^2 {\operatorname{E}}[ \tfrac{1}{S_i/n} {\operatorname{\mathds{1}}}_{\{ S_i/n > 0 \}} ] \\
&{\leqslant}-u_i + 2{\mathbb{P}}[S_i = 0] + n u_i^2 {\operatorname{E}}[ \tfrac{1}{S_i} {\operatorname{\mathds{1}}}_{\{ S_i {\geqslant}1 \}} ].\end{aligned}$$ Recall that $u_i \in [{\varepsilon}_n, 1-{\varepsilon}_n]$ and thus ${\mathbb{P}}[S_i = 0] {\leqslant}(1-{\varepsilon}_n)^n {\leqslant}\exp(-n{\varepsilon}_n) = n^{-1} = {\mathrm{o}}({\varepsilon}_n)$. Further, the expectation of the reciprocal of a binomial random variable is treated in Lemma \[lem:binom\]. Note that $n {\varepsilon}_n = {\ln}n \to \infty$. We find $$\sup_{{\bm }u \in [{\varepsilon}_n, 1 - {\varepsilon}_n]^d} \max_{i=1,\ldots,d}
\int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i(1-w_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
= {\mathrm{O}}(n^{-1}) = {\mathrm{o}}({\varepsilon}_n), \qquad n \to \infty.$$ The proof is complete.
\[lem:bias\] Suppose Condition \[cond:second\] holds. Then as $n\to\infty$, for $1/\{2(1-\omega)\} < \gamma < 1$, $$\sup_{{\bm }u \in \{ g > n^{-\gamma}\}}
\left\lvert
\int_{[0,1]^d} \sqrt n \frac{C({\bm }w) - C({\bm }u)}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
\right\rvert
= {\mathrm{o}}(1),
\qquad n \to \infty.$$
[ ]{} First, suppose ${\bm }u \in [n^{-\gamma}, 1 - n^{-\gamma}]^d$. Note that $g({\bm }u) {\geqslant}n^{-\gamma}$ for such ${\bm }u$. Recall that $\mu_{n, {\bm }u}$ is the joint distribution of the random vector $(S_1/n, \ldots, S_d/n)$, where $S_1, \ldots, S_d$ are independent random variables with $S_j \sim {\operatorname{Bin}}(n, u_j)$.
By the first part of Condition \[cond:second\] and the fact that $\int_{[0, 1]^d} w_j {{\,\mathrm{d}}}\mu_{n, {\bm }u}({{\,\mathrm{d}}}w) = {\operatorname{E}}[S_j / n] = u_j$ for all $j \in \{1, \ldots, d\}$, we have $$\begin{aligned}
\lefteqn{
\int_{[0,1]^d}
\sqrt{n} \frac{C({\bm }w) - C({\bm }u)}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
} \\
&=
\frac{\sqrt{n}}{g({\bm }u)^\omega}
\int_{[0,1]^d}
\sum_{j=1}^d
(w_j - u_j)
\int_0^1
\dot{C}_j \bigl( {\bm }u + t( {\bm }w - {\bm }u) \bigr)
{{\,\mathrm{d}}}t
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
&=
\frac{\sqrt{n}}{g({\bm }u)^\omega}
\sum_{j=1}^d
\int_{[0, 1]^d}
(w_j - u_j)
\int_0^1
\bigl\{ \dot{C}_j \bigl( {\bm }u + t( {\bm }w - {\bm }u) \bigr) - \dot{C}_j( {\bm }u ) \bigr\}
{{\,\mathrm{d}}}t
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w).\end{aligned}$$ Since $0 {\leqslant}\dot{C}_j {\leqslant}1$ and since ${\mathbb{P}}[S_j \in \{0, n\}] {\leqslant}2 (1 - n^{-\gamma})^n {\leqslant}2 \exp(-n^{1-\gamma})$ for $u_j \in [n^{-\gamma}, 1-n^{-\gamma}]$, we obtain that we can replace the integration domain $[0, 1]^d$ by $(0, 1)^d$: $$\begin{gathered}
\int_{[0,1]^d}
\sqrt{n} \frac{C({\bm }w) - C({\bm }u)}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w) \\
=
\frac{\sqrt{n}}{g({\bm }u)^\omega}
\sum_{j=1}^d
\int_{(0, 1)^d}
(w_j - u_j)
\int_0^1
\bigl\{ \dot{C}_j \bigl( {\bm }u + t( {\bm }w - {\bm }u) \bigr) - \dot{C}_j( {\bm }u ) \bigr\}
{{\,\mathrm{d}}}t
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
+
R_n({\bm }u)\end{gathered}$$ where $\sup \{ \lvert R_n({\bm }u) \rvert : {\bm }u \in [n^{-\gamma}, 1-n^{-\gamma}]^d \} = {\mathrm{o}}(1)$ as $n \to \infty$. Let $t \in (0, 1)$ and ${\bm }w \in (0, 1)^d$. By the second part of Condition \[cond:second\], we have $$\dot{C}_j \bigl( {\bm }u + t( {\bm }w - {\bm }u) \bigr) - \dot{C}_j( {\bm }u )
=
\sum_{k=1}^d t (w_k - u_k) \int_0^1 \ddot{C}_{jk} \bigl( {\bm }u + st ({\bm }w - {\bm }u) \bigr) {{\,\mathrm{d}}}s.$$ We apply the bound in to $\ddot{C}_{jk}( {\bm }u + st ({\bm }w - {\bm }u) )$. We have $\min(a^{-1}, b^{-1}) {\leqslant}(ab)^{-1/2}$, the upper bound being a convex function of $(a, b) \in (0, \infty)^2$. Moreover, the point ${\bm }u + st ({\bm }w - {\bm }u)$ is located on the line segment connecting ${\bm }u$ and ${\bm }w$. Therefore, $$\begin{aligned}
\bigl\lvert \ddot{C}_{jk}( {\bm }u + st ({\bm }w - {\bm }u) ) \bigr\rvert
{\leqslant}K \bigl[ \{u_j(1-u_j)u_k(1-u_k)\}^{-1/2} + \{w_j(1-w_j)w_k(1-w_k)\}^{-1/2} \bigr].\end{aligned}$$ We obtain that, for ${\bm }u \in [n^{-\gamma}, 1-n^{-\gamma}]^d$, $$\begin{gathered}
\left\lvert
\int_{[0,1]^d}
\sqrt{n} \frac{C({\bm }w) - C({\bm }u)}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
\right\rvert \\
\shoveleft
{\leqslant}K n^{1/2 + \gamma \omega}
\sum_{j=1}^d \sum_{k=1}^d
\int_{(0, 1)^d}
\biggl[
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}} \\
+
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{w_j(1-w_j)w_k(1-w_k)\}^{1/2}}
\biggr]
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
+ {\mathrm{o}}(1), \qquad n \to \infty,\end{gathered}$$ the ${\mathrm{o}}(1)$ term being uniform in ${\bm }u \in [n^{-\gamma}, 1-n^{-\gamma}]^d$. Fix $j, k \in \{1, \ldots, d\}$. By the Cauchy–Schwarz inequality and the fact that ${\operatorname{E}}[ (S_i/n - u_i)^2 ] = n^{-1} u_i(1-u_i)$ for all $i \in \{1, \ldots, d\}$, we have $$\begin{gathered}
\lefteqn{
\int_{(0, 1)^d}
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
} \\
{\leqslant}\prod_{i \in \{j,k\}} \left\{ \int_{(0, 1)^d} \frac{(w_i - u_i)^2}{u_i(1-u_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \right\}^{1/2}
{\leqslant}n^{-1}.\end{gathered}$$ This bound is sharp enough since $1/2 + \gamma \omega - 1 < 0$ as $0 < \omega < 1/2$ and $0 < \gamma < 1$. Again the Cauchy–Schwarz inequality, $$\begin{gathered}
\lefteqn{
\int_{(0, 1)^d}
\frac{\lvert (w_j - u_j) (w_k - u_k) \rvert}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}}
{{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
} \\
{\leqslant}\prod_{i \in \{j,k\}} \left\{ \int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i(1-w_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \right\}^{1/2}\end{gathered}$$ Each of the two integrals, and therefore their geometric mean, will be bounded by the same quantity. Note that $\tfrac{1}{w_i(1-w_i)} = \tfrac{1}{w_i} + \tfrac{1}{1-w_i}$ and that the integral involving $\tfrac{1}{1-w_i}$ is equal to the one involving $\tfrac{1}{w_i}$ when $u_i$ is replaced by $1-u_i$. Further, $$\begin{aligned}
\int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
&{\leqslant}\int_{[0, 1]^d} {\operatorname{\mathds{1}}}_{(0, 1]}(w_i) \frac{(w_i - u_i)^2}{w_i} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
&=
\int_{[0, 1]^d} {\operatorname{\mathds{1}}}_{(0, 1]}(w_i) \bigl( w_i - 2u_i + \tfrac{u_i^2}{w_i} \bigr) {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w) \\
&=
u_i - 2u_i {\mathbb{P}}[S_i/n > 0] + u_i^2 {\operatorname{E}}[ \tfrac{1}{S_i/n} {\operatorname{\mathds{1}}}_{\{ S_i/n > 0 \}} ] \\
&{\leqslant}-u_i + 2{\mathbb{P}}[S_i = 0] + n u_i^2 {\operatorname{E}}[ \tfrac{1}{S_i} {\operatorname{\mathds{1}}}_{\{ S_i {\geqslant}1 \}} ].\end{aligned}$$ Recall that $u_i \in [n^{-\gamma}, 1-n^{-\gamma}]$ and thus ${\mathbb{P}}[S_i = 0] {\leqslant}(1-n^{-\gamma})^n {\leqslant}\exp(-n^{1-\gamma})$. Further, the expectation of the reciprocal of a binomial random variable is treated in Lemma \[lem:binom\]. We find $$\sup_{{\bm }u \in [n^{-\gamma}, 1-n^{-\gamma}]^d} \max_{i=1,\ldots,d}
\int_{(0, 1)^d} \frac{(w_i - u_i)^2}{w_i(1-w_i)} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
= {\mathrm{O}}(n^{-1}), \qquad n \to \infty.$$
By Condition \[cond:second\] we can rewrite the bias term as follows $$\begin{gathered}
\int_{[0,1]^d}\sqrt n \frac{C({\bm }w ) - C({\bm }u)}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w)
= ~
\sqrt n \sum_{j=1}^d \frac{\dot C_j ({\bm }u)} {g({\bm }u)^\omega}\int_{[0,1]^d} (w_j -u_j) {{\,\mathrm{d}}}\mu_{n, {\bm }u} ({\bm }w)\\
+ ~
\sum_{j=1}^d \sum_{k=1}^d \sqrt n \int_{[0,1]^d} \frac{(w_j-u_j)(w_k - u_k)}{g({\bm }u)^\omega} \int_0^1 \int_0^1 t ~ \ddot C_{jk}({\bm }v(st)){{\,\mathrm{d}}}s{{\,\mathrm{d}}}t {{\,\mathrm{d}}}\mu_{n, {\bm }u} ({\bm }w)
,
\end{gathered}$$ with ${\bm }v(r) = {\bm }u +r({\bm }w- {\bm }u)$ fir $r \in [0,1]$. By the definition of $\mu_{n,{\bm }u}$, the first integral on the right-hand side is equal to zero. Let $$S_{j,k}({\bm }u)=\sqrt n \int_{[0,1]^d} \frac{(w_j-u_j)(w_k - u_k)}{g({\bm }u)^\omega} \int_0^1 \int_0^1 t ~ \ddot C_{jk}({\bm }v(st)){{\,\mathrm{d}}}s{{\,\mathrm{d}}}t {{\,\mathrm{d}}}\mu_{n, {\bm }u} ({\bm }w).$$ The result follows if we can show that $\sup_{{\bm }u \in\{ g > n^{-\gamma}\}} \lvert S_{j,k}({\bm }u) \rvert = {\mathrm{o}}(1)$ as $n\to\infty$.
By Condition \[cond:second\], we have $$\lvert \ddot C_{jk} ({\bm }v(st)) \rvert
{\leqslant}K \left( \frac{1}{\{u_j(1-u_j)u_k(1-u_k)\}^{1/2}} + \frac{1}{\{w_j(1-w_j)w_k(1-w_k)\}^{1/2}} \right).$$ To see this, use convexity of the upper bound in Condition \[cond:second\] as a function of ${\bm }u$ together with the inequality $a \wedge b {\leqslant}\sqrt{a b}$ for $a, b {\geqslant}0$. This upper bound for the absolute value of the second-order partial derivatives and the Cauchy–Schwarz inequality lead to $$\begin{gathered}
\label{eq:sjk}
\vert S_{j,k}({\bm }u) \vert
{\leqslant}~
\frac{\sqrt n }{g({\bm }u)^\omega}\Big (\int _{[0,1]^d} \frac{(w_j-u_j)^2}{u_j(1-u_j)}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) \Big )^{1/2}\Big (\int _{[0,1]^d} \frac{(w_k-u_k)^2}{u_k(1-u_k)}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) \Big )^{1/2}\\
+~
\frac{\sqrt n }{g({\bm }u)^\omega}\Big (\int _{[0,1]^d} \frac{(w_j-u_j)^2}{w_j(1-w_j)}{\operatorname{\mathds{1}}}_{\{ w_j \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) \Big )^{1/2}\\
\times
\Big (\int _{[0,1]^d} \frac{(w_k-u_k)^2}{w_k(1-w_k)}{\operatorname{\mathds{1}}}_{\{ w_k \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) \Big )^{1/2}
+ {\mathrm{O}}((1-n^{-\gamma})^n).\end{gathered}$$ [ ]{} To see this, note that ${\mathbb{P}}(S_j/n=0)= (1-u_j)^n$ and ${\mathbb{P}}(S_j/n=1)= u_j^n$ which vanishes at least with rate $O((1-n^{-\gamma})^n)$ for ${\bm }u \in[1/n^\gamma,1-1/n^\gamma]^d$. Since $$\int _{[0,1]^d}(w_j-u_j)^2{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) = {\operatorname{Var}}(S_j/n) =n^{-1} u_j(1-u_j)$$ the first term on the right-hand side of is equal to $n^{-1/2}/g({\bm }u)^\omega$, which in turn vanishes uniformly in ${\bm }u \in [1/n^\gamma,1-1/n^\gamma]^d$, as $n \to \infty$. Now consider the second term on the right-hand side. First note, that $$\frac{1}{u(1-u)}=\frac{1}{u}+\frac{1}{1-u}$$ and hence $$\begin{gathered}
\int _{[0,1]^d} \frac{(w_j-u_j)^2}{w_j(1-w_j)}{\operatorname{\mathds{1}}}_{\{ w_j \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w)
=~
\int _{[0,1]^d} \frac{(w_j-u_j)^2}{w_j}{\operatorname{\mathds{1}}}_{\{ w_j \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w)\\
+~
\int _{[0,1]^d} \frac{(w_j-u_j)^2}{1-w_j}{\operatorname{\mathds{1}}}_{\{ w_j \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w).\end{gathered}$$ We will only consider the first integral on the right-hand side, the second one can be treated analogously. Note that $$\int _{[0,1]^d} \frac{(w_j-u_j)^2}{w_j}{\operatorname{\mathds{1}}}_{\{ w_j \not\in \{0,1 \} \}}{{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w)
=
-u_j+u_j^2 {\operatorname{E}}\Big [\frac{n}{S_j} {\operatorname{\mathds{1}}}_{\{S_j {\geqslant}1\}} \Big] +O((1-n^{-\gamma})^n),$$ which by Lemma \[lem:binom\] is bounded by $ O(n^{-1}) +O((1-n^{-\gamma})^n)= O(n^{-1})$, uniformly in $u_j \in [1/n^\gamma,1-1/n^\gamma]$. Hence, uniformly in ${\bm }u \in [1/n^\gamma,1-1/n^\gamma]^d$ the second term on the right-hand side of is of the order $O(n^{-1+1/2+\omega})=o(1)$, as $ n \to \infty$.\
Finally, we need to consider ${\bm }u\in \{{\bm }v \vert g({\bm }v) > n^{-\gamma}\}\setminus[1/n\gamma,1-1/n^\gamma]^d$, i.e., it exists at least one (and at most $d-2$) component $j$ such that $1-u_j<n^{-\gamma}$. Suppose $j=1$. Then define ${\bm }u_{-1}=(1,u_2, \dots,u_d)$ and proceed as follows $$\begin{gathered}
\int_{[0,1]^d} \sqrt n \frac{C({\bm }u) - C({\bm }w)}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w) \\
=~
\int_{[0,1]^d} \sqrt n \frac{C({\bm }u) - C({\bm }u_{-1})+C({\bm }u_{-1}) - C ({\bm }w_{-1})+C({\bm }w_{-1})- C({\bm }w)}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u}({\bm }w)\\
=~
\int_{[0,1]^{d-1}} \sqrt n \frac{C({\bm }u_{-1}) - C({\bm }w_{-1})}{g({\bm }u_{-1})^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u_{-1}}({\bm }w_{-1}) \\
+ O(2n^{-\gamma+1/2+\gamma \omega})+ n^{1/2+\gamma \omega}\int_{[0,1]} \vert u_1 - w_1\vert {{\,\mathrm{d}}}\mu_{n, u_1}(w_1).\end{gathered}$$ Since, $ \int_{[0,1]} \vert u_1 - w_1\vert {{\,\mathrm{d}}}\mu_{n, u_1}(w_1) {\leqslant}\sqrt{{\operatorname{Var}}(S_1/n)} = \sqrt{u_1(1-u_1)/n}{\leqslant}n^{-1/2-\gamma/2}$ the last integral vanishes as $n \to \infty$, and the remaining term can be treated exactly as the bias of a $(d-1)$-dimensional empirical beta copula.
\[lem:int\] For any $1/\{2(1-\omega)\} < \gamma < 1$, we have $$\sup_{{\bm }u \in \{g {\geqslant}n^{-\gamma}\}}
\left\lvert
\int_{[0,1]^d}
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
- 1
\right\rvert
=
{\mathrm{O}}\left\{ n^{-(1-\gamma)/2} {\ln}(n) \right\},
\qquad n \to \infty.$$
Since $g(\frac{S_1}{n}, \dots, \frac{S_d}{n})$ is a random variable taking values in $[0, 1]$, we can write $$\begin{aligned}
\nonumber
\int_{[0,1]^d}
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
&=
\frac{1}{g({\bm }u)^\omega}
{\operatorname{E}}\Bigl[
g \bigl( \tfrac{S_1}{n}, \dots, \tfrac{S_d}{n} \bigr)^\omega
\Bigr] \\
\label{eq:int:aux}
&=
\frac{1}{g({\bm }u)^\omega}
\int_0^1
{\mathbb{P}}\Bigl\{
g \bigl(\tfrac{S_1}{n}, \dots, \tfrac{S_d}{n} \bigr) > t^{1/\omega}
\Bigr\}
{{\,\mathrm{d}}}t.\end{aligned}$$ Split the integral into two pieces, $\int_0^{a_{n, \pm}} + \int_{a_{n, \pm}}^1$, where $a_{n, \pm} = a_{n, \pm}({\bm }u) = g({\bm }u)^\omega (1 \pm {\varepsilon}_n)^\omega$. Write ${\varepsilon}_n = n^{-(1-\gamma)/2} {\ln}n$. Recall that $0 < \omega < 1/2$. On the one hand, we find $$\begin{aligned}
\int_{[0,1]^d}
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
&{\leqslant}\frac{a_{n, +}}{g({\bm }u)^\omega}
+
\frac{1 - a_{n,+}}{g({\bm }u)^\omega}
{\mathbb{P}}\Bigl\{
g \bigl(\tfrac{S_1}{n}, \dots, \tfrac{S_d}{n} \bigr) > a_{n,+}^{1/\omega}
\Bigr\} \\
&{\leqslant}(1 + {\varepsilon}_n)^\omega
+
g( {\bm }u )^{-\omega}
{\mathbb{P}}\Bigl\{
g \bigl(\tfrac{S_1}{n}, \dots, \tfrac{S_d}{n} \bigr)
> g({\bm }u) (1 + {\varepsilon}_n)
\Bigr\} \\
&{\leqslant}1 + {\varepsilon}_n
+
g( {\bm }u )^{-\omega}
2d \exp \{ - n g({\bm }u) h(1+{\varepsilon}_n) \}\end{aligned}$$ where we used in the last step. Since $h(1+{\varepsilon}_n) {\geqslant}\frac{1}{3} {\varepsilon}_n^2$ for $0 {\leqslant}{\varepsilon}_n {\leqslant}1$ and since $g({\bm }u) {\geqslant}n^{-\gamma}$, the upper bound is bounded by $$1 + {\varepsilon}_n + 2d n^{\gamma \omega} \exp \{ - \tfrac{1}{3} ({\ln}n)^2 \}
=
1 + {\varepsilon}_n + {\mathrm{o}}( {\varepsilon}_n ),
\qquad n \to \infty.$$
On the other hand, restricting the integral in to $[0, a_{n,-}]$, we have $$\begin{aligned}
\int_{[0,1]^d}
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
&{\geqslant}\frac{a_{n,-}}{g({\bm }u)^\omega}
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
>
a_{n,-}^{1/\omega}
\Bigr\} \\
&=
(1 - {\varepsilon}_n)^\omega
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
>
g({\bm }u) (1 - {\varepsilon}_n)
\Bigr\} \\
&{\geqslant}(1-{\varepsilon}_n)^\omega [1 - 4d\exp \{ - n g({\bm }u) h(1+{\varepsilon}_n) \}],\end{aligned}$$ where we used in the last step. Since $0 {\leqslant}{\varepsilon}_n \to 0$ and $g({\bm }u) {\geqslant}n^{-\gamma}$, the lower bound is bounded from below by $$\begin{aligned}
(1-{\varepsilon}_n)^\omega [1 - 4d\exp \{ - n g({\bm }u) h(1+{\varepsilon}_n) \}]
&{\geqslant}(1-{\varepsilon}_n) [1 - 4d \exp \{ - \tfrac{1}{3} ({\ln}n)^2 \} ] \\
&{\geqslant}1 - {\varepsilon}_n - {\mathrm{o}}( {\varepsilon}_n ),
\qquad n \to \infty.\end{aligned}$$
First note that, since $g(\frac{S_1}{n}, \dots, \frac{S_d}{n})$ is a positive random variable, we can write $$\begin{gathered}
\label{eq:intdecomp}
\int_{[0,1]^d}\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
=~
\int_0^1 \frac{ {\mathbb{P}}\big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})>t^{1/\omega} \big) }{ g({\bm }u)^\omega } {{\,\mathrm{d}}}t \\
=~
\int_0^{g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega} f_{n}(t){{\,\mathrm{d}}}t
+
\int_{g({\bm }u)^\omega(1+\frac{1}{{\ln}(n)})^\omega}^1 f_{n}(t) {{\,\mathrm{d}}}t
+R_n,
$$ where we define $$f_{n}(t)
=
\frac{ {\mathbb{P}}\big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})>t^{1/\omega} \big)}{g({\bm }u)^\omega} .$$ and where $$\begin{gathered}
R_n
= \int_{g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega}^{g({\bm }u )^\omega(1+\frac{1}{{\ln}(n)})^\omega}\frac{ {\mathbb{P}}\big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})>t^{1/\omega} \big) }{ g({\bm }u)^\omega }{{\,\mathrm{d}}}t
{\leqslant}(1+ \tfrac{1}{{\ln}n})^\omega - (1- \tfrac{1}{{\ln}n})^\omega\\
{\leqslant}\omega (1- \tfrac{1}{{\ln}n})^{\omega -1} \tfrac{2}{{\ln}n}
= {\mathrm{O}}( \tfrac{1}{{\ln}n} ),
\qquad n \to \infty.\end{gathered}$$
First consider the integral over $f_{n,1}$. Note that for $t {\leqslant}g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega$ $$1/{g({\bm }u)^\omega}
{\geqslant}~ f_{n}(t)
{\geqslant}~ \frac{ {\mathbb{P}}\big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})> g ({\bm }u ) \{ 1-\frac{1}{{\ln}(n)}\} \big)}{g({\bm }u)^\omega} .$$
From in Lemma \[lem:bounds\] we obtain $${\mathbb{P}}\Big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})> g ({\bm }u ) \big \{ 1-\frac{1}{{\ln}(n)} \big \} \Big)
{\geqslant}1 - K_2 \exp\big (-n g({\bm }u) h \{ {\ln}(n)^{-1}+1\} \big ).$$ This in turn implies that for $t {\leqslant}g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega$ $$1/{g({\bm }u)^\omega}
{\geqslant}~ f_{n}(t)
{\geqslant}~ 1/{g({\bm }u)^\omega} \{ 1- K_2 \exp\big (-n g({\bm }u) h \{ {\ln}(n)^{-1}+1\} \big ) \}$$ and hence $$\begin{gathered}
\{1- {\ln}(n^{-1})\} ^\omega
{\geqslant}~ \int_0^{g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega} f_{n}(t){{\,\mathrm{d}}}t \\
{\geqslant}~ \{ 1- {\ln}(n^{-1}) \}^\omega \{ 1-K_2 \exp\big (-n g({\bm }u) h \{ {\ln}(n)^{-1}+1\} \big ) \} .\end{gathered}$$ Since $g({\bm }u) {\geqslant}n^{- \gamma}$ for any ${\bm }u \in \{{\bm }v \vert g({\bm }v) {\geqslant}n^{-\gamma}\}$, this implies that $$\sup_{{\bm }u \in \{{\bm }v \vert g({\bm }v) {\geqslant}n^{-\gamma}\} } \Big \vert \int_0^{g({\bm }u)^\omega(1-\frac{1}{{\ln}(n)})^\omega} f_{n}(t){{\,\mathrm{d}}}t -1 \Big \vert = {\mathrm{O}}({\ln}(n)^{-1}).$$
Now consider the second integral in . Note that for $t {\geqslant}g({\bm }u)^\omega(1+\frac{1}{{\ln}(n)})^\omega$ $$0
{\leqslant}~ f_{n}(t)
{\leqslant}~ \frac{ {\mathbb{P}}\big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})> g ({\bm }u ) \{ 1+\frac{1}{{\ln}(n)}\} \big)}{g({\bm }u)^\omega} .$$ Furthermore, by in Lemma \[lem:bounds\] we have that $${\mathbb{P}}\Big (g(\frac{S_1}{n}, \dots, \frac{S_d}{n})> g ({\bm }u ) \{ 1+\frac{1}{{\ln}(n)}\} \Big)
{\leqslant}~ K_1 \exp\big (-n g({\bm }u) h \{ {\ln}(n)^{-1}+1\} ),$$ which is why the second integral in is bounded as follows $$0
{\leqslant}~ \int_{g({\bm }u)^\omega(1+\frac{1}{{\ln}(n)})^\omega}^1 f_{n}(t){{\,\mathrm{d}}}t ~\\
{\leqslant}~ g ({\bm }u )^{- \omega }K_1 \exp\big (-n g({\bm }u) h \{ {\ln}(n)^{-1}+1\} ).$$ Thus, the result follows by noticing that $$\sup_{{\bm }u \in \{{\bm }v \vert g({\bm }v) {\geqslant}n^{-\gamma}\} } \Big \vert \int_{g({\bm }u)^\omega(1+\frac{1}{{\ln}(n)})^\omega}^1 f_{n}(t){{\,\mathrm{d}}}t \Big \vert = {\mathrm{O}}(n^{-m}),$$ fo any $m>0$.
\[lem:bias2\] As $n \to\infty$, we have, for any $\gamma \in (1/(2(1-\omega)),1)$ $$\sup_{{\bm }u \in \{g {\geqslant}n^{-\gamma}\}}
\left\lvert
\int_{[0,1]^d}
\biggl\{
\frac{\hat {{\mathbb{C}}}_n({\bm }w)}{g({\bm }w )^\omega}
-
\frac{\hat{{\mathbb{C}}}_n({\bm }u)}{g({\bm }u)}
\biggr\}
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega}
{{\,\mathrm{d}}}\mu_{n,{\bm }u} ({\bm }w)
\right\rvert
= {\mathrm{o}}_{\mathbb{P}}(1).$$
Let $\delta_n = 1 / {\ln}(n)$. Write $\lVert \hat{{\mathbb{C}}}_n \rVert_\infty = \sup \{ \lvert \hat{{\mathbb{C}}}_n( {\bm }v) \rvert : {\bm }v \in [0, 1]^d \}$. We have $\lVert \hat{{\mathbb{C}}}_n \rVert_\infty = {\mathrm{O}}_p(1)$ as $n \to \infty$ by weak convergence of $\hat{{\mathbb{C}}}_n$ in $\ell^\infty([0, 1]^d)$.
We split the integral over ${\bm }w \in [0, 1]^d$ into two pieces: the integral over the domain $$A_{n, {\bm }u} =
\{ {\bm }w \in [0, 1]^d : \lvert {\bm }w - {\bm }u \rvert_\infty > \delta_n \} \cup \{ g < n^{-\gamma}(1-\delta_n) \}$$ and the integral over its complement; here $\lvert {\bm }x \rvert_\infty = \max\{ \lvert x_j \rvert : j = 1, \ldots, d \}$.
For all ${\bm }w \in [0, 1]^d$ and all ${\bm }u \in \{g {\geqslant}n^{-\gamma} \}$, we have $$\begin{aligned}
R_n({\bm }u, {\bm }w)
:=
\biggl\lvert
\frac{\hat{{\mathbb{C}}}_n({\bm }w)}{g({\bm }w)^\omega}
-
\frac{\hat{{\mathbb{C}}}_n({\bm }u)}{g({\bm }u)^\omega}
\biggr\rvert
\frac{g({\bm }w)^\omega}{g({\bm }u)^\omega} {\leqslant}\frac{\lvert \hat{{\mathbb{C}}}_n({\bm }w) \rvert}{g({\bm }u)^\omega}
+
\frac{ \lvert \hat{{\mathbb{C}}}_n({\bm }u) \rvert}{g({\bm }u)^{2\omega}} {\leqslant}2 \lVert \hat{{\mathbb{C}}}_n \rVert_\infty n^{2\gamma\omega}.\end{aligned}$$ Moreover, for all ${\bm }u \in \{g {\geqslant}n^{-\gamma}\}$, using Chebyshev’s inequality and the concentration inequality , we have $$\begin{aligned}
\mu_{n,{\bm }u} \bigl( A_{n, {\bm }u} \bigr)
&{\leqslant}\sum_{j=1}^d {\mathbb{P}}\Bigl\{ \bigl\lvert \tfrac{S_j}{n} - u_j \bigr\rvert > \delta_n \Bigr\}
+
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
< g({\bm }u) (1 - \delta_n)
\Bigr\} \\
&{\leqslant}d n^{-1} \delta_n^{-2} + 4d \exp \{ - n^{1-\gamma} h(1+\delta_n) \}.\end{aligned}$$ Since $0 < \omega < 1/2$, $0 < \gamma < 1$, $\delta_n = 1/{\ln}(n)$ and $h(1+\delta_n) {\geqslant}\tfrac{1}{3} \delta_n^2$, it follows that $$\sup_{{\bm }u \in \{g {\geqslant}n^{-\gamma}\}}
\int_{A_{n, {\bm }u}}
R_n( {\bm }u, {\bm }w )
{{\,\mathrm{d}}}\mu_{n, {\bm }u}( {\bm }w )
{\leqslant}n^{2\gamma\omega} [n^{-1} \delta_n^{-2} + \exp \{ - n^{1-\gamma} h(1+\delta_n) \}] \, {\mathrm{O}}_p(1)
=
{\mathrm{o}}_p(1), \qquad n \to \infty.$$ It remains to consider the integral over ${\bm }w \in [0, 1]^d \setminus A_{n, {\bm }u}$, i.e., $\lvert {\bm }w - {\bm }u \rvert_\infty {\leqslant}\delta_n$ and $g( {\bm }w ) {\geqslant}n^{-\gamma}(1 - \delta_n) > n^{-1}$, at least for sufficiently large $n$. By Lemma 4.1 in [@BerBucVol17], we have $$\label{eq:modcont}
\sup_{\substack{ {\bm }u, {\bm }w \in \{ g {\geqslant}n^{-1} \} \\ \lvert {\bm }u - {\bm }w \rvert_\infty {\leqslant}\delta_n }}
\left|
\frac{\hat{{\mathbb{C}}}_n({\bm }w)}{g({\bm }w )^\omega}
-
\frac{\hat{{\mathbb{C}}}_n({\bm }u)}{g({\bm }u)}
\right|
= {\mathrm{o}}_{{\mathbb{P}}}(1), \qquad n \to \infty.$$ In view of Lemma \[lem:int\], we obtain that $$\sup_{{\bm }u \in \{g {\geqslant}n^{-\gamma}\}}
\int_{[0, 1]^d \setminus A_{n, {\bm }u}}
R_n( {\bm }u, {\bm }w )
{{\,\mathrm{d}}}\mu_{n, {\bm }u}( {\bm }w )
{\leqslant}{\mathrm{o}}_{{\mathbb{P}}}(1) \int_{[0, 1]^d} \frac{g({\bm }w)^\omega}{g({\bm }u)^\omega} {{\,\mathrm{d}}}\mu_{n, {\bm }u}({\bm }w)
= {\mathrm{o}}_{{\mathbb{P}}}(1),$$ as $n \to \infty$. The stated limit relation follows by combining the assertions on the integral over $A_{n, {\bm }u}$ and the one over its complement.
Note that in Lemma 4.1 in [@BerBucVol17], the supremum in is taken over $[1/n,1-1/n]^d$ instead of over $\{g > n^{-1}\}$. But it can be seen in the proof of that statement that the result can be extended to the set $\{ g {\geqslant}n^{-1} \}$. Furthermore, in the latter reference, the pseudo-observations are defined as $\hat U_{i,j}=\frac{1}{n+1} R_{i,j}$. However, this does not affect the above proof, since the difference of the two empirical copulas is at most $d/n$, almost surely. This gives an additional error term on the event $\{ g {\geqslant}n^{-1} \}$ which is of the order ${\mathrm{O}}_{\mathbb{P}}(n^{\omega+1/2-1}) = {\mathrm{o}}_{\mathbb{P}}(1)$, as $n \to \infty$.
On the expectation of the reciprocal of a binomial random variable
------------------------------------------------------------------
Let $0 < u {\leqslant}1$ and let $n {\geqslant}2$ be integer. If $S \sim {\operatorname{Bin}}(n, u)$ and $T \sim {\operatorname{Bin}}(n-1, u)$, then $$\label{eq:Bin:invMoment}
{\operatorname{E}}\left[ \frac{1}{S} {\operatorname{\mathds{1}}}_{ \{ S {\geqslant}1 \} } \right]
=
nu {\operatorname{E}}\left[ \frac{1}{(1+T)^2} \right]
=
nu \int_0^1 (1-u+us)^{n-1} (-{\ln}s) \, {{\,\mathrm{d}}}s.
$$
For $k\in\{1,\dots,n\}$, we have $$\frac{{\mathbb{P}}(S=k)}{{\mathbb{P}}(T+1=k)}
=
\frac{\binom{n}{k}u^k(1-u)^{n-k}}{\binom{n-1}{k-1}u^{k-1}(1-u)^{(n-1)-(k-1)}}
=
\frac{nu}{k}.$$ We obtain that $$\begin{aligned}
{\operatorname{E}}\left[ \frac{1}{S} {\operatorname{\mathds{1}}}_{\{S{\geqslant}1\}} \right]
=
\sum_{k=1}^n \frac{1}{k} {\mathbb{P}}(S = k)
=
\sum_{k=1}^n \frac{1}{k} \frac{nu}{k} {\mathbb{P}}(T+1=k)
=
(nu) {\operatorname{E}}\left[ \frac{1}{(1+T)^2} \right].\end{aligned}$$ Now we apply a trick due to [@ChaStr72]: we have $$\frac{1}{(1+T)^2}
=
\int_{t=0}^1 \frac{1}{t} \int_{s=0}^t s^T \, {{\,\mathrm{d}}}s \, {{\,\mathrm{d}}}t
=
\int_{s=0}^1 s^T \int_{t=s}^1 \frac{{{\,\mathrm{d}}}t}{t} \, {{\,\mathrm{d}}}s
=
\int_0^1 s^T (- {\ln}s) \, {{\,\mathrm{d}}}s.$$ Taking expectations and using Fubini’s theorem, we obtain $${\operatorname{E}}\left[ \frac{1}{(1+T)^2} \right]
=
\int_0^1 {\operatorname{E}}(s^T) \, (- {\ln}s) \, {{\,\mathrm{d}}}s
=
\int_0^1 (1-u+us)^{n-1} (-{\ln}s) \, {{\,\mathrm{d}}}s,$$ as required.
\[lem:binom:aux\] Let $0 < u_n {\leqslant}1$ and let $S_n \sim {\operatorname{Bin}}(n, u_n)$. If $nu_n \to \infty$, then $$(nu_n^2) {\operatorname{E}}\left[ \frac{1}{S_n} {\operatorname{\mathds{1}}}_{ \{ S_n {\geqslant}1 \} } \right]
=
u_n+ {\mathrm{O}}\bigl( n^{-1} \bigr),
\qquad n \to \infty.$$
We start from : $$(nu_n^2) {\operatorname{E}}\left[ \frac{1}{S_n} {\operatorname{\mathds{1}}}_{ \{ S_n {\geqslant}1 \} } \right]
=
n^2 u_n^3 \int_0^1 (1-u_n+u_ns)^{n-1} (-{\ln}s) \, {{\,\mathrm{d}}}s.$$ We split the integral in two parts, cutting at $s = 1/2$.
First we consider the case $s {\leqslant}1/2$. For some positive constant $K$, we have $$\begin{aligned}
n^2u_n^3 \int_0^{1/2} (1-u_n+u_ns)^{n-1} (-{\ln}s) \, {{\,\mathrm{d}}}s
&{\leqslant}n^2u_n^3 \, (1-u_n/2)^{n-1} \int_0^{1/2} (-{\ln}s) \, {{\,\mathrm{d}}}s \\
&{\leqslant}K \, n^2u_n^3 \, (1-u_n/2)^n \\
&{\leqslant}K \, n^2u_n^3 \exp(-nu_n/2).\end{aligned}$$ For any $m > 0$, this expression is ${\mathrm{O}}(u_n (nu_n)^{-m}) = {\mathrm{o}}(n^{-1})$ as $n \to \infty$, hence by choosing $m=1$ it is ${\mathrm{O}}(n^{-1})$ as $n \to \infty$.
Second we consider the case $s {\geqslant}1/2$. The substitution $s = 1 - v / (nu_n)$ yields $$\label{eq:Bin:invMoment:aux}
n^2 u_n^3 \int_{1/2}^1 (1-u_n+u_ns)^{n-1} (-{\ln}s) \, {{\,\mathrm{d}}}s
=
u_n \int_0^{(nu_n/2)} (1-v/n)^{n-1} [-(nu_n) {\ln}\{ 1 - v/(nu_n) \}] \, {{\,\mathrm{d}}}v.$$ We need to show that this integral is $u_n + {\mathrm{O}}(n^{-1})$ as $n \to \infty$.
For facility of writing, put $k_n = nu_n$. Recall that $k_n \to \infty$ as $n \to \infty$ by assumption. The inequalities $x {\leqslant}- {\ln}(1-x) {\leqslant}x/(1-x)$ for $0 {\leqslant}x < 1$ imply that $$0 {\leqslant}-k_n {\ln}(1 - v/k_n) - v {\leqslant}\frac{v^2}{k_n-v} {\leqslant}\frac{2 v^2}{k_n}, \qquad v \in [0, k_n/2].$$ As $(1-v/n)^{n-1} {\leqslant}(1-k_n/(2n))^{-1} (1-v/n)^{-n} {\leqslant}2 \exp(-v)$ for $v \in [0, k_n/2]$, we find $$\begin{aligned}
u_n \int_0^{k_n/2} (1-v/n)^{n-1} \, \left\lvert -k_n {\ln}( 1 - v/k_n ) - v \right\rvert \, {{\,\mathrm{d}}}v
&{\leqslant}\frac{4 u_n}{k_n} \int_0^{k_n/2} \exp(-v) \, v^2 \, {{\,\mathrm{d}}}v \\
&=
{\mathrm{O}}( n^{-1} ),
\qquad n \to \infty.\end{aligned}$$ As a consequence, replacing $-k_n {\ln}(1-v/k_n)$ by $v$ in produces an error of the required order ${\mathrm{O}}(n^{-1})$.
It remains to consider the integral $$u_n\int_0^{k_n/2} (1-v/n)^{n-1} \, v \, {{\,\mathrm{d}}}v.$$ Via the substitution $x = 1-v/n$, this integral can be computed explicitly. After some routine calculations, we find it is equal to $$u_n \frac{n}{n+1} [ 1 - \{1 - k_n/(2n)\}^{n} (1 + k_n/2) ].$$ Since $\{1 - k_n/(2n)\}^{n} {\leqslant}\exp(-k_n/2)$, the previous expression is $$u_n + {\mathrm{O}}(u_n n^{-1}) + {\mathrm{O}}(u_n \exp(-k_n/2) k_n ), \qquad n \to \infty,$$ The error term is ${\mathrm{O}}(n^{-1})$, as required.
\[lem:binom\] If $0 < u_n {\leqslant}1$ is such that $nu_n \to \infty$ as $n \to \infty$, then $$\sup_{u_n {\leqslant}u {\leqslant}1} \left\lvert nu^2 {\operatorname{E}}\left[ \frac{1}{S} {\operatorname{\mathds{1}}}_{ \{ S {\geqslant}1 \} } \right] - u \right\rvert
=
{\mathrm{O}}( n^{-1} ),
\qquad n \to \infty,$$ where the expectation is taken for $S \sim {\operatorname{Bin}}(n,u)$.
The function sending $u \in [u_n, 1]$ to $\lvert nu^2 {\operatorname{E}}[ S^{-1} {\operatorname{\mathds{1}}}_{ \{ S {\geqslant}1 \} } ] - 1 \rvert$, with $S \sim {\operatorname{Bin}}(n, u)$, is continuous and therefore attains its supremum at some $v_n \in [u_n, 1]$. Since $nv_n {\geqslant}nu_n \to \infty$ as $n \to \infty$, we can apply Lemma \[lem:binom:aux\] to find that the supremum is ${\mathrm{O}}(n^{-1})$ as $n \to \infty$.
Inequalities for binomial random variables
------------------------------------------
If $S \sim {\operatorname{Bin}}(n, u)$ is a binomial random variable with succes probability $0 < u < 1$, then Bennett’s inequality states that $$\Pr \Bigl(
\sqrt{n} \bigl\lvert \tfrac{S}{n} - u \bigr\rvert {\geqslant}\lambda
\Bigr)
{\leqslant}2 \exp \Bigl\{
- \frac{\lambda^2}{2u}
\psi \Bigl( \frac{\lambda}{\sqrt{n} u} \Bigr)
\Bigr\}
=
2 \exp \Bigl\{
- nu \, h \Bigl( 1 + \frac{\lambda}{\sqrt{n} u} \Bigr)
\Bigr\}$$ for $\lambda > 0$, where $\psi(x) = 2 \, h(1+x)/x^2$ and $h(x) = x({\ln}x - 1) + 1$; see for instance @VanWel96 [Proposition A.6.2]. Setting $\lambda = \sqrt{n} u \delta$, we find $$\label{eq:bennett}
\Pr \Bigl(
\big\lvert \tfrac{S}{n} - u \big\rvert {\geqslant}u \delta
\Bigr)
{\leqslant}2 \exp \{ - nu \, h(1+\delta) \},
\qquad \delta > 0.$$ Note that $h(1+\delta) = \int_0^\delta {\ln}(1+t) {{\,\mathrm{d}}}t {\geqslant}\int_0^\delta (t - \tfrac{1}{2} t^2) {{\,\mathrm{d}}}t = \tfrac{1}{2} \delta^2 (1 - \tfrac{1}{3} \delta)$ for $\delta {\geqslant}0$ and thus $h(1+\delta) {\geqslant}\tfrac{1}{3} \delta^2$ for $0 {\leqslant}\delta {\leqslant}1$. We extend to a vector of independent binomial random variables and in terms of the weight function $g$ in .
\[lem:bounds\] If $S_1, \dots, S_d$ are independent random variables with $S_j \sim {\operatorname{Bin}}(n,u_j)$ and $0 < u_j < 1$ for all $j \in \{1, \ldots, d\}$, then, for $\delta > 0$, $$\begin{aligned}
\label{eq:upper}
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
{\geqslant}g({\bm }u ) ( 1 + \delta )
\Bigr\}
&{\leqslant}2d \exp\bigl\{ - n g({\bm }u) h(1+\delta) \bigr\}, \\
\label{eq:lower}
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
{\leqslant}g({\bm }u ) ( 1 - \delta )
\Bigr\}
&{\leqslant}4d \exp \bigl\{ - n g({\bm }u) h(1+\delta) \bigr\},\end{aligned}$$ with $h$ as above; in particular, $h(1+\delta) {\geqslant}\tfrac{1}{3} \delta^2$ for $0 < \delta {\leqslant}1$.
Let us start with . The definition of the weight function $g$ in yields $${\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \dots , \tfrac{S_d}{n} \bigr)
{\leqslant}g({\bm }u ) ( 1 - \delta )
\Bigr\}
{\leqslant}\sum_{j=1}^d
\Bigl[
{\mathbb{P}}\Bigl\{
\tfrac{S_j}{n} {\leqslant}g({\bm }u ) (1 - \delta)
\Bigr\}
+
{\mathbb{P}}\Bigl\{
\max_{k \ne j} \bigl( 1 - \tfrac{S_{k}}{n} \bigr)
{\leqslant}g({\bm }u ) ( 1 - \delta)
\Bigr\}
\Bigr].
$$ Let us first consider the first term on the right-hand side, i.e., ${\mathbb{P}}\{ \tfrac{S_j}{n} {\leqslant}g({\bm }u ) ( 1 - \delta ) \}$. By definition of the weight function we have $g ({\bm }u) {\leqslant}u_j$. By Bennett’s inequality , $$\begin{aligned}
{\mathbb{P}}\Bigl\{ \tfrac{S_j}{n} {\leqslant}g({\bm }u) ( 1 - \delta ) \Bigr\}
&{\leqslant}{\mathbb{P}}\Bigl\{ \tfrac{S_j}{n} {\leqslant}u_j ( 1 - \delta ) \Bigr\} \\
&{\leqslant}{\mathbb{P}}\Bigl\{ \big\lvert \tfrac{S_j}{n} - u_j \big\rvert
{\geqslant}u_j \delta \Bigr\} \\
&{\leqslant}2 \exp \{ - n u_j h(1 + \delta) \} \\
&{\leqslant}2 \exp \{ - n g({\bm }u) h(1 + \delta) \} .\end{aligned}$$ Second, consider the term ${\mathbb{P}}\{ \max_{k \ne j} ( 1 - \tfrac{S_{k}}{n} ){\leqslant}g({\bm }u ) ( 1 - \delta) \} $. Suppose $j=1$; the other cases can be treated exactly along the same lines. We have $g({\bm }u) {\leqslant}\max_{k \neq 1}(1 - u_k)$. Assume without loss of generality that $\max_{k \neq 1} (1 - u_k) = 1 - u_2$. Then we obtain $g({\bm }u) {\leqslant}1 - u_2$ and, by Bennett’s inequality applied to $n - S_2 \sim {\operatorname{Bin}}(n, 1-u_2)$, $$\begin{aligned}
{\mathbb{P}}\Bigl\{
\max_{k \ne 1} \bigl(1-\tfrac{S_{k}}{n}\bigr)
{\leqslant}g({\bm }u ) (1 - \delta)
\Bigr\}
&{\leqslant}{\mathbb{P}}\Bigl\{
\max_{k \ne 1} \bigl(1-\tfrac{S_{k}}{n}\bigr)
{\leqslant}(1 - u_2) (1 - \delta)
\Bigr\} \\
&{\leqslant}{\mathbb{P}}\Bigl\{
1 - \tfrac{S_{2}}{n}
{\leqslant}(1- u_2) (1 - \delta)
\Bigr\} \\
&{\leqslant}{\mathbb{P}}\Bigl\{
\big\lvert 1 - \tfrac{S_2}{n} - (1- u_2) \big\rvert
{\geqslant}(1- u_2 ) \delta
\Bigr\} \\
&{\leqslant}2 \exp \{ -n (1-u_2) h(1 + \delta) \} \\
&{\leqslant}2 \exp \{ -n g({\bm }u) h(1 + \delta) \}.\end{aligned}$$ Let us now show . First suppose $g({\bm }u) = u_1$. Since $g(\frac{S_1}{n} , \dots \frac{S_d}{n }) {\leqslant}\frac{S_1}{n} $ we have, by Bennett’s inequality , $$\begin{aligned}
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \dots, \tfrac{S_d}{n} \bigr)
{\geqslant}g({\bm }u) (1 + \delta)
\Bigr\}
&{\leqslant}{\mathbb{P}}\Bigl\{
\tfrac{S_1}{n} {\geqslant}u_1 (1 + \delta)
\Bigr\} \\
&{\leqslant}{\mathbb{P}}\Bigl\{
\bigl\lvert \tfrac{S_1}{n} - u_1 \bigr\rvert {\geqslant}u_1 \delta
\Bigr\} \\
&{\leqslant}2 \exp \bigl\{ - nu_1 h(1+\delta) \bigr\}
=
2 \exp \bigl\{ - n g({\bm }u) h(1+\delta) \bigr\}.\end{aligned}$$ Finally, suppose that $g({\bm }u) = 1 - u_1 {\geqslant}1 - u_k$, for $k = 3, \ldots, d$. Note that $g (\tfrac{S_1}{n}, \dots, \tfrac{S_d}{n }) {\leqslant}\max_{k \neq 2} (1 - \tfrac{S_j}{n}) $, which yields $$\begin{aligned}
{\mathbb{P}}\Bigl\{
g \bigl( \tfrac{S_1}{n}, \ldots, \tfrac{S_d}{n} \bigr)
{\geqslant}g({\bm }u ) (1 + \delta)
\Bigr\}
&{\leqslant}{\mathbb{P}}\Bigl\{
\max_{k \neq 2} (1- \tfrac{S_k}{n}) {\geqslant}(1- u_1) (1 + \delta)
\Bigr\} \\
&{\leqslant}\sum_{k \neq 2} {\mathbb{P}}\Bigl\{
1 - \tfrac{S_k}{n} {\geqslant}(1- u_1)(1 + \delta)
\Bigr\}.
$$ By Bennett’s inequality applied to $n - S_k \sim {\operatorname{Bin}}(n, 1-u_k)$ for every $k \ne 2$, we have, since $(1-u_1)/(1-u_k) {\geqslant}1$, $$\begin{aligned}
{\mathbb{P}}\Bigl\{
1 - \tfrac{S_k}{n} {\geqslant}(1- u_1)(1 + \delta)
\Bigr\}
&{\leqslant}{\mathbb{P}}\Bigl\{
\bigl\lvert 1 - \tfrac{S_k}{n} - (1-u_k) \bigr\rvert
{\geqslant}(1 - u_1)(1+\delta) - (1 - u_k)
\Bigr\} \\
&{\leqslant}2 \exp \Bigl\{
- n (1 - u_k) h \Bigl( \tfrac{1 - u_1}{1 - u_k} (1 + \delta) \Bigr)
\Bigr\}.\end{aligned}$$ For $a {\geqslant}1$ and $\delta {\geqslant}0$, a direct calculation[^1] shows that $h(a(1+\delta)) - a \, h(1+\delta) {\geqslant}h(a) {\geqslant}0$ and thus $h(a(1+\delta)) {\geqslant}a \, h(1+\delta)$. Apply this inequality to $a = (1-u_1) / (1-u_k)$ to find $$\begin{aligned}
{\mathbb{P}}\Bigl\{
1 - \tfrac{S_k}{n} {\geqslant}(1- u_1)(1 + \delta)
\Bigr\}
&{\leqslant}2 \exp \Bigl\{ - n (1 - u_k) \tfrac{1 - u_1}{1 - u_k} h(1+\delta) \Bigr\} \\
&=
2 \exp \{ - n (1 - u_1) h(1+\delta) \}
=
2 \exp \{ - n g({\bm }u) h(1+\delta) \}. \qedhere\end{aligned}$$
Extensions of results in [@BerBucVol17] {#sec:BerBucVol17}
---------------------------------------
For any sequence $\delta_n>0 $ that converges to zero as $n \to \infty$, Lemma 4.10 in [@BerBucVol17] can be extended to $$\label{eq:extend1}
\sup \left\{
\left|
\frac{{\mathbb{C}}_n({\bm }u )}{g({\bm }u)^\omega}
-
\frac{{\mathbb{C}}_n({\bm }u')}{g({\bm }u')^\omega}
\right|
\; : \;
g({\bm }u) {\geqslant}c/n, \,
g({\bm }u') {\geqslant}c/n, \,
\lvert {\bm }u - {\bm }u' \rvert {\leqslant}\delta_n
\right\}
= {\mathrm{o}}_{\mathbb{P}}(1), \qquad n \to \infty.$$ Here, ${\mathbb{C}}_n = \sqrt{n} ( \tilde{C}_n - C )$ and $\tilde{C}_n$ is the empirical copula based on the generalized inverse function of the marginal empirical distribution functions [@BerBucVol17 beginning of Section 4.2]. Furthermore, Theorem 4.5 in the same reference can be extended to $$\label{eq:extend2}
\sup \left\{
\left|
\frac{\hat {\mathbb{C}}_n({\bm }u)}{g({\bm }u)^\omega} - \frac{\bar {\mathbb{C}}_n({\bm }u)}{g({\bm }u )^\omega}
\right|
\; : \;
g({\bm }u) {\geqslant}c/n
\right\}
= {\mathrm{o}}_{\mathbb{P}}(1), \qquad n \to \infty.$$
Let us start with . The result is similar to the result in Lemma 4.10, in particular Equation (4.1), in [@BerBucVol17]. A look at the proof of the result shows that the restriction ${\bm }u , {\bm }u' \in [c/n, 1-c/n]^d$ instead of $ {\bm }u , {\bm }u' \in \{ g {\geqslant}c/n \}$ is not needed. The proof of Equation (4.1) in Lemma 4.10 in [@BerBucVol17] is based on Lemma 4.7, 4.8 and Equations (4.8) and (4.8) which are all valid on sets of the form $N(c_{n1},c_{n2})= \{ g \in (c_{n1}, c_{n2}] \}$. Hence, in the proof, all suprema can be taken over $ {\bm }u , {\bm }u' \in \{ g {\geqslant}c/n \}$ instead of ${\bm }u , {\bm }u' \in [c/n, 1-c/n]^d$, which gives us exactly .
For the proof of note that for any ${\bm }u \in \{ g {\geqslant}c/n \}$ we can find ${\bm }u' \in \{ g {\geqslant}n^{-1/2} \}$ such that $\lvert {\bm }u - {\bm }u' \rvert {\leqslant}d n^{-1/2}$. To find such a ${\bm }u'$ is all that it is needed to extend the proof of Theorem 4.5 in [@BerBucVol17] to obtain .
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors gratefully acknowledge the editor-in-chief, the associate editor, and the referees for additional references, for suggesting the idea of a weighted test of independence, and for various suggestions concerning the numerical experiments on the estimation of the Pickands dependence function.
Betina Berghaus gratefully acknowledges support by the Collaborative Research Center “Statistical modeling of nonlinear dynamic processes” (SFB 823) of the German Research Foundation (DFG).
Johan Segers gratefully acknowledges funding by contract “Projet d’Actions de Recherche Concertées” No. 12/17-045 of the “Communauté française de Belgique” and by IAP research network Grant P7/06 of the Belgian government.
[^1]: Or, since $h(x) = \int_1^x {\ln}(t) \, {{\,\mathrm{d}}}t$, we have $h(a(1+\delta)) = \int_1^{a(1+\delta)} {\ln}(t) \, {{\,\mathrm{d}}}t = a \int_{1/a}^{1+\delta} {\ln}(as) \, {{\,\mathrm{d}}}s {\geqslant}a \int_1^{1+\delta} {\ln}(s) \, {{\,\mathrm{d}}}s = a \, h(1+\delta)$ for $a {\geqslant}1$ and $\delta {\geqslant}0$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The enormous theoretical potential of Quantum Information Processing (QIP) is driving the pursuit for its practical realization by various physical techniques. Currently Nuclear Magnetic Resonance (NMR) has been the forerunner by demonstrating a majority of quantum algorithms. In NMR, spin systems consisting of coupled nuclear spins are utilized as qubits. In order to carry out QIP, a spin system has to meet two major requirements: (i) qubit addressability and (ii) mutual coupling among the qubits. It has been demonstrated that the magnitude of the mutual coupling among qubits can be increased by orienting the spin-systems in a liquid crystal matrix and utilizing the residual dipolar couplings. While utilizing residual dipolar couplings may be useful to increase the number of qubits, nuclei of same species (homonuclei) might become strongly coupled. In strongly coupled spin-systems, spins loose their individual identity of being qubits. We propose that even such strongly coupled spin-systems can be used for QIP and the qubit-manipulation can be achieved by transition-selective pulses. We demonstrate experimental preparation of pseudopure states, creation of maximally entangled states, implementation logic gates and implementation of Deutsch-Jozsa (DJ) algorithm in strongly coupled 2,3 and 4 spin systems. The energy levels of the strongly coupled 3 and 4 spin systems were obtained by using a Z-COSY experiment.'
author:
- |
T.S. Mahesh$^{\dagger 1}$, Neeraj Sinha$^{\dagger 2}$, Arindam Ghosh$^{\dagger}$, Ranabir Das$^{\dagger}$, N. Suryaprakash$^{\ddagger}$\
Malcom H. Levitt$^{\top}$, K.V. Ramanathan$^{\ddagger}$ and Anil Kumar$^{\dagger \ddagger *}$\
$^{\dagger}$ [*Department of Physics, Indian Institute of Science, Bangalore, India*]{}\
$^{\ddagger}$ [*Sophisticated Instruments Facility, Indian Institute of Science, Bangalore, India*]{}\
$^{\top}$ [*Department of Chemistry, University of Southampton, Southampton SO17 1BJ, England*]{}\
title: Quantum Information Processing by NMR using strongly coupled spins
---
=-0.5cm
[^1]
Introduction
============
The theoretical success of exploiting the quantum nature of physical systems in certain information processing tasks like prime factorization [@shor] and unsorted database search [@grover] has motivated the pursuit for the practical realization of Quantum Information Processing (QIP) [@preskill; @chuangbook; @bouwmeester]. With the demonstration of many quantum algorithms, Nuclear Magnetic Resonance (NMR) is now considered as a suitable test-bed for QIP. One of the main challenges for the progress of NMR QIP is “how to increase the number of qubits?". In this direction several attempts are being made, such as (i) find molecules with different chemical shifts and J-couplings, and (ii) use of dipolar and quadrupolar couplings. This paper concentrates on one aspect: “how to use dipolar couplings among homonuclear spins?". This problem is outlined in the following paragraphs.
In NMR, systems consisting of coupled spin-1/2 nuclei form qubits. In order to carry out QIP, the spin-system has to meet two main requirements: qubit addressability and mutual coupling among the qubits. In liquid state NMR using isotropic fluids, the qubit addressability is normally provided by the differences in Larmor frequencies of the various spin-1/2 nuclei, while the mutual coupling is normally provided by the scalar (J) coupling among the nuclei connected by covalent bonds. The Hamiltonian for a J-coupled spin system is [@ernstbook], $$\begin{aligned}
{\mathcal H}&=&{\mathcal H}_Z+{\mathcal H}_J \nonumber \\
&=& \sum_i \omega_i I_{iz} +\sum_{i,j (i<j)} 2\pi J_{ij}(I_{iz}I_{jz}+I_{ix}I_{jx}+I_{iy}I_{iy}) \end{aligned}$$ where ${\mathcal H}_Z$ is the Zeeman Hamiltonian, and ${\mathcal H}_J$ is the coupling Hamiltonian. When $2\pi J_{ij} \ll \vert \omega_i-\omega_j \vert$, the system is said to be weakly coupled, and the Hamiltonian can be approximated to [@ernstbook], $$\begin{aligned}
{\mathcal H}= \sum_i \omega_i I_{iz} +\sum_{i,j (i<j)} 2\pi J_{ij}I_{iz}I_{jz}\end{aligned}$$
For qubit addressability, all $\omega_i$ should be sufficiently dispersed and all $J_{ij}$ should be non-negligible ($>$ 1Hz) and unequal in magnitude. In such a circumstance each spin can be treated as a qubit and the coupled nuclei as several qubits. The values of $J_{ij}$ depend on the covalent bonds connecting spins i and j, and normally has a small range ($< 10^2$ Hz) and becomes too small ($< 1$ Hz) if the spins are connected by more than 4-5 covalent bonds. This places a natural limit on the number of qubits reachable by liquid state NMR using J-couplings alone. To overcome this limitation the possibility of using dipolar couplings was considered. The truncated Hamiltonian for dipolar interaction is [@ernstbook; @slic], $$\begin{aligned}
{\mathcal H}_D= \sum_{i,j (i<j)} 2\pi D_{ij}(3I_{iz}I_{jz}-{\mathrm{\bf I_i.I_j}})\end{aligned}$$ The dipolar coupling $D_{ij}$, between spins of gyromagnetic ratios $\gamma_i$ and $\gamma_j$ whose inter-distance vector $r_{ij}$ makes an angle $\theta_{ij}$ with the Zeeman magnetic field is of the form [@ernstbook; @slic], $$\begin{aligned}
D_{ij}= \frac{\gamma_i \gamma_j \hbar}{4 \pi r^3_{ij}}(1-3cos^2\theta_{ij})\end{aligned}$$
Dipolar couplings among common nuclear species have larger magnitudes ($\sim$ $10^3$ Hz) and longer range than the scalar couplings. However, in isotropic liquids the time average of $D_{ij}$ vanishes, while in solids there are too many dipolar couplings resulting in broad unresolved lines and loss of qubit addressability. In molecules oriented in a liquid crystal matrix, while the intermolecular dipolar couplings are vanishingly small, the intra molecular dipolar couplings survive, scaled down by the order parameter $(S_{ij})$ of the liquid crystal [@slic; @khetrapal], $$\begin{aligned}
D_{ij}^{ori}= \frac{\gamma_i \gamma_j \hbar}{4 \pi r^3_{ij}} < 1-3cos^2\theta_{ij} >
=- \frac{\gamma_i \gamma_j \hbar}{4 \pi r^3_{ij}} S_{ij}\end{aligned}$$ In such systems one obtains a finite number of sharp well resolved spectral lines making it possible to use such systems for NMR-QIP. In the NMR-QIP experiments implemented so far, the systems have been chosen such that either (i) $2\pi(J_{ij} + 2D_{ij}) \ll \vert (\omega_i - \omega_j) \vert$, yielding weakly coupled spin-systems which is the case for the heteronuclear spin-systems [@chuang00b; @marka] or (ii) the coupling $2(J_{ij} + 2D_{ij})$ is finite and $\vert \omega_i - \omega_j \vert$ = 0, i.e., equivalent-spins case [@fung00; @mahesh02]. In the latter case, the symmetry filtering of energy levels become increasingly difficult for higher number of qubits [@mahesh02]. Even though the heteronuclear spins oriented in liquid crystal matrix are excellent for QIP since they provide very good qubit addressability as well as large mutual coupling, the use of more than 3 to 4 heteronuclear spins is limited by the extensive hardware requirements. Therefore for reaching larger number of qubits, one needs to utilize homonuclear (nuclei of same species having same $\gamma$ but different chemical shifts ) spins oriented in a liquid crystal.
Homonuclear spins oriented in a liquid crystal generally become strongly coupled since the dipolar couplings become comparable to or more than the differences in Larmor frequencies $\vert \omega_i - \omega_j \vert$. In such a situation, the Zeeman and the coupling parts of the Hamiltonian do not commute. Therefore the eigenstates of strongly coupled spins are obtained as the linear combinations of product states of various spins and the individual spins can no more be treated as qubits. We propose and demonstrate here that, the $2^{N}$ eigenstates of a coupled N-spin 1/2 system can be treated as an N qubit system even in the presence of strong coupling. Similar idea has already been used in demonstration of QIP using quadrupolar (S $>$1/2) nuclei oriented in high magnetic field, where the 2S+1 non-equidistant energy levels have been treated as N-qubit systems, where $2^N=2S+1$. So far S=3/2 and 7/2 have been utilized respectively as 2 and 3 qubit systems [@fun; @mulf; @neeraj01; @mur; @ranapra], for various NMR-QIP.
While substantial work has been carried out in NMR-QIP using weekly coupled spin systems [@jonescrit], till now the use of the strongly coupled spin systems for QIP has not been experimentally demonstrated, presumably because (i) spin-selective pulses are not defined in the case of strongly coupled spins [@ernstbook], and (ii) the difficulty in constructing a general unitary operator using the evolution under scalar coupling [@cory98b]. The problem of using scalar coupling evolution of a strongly coupled two-spin system for a general unitary transform has recently been addressed theoretically, but extending to a N-spin system is complicated [@benjamin]. However, we note that unlike spin-selective pulses, the transition-selective pulses are well defined even in strongly coupled spin-systems [@ernstbook] and hence it is possible to construct a unitary transformation using transition selective pulses. In section \[2spinstrong\] we demonstrate NMR-QIP on a strongly coupled two spin-system in isotropic medium by preparing pseudopure states, implementing DJ algorithm, by creating Einstein-Podolsky-Rosen (EPR) state and by implementing logic gates. Section \[3spinstrong\] describes the creation of Greenberger-Horne-Zeilinger (GHZ) states and implementation of two-qubit DJ algorithm on a strongly coupled three spin system in an oriented medium after labeling the transitions using the Z-COSY experiment. Labeling of transitions, preparation of pseudopure states, and implementation of gates on a four-spin strongly coupled system are demonstrated in section \[4qstrong\].
In this study the strongly coupled systems used are:
Sl. No. Sample Solvent Isotropic/ Oriented No. of qubits
--------- ----------------------------- ---------- --------------------- ---------------
1 Trisodium citrate $D_2$O Isotropic 2
2 Organometallic compound (I) CDCl$_3$ Isotropic 2
3 1-bromo-2,3-dichlorobenzene ZLI-1132 Oriented 3
4 2-chloroiodobenzene ZLI-1132 Oriented 4
The experiments have been carried out on a Bruker DRX-500 NMR spectrometer at 300K temperature.
Two-spin strongly coupled system {#2spinstrong}
================================
The four eigenstates of a strongly coupled two-spin system (of spin 1/2 nuclei; AB spin-system) in isotropic medium are, $\vert \alpha \alpha \rangle$, $\cos\Theta \vert \alpha \beta \rangle+\sin\Theta \vert \beta \alpha \rangle$, $\cos\Theta \vert \beta \alpha \rangle-\sin\Theta \vert \alpha \beta \rangle$, and $\vert \beta\beta \rangle$ (Figure 1(a)), where $\Theta=\frac{1}{2}\tan^{-1}(2 \pi J_{AB}/(\omega_A-\omega_B))$ [@ernstbook]. These eigenstates are labeled respectively as $\vert 00 \rangle$, $\vert 01 \rangle$, $\vert 10 \rangle$, and $\vert 11 \rangle$, thus forming a two-qubit system (Figure 1(b)). To demonstrate QIP on such a system, we have taken the strongly coupled $^1$H spins of trisodium citrate (Figure 1(c)). In this system, the scalar coupling (J) is 15 Hz, the difference in Larmor frequencies ($\Delta f$) is 55.5 Hz, and the strong coupling parameter ($\Theta$) is $7.6^\circ$. The equilibrium spectrum of the system is shown Figure 1(d).
Preparation of Pseudopure states
--------------------------------
In QIP, the computation normally begins from a definite initial state known as a pure state [@preskill; @chuangbook]. In NMR however because of the small energy gaps, it is not possible to realize a pure state, wherein the whole population is in one energy level, since it requires very low temperatures as well as very high magnetic fields. However, an alternate solution was discovered to overcome this problem [@cory97; @chuang97]. In thermal equilibrium, NMR density matrix can be written as $$\rho=2^{-N} \{I + \epsilon \rho_{dev}\}.$$ The first part is a normalized unit matrix which corresponds to a uniform population background. The second part containing the traceless deviation density matrix $\rho_{dev}$ (with a small coefficient $\epsilon \sim 10^{-5}$) evolves under various NMR Hamiltonians, and gives measurable signal. It was observed independently by Cory et al [@cory97] and Chuang et al [@chuang97], that by applying certain pulse sequence to the system in equilibrium, we can prepare the so called pseudopure density matrix, $$\rho_{pps}=2^{-N} \{(1-\epsilon'/2^N)I + \epsilon' \rho_{pure}\}.$$ The first part is again a scaled unit matrix, but the second part corresponds to a pure state. Such pseudopure states mimic pure states [@cory97; @chuang97]. Many methods have been proposed for the preparation of pseudopure states including spatial averaging [@cory97; @cory98], temporal averaging [@knill98], logical labeling [@chuang97; @chuang98; @kavita00], and spatially averaged logical labeling [@maheshpra01]. Some of the other methods include the preparation of pseudopure states via cat states [@knillnature] and preparation of pair of pseudopure states [@fung00]. Both J-evolution and transition selective pulse methods have been utilized for preparation of pseudopure states [@cory97; @knill98; @ernst; @kavita00; @neeraj01; @peng].
We have adopted the method of spatial averaging using transition selective pulses. The Boltzmann distribution of populations at high-temperature approximation is linear with energy gap. The equilibrium populations (in excess of a large uniform background population) of a homonuclear two spin system is given in Fig 2(a). For creating a pseudopure state all the populations except one of the states have to be equalized. This distribution can be achieved by a sequence of transition selective pulses intermittent with field gradient pulses to destroy any coherence created in the process.
A transition-selective pulse of nutation angle $\theta$ and of any transverse phase between states $(i,j)$, changes the populations as follows: $$\begin{aligned}
p_i^\prime &=& p_i \cos^2(\theta/2)+p_j \sin^2(\theta/2) \nonumber \\
p_j^\prime &=& p_j \cos^2(\theta/2)+p_i \sin^2(\theta/2).
\label{popurf}\end{aligned}$$ To prepare the $\vert 00 \rangle$ pseudopure state (Fig 2(b)) from equilibrium (Fig 2(a)), we use a sequence $[(\theta)^{\vert 10 \rangle \leftrightarrow \vert 11 \rangle}-G_z
-(90^\circ)^{\vert 01 \rangle \leftrightarrow \vert 11 \rangle}-G_z^{\prime}]$ (pulses are applied from left to right), and it is inferred from the deviation populations of Fig 2(a) and (b), that $\theta$ should be such that $\cos^2(\theta/2)$=2/3. This yields $\theta=70.5^\circ$. $\theta=90^\circ$ on the other hand equalizes the populations of the two levels to an average value. Thus, the $\vert 00 \rangle$ pseudopure state is prepared by the pulse sequence $[(70.5^\circ)^{\vert 10 \rangle \leftrightarrow \vert 11 \rangle}-G_z
-(90^\circ)^{\vert 01 \rangle \leftrightarrow \vert 11 \rangle}-G_z^{\prime}]$ and the corresponding spectrum is given in Fig. 2(b). The $\vert 01 \rangle$ and $\vert 10 \rangle$ pseudopure states (Figure 2(c), 2(d)) are respectively prepared by applying respectively $(180^\circ)^{(\vert 00 \rangle \leftrightarrow \vert 01 \rangle)}$ and $(180^\circ)^{(\vert 00 \rangle \leftrightarrow \vert 10 \rangle)}$ pulse after creating the $\vert 00 \rangle$ pseudopure state. The $\vert 11 \rangle$ pseudopure state (Figure 2(e)) is prepared by the pulse sequence $(70.5^\circ)^{(\vert 00 \rangle \leftrightarrow \vert 01 \rangle}-
G_z-(90^\circ)^{\vert 00 \rangle \leftrightarrow \vert 10 \rangle}-G_z^{\prime}$. The observed intensities in the spectra on the left hand side of Fig. 2 correspond to the created population distribution and hence confirm the creation of the pseudopure states.
Deutsch-Jozsa Algorithm
-----------------------
Deutsch-Jozsa (DJ) algorithm is one of the first quantum algorithms which successfully demonstrated the power of QIP [@deu; @cleve]. The task of DJ algorithm is to distinguish between two classes of many-input-one-output functions, constant and balanced. Constant functions are those functions in which all the outputs are same independent of inputs; and balanced functions are those in which half the number of inputs give one output and the other half gives another output. Classically, given a function of n input bits, it takes $2^{n-1}+1$ function-calls on an average, to determine whether the function is constant or balanced, whereas DJ algorithm needs only one function-call for any number of qubits. DJ algorithm has been implemented in NMR using scalar coupling evolution as well as using spin and transition selective pulses [@kavita00; @djchu; @djjo; @free]. We have followed Cleve’s version of DJ algorithm which requires one extra work qubit [@cleve]. The circuit diagram and the NMR pulse sequence for implementing 1-qubit DJ algorithm are shown in Figures 3(a) and 3(b) respectively. The experiment begins with the $\vert 00 \rangle$ pseudopure state. An initial $(\pi/2)_{-y}$ pulse (the pseudopure Hadamard operation [@djjo]) on all the qubits creates a superposition. It may be noted that unlike weakly coupled spins the superposition created here is not uniform in the eigenbasis, since the coefficients of various eigenstates are different. However, as is shown here it is still possible to distinguish between the different classes of functions. The Hadamard operation is followed by an unitary operator $U_f$ corresponding to the given function $f$. The unitary operator carries out the transformation $\vert r,s \rangle \stackrel{U_f}{\rightarrow} \vert f(s) \oplus r,s \rangle$, where $\vert r \rangle $ and $\vert s \rangle $ are the states of the work-qubit and the input-qubit respectively. The four different one-qubit functions and corresponding unitary operator as well as r.f. pulses are listed in Table 1. The test for a balanced function is that the transitions of input-qubit will gain opposite phases at the end of the algorithm. The experimental results corresponding to all the four functions $f_1$, $f_2$, $f_3$, and $f_4$ along with their corresponding simulated spectra are given in Figures 3(c)-3(f). From the spectra we can identify that functions $f_1$ and $f_2$ are constant since the transitions of 1 and 2 are of same phase, whereas $f_3$ and $f_4$ are balanced since the transitions of 1 and 2 are of opposite phase.
Creation of an EPR state
------------------------
Einstein-Podolsky-Rosen (EPR) pairs are the maximally entangled pairs of the form $$\begin{aligned}
&& \frac{1}{\sqrt{2}}(\vert 00 \rangle \pm \vert 11 \rangle) \;\;\;\;
{\mathrm {or}} \nonumber \\
&& \frac{1}{\sqrt{2}}(\vert 01 \rangle \pm \vert 10 \rangle)\end{aligned}$$ which are not reducible into product states of individual qubits [@preskill]. The non-local correlation exhibited by these pairs have no classical equivalence, and are exploited in many branches of quantum information processing including quantum computation and quantum teleportation [@preskill; @chuangbook]. EPR states of a pair of weakly coupled nuclear spins have been earlier created by NMR using spin selective pulses and evolution of coupling [@chuang00b].
Here we demonstrate the creation of EPR state on the above strongly coupled two-spin system using transition selective pulses and tomograph the result using non-selective pulses. Starting from $\vert 00 \rangle$ pseudopure state, the EPR state $(\vert 00 \rangle + \vert 11 \rangle)/\sqrt{2}$ can be created by applying the pulse sequence (pulses are to be applied from left to right) $$\left [
\left (\frac{\pi}{2} \right )^k_{\phi_1}
\cdot
(\pi)^l_{\phi_2}
\right ],
\label{eprpulex}$$ where $k$ and $l$ are the transitions, and $\phi_1$ and $\phi_2$ are the phases as shown in Table 2. For example, the unitary operator U for the pulse sequence $\left [
\left (\frac{\pi}{2} \right )^2_x
\cdot
(\pi)^3_{-x}
\right ]
$ is $$\begin{aligned}
U &=& \exp \left ( i \pi
I_{x}^{\vert 10 \rangle \leftrightarrow \vert 11 \rangle}
\right )
\cdot
\exp \left ( -i \frac{\pi}{2}
I_{x}^{\vert 00 \rangle \leftrightarrow \vert 10 \rangle}
\right ) \nonumber \\
&=& \frac{1}{\sqrt{2}} \left (
\begin{array}{cccc}
1 & 0 & -i & 0 \\
0 & \sqrt{2} & 0 & 0 \\
0 & 0 & 0 & i\sqrt{2} \\
1 & 0 & i & 0 \\
\end{array}
\right ),\end{aligned}$$ where the operators are however applied from right to left. EPR state is obtained by applying $U$ on $\vert 00 \rangle$ pseudopure state $$\begin{aligned}
U \cdot
\pmatrix{1 & 0 & 0 & 0 \cr 0 & 0 & 0 & 0 \cr 0 & 0 & 0 & 0 \cr 0 & 0 & 0 & 0}
\cdot U^{\dagger} =
\frac{1}{2}
\pmatrix{1 & 0 & 0 & 1 \cr 0 & 0 & 0 & 0 \cr 0 & 0 & 0 & 0 \cr 1 & 0 & 0 & 1}\end{aligned}$$ A phase-cycle over different combinations given in Table 2 helps to reduce errors in the off-diagonal elements. Figure 4(a) shows the experimental equilibrium spectrum and 4(b) shows the experimental spectrum after creating the EPR state. Since EPR state does not consist of any single quantum coherence, no signal is obtained (Figures 4(b)). The corresponding simulated spectrum are also shown on the right hand side in each figure.
To verify the creation of the EPR state it is necessary to tomograph the complete density matrix. Tomography in NMR is normally carried out using spin-selective pulses obtaining a series of one-dimensional NMR experiments each giving a linear equation of different elements of the density matrix [@chuang98]. However, in the case of strongly coupled systems, all operations including tomography excludes the use of spin selective pulses and demands either non-selective pulses or transition selective pulses or both. Recently, a robust method for tomography was suggested based on two-dimensional Fourier spectroscopy, which utilizes only non-selective pulses [@ranabirtomo]. This method involves:\
(i) an one-dimensional experiment for measuring diagonal elements: $ [ G_z - 10^\circ_x-t]$, and\
(ii) a two-dimensional multiple quantum experiment for measuring all off-diagonal elements: $[t_1-(\pi/2)_y-G_z-(\pi/4)_{-y}-t_2]$. $$\begin{aligned}
\end{aligned}$$ The scheme involves only non-selective pulses and therefore it is not only simple and accurate but also applicable to strongly coupled systems. The result of measurement of diagonal elements of the EPR state \[experiment (i)\] is shown in Figure 4(c) and the corresponding simulated spectrum is shown in Figure 4(h). Figure 5 shows the complete pulse-sequence for creation of EPR state followed by measurement of the off-diagonal elements \[experiment (ii)\]. The resulting 2D spectrum of experiment is given in 4(k), which clearly shows the double quantum peaks corresponding to the EPR state. No zero quantum or single quantum peaks are observed. [@ranabirtomo].
Since the diagonal and off-diagonal terms are measured by two different schemes, it is necessary also to determine the scaling between the two measurements. Normally this is achieved by comparing single quantum terms in the two-dimensional experiment with the spectrum obtained by direct detection of the density matrix [@ranabirtomo]. However since no single quantum coherence is present in the EPR density matrix, we carried out two additional one-dimensional experiments: (iii)\[ $(\pi/4)_y-t$\] and (iv) \[$(\pi/4)_x-t$\] after creation of EPR pair. Experimental spectra corresponding to (iii) and (iv) are shown in Figures 4(d) and 4(e) respectively, and corresponding simulated spectra are shown in Figures 4(i) and 4(j) respectively. The signals in 4(d) are proportional to the sum of the amplitudes of diagonal and double quantum coherences of the EPR state while those in 4(e) are proportional to the differences (when single quantum coherences are not present as in the present case). Since in a perfect EPR state the diagonal elements and double quantum coherences are equal, the spectrum of Fig 4(e) should have no signal as is evident from the simulated spectrum of Fig. 4(j). The signals of Fig. 4(e) compared to Fig 4(d) are measures of experimental errors, which in the present case are estimated to be less than 15$\%$ [@mahthesis]. The complete density matrices corresponding to the theoretical and experimentally obtained EPR state are shown in Figures 4(l) and 4(m).
Implementation of Logic gates
-----------------------------
Logic gates have been implemented earlier by one and two-dimensional NMR using weakly coupled spin-1/2 nuclei as well as quadrupolar nuclei [@cory98; @chuang97b; @ernst; @mahesh01; @neeraj01]. We demonstrate here the first implementation of a complete set of 24 one-to-one logic gates in a 2-qubit system using strongly coupled spin-1/2 nuclei. The system chosen for this purpose is the two phosphorus nuclei of the organometallic compound (I) shown in Fig. 6(a). The energy level diagram (Fig. 6(b)) and the equilibrium phosphorus spectrum of this molecule in isotropic liquid state is given in Fig. 6(c). Starting from equilibrium, the logic gates were implemented using sequences of transition selective pulses. The final populations were mapped by a small angle $(10^o)$ non-selective pulse. The spectra corresponding to final populations of all the 24 logic gates are given in Fig 7. The unitary transforms and pulse sequences for implementation of these gates are given in reference [@mahesh01] with the modification that the r.f. power has been adjusted for given angle of flip for the two inner versus the two outer transitions.
Three-spin strongly coupled system {#3spinstrong}
==================================
The system and labeling of transitions
--------------------------------------
The system chosen is the three strongly coupled protons of 3-bromo-1,2-dichlorobenzene (Figure 8) oriented in the nematic liquid crystal ZLI-1132. The equilibrium spectrum of the system at 300 K obtained from DRX 500 MHz spectrometer is shown in Figure 8. There are only nine out of a total 15 possible single quantum transitions with observable intensity, in this spin system. Construction of energy level diagram and labeling of transitions were performed using a Z-COSY experiment [@zcosy1; @zcosy2]. The Z-COSY spectrum along with cross-sections parallel to $\omega_2$ axis at various transitions is given in Figure 9. The zero-quantum artifacts were suppressed in the Z-COSY experiment by incrementing a delay synchronized with t$_1$ increment [@zcosy1; @zcosy2]. The connectivity matrix is obtained by the MATLAB assisted automation $$\begin{aligned}
\begin{array}{r|c}
&
\begin{array}{rrrrrrrr}
\;\;\;\;1\;\;\; & 2\;\;\; & 3\;\;\; & 4\;\;\; & 5\;\;\; & 6\;\;\; & 7\;\;\; & 8\;\;\;
\end{array}
\\
\hline
\\
\begin{array}{c}
1 \\
2 \\
3 \\
4 \\
5 \\
6 \\
7 \\
8
\end{array}
&
\left [
\begin{array}{rrrrrrrr}
0 & 0 & -1 & 1 & 0 & 0 & 1 & 0 \\
0 &0 & 1 & -1 & -1 & 0 & 0 & 1 \\
-1 & 1 &0 & 0 & 1 & 0 & 0 & 0 \\
1 & -1 & 0 &0 & 0 & 0 & -1 & 1 \\
0 & -1 & 1 & 0 &0 & 1 & -1 & 0 \\
0 & 0 & 0 & 0 & 1 &0 & 1 & -1\\
1 & 0 & 0 & -1 & -1 & 1 &0 & 0 \\
0 & 1 & 0 & 1 & 0 & -1 & 0 & 0
\end{array}
\right ]
\end{array}.\end{aligned}$$ The constructed energy level diagram for the above connectivity matrix is shown in Figure 10. The ninth transition shown by dashed line belongs to the transition 011$\leftrightarrow$100, and is not connected to any other observed transitions. Therefore the transition did not show any connectivity to other transitions in Z-COSY experiment (Figure 9) and is marked as \* in Figure 8. It turns out that these nine transitions are sufficient to carry out certain QIP operations as shown in section (B),(C) and (D).
Preparation of Pseudopure state
-------------------------------
We have used the method of “POPS" to prepare a pair of pseudopure states on this three-spin strongly coupled system [@fung00]. POPS requires only two population distributions: (i) Equilibrium populations (Fig. 11(a)) and (ii) Equilibrium populations changed by a single transition selective $\pi$ pulse on a given transition (Fig. 11(b)). Subtraction of (ii) from (i) yields effectively a pair of pseudopure states $\vert 000 \rangle \langle 000 \vert - \vert 001 \rangle \langle 001 \vert$ (Fig. 11(c)).
Creation of $\vert {\sf GHZ} \rangle \langle
{\sf GHZ} \vert - \vert 001 \rangle \langle 001 \vert$ state {#ghzcreation}
------------------------------------------------------------
Entanglement between many particles is essential for most quantum communication schemes, including error-correction schemes and secret key distribution network [@bouwmeester]. Greenberger-Horne-Zeilinger (GHZ) states are three-spin entangled states of the form [@ghz1; @ghz2] $$\vert {\sf GHZ} \rangle = \frac{1}{\sqrt{2}}
\left(
\vert 000 \rangle + \vert 111 \rangle
\right )$$ Three particles in GHZ state exhibit one of the strangest correlations that can not be explained by any hidden variable theory [@ghz1; @ghz2]. A set of measurements carried out on three particles in GHZ state prepare the particles in a classically impossible correlated state [@bouwmeester]. In NMR, the GHZ state was first created by Laflamme et al [@laflamme98]. The correlations of the GHZ state has been studied using NMR by Nelson et al [@nelson00].
The preparation of GHZ state requires preparing a pseudopure initial state, like $\vert 000 \rangle$. However, since we have prepared pairs of pseudopure states as the initial state, we will be actually preparing a state $$\vert {\sf GHZ} \rangle \langle {\sf GHZ} \vert -
\vert 001 \rangle \langle 001 \vert.$$ This state differs from GHZ state only in the diagonal elements and therefore retains the essential correlations of the GHZ state.
The GHZ state can be created from the $\vert 000 \rangle$ pseudopure state using a cascade of three transition selective pulses (i) $(\pi/2)_{\phi_1}$ pulse on the transition [**8**]{}, (ii) $(\pi)_{\phi_2}$ pulse on the transition [**4**]{} and (iii) $(\pi)_{\phi_3}$ pulse on the transition [**1**]{}. The pulse sequence is, $$\begin{aligned}
[ \left ( \frac{\pi}{2} \right )_{\phi_1}^{(8)} \cdot
\pi_{\phi_2}^{(4)} \cdot \pi_{\phi_3}^{(1)} ],
\label{ghzpulses}\end{aligned}$$ where pulses are to be applied from left to right. The phases $\phi_1$, $\phi_2$ and $\phi_3$ of the pulses can be any one of the 16 possible combinations, as shown in Table 3. A phase-cycle over these combinations helps to reduce the errors in the off-diagonal elements.\
The unitary operator for the above pulse-sequence can be written as, $$\begin{aligned}
U_{\sf GHZ} &=&
\exp(-i \pi I^{(1)}_{\phi_3})
\cdot
\exp(-i \pi I^{(4)}_{\phi_2})
\cdot
\exp\left (-i \frac{\pi}{2}I^{(8)}_{\phi_1} \right ),\end{aligned}$$ where $I^{(1)}_{\phi_3}$, $I^{(4)}_{\phi_2}$ and $I^{(8)}_{\phi_1}$ are the single transition operators on transitions [**1**]{}, [**4**]{} and [**8**]{} respectively. Spectrum corresponding to the diagonal part of $\vert {\sf GHZ} \rangle \langle {\sf GHZ} \vert
- \vert 001 \rangle \langle 001 \vert$ state is shown in Figure 11(d). It is clearly seen that in the GHZ state (Eq. 15) transitions 1,3 and 8 have approximately half the intensity of the equilibrium spectrum. Transitions 5,6 and 7 appear due to POPS. However to confirm the creation of GHZ state a complete tomography of the created GHZ state is needed. The pulse sequence for preparation of POPS, creation of $\vert {\sf GHZ} \rangle \langle {\sf GHZ} \vert- \vert 001 \rangle \langle 001 \vert$ state followed by tomography (using experiment (ii) of expression (13)) is given in the Figure 12(a). The 2D spectrum corresponding to the measurement of all off-diagonal elements of GHZ state is shown in Figure 12(b). Presence of only triple quantum coherence and absence of all other coherences confirms the creation of GHZ state. The axial peaks at zero frequency are $\Omega_1$ dimension is due to the longitudinal relaxation during the $t_1$ period and imperfections of the $90^o$ pulse (see caption of Fig. 4).
Two-qubit DJ using two-dimensional NMR
--------------------------------------
Implementation of DJ-algorithm on two qubits requires three qubits including one work qubit. The algorithm can be described as
$$\vert r \rangle \vert s \rangle \vert t \rangle \stackrel{U_f}{\rightarrow} \vert r \rangle \vert s \rangle \vert t\oplus f(r,s) \rangle$$
$\vert r \rangle \vert s \rangle$ and $\vert t \rangle$ being the states of the two input qubits ($I_1$,$I_2$) and the work qubit ($I_0$) respectively. There are eight possible two-bit binary functions (f) of which two are constant and six balanced. The transformations corresponding to the constant functions $f_1$ and $f_2$ are respectively unity operator and a $\pi$ pulse on all the transitions of the work qubit. The unitary transformations encoding the remaining six balanced functions $f_3-f_8$ are acheived by transition selective pulses on different transitions (3,4,9(\*),6) of the work qubit as $[0,0,\pi,\pi]$,$[\pi,\pi,0,0]$,$[\pi,0,\pi,0]$,$[0,\pi,0,\pi]$, $[\pi,0,0,\pi]$ and $[0,\pi,\pi,0]$;[@mahesh01].\
The pulse sequence for two-dimensional DJ algorithm is $[(\frac{\pi}{2})^{I_0 ,I_1 ,I_2} - t_1 - U_f - Det (t_2)]$. The transitions of $I_1$ and $I_2$ qubits are frequency labelled during the $t_1$ period and detected during the $t_2$ period. Fourier transformation with respect to $t_1$ and $t_2$ yields the desired two-dimensional spectrum [@mahesh01].\
The experimental result of the above operations on the strongly coupled 3-qubit system of 3-bromo-1,2-dichloro benzene is given in Figure 13. The experimental results match the expected theoretical results, confirming that two-dimensional DJ algorithm can also be carried out in a strongly coupled three spin system.
Strongly coupled 4-spin system {#4qstrong}
==============================
Labeling of transitions
-----------------------
The system chosen is 2-chloroiodobenzene dissolved in ZLI-1132. The four aromatic protons form a strongly coupled 4-spin system. Equilibrium spectrum is shown in Figure 14. The transitions are labeled according to the descending order of their intensities. The total number of transitions for a 4-spin strongly coupled system is $^{2\times4}C_{4-1} = 56$ of which only 30 transitions have been observed with sufficient intensity. There are other transitions of smaller intensity comparable to that of $^{13}$C satellites. In order to avoid any interference due to $^{13}$C satellites we have used $^{13}$C decoupling in the experiment. Decoupling in $t_1$ dimension is achieved by a single $\pi$ pulse on $^{13}$C channel in the middle of $t_1$ period while that in $t_2$ dimension is carried out by multi-pulse decoupling sequences on the $^{13}$C channel. The Z-COSY spectrum obtained after $^{13}$C decoupling for the present system is shown in Figure 15. The spectrum consists of more than 2000 peaks. The MATLAB analysis of the spectrum is carried out. Figure 16(a) shows the connectivity and labeling of the 30 transitions of Figure 14.
4-Qubit gates and pseudopure states {#4qstrongcnot}
-----------------------------------
Figure 16 shows implementation of gates and preparation of pairs of pseudopure states using the 4-qubit system 2-chloroiodobenzene. The labeling scheme for the energy levels is shown in Figure 16(a). Figure 16(b) shows the equilibrium spectrum. A C$^3$-[NOT]{} gate can be implemented using a single $\pi$ pulse on the transition 4, The spectrum corresponding to the C$^3$-[NOT]{} gate obtained using a small angle detection pulse is shown in Figure 16(c). A pair of pseudopure states, namely, $\vert 1111 \rangle \langle 1111 \vert - \vert 1110 \rangle \langle 1110 \vert$ is prepared by subtracting the spectrum 16(c) from the equilibrium spectrum shown in 16(b). The resultant spectrum is shown in Figure 16(d). Similarly the pair of pseudopure states $\vert 1110 \rangle \langle 1110 \vert - \vert 1010 \rangle \langle 1010 \vert$ (Figure 16(e)) is prepared by inverting the transition 1 and subtracting the obtained spectrum from the equilibrium spectrum. Figure 16(f) demonstrates the implementation of C$^2$-[SWAP]{} gate after preparing the pair of pseudopure states $\vert 1110 \rangle \langle 1110 \vert
- \vert 1010 \rangle \langle 1010 \vert$. The action of C$^2$-[SWAP]{} gate is to interchange the states $\vert 1110 \rangle$ and $\vert 1101 \rangle$. This is achieved by three transition selective pulses $$\begin{aligned}
\left [
\pi^{(4)} \cdot \pi^{(14)} \cdot \pi^{(4)}
\right ],
\label{ccswapex}\end{aligned}$$ where the superscripts indicate the transition numbers. Since the states $\vert 1101 \rangle$ and $\vert 1110 \rangle$ have almost same populations in the present case, the spectrum after applying C$^2$-[SWAP]{} gate on equilibrium input state will not be very much different from the equilibrium spectrum. However, if one starts with a pair of pseudopure states $\vert 1110 \rangle \langle 1110 \vert
- \vert 1010 \rangle \langle 1010 \vert$ as the input, the output will be different pair of pseudopure states, $$\begin{aligned}
\vert 1110 \rangle \langle 1110 \vert
- \vert 1010 \rangle \langle 1010 \vert
\stackrel{{\mathrm C}^2-{\sf SWAP}}{\longrightarrow}
\vert 1101 \rangle \langle 1101 \vert
- \vert 1010 \rangle \langle 1010 \vert.\end{aligned}$$ The spectrum corresponding to the state $\vert 1101 \rangle \langle 1101 \vert
- \vert 1010 \rangle \langle 1010 \vert$ obtained by applying C$^2$-[SWAP]{} gate on $\vert 1110 \rangle \langle 1110 \vert
- \vert 1010 \rangle \langle 1010 \vert$ is shown in Figure 16(f). The pair of pseudopure state $\vert 1101 \rangle \langle 1101 \vert
- \vert 1010 \rangle \langle 1010 \vert$ can also be prepared by inverting the transition 15 and subtracting the spectrum obtained from the equilibrium spectrum, as shown in Figure 16(g). The spectra in figures 16(f) and 16(g) match fairly well, indicating good implementation of the C$^2$-[SWAP]{} in Figure 16(f).
Conclusions
===========
Increasing the number of qubits in NMR calls for the use of the dipolar couplings in oriented homonuclear systems which are generally strongly coupled. Spin-selective pulses are not defined in the case of strongly coupled systems and qubit-addressability in such a scenario is achieved through transition selective pulses. It has been demonstrated earlier on weakly coupled systems, that using only non-selective pulses and transition-selective pulses one can implement logic gates [@kavita00; @mahesh01; @cory98b], and algorithms such as Grover’s algorithm and Quantum Fourier Transform [@ranajmr]. Efforts are ongoing to implement these algorithms in such systems as well as to realize higher qubit systems using nuclear spins oriented in liquid crystal matrices.
Acknowledgments
===============
The authors thank Dr. Swadhin K. Mandal and Prof. S.S. Krishnamurthy of the Department of Inorganic and Physical Chemistry, Indian Institute of Science, for the organometallic compound (I). The use of DRX-500 NMR spectrometer funded by the Department of Science and Technology, New Delhi, at the Sophisticated Instuments Facility, Indian Institute of Science, Bangalore, is also gratefully acknowledged.
[99]{} P.W.Shor, [*Polynomial-time algorithms for prime factorization and discrete algorithms on quantum computer*]{}, SIAM Rev. [**41**]{}, 303-332 (1999). L.K. Grover, [*Quantum Mechanics helps in searching for a needle in a haystack*]{}, Phys. Rev. Lett. [**79**]{}, (1997) 325. J. Preskill, [*Lecture notes for Physics 229: Quantum information and Computation*]{}, http://theory.caltech.edu/people/preskill/. M.A. Nielsen and I.L. Chuang, [*Quantum Computation and Quantum Information*]{}, Cambridge University Press 2000. D. Bouwmeester, A. Ekert and A. Zeilinger, [*The Physics of Quantum Information*]{}, Springer, 2000. R.R. Ersnt, G. Bodenhausen, and A. Wokaun, [*Principles of Nuclear Magnetic Resonance in One and Two Dimensions*]{}, Oxford university press 1987. C.P. Slichter, [*Principles of Magnetic Resonance*]{}, Harper and Row, New York, 1978. P. Dhiel and C.L. Khetrapal, [*NMR-Basic Principles and Progress*]{}, Springer-Verlog, New York, 1996, Vol. 1, 1995. C.S. Yannoni, M.H. Sherwood, D.C. Miller, I.L. Chuang, L.M.K. Vandersypen and M.G. Kubinec, [*Nuclear magnetic resonance quantum computing using liquid crystal solvents*]{}, Appl. Phys. Lett. [**75**]{}, 3563-3565 (1999). M. Marjanska, I.L. Chuang and M.G. Kubinec, [*Demonstration of quantum logic gates in liquid crystal nuclear magnetic resonance*]{}, J. Chem. Phys. [**112**]{}, 5095 (2000). B. M. Fung, [*Using pair of pseudopure states for NMR quantum computing*]{}, Phys. Rev. A [**63**]{}, 022304 (2001). T.S. Mahesh, Neeraj Sinha, K.V. Ramanathan and Anil Kumar, [*Ensemble quantum-information processing by NMR: Implementation of gates and creation of pseudopure states using dipolar coupled spins as qubits*]{}, Phys. Rev. A. [**66**]{}, 022313 (2002). Z.L. Madi, R. Bruschweiler and R.R. Ernst, [*One- and two-dimensional ensemble quantum computing in Spin Liouville space*]{}, J. Chem. Phys. [**109**]{}, 10603 (1998). A.K. Khitrin and B.M. Fung, [*Nuclear magnetic resonance quantum logic gates using quadrupolar nuclei*]{} J. Chem. Phys. [**112**]{}, 6963 (2000). A. Khitrin, H. Sun, and B.M. Fung, [*Method of multifrequency excitation for creating pseudopure states for NMR quantum computing*]{}, Phys. Rev. A [**63**]{}, 020301(R) (2001). Neeraj Sinha, T.S. Mahesh, K.V. Ramanathan and Anil Kumar, [*Toward quantum information processing by nuclear magnetic resonance: Pseudopure states and logical operations using selective pulses on an oriented spin-3/2 nucleus*]{}, J. Chem. Phys. [**114**]{}, 4415 (2001). K.V.R.M. Murali, Neeraj Sinha, T.S. Mahesh, Malcom Levitt, K.V. Ramanathan, and Anil Kumar, [*Quantum information processing by nuclear magnetic resonance: Experimental implementation of half-adder and subtractor operations using an oriented spin-7/2 system*]{}, Phys. Rev. A [**66**]{}, 022313 (2002). Ranabir Das and Anil Kumar, [*Use of Quadrupolar Nuclei for Quantum Information processing by Nuclear Magnetic Resonance: Implementation of a Quantum Algorithm*]{}, Phys. Rev. A. (accepted). J. A. Jones, [*NMR quantum computation: a critical evaluation*]{}, LANL Archive quant-ph/0002085 D.G. Cory, A.E. Dunlop, T.F. Havel, S.S. Somaroo and W. Zhang, [*The effective Hamiltonian of the Pound-Overhauser controlled-NOT gate*]{}, LANL Archive: quant-ph/9809045. S.C. Benjamin, [*Simple pulses for universal quantum computation with a Heisenberg ABAB chain*]{}, Phys. Rev. A. [**64**]{}, 054303 (2001). D.G. Cory, A.F. Fahmy and T.F. Havel, [*Ensemble quantum computing by NMR spectroscopy*]{}, Proc.Natl.Acad.Sci. USA [**94**]{}, 1634 (1997). N. Gershenfeld and I.L. Chuang, [*Bulk spin-resonance quantum computation*]{}, Science [**275**]{}, 350 (1997). D.G. Cory, M.D. Price and T.F. Havel, [*Nuclear magnetic resonance spectroscopy: An experimentally accessible paradigm for quantum computing*]{}, Physica D [**120**]{}, 82 (1998). E. Knill, I.L. Chuang and R. Laflamme, [*Effective pure states for bulk quantum computation*]{}, Phys. Rev. A. [**57**]{}, 3348 (1998). I.L. Chuang, N. Gershenfeld, M.G. Kubines and D.W. Leung, [*Bulk quantum computation with nuclear magnetic resonance*]{} Proc.Roy.Soc.Lond. A [**454**]{}, 447-467 (1998). Kavita Dorai, Arvind and Anil Kumar, [*Implementing quantum-logic operations, pseudopure states, and the Deutsch-Jozsa algorithm using noncommuting selective pulses in NMR*]{},Phys. Rev. A. [**61**]{}, 042306 (2000). T.S. Mahesh and Anil Kumar, [*Ensemble quantum-information processing by NMR: Spatially averaged logical labeling technique for creating pseudopure states*]{}, Phys. Rev. A. [**64**]{}, 012307 (2001). E. Knill, R. Laflamme, R. Martinez and C.H. Tseng, [*An algorithmic benchmark for quantum information processing*]{} Nature, [**404**]{}, 368 (2000). X. Peng, X. Zhu, X. Fang, M. Feng, K. Gao and M. Liu, [*Preparation of pseudopure states by line selective pulses in nuclear magnetic resonanace*]{}, Phys. Rev. A [**65**]{}, 042315 (2002). T.S. Mahesh, [*Quantum Information Processing by Nuclear Magnetic Resonance using transition-selective pulses*]{} PhD thesis, Indian Institue of Science, 2003 (unpublished). D. Deutsch and R. Jozsa, [*Rapid solution of problems by quantum computation*]{}, Proc. R. Soc. Lond. A [**400**]{}, 97 (1985). R.Cleve, A. Ekert, C. Macchiavello, and M. Mosca, [*Quantum algorithms revisited*]{}, Proc. Roy. Soc. Lond. A [**454**]{}, 339 (1998).
I.L.Chuang, L. M. K. Vanderspyen, X. Zhou, D.W. Leung, and S. Llyod, [*Experimental realization of a Quantum algorithm*]{}, Nature (London) [**393**]{}, 1443 (1998). J.A. Jones and M. Mosca, [*Implementation of a quantum algorithm on a nuclear magnetic resonance quantum computer*]{} J. Chem. Phys. [**109**]{}, 1648 (1998). N. Lindan, H. Barjat, and R. Freeman, [*An implementation of Deutsch-Jozsa algorithm on a three-qubit NMR quantum computer*]{}, Chem. Phys. Lett. [**296**]{}, 61 (1998). Ranabir Das, T.S. Mahesh and Anil Kumar, [*Efficient quantum state tomography for quantum information processing using a two-dimensional Fourier transform technique*]{}, Phys. Rev. A. [**67**]{}, 062304 (2003). M.A. Nielsen and I.L. Chuang,[*Programable Quantum Gate Arrays*]{}, Phys. Rev. Lett. [**79**]{}, 321 (1997). T.S. Mahesh, Kavita Dorai, Arvind and Anil Kumar, [*Implementing logic gates and the Deutsch-Jozsa quantum alorithm by two-dimensional NMR using spin- and transition-selective pulses*]{}, J. Magn. Reson. [**148**]{}, 95 (2001). M. rance, G. Bodenhausen, G. Wagner, K. Wuthrich and R.R. Ernst, [*A systematic appraoch to the suppression of J cross peaks in 2D exchange and 2D NOE spectroscopy*]{}, J.Mag. Reson. [**62**]{}, 497 (1985). R.C.R. Grace and Anil Kumar, [*Flip-angle dependence of nonoequilibrium states yielding information on connectivity of transitions and energy levels of oriented molecules*]{}, J. Mag. Reson. [**99**]{}, 81 (1992). R. Laflamme, E. Knill, W.Zurek, P. Catasti and S. Mariappan, [*NMR GHZ*]{}, Philos. Trans. R. Soc. London, Ser. A [**356**]{} 1941 (1998). Richard J. Nelson, D.G. Cory and Seth Llyod, [*Experimental demonstration of Greenberger-Horne-Zeilinger correlations using nuclear magnetic resonanace*]{} Phys. Rev. A [**61**]{}, 022106 (2000). D.M. Greenberger, M. Horne and A. Zeilinger, [*Bell’s theoram, Quantum theory and Conceptions of the Universe*]{}, pp.73-76 (Kluwer, Dordrecht, 1989). D.M. Greenberger, M. Horne, A. Shimony and A. Zeilinger, [*Bell’s theoram without inequalities*]{}, Am. J. Phys. [**58**]{}, 1131 (1990). Ranabir Das, T.S. Mahesh and Anil Kumar, [*Implementation of Conditional Phase Shift Gates for quantum information processing by NMR, using transition selective pulses*]{}, J. Magn. Reson. [**159**]{}, 46 (2002).
\
\
Table 1. Functions, unitary operators and r.f. pulses for one-qubit DJ algorithm.\
------- -------------- ----------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------- -------------------------------------------------------------
$f_1$ $f_2$ $f_3$ $f_4$
0 $0$ $1$ $0$ $1$
1 $0$ $1$ $1$ $0$
$U_f$ [$\pmatrix{0&1&0&0 \cr 1&0&0&0 \cr 0&0&0&1 \cr 0&0&1&0}$]{} [$\pmatrix{0&1&0&0 \cr 1&0&0&0 \cr 0&0&1&0 \cr 0&0&0&1}$]{}
Pulse $no~~ pulse$ $(\pi)^{\vert 00 \rangle \leftrightarrow \vert 01 \rangle}-(\pi)^{\vert 10 \rangle \leftrightarrow \vert 11 \rangle}$ $(\pi)^{\vert 10 \rangle \leftrightarrow \vert 11 \rangle}$ $(\pi)^{\vert 00 \rangle \leftrightarrow \vert 01 \rangle}$
------- -------------- ----------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------- -------------------------------------------------------------
Table 2. Transition numbers of expression 10 and corresponding phases of the r.f. pulses for creating $\vert 00 \rangle + \vert 11 \rangle$ EPR state. The transition numbers are from Figure 1(b).\
------- ----- ----- ---------- ---------- ------- ----- ----- ---------- ----------
Expt. $k$ $l$ $\phi_1$ $\phi_2$ Expt. $k$ $l$ $\phi_1$ $\phi_2$
No. No.
1 2 3 x -x 5 4 1 x -x
2 2 3 -x x 6 4 1 -x x
3 2 3 y y 7 4 1 y y
4 2 3 -y -y 8 4 1 -y -y
------- ----- ----- ---------- ---------- ------- ----- ----- ---------- ----------
Table 3. The 16-step phase-cycle for preparing the GHZ state. The pulse-sequence is given in expression \[ghzpulses\].\
$\phi_1$ $\phi_2$ $\phi_3$ $\phi_1$ $\phi_2$ $\phi_3$ $\phi_1$ $\phi_2$ $\phi_3$ $\phi_1$ $\phi_2$ $\phi_3$
----- ---------- ---------- ---------- ------ ---------- ---------- ---------- ------- ---------- ---------- ---------- ------- ---------- ---------- ----------
1 y y y 5 y x -x 9 x y -x 13 x x -y
2 y -y -y 6 y -x x 10 x -y x 14 x -x y
3 -y y -y 7 -y x x 11 -x y x 15 -x x y
4 -y -y y 8 -y -x -x 12 -x -y -x 16 -x -x -y
[**FIGURE CAPTIONS**]{}
Figure 1. (a) Eigenstates of a strongly coupled two-spin system, (b) qubit labeling, (c) trisodiumcitrate yielding a strongly coupled two-spin system, and (d) the equilibrium 500MHz $^1$H spectrum of (c).\
\
Figure 2. (a) Equilibrium spectrum of trisodiumcitrate, and spectra corresponding to various pseudopure states (b) $\vert 00 \rangle$, (c) $\vert 01 \rangle$, (d) $\vert 10 \rangle$ and (e) $\vert 11 \rangle$. Numbers above the enrgy levels indicate populations and the binary numbers below the levels indicate the labels. Transitions are also indentified as 1, 2 , 3 and 4 in (a). The pulse sequences applied to prepare each pseudopure state is explained in the text. Transition selective pulses used were of length 100 ms and the gradient pulse was of length 1 ms and strength 10 G/cm. Each spectrum was obtained by a non-selective high power pulse of duration 1 $\mu$s corresponding to a flip-angle of 10$^{\circ}$.\
\
Figure 3. Implementation of Deutsch-Jozsa algorithm. (a) Quantum circuit of and (b) experimental scheme for the implementation of DJ algorithm. (c) and (d) are the experimental spectra corresponding to the two constant functions of $U_1$ and $U_2$ respectively, and (e) and (f) corresponding to the two balanced functions of $U_3$ and $U_4$ respectively. The expected spectrum for all the four functions are given on the right hand side.\
\
Figure 4. Creation and tomography of $(\vert 00 \rangle + \vert 11 \rangle)/\sqrt2$ EPR state. Spectra in (a-e) are experimental and in (f-g) are simulated corresponding to various steps of creation and tomography. (a,f) The equilibrium spectra, (b,g) spectra after creation EPR state (no single quantum signal observed); (c,h) the spectra of diagonal part measured by the pulse sequence $(EPR)-G_z-10^{\circ}_x$; (d,i) spectra obtained by the sequence $(EPR)-(\pi/4)_x-10^\circ_x$, (e,j) spectra obtained by $(EPR)-(\pi/4)_y-10^\circ_x$. (k) The two dimensional spectrum to measure the off-diagonal elements. This spectrum clearly shows the double quantum coherence present in the EPR state. The axial peaks at zero frequency originate from the longitudinal relaxation of EPR state during $t_1$ period which is detected due to the imperfection of the $90^o$ r.f. pulse following the $t_1$ period. The theoretical (l) and, and experimental (m) density matrices of EPR state. Spectra in (d,e) are used for calculating the scaling between diagonal and off-diagonal measurements. While plotting, the spectra shown in (d,e,i,j) are scaled up by a factor of 4. Pseudopure state and EPR state are created by using transition selective pulses of length 100 ms. An eight step cycle (shown in Table 2) was employed to minimize the errors.\
\
Figure 5. Pulse sequence for creation and tomography of EPR state. Numbers inside the parenthesis indicate transition numbers as shown in Figure 1(d). ${\sf U_{EPR}}$ is applied on transitions \[(2),(3)\] or \[(4),(1)\] as described in the expression \[eprpulex\] and Table 2. $G_1$, $G_2$ and $G_3$ are field gradient pulses of different strengths along $\hat{z}$-direction.\
\
Figure 6. (a) Organometallic compound (I) in which the two Phosphorus ($^{31}$P) nuclei constitute a two spin system. (b) energy level diagram of the two spin system and (c) the equilibrium phosphorus spectrum recorded at 202 MHz in a magnetic field of 11.7 Tesla.\
\
Figure 7. Implementation of 24 one-to-one logic gates. Starting from equilibrium all the gates were implemented using sequences of transition selective pulses and non-selective pulses. The unitary transforms and pulse sequences for implementation of these gates are given in reference [@mahesh01]. Gaussian shaped pulses of 100ms duration were used as selective pulses. The r.f. power has been calibrated for given angle of flip for the two inner versus the two outer transitions. A sine-bell shaped gradient was applied after implementation of each selective pulse, to kill any coherences created due to imperfection of pulses. The final populations were mapped by a small angle $(10^o)$ non-selective pulse.\
\
Figure 8. Equilibrium proton spectrum of 3-bromo-1,2-dichlorobenzene oriented in ZLI-1132 at 500 MHz. The transitions are labeled from left to right. The ninth transition marked \* did not show connectivity to other transitions (Figure 9), and belongs to the lone transition between 011 and 100 (markedby dashed line in Figure 10).\
\
Figure 9. Z-COSY spectrum of oriented 3-bromo-1,2-dichlorobenzene. The equilibrium spectrum and the cross-sections of the Z-COSY spectrum are shown on the right-hand side.\
\
Figure 10. Energy level diagram of oriented 3-bromo-1,2-dichlorobenzene constructed using the Z-COSY spectrum shown in Figure 9. The transitions are labeled as in Figure 8. Only nine transitions are assigned. The remaining transitions are having very low intensity.\
\
Figure 11. Preparation of pseudopure states on the three spin strongly coupled system of Figure 8. Energy levels, transitions and representative deviation populations (numbers inside the circles) are shown on the left hand side and the corresponding spectra are shown on the right hand side. (a) Equilibrium deviation populations and the corresponding spectrum, (b) deviation populations and spectrum obtained after inverting the transition [**6**]{}, and (c) the deviation populations and the spectrum obtained by subtracting (b) from (a). The deviation populations and the spectrum in (c) correspond to the pair of pseudopure states: $\vert 000 \rangle \langle 000 \vert -
\vert 001 \rangle \langle 001 \vert$. (d) Population distribution and spectrum corresponding to the state $\vert {\sf GHZ} \rangle \langle {\sf GHZ} \vert- \vert 001 \rangle \langle 001 \vert$. All spectra were recorded using a final small angle $(10^o)$ detection pulse to maintain linear response such that the intensities are proportional to the population differences of the two involved levels only.\
\
Figure 12. (a) Pulse sequence for creation and tomography of $\vert {\sf GHZ} \rangle \langle {\sf GHZ} \vert
- \vert 001 \rangle \langle 001 \vert$ state. Numbers inside the parenthesis indicate transition numbers as shown in Figure 8. $G_1$ and $G_2$ are field gradient pulses of different strengths along $\hat{z}$-direction. (b) The two-dimensional spectrum obtained by using the pulse sequence (a). Pure triple quantum coherence at $\omega_1+\omega_4+\omega_8$ (where, $\omega_k$ is the frequency of the transition $k$) confirms the creation of GHZ state.\
\
Figure 13. The result of DJ algorithm on 3-bromo-1,2-dichlorobenzene desolved in ZLI-1132 for various functions $f_1-f_8$. Only the expansions of the transitions of the input qubits ($I_1$ and $I_2$) are shown. Expected patterns obtained by GAMMA simulation are also shown against each spectra. Transitions 3,4,9(\*) and 6 are used as the four work qubit transitions. Interchange of labels 011$\leftrightarrow$101 allows us to identify transitions 4 and 9 as transitions belonging to the 3rd qubit. This does not affect other operations and the four 3rd qubit transitions namely 3,4,9 and 6 remain unconnected. However transition 1 now belongs to 1st qubit (011$\leftrightarrow$111) along with 2 (010$\leftrightarrow$110) and transition 7 to 2nd qubit (001$\leftrightarrow$011) along with 8 (000$\leftrightarrow$010). All the experiments were carried out on a Bruker DRX-500 spectrometer at 300K. The transition selective pulses were 1.5,7.4,20 and 1.5 ms long respectively for transitions 3,4,9 and 6. The pulse power was then adjusted to make the flip angle of each pulse as $\pi$. A phase cycle of $(x,-x)$ was used to minimize the error of the $\pi$-pulses during computation. The extra peaks in $f_2$ (shown by “$\leftarrow$" marks) in the experimental spectra were originated due to undesired coherence transfer during computation. All the experiments were done using 2048 $t_2$ and 128 $t_1$ data points. All plots are shown in magnitude mode. The resonance frequencies of various transitions (1,2,7 and 8) in $\omega_2$ domain are schematically identified in the bottom line of the Figure. The $\pi$ pulses applied to various work qubit transitions are indicated for each f, with the transitions identified in $f_2$. The same order follows for other f’s. For example, $f_2=(\pi,\pi,\pi,\pi)$ means $\pi$ pulses are applied to all the transitions of work qubit and $f_5=(\pi,0,\pi,0)$ means $\pi$ pulses are applied to transitions 3 and 9 and no pulses to transitions 4 and 6.\
\
Figure 14. One dimensional 500 MHz proton spectrum of 2-chloroiodobenzene oriented in liquid crystal ZLI-1132, at 300 K forming a 4-qubit system. The transitions are labeled according to descending order of their intensities.\
\
Figure 15. Z-COSY spectrum of the 4-spin strongly coupled system shown in Figure 13. The spectrum consists of more than 2000 desired peaks.\
\
Figure 16. (a) Labeling scheme for the states of the 4-qubit system of Figure 13. (b) Equilibrium spectrum obtained using a small angle (10$^o$) pulse \[the spectra of Fig. 13 was obtained using a 90$^o$ detection pulse\]. (c) Spectrum corresponding to C$^3$-[NOT]{} gate obtained by selective inversion of transition number 4 ($\vert 1110 \rangle \leftrightarrow \vert 1111 \rangle$). (d) Spectrum corresponding to the pair of pseudopure states $\vert 1111 \rangle \langle 1111 \vert
- \vert 1110 \rangle \langle 1110 \vert$ obtained by subtracting (c) from (b), named POPS(4). (e) Spectrum corresponding to the pair of pseudopure states $\vert 1110 \rangle \langle 1110 \vert
- \vert 1010 \rangle \langle 1010 \vert$ obtained by inverting the transition 1 and subtracting the obtained spectrum from the equilibrium spectrum (b), named POPS(1). (f) Spectrum corresponding the pair of pseudopure states $\vert 1101 \rangle \langle 1101 \vert - \vert 1010 \rangle \langle 1010 \vert$ obtained by applying C$^2$-[SWAP]{} gate on (e). The pulse sequence for C$^2$-[SWAP]{} gate is given in the expression \[ccswapex\]. (g) Spectrum corresponding the pair of pseudopure states $\vert 1101 \rangle \langle 1101 \vert
- \vert 1010 \rangle \langle 1010 \vert$ obtained by inverting the transition 15 and subtracting the spectrum for the equilibrium spectrum (b), named POPS(15). Spectra (f) and (g) match fairly well, indicating good implementation of the C$^2$-[SWAP]{}. All spectra were recorded using a final small angle $(10^o)$ detection pulse to maintain linear response such that the intensities are proportional to the population differences of the two involved levels only.
\[chp5enlevels\]
[Figure 1]{}
\[chp5pps\]
[Figure 2]{}
[Figure 3]{}
\[chp5epr\]
[Figure 4]{}
\[eprpulseq\]
[Figure 5]{}
[Figure 6]{}
\[phos\]
[Figure 7]{}
\[cl2br1d\]
[Figure 8]{}
\[cl2brzcosy\]
[Figure 9]{}
\[cl2brenlev\]
[Figure 10]{}
\[cl2brpps\]
[Figure 11]{}
[Figure 12]{}
\[3qdj2d\]
[Figure 13]{}
\[cli1dlabeling\]
[Figure 14]{}
\[cli4qzcosy\]
[Figure 15]{}
\[4qgatesstrong\]
[Figure 16]{}
[^1]: Present address: [*Francis Bitter Magnet Lab, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA*]{}\
$^{2}$ Present address: [*Department of Chemistry, Iowa state University, Iowa 50011-3111, USA*]{}\
$^*$For correspondence (e-mail:[*[email protected]*]{})
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we propose and analyze a new fractional order malaria model with temporary immunity. We extend the existing malaria model to include fractional derivatives. Basic reproduction number is computed using the next generation matrix method. Model equilibria are determined and their local asymptotic stability analysis are considered. An efficient Adams-type predictor-corrector method is used to solve the initial value problem. Finally, phase plane portraits are generated to describe the dynamics between susceptible fractions of human hosts and infected fractions of mosquito vectors.'
author:
- 'E. Okyere, F. T. Oduro, S. K. Amponsah, I. K. Dontwi, N. K. Frempong'
bibliography:
- 'sample1.bib'
date: 'March, 2016'
title: Fractional Order Malaria Model With Temporary Immunity
---
Introduction
============
Malaria is one of the deadly vector-borne diseases affecting the developing part of the world and mostly in Africa. The reports by the World Health Organization [see, e.g, @WHO2005; @WHO2012; @WHO2014] indicates that, malaria is a serious threat to human life and remains a dangerous infectious disease. Its economic burden on the affected countries are extremely huge and needs serious public health attention. Since its discovery, many researchers have studied the disease dynamics mathematically to capture and understand the complex interaction between infected mosquitoes and the susceptible human population.\
Mathematical modeling of infectious diseases has played a key role in understanding the dynamical processes of disease transmission and control strategies. The first malaria epidemiological model was constructed and formulated by [@Ross1911]. He used deterministic compartmental models to describe malaria infection dynamics between human and mosquito populations. [@shu2000] proposed a malaria model and considered global stability analysis for the disease free equilibrium point. [@Mandal2011] considered an extensive survey and review on modeling of malaria infection. Backward bifurcation analysis is very important in mathematical modeling of infectious diseases and has been investigated by many authors in malaria infection models [see, e.g, @Keegan2013; @Gosh2013; @Chiyaka2008; @Chitnis2005; @Chitnis2006; @Ngonghala2012]. [@Aguas2012] developed an age structure model to study the effects of relapse in malaria parasites. A periodic model based on the Ross-MacDonald model was considered by [@Gao2014]. They studied the effects of spatial heterogeneity and temporal on disease dynamics. Their proposed model incorporated periodic behavior in mosquito ecology and seasonal human movement to capture variations of malaria spread among different regions.\
[@Tumwiine2010] proposed and analyzed a malaria infection model with infective immigrants. [@Tumwiine2007b] formulated an epidemic model with standard incidence for the dynamics of malaria infection. They applied additive compound matrices approach to show global stability analysis of the endemic equilibrium. Using the model in [@Tumwiine2007b], the same authors [@Tumwiine2007a] studied the oscillatory behavior of malaria disease in a population with temporary immunity. Their numerical results shows that the endemic equilibrium converges to a steady state. [@Okosun2011] developed a deterministic compartmental model to investigate the impact of drug resistance in malaria infection and the same authors [@makinde2014] formulated an optimal control model for malaria and cholera co-infection.\
[@Abdullahi2013] have assessed the effectiveness of drugs in a malaria transmission model. [@Barley2012], [@Mukandavire2009] and [@Abu-Raddad2006] formulated a model to study infectious disease co-infection of HIV and malaria. [@Okosun2013] studied the spread of malaria infection with an optimal control model. [@Dondrop2008], they studied the spread of anti-malaria resistance. [@lawi2011] studied co-infection model for malaria and meningitis among children. They observed that, when the threshold parameter $R_{0}<1$, the disease-free equilibrium might exhibit some instabilities in the global sense. They further applied centre manifold theorem to investigate asymptotic stability of the endemic equilibrium. [@Anita2008] studied the dynamics of acute malaria transmission using an age-structured models. A mathematical model for malaria infection which incorporates weather has been formulated and studied by [@Moshe2004]. [@Koella2003] developed an epidemic model to explore the incidence of anti-malarial resistance.\
Mathematical modeling of infectious diseases dynamics and other important areas of studies such as economics, finance and engineering has extensively been explored using the theory and applications of the classical differential equations. But in recent times, the theory and applications of fractional calculus has become extremely useful and important in modeling of biological processes and other areas of studies due to the memory property of fractional derivatives. Many authors have contributed significantly in compartmental modeling of infectious disease dynamics using fractional differential equations. [@wow7] have developed a mosquito-transmitted disease model with fractional differential equations. [@wow17] described fractional order models for malaria infection. They modified the integer order compartmental model formulated by [@Chiyaka2008] to numerically study control strategies in malaria dynamics. Fractional order models for HIV and TB co-infection have been analysed by [@Pinto2014]. Fractional order models for HIV infection have been considered and analysed by several authors [see, e.g, @wow24; @wow25; @wow22; @wow15; @wow12; @wow13; @Liu2014]. [@wow21] constructed fractional order competition model for love triangle. [@wow20] proposed a mathematical model for love using fractional order systems. [@wow16] investigated backward bifurcation in a fractional order vaccination model.\
[@Javidi2014] developed fractional order model for cholera infection. [@Arafa2012] applied homotopy analysis method and fourth order Runge-Kutta method to generate analytical and numerical results for fractional order childhood disease models. An epidemic model with fractional order for influenza A have been formulated and analysed [see, e.g, @El-Shahed2011; @Gonzalez2014]. [@wow23] studied deterministic fractional order SIRC model with Salmonella bacteria infection. [@wow14] presented and discussed fractional order systems for Hepatitis C virus. The deadly ebola disease which killed a lot of people in some part of Africa has been modelled by a system of fractional order derivatives [@wow19]. An endemic model with constant population has been analytically and numerically studied by [@Okyere2016a] using Caputo fractional derivatives.\
In this work, we formulate a fractional order model for malaria infection with temporary immunity. We consider the malaria disease model studied and analysed by [@Tumwiine2007a; @Tumwiine2007b]. The authors developed their model using deterministic integer order differential equations. Our new model will be constructed using fractional order derivatives. We are motivated by this method of mathematical formulation in epidemiological modeling due to the effective nature of fractional derivatives.\
The rest of the article is as follows. In section 2, we formulate the fractional order malaria model. Section 3, deals with non-negativity of model solutions. In section 4, we compute model equilibria and the basic reproduction number and then investigate local asymptotic stability analysis. We consider numerical simulations in section 5. In section 6, we discuss simulation results and finally conclude the paper in section 7.
Model Derivation
================
In this section, we extend and modify the integer order malaria model by [@Tumwiine2007a; @Tumwiine2007b] to become fractional order malaria model. The theory and applications of fractional calculus has become extremely useful and important in modeling of biological processes and other areas of studies due to the memory property of fractional derivatives. Many authors have contributed significantly in compartmental modeling of infectious disease dynamics using fractional differential equations. There are several definitions of fractional derivatives and integrals in fractional calculus [see, e.g, @wow2; @wow], but in this work, our model construction will be based on Caputo derivative. Our motivation for this type of fractional derivative is based on the fact that, it has an advantage on initial value problems.
[@wow; @wow2] Fractional integral of order $\alpha$ is defined as $$I^{\alpha}g(t)=\frac{1}{\Gamma(\alpha)}\int_0^t \! \frac{g(x)}{(t-x)^{1-\alpha}} \, \mathrm{d}x$$
for $0<\alpha<1,\ t>0.$
[@wow; @wow2] Caputo fractional derivative is defined as $$D^{\alpha}g(t)=\frac{1}{\Gamma(k-\alpha)}\int_0^t \! \frac{g^{k}(x)}{(t-x)^{\alpha+1-k}} \, \mathrm{d}x.$$
for $k-1<\alpha<k.$\
In this article, we consider the scaled malaria model model presented by\
[@Tumwiine2007a; @Tumwiine2007b]. The authors, transformed their compartmental malaria model ($S_H I_H R_H- S_V I_V$) into systems of proportions ($s_h i_h r_h- s_v i_v$) for the human and mosquito populations describe by systems of ordinary differential equations.
$$\label{SIRmalaria1}
\begin{array}{lllll}
\displaystyle \frac{d s_{h}}{dt} &=& \displaystyle {\lambda_{h}(1-s_{h})-abm s_h i_v+\nu i_{h}+\gamma r_h+\delta s_h i_h} ,\qquad \\[15pt]
\displaystyle \frac{d i_{h}}{dt} &=& \displaystyle {abm s_h i_v-(\nu +r+\lambda_{h}+\delta )i_h+\delta i^2_{h}}, \qquad \\[15pt]
\displaystyle \frac{d{r_{h}}}{dt} &=& \displaystyle {r i_{h}-(\gamma+\lambda_{h})r_h+\delta i_h r_h}, \qquad \\[15pt]
\displaystyle \frac{d s_{h}}{dt} &=& \displaystyle {\lambda_{v}(1-s_v)-ac i_h s_v} ,\qquad \\[15pt]
\displaystyle \frac{d{i_{v}}}{dt} &=& \displaystyle {ac s_v i_h-\lambda_{v}i_v}, \qquad
\end{array}$$
with $s_h(t)+i_h(t)+r_h (t)=1$ and $s_v(t)+i_v(t)=1$,\
where
$s_h$: proportion of susceptible human hosts at time $t$
$i_h$: proportion of infected human hosts at time $t$
$r_h$: proportion of immune human hosts at time $t$
$s_v$: proportion susceptible mosquito vectors at time $t$
$i_v$: proportion of infected mosquito vectors at time $t$\
\
$a$: average daily biting rate on man by a single mosquito (infection rate)
$b$: proportion of bites on man that produce an infection
$\nu$ recovery rate of human hosts from the disease
$c$: probability that a mosquito becomes infectious
$\gamma$: per capita rate of loss of immunity in human hosts
$r$: rate at which human hosts acquire immunity
$\delta$: per capita death rate of infected human hosts due to the disease
$m=\frac{N_V}{N_H}$: number of female mosquitoes per human host
$N_H$: total human population
$N_V$: total mosquito population
$\mu_h$: per capita natural death rate of humans
$\mu_v$: per capita natural death rate of mosquitoes
$\lambda_{h}$: per capita natural birth rate of humans
$\lambda_{v}$: per capita natural birth rate of mosquitoes
The new model we consider in this paper is the system of non-linear fractional differential equations in the sense of Caputo fractional derivatives.
$$\label{SIRmalaria2}
\begin{array}{lllll}
\displaystyle D^{\alpha}s_h &=& \displaystyle {\lambda^{\alpha}_{h}(1-s_{h})-a^{\alpha}bm s_h i_v+\nu^{\alpha} i_{h}+\gamma^{\alpha} r_h+\delta^{\alpha} s_h i_h} ,\qquad \\[15pt]
\displaystyle D^{\alpha}i_h &=& \displaystyle {a^{\alpha}bm s_h i_v-(\nu^{\alpha} +r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha} )i_h+\delta^{\alpha} i^2_{h}}, \qquad \\[15pt]
\displaystyle D^{\alpha}r_h &=& \displaystyle {r^{\alpha}i_{h}-(\gamma^{\alpha}+\lambda^{\alpha}_{h})r_h+\delta^{\alpha} i_h r_h}, \qquad \\[15pt]
\displaystyle D^{\alpha}s_v&=& \displaystyle {\lambda^{\alpha}_{v}(1-s_v)-a^{\alpha}c i_h s_v} ,\qquad \\[15pt]
\displaystyle D^{\alpha}i_v &=& \displaystyle {a^{\alpha}c s_v i_h-\lambda^{\alpha}_{v}i_v}, \qquad
\end{array}$$
with $s_h(t)+i_h(t)+r_h (t)=1$ and $s_v(t)+i_v(t)=1$.\
It is important to remark that when the fractional order $\alpha\rightarrow 1$, the fractional order malaria model (\[SIRmalaria2\]) with the model restrictions $s_h(t)+i_h(t)+r_h (t)=1$ and $s_v(t)+i_v(t)=1$, becomes the integer order malaria model (\[SIRmalaria1\]) studied by [@Tumwiine2007a; @Tumwiine2007b].
Non-negative Solutions
======================
Let $\mathbf{R^{5}_{+}}=\{X\in\mathbf{R^{5}}:X\geq 0\}$, where $X=(s_h, i_h, r_h, s_v, i_v )^T$. We apply the following Lemma in [@wow8] to show the theorem about the non-negative solutions of the model.
[@wow8]\[lema\] (Generalized Mean Value Theorem).\
Suppose that $w(x)\in C[a, b]$ and $D^{\alpha}w(x)\in C(a, b]$ for $0<\alpha\leq 1$, then we have $$w(x)=w(a)+\frac{1}{\Gamma (\alpha)}D^{\alpha}w(\xi)(x-a)^{\alpha}$$ with $a\leq \xi\leq x, \ \forall x\in (a, b].$
Assume that $w(x)\in C[a,b]$ and $D^{\alpha}w(x)\in C(a,b]$, for $0<\alpha\leq 1$. It follows from Lemma \[lema\] that if $D^{\alpha}w(x)\geq 0, \forall x\in (a,b)$ then $w(x)$ is non-decreasing $\forall x\in [a,b]$ and if $D^{\alpha}w(x)\leq 0, \forall x\in (a, b)$, then $w(x)$ is non-increasing $\forall x\in [a,b].$
The malaria disease model (\[SIRmalaria2\]) has a unique solution and it remains in $\mathbf{R^{5}_{+}}$.
By applying Theorem $3.1$ and Remark $3.2$ in [@wow10], the existence and uniqueness of the solution of equation (\[SIRmalaria2\]) in $(0, \infty)$ follows. The domain $\mathbf{R^{5}_{+}}$ for the model problem is positively invariant, since
$$\label{SIRmalaria2a}
\begin{array}{lllll}
\displaystyle D^{\alpha}s_h|_{s_h=0}&=& \displaystyle {\lambda^{\alpha}_{h}+\nu^{\alpha} i_{h}+\gamma^{\alpha} r_h\geq 0} ,\qquad \\[15pt]
\displaystyle D^{\alpha}i_h|_{i_h=0} &=& \displaystyle {a^{\alpha}bm s_h i_v\geq 0}, \qquad \\[15pt]
\displaystyle D^{\alpha}r_h|_{r_h=0} &=& \displaystyle {r^{\alpha}i_{h}\geq 0}, \qquad \\[15pt]
\displaystyle D^{\alpha}s_v|_{s_v=0}&=& \displaystyle {\lambda^{\alpha}_{v}\geq 0} ,\qquad \\[15pt]
\displaystyle D^{\alpha}i_v|_{i_v=0} &=& \displaystyle {a^{\alpha}c s_v i_h\geq 0}, \qquad
\end{array}$$
on each hyperplane bounding the non-negative orthant, the vector filed points into $\mathbf{R^{5}_{+}}$.
Analysis of Model Equilibria
============================
In order to determine the model equilibria, we proceed this way.\
Let
$$\label{eqm}
\begin{cases}
D^{\alpha}s_h=0 &\\
D^{\alpha}i_h=0 & \\
D^{\alpha}r_h=0 &\\
D^{\alpha}s_v=0 & \\
D^{\alpha}i_v=0
\end{cases}$$
By solving equation (\[eqm\]) and ignoring the steps involved, the disease-free equilibrium is given by $F_o=(1, 0, 0, 1, 0)$ and the endemic equilibrium as\
$F_{*}=(s^{*}_{h}, i^{*}_{h}, r^{*}_{h}, s^{*}_{v}, i^{*}_{v})$, where $s^{*}_{h}, r^{*}_{h}, s^{*}_{v}$ and $i^{*}_{v}$ are expressed in terms $i^{*}_{h}$ as follows:
$$\label{SIRmalaria3}
\begin{array}{lllll}
\displaystyle s^{*}_{h} &=& \displaystyle {\dfrac{(\lambda^{\alpha}_{v}+a^{\alpha}c i^{*}_{h})[(\lambda^{\alpha}_{h}+\gamma^{\alpha}-\delta^{\alpha}i^{*}_{h})(\lambda^{\alpha}_{h}+\nu i^{*}_{h})+\gamma^{\alpha}r^{\alpha} i^{*}_{h}]}{(\lambda^{\alpha}_{h}+\gamma^{\alpha}-\delta^{\alpha}i^{*}_{h})[(\lambda^{\alpha}_{h}-\delta^{\alpha }i^{*}_{h})(\lambda^{\alpha}_{v}+a^{\alpha}c i^{*}_{h})+a^{2\alpha}bmci^{*}_{h}]}} ,\qquad \\[15pt]
\displaystyle r^{*}_{h} &=& \displaystyle {\dfrac{r^{\alpha}i^{*}_{h}}{(\lambda^{\alpha}_{h}+\gamma^{\alpha}-\delta^{\alpha}i^{*}_{h})}}, \qquad \\[15pt]
\displaystyle s^{*}_{v} &=& \displaystyle {\dfrac{\lambda^{\alpha}_{v}}{(\lambda^{\alpha}_{v}+a^{\alpha}c i^{*}_{h})}} ,\qquad \\[15pt]
\displaystyle i^{*}_{v}&=& \displaystyle {\dfrac{a^{\alpha}c i^{*}_{h}}{(\lambda^{\alpha}_{v}+a^{\alpha}c i^{*}_{h})}}, \qquad
\end{array}$$
Next, we compute the basic reproduction number $R_o$ using the next generation method on the fractional order malaria model (\[SIRmalaria2\]). By using the method and the notations by [@van2002], the special matrices $F$( for new infections terms) and $V$ (for the remaining transitions terms) related to the model problem (\[SIRmalaria2\]) are respectfully given by $$F=\begin{bmatrix}
0 & & & a^{\alpha}bm \\ \\
a^{\alpha}c & & &0
\end{bmatrix}$$
and
$$V=\begin{bmatrix}
\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha}& & & 0 \\ \\
0 & & &\lambda^{\alpha}_{v}
\end{bmatrix}$$
The basic reproduction number is the spectral radius of the generation matrix $F{V^{-1}}$, where\
$$F{V^{-1}}=\begin{bmatrix}
0& & &\frac{a^{\alpha}bm}{\lambda^{\alpha}_{v}} \\ \\
\dfrac{a^{\alpha}c}{\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha}}& & &0
\end{bmatrix}$$
Therefore we have\
$R_{o}=\sqrt{\dfrac{a^{2\alpha}bmc}{{\lambda^{\alpha}_{v}}(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha})}}$
Local Asymptotic Stability Analysis of Disease-Free Equilibrium
---------------------------------------------------------------
The equilibrium $F_{o}$ of the malaria disease model (\[SIRmalaria2\]) is locally asymptotically stable if $R_o<1$ and unstable if $R_o>1.$
The Jacobian matrix $J(F_{o})$ for the malaria disease model (\[SIRmalaria2\]) evaluated at $F_{o}$ is given by
$$\label{dfe}
J(F_{o})=\begin{bmatrix}
-\lambda^{\alpha}_h & \nu^{\alpha}+\delta^{\alpha} & \gamma^{\alpha} & 0 & -a^{\alpha}bm \\
0 & -(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}) & 0 & 0 & a^{\alpha}bm \\
0& r^{\alpha}&-(\lambda^{\alpha}_h+\gamma^{\alpha}) &0 & 0 \\
0 &-a^{\alpha}c & 0 & -\lambda{\alpha}_{v} & 0\\
0 &a^{\alpha}c & 0 & 0 & -\lambda^{\alpha}_{v}\\
\end{bmatrix}$$
The disease-free equilibrium point $F_o$ is locally asymptotically if all the eigenvalues $\lambda_{i}, i=1, 2, 3, 4, 5$ of $J(F_o)$ satisfy the following condition \[[@wow5], [@wow6]\]: $ \abs{arg(\lambda_{i })} >\frac{\alpha \pi}{2}$.\
From the square matrix $J(F_{o})$, it is obvious that, three of the eigenvalues are given by $-\lambda^{\alpha}_h, -(\lambda^{\alpha}_h+\gamma^{\alpha})$ and $-\lambda^{\alpha}_{v}$. It can be seen that all the three eigenvalues are negative and hence satisfy the stability condition: $ \abs{arg(\lambda_{i })} >\frac{\alpha \pi}{2}$.\
The rest of the eigenvalues can determined from the sub-matrix given as:
$$\label{dmt}
B=\begin{bmatrix}
-(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}) & & & a^{\alpha}bm \\ \\
a^{\alpha}c & & &-\lambda^{\alpha}_{v}
\end{bmatrix}$$
The trace of matrix $B$ is given by $$tr(B)=-(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}+\lambda^{\alpha}_{v})<0$$
The determinant of $B$ yields :
$$\begin{aligned}
det B &= \lambda^{\alpha}_{v}(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha})-a^{2\alpha}bmc \\ \\
&= \lambda^{\alpha}_{v}(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha})\left[1-\dfrac{a^{2\alpha}bmc}{\lambda^{\alpha}_{v}(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha})} \ \right]\\ \\
&=\lambda^{\alpha}_{v}(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha})\left[1-R^2_{o} \right]\end{aligned}$$
Therefore $det B >0 $ if $R_0<1$.\
Since $det B >0 $ if $R_0<1$ and $tr(B)<0$, it follows that the two remaining eigenvalues will have negative real parts and we can therefore conclude that, the disease equilibrium is locally asymptotically stable, otherwise unstable.
Local Asymptotic Stability Analysis of Endemic Equilibrium
----------------------------------------------------------
In this subsection, we discuss the local stability of the endemic equilibrium point. Our analysis is based on the construction by [@Ahmed2006; @wow5]. From the model constraints $s_v+i_v=1$ and $s_h+i_h+r_h=1$, we can reduce the dimension of the malaria model in order to simplify and reduce the complexity of the analysis.\
As in [@Tumwiine2007a; @Tumwiine2007b], let $s_v=1-i_v$ and $r_h=1-s_h-i_h$ then the 5-dimensional fractional order model (\[SIRmalaria2\]) reduces to a 3-dimensional problem with Jacobian matrix evaluated the at endemic equilibrium $F_{*}$ as:
$$J(F_{*})=\begin{bmatrix}
-(\lambda^{\alpha}_h +\gamma^{\alpha}+a^{\alpha}bmi^{*}_{v}-\delta^{\alpha}i^{*}_{h})& \nu^{\alpha}+\delta^{\alpha}s^{*}_{h}-\gamma^{\alpha} &- a^{\alpha}bms^{*}_{h}\\ \\
a^{\alpha}bmi^{*}_{v} & -(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}-2\delta^{\alpha} i^{*}_{h}) & a^{\alpha}bm s^{*}_{h} \\ \\
0& a^{\alpha}c(1-i^{*}_v)&-(\lambda^{\alpha}_v+a^{\alpha}ci^{*}_h)
\end{bmatrix}$$
The characteristics equation of the Jacobian matrix, $J(F_{*})$ is $$\label{cequation}
\lambda^{3}+b_1\lambda^{2}+b_2\lambda+b3=0$$
where
$$\begin{split}
b_1= & \lambda^{\alpha}_{v}+a^{\alpha}ci^{*}_{h}+ \lambda^{\alpha}_{h}+\gamma^{\alpha}_{v}+a^{\alpha}bmi^{*}_{v}-\delta^{\alpha}i^{*}_{h}+\nu^{\alpha}+r^{\alpha}+ \lambda^{\alpha}_{h}+\delta^{\alpha}-2\delta^{\alpha}i^{*}_{h}\\
\\
b_2= & (\lambda^{\alpha}_v+a^{\alpha}ci^{*}_h)(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}-2\delta^{\alpha} i^{*}_{h})-a^{\alpha}bmc(1-i^{*}_v)s^{*}_h+(\gamma^{\alpha}-\nu^{\alpha}-\delta^{\alpha}s^{*}_h)a^{\alpha}bmi^{*}_v\\
& +(\lambda^{\alpha}_{h}+\gamma^{\alpha} +a^{\alpha}bmi^{*}_v -\delta^{\alpha} i^{*}_{h})(\lambda^{\alpha}_v+a^{\alpha}ci^{*}_h+\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}-2\delta^{\alpha} i^{*}_{h})\\ \\
b_3= & (\lambda^{\alpha}_{h}+\gamma^{\alpha}+a^{\alpha}bmi^{*}_{v}-\delta^{\alpha}i^{*}_{h})[(\lambda^{\alpha}_{v}+a^{\alpha}ci^{*}_{h})(\nu^{\alpha}+r^{\alpha}+\lambda^{\alpha}_h+\delta^{\alpha}-2\delta^{\alpha} i^{*}_{h})-a^{2\alpha}bmc(1-i^{*}_v)s^{*}_{h}]\\
& +a^{3\alpha}b^{2}mc(1-i^{*}_v)s^{*}_{h}+a^{\alpha}bmi^{*}_v(\lambda^{\alpha}_{v}+a^{\alpha}ci^{*}_h)(\gamma^{\alpha}-\nu^{\alpha}-\delta^{\alpha}s^{*}_{h})
\end{split}$$
Let $D(g)$ represent the discriminant of the polynomial function
$$g(x)=x^{3}+b_1x^{2}+b_2x+b3$$\
where $$D(g)=-\left|
\begin{array}{cccccccc}
1 & &b_1 && b_2 & b_3& &0\\
0 & &1& &b_1 & b_2 &&b_3\\
3 & &2 b_1& &b_2 & 0&&0\\
0 & &3& &2b_1&b_2 &&0\\
0 & &0& &3 &2b_1&&b_2
\end{array}
\right|=18b_{1}b_{2}b_{3}+(b_1 b_2)^2-4b_3 b^{3}_{1}-4b^{3}_{2}-27b^{2}_{3}$$
By following the results in [@Ahmed2006; @wow5] and similar applications of their method in [@El-Shahed2011; @OZLAP2011; @wow40], we obtain the following proposition.
One assume that $F_{*}$ exist in $\mathbf{R^{3}_{+}}$.\
(i) If the discriminant of $g(x)$, $D(g)$ is positive and Routh-Hurwitz conditions are satisfied , i.e. $D(g)>0,\ b_1>0, \ b_3>0$, then the endemic equilibrium $F_{*}$ locally asymptotically stable.
\(ii) If $D(g)<0,\ b_1\geq 0,\ b_2\geq0, \ b_3>0, \ \alpha<2/3$, then the endemic equilibrium $F_{*}$ locally asymptotically stable.
(iii)If $D(g)<0,\ b_1>0,\ b_2>0, \ b_1b_2=b_3, \ \alpha\in (0,1),$ then the endemic equilibrium $F_{*}$ locally asymptotically stable.
\(iv) If $D(g)<0,\ b_1<0,\ b_2<0, \ \alpha>2/3$, then the endemic equilibrium $F_{*}$ is unstable.
Numerical Simulations
=====================
Analytical solutions of nonlinear coupled fractional order differential equations are as difficult as the classical integer order system of nonlinear equations. Therefore, there is need to consider numerical discretizations and approximations for such systems or models. Over the past years, several numerical schemes have been developed to solve fractional order models. One of the powerful approximation methods that has been considered by many authors [see, e.g, @Ahmed2007logistic; @OZLAP2011; @wow40; @sheen2015; @Okyere2016a] is the efficient Adams-type predictor-corrector method developed by [@wow3; @wow4]. [@Tao2009] have considered a rigorous error analysis of this numerical scheme.\
For our numerical approximations, we re-define the variables in the model problem (\[SIRmalaria2\]) as:
Let $s_h=X, \ i_h=Y, \ r_h=Z, \ s_v=S, \ i_v=I.$
Therefore following the approach in [@wow3; @wow4; @sheen2015; @wow40] and ignoring the details in the derivation of the scheme, we obtain the discretized system corresponding to the fractional order malaria model (\[SIRmalaria2\])as follows:
$$\begin{split}
X_{k+1}= & X_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}a_{_{j,k+1} }\biggl[\lambda^{\alpha}_{h}(1-X_j)-a^{\alpha}bm X_{j} I_j+\nu^{\alpha} Y_j+\gamma^{\alpha} Z_j+\delta^{\alpha} X_{j} Y_{j}\biggr]\\ + & \frac{1}{\Gamma (\alpha)}\biggl[\lambda^{\alpha}_{h}(1-X^{p}_{k+1})-a^{\alpha}bm X^{p}_{k+1} I^{p}_{k+1}+\nu^{\alpha} Y^{p}_{k+1}+\gamma^{\alpha} Z^{p}_{k+1}+\delta^{\alpha} X^{p}_{k+1} Y^{p}_{k+1}\biggr]\\
\\
Y_{k+1}= & Y_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}a_{_{j,k+1} }\biggl[a^{\alpha}bm X_j I_j-(\nu^{\alpha} +r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha} )Y_j-\delta^{\alpha} Y^2_{j}\biggr]\\ + & \frac{1}{\Gamma (\alpha)}\biggl[a^{\alpha}bm X^{p}_{k+1} I^{p}_{k+1}-(\nu^{\alpha} +r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha} )Y^{p}_{k+1}+\delta^{\alpha} (Y^{P}_{K+1})^2\biggr]\\ \\
Z_{k+1}= & Z_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}a_{_{j,k+1} }\biggl[r Y_{j}-(\gamma^{\alpha}+\lambda^{\alpha}_{h})Z_j+\delta Y_j Z_j\biggr] + \frac{1}{\Gamma(\alpha)}\biggl[r Y^{p}_{k+1}-(\gamma^{\alpha}+\lambda^{\alpha}_{h})Z^{p}_{k+1}+\delta Y^{p}_{k+1} Z^{p}_{k+1}\biggr]\\\\
S_{k+1}= & S_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}a_{_{j,k+1} }\biggl[\lambda^{\alpha}_{v}(1-S_J)-a^{\alpha}c Y_j S_j\biggr] + \frac{1}{\Gamma (\alpha)}\biggl[\lambda^{\alpha}_{v}(1-S^{p}_{k+1})-a^{\alpha}c Y^{p}_{k+1} S^{p}_{k+1}\biggl]\\\\
I_{k+1}= & I_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}a_{_{j,k+1}}\biggl[a^{\alpha}c S_j Y_j-\lambda^{\alpha}_{v}I_j\biggr] + \frac{1}{\Gamma (\alpha)}\biggl[a^{\alpha}c S^{p}_{k+1} Y^{p}_{k+1}-\lambda^{\alpha}_{v}I^{p}_{k+1}\biggl]
\end{split}$$
where $$\begin{split}
X^{p}_{k+1}= & X_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}b_{_{j,k+1} }\biggl[\lambda^{\alpha}_{h}(1-X_j)-a^{\alpha}bm X_{j} I_j+\nu^{\alpha} Y_j+\gamma^{\alpha} Z_j+\delta^{\alpha} X_{j} Y_{j}\biggr]\\
\\
Y^{p}_{k+1}= & Y_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}b_{_{j,k+1} }\biggl[a^{\alpha}bm X_j I_j-(\nu^{\alpha} +r^{\alpha}+\lambda^{\alpha}_{h}+\delta^{\alpha} )Y_j-\delta^{\alpha} Y^2_{j}\biggr]\\
\\
Z^{p}_{k+1}= & Z_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}b_{_{j,k+1} }\biggl[r Y_{j}-(\gamma^{\alpha}+\lambda^{\alpha}_{h})Z_j+\delta Y_j Z_j\biggr] \\\\
S^{p}_{k+1}= & S_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}b_{_{j,k+1} }\biggl[\lambda^{\alpha}_{v}(1-S_J)-a^{\alpha}c Y_j S_j\biggr] \\\\
I^{p}_{k+1}= & I_0+\frac{1}{\Gamma (\alpha)} \sum_{j=0}^{k}b_{_{j,k+1} }\biggl[a^{\alpha}c S_j Y_j-\lambda^{\alpha}_{v}I_j\biggr]
\end{split}$$
$ a_{_{j,k+1} }=\dfrac{h^{\alpha}}{\alpha (\alpha +1)}
\begin{cases}
k^{\alpha+1}-(k-\alpha)(k+1)^{\alpha} & \quad \text{if } j=0\\
(k-j+2)^{\alpha+1}+ (k-j)^{\alpha+1}-2(k-j+1)^{\alpha+1} & \quad \text{if } 1\leq j\leq k\\
1 & \quad \text{if } j=k+1
\end{cases}
$ and
$b_{_{j,k+1} }=\frac{h^{\alpha}}{\alpha}\left[(k-j+1)^{\alpha}-(k-j)^{\alpha}\right], \ \ \ 0\leq j\leq k $\
We then implement our numerical scheme in Matlab and the simulation results are discussed in the next section.
Results and Discussion
======================
In this section, we discuss the endemic trajectories of our numerical solutions with model parameter values adapted from the work by [@Tumwiine2007a] and basic reproduction number $R_o=1.5$. The simulated solutions for the scaled human population $s_h(t), \ i_h(t)$ and $r_h (t)$ are shown in Figure \[fg1\] and that of the scaled mosquito populations $s_v (t), \ i_v (t)$ are shown in Figure \[fg2\]. In Figures \[fg3\] and \[fg4\], we considered four different values of the fractional $\alpha$, $\alpha=1,\ 0.99, \ 0.95\ 0.90$. The oscillatory behavior of the fractional order malaria model (\[SIRmalaria2\]) are interesting and can be compared to the research work be [@Tumwiine2007a]. It is important to observe that when $\alpha=1$, the model problem (\[SIRmalaria2\]) is equivalent to the classical initial value problem (\[SIRmalaria2\]) and that as the fractional order $\alpha$ increases the behaviour of the fractional order model solutions approaches that of the integer order model. Phase plane portrait is displayed in Figure \[fg5\] to describe the dynamics between susceptible fractions of human hosts and infected fractions of mosquito vectors.
![Phase portrait for fractional order model with $\alpha=1, 0.99, 0.95, 0.90$.[]{data-label="fg5"}](phaseportrait.pdf)
Conclusion
==========
In this research, we have proposed and studied fractional order model for the integer order model developed and studied by [@Tumwiine2007a; @Tumwiine2007b]. We determined the basic reproduction number $R_o$ using the method and the notations by [@van2002]. We have investigated local asymptotic stability analysis of the model equilibria. Furthermore, we have numerically describe the trajectories of the model using Adams-type predictor-corrector method. We have extended the existing malaria model propose by [@Tumwiine2007a; @Tumwiine2007b] to include fractional derivatives which is very interesting and important in the field of mathematical biology.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'NASA’s Parker Solar Probe (PSP) mission is currently investigating the local plasma environment of the inner-heliosphere ($< $0.25$R_\odot$) using both [*[in-situ]{}*]{} and remote sensing instrumentation. Connecting signatures of microphysical particle heating and acceleration processes to macro-scale heliospheric structure requires sensitive measurements of electromagnetic fields over a large range of physical scales. The FIELDS instrument, which provides PSP with [*[in-situ]{}*]{} measurements of electromagnetic fields of the inner heliosphere and corona, includes a set of three vector magnetometers: two fluxgate magnetometers (MAGs), and a single inductively coupled search-coil magnetometer (SCM). Together, the three FIELDS magnetometers enable measurements of the local magnetic field with a bandwidth ranging from DC to 1 MHz. This manuscript reports on the development of a merged data set combining SCM and MAG (SCaM) measurements, enabling the highest fidelity data product with an optimal signal to noise ratio. On-ground characterization tests of complex instrumental responses and noise floors are discussed as well as application to the in-flight calibration of FIELDS data. The algorithm used on PSP/FIELDS to merge waveform observations from multiple sensors with optimal signal to noise characteristics is presented. In-flight analysis of calibrations and merging algorithm performance demonstrates a timing accuracy to well within the survey rate sample period of $\sim340 \mu s$.'
bibliography:
- 'references.bib'
title: 'A Merged Search-Coil and Fluxgate Magnetometer Data Product for Parker Solar Probe FIELDS'
---
Introduction
============
The [*[in-situ]{}*]{} measurements of coronal heating, solar wind acceleration, and energetic particle transport made by NASA’s Parker Solar Probe (PSP) will likely answer many fundamental questions relating to the heliosphere and astrophysical plasmas [@Fox2016]. The FIELDS instrument on PSP provides [*[in-situ]{}*]{} measurements of the electric and magnetic fields [@Bale2016] required to achieve the principal objectives. The measurements made by FIELDS are complemented by [*[in-situ]{}*]{} measurements of the solar wind and coronal plasma through the Solar Wind Electrons Alphas and Protons investigation; energetic particles through the Integrated Science Investigation of the Sun ; as well as white light images from the Wide-Field Imager for Solar Probe .
Accomplishing the scientific objectives of PSP requires [*[in-situ]{}*]{} observations of magnetic and electric fields over a wide bandwidth and large dynamic range. FIELDS measures electric field and potential fluctuations ranging from DC-19.2 MHz with a diverse combination of survey mode waveform, burst mode waveform, and spectral observations made by the FIELDS Radio Frequency Spectrometer , Digital Fields Board , and Time Domain Sampler (TDS) [@Bale2016]. Magnetic field instrumentation consists of a suite of three magnetometers, two vector fluxgate magnetometers (MAGs) and a single vector search-coil magnetometer (SCM) located on boom extending behind the spacecraft and within the umbra of the PSP thermal protection system (TPS).
The two fluxgate magnetometers, built at Goddard Space Flight Center (GSFC), provide vector measurements of DC and low frequency magnetic fields with a maximum survey sample (Sa) rate of $f_{svy}^{max}$=292.969 Sa/s. These low frequency measurements are accompanied by measurements from the FIELDS SCM low frequency (LF) windings sensitive from $\approx$3 Hz-20 kHz. The SCM sensor $x$-axis additionally contains a second mid-frequency (MF) winding sensitive from $\approx$10 kHz-1 MHz. During the perihelion encounters the SCM is continuously sampled by the DFB typically at $f_{svy}$=292.969 Sa/s, though a maximum rate of 18,750 Sa/s is theoretically possible [@Malaspina2016]. When possible, data is additionally acquired during the aphelion cruise phases at reduced sample rates in accordance with allowed telemetry and spacecraft operations. Burst mode waveforms for the LF or MF are sampled up to a maximum rate of 150 kSa/s by the DFB; the MF channel can be sampled at 1.92 MSa/s by the TDS and incorporated in the spectral products generated by the RFS. Generally, the DFB, RFS, and TDS, are highly configurable and generate a variety of waveform and cross and auto-spectral matrix data products at various cadences.
Observational signatures of physical processes occurring in astrophysical plasmas, such as the solar wind and corona, are commonly sensitive to properties of the mean magnetic field: e.g. electromagnetic wave vector polarizations [@PodestaGary2011a; @He2011]; magnetic compressibility [@Bale2009; @Alexandrova2013]; variance and wave-vector anisotropies [@Horbury2008; @Chen2010; @Horbury2012]; MHD Elsässer and Poynting flux [@Balogh1999; @McManus2019]; helicical signatures of turbulence at kinetic scales [@Leamon1998; @HowesQuataert2010; @Woodham2018] magnetic reconnection [@Phan2018]. Though microphysical processes in the solar wind are sensitive to the low frequency mean field behavior, they frequently occur on fast time scales with small amplitude magnetic signatures. Accomplishing the scientific objectives of PSP thus inherently requires [*[in-situ]{}*]{} observations of magnetic fields over large bandwidths and dynamic ranges. Through combining both fluxgate and search-coil measurements, FIELDS is capable of observing the [*[in-situ]{}*]{} magnetic field of the inner heliosphere with a bandwidth from DC-1 MHz and 115 dB of dynamic range [@Bale2016].
Recently, missions with multi-sensor observational suites have moved towards merged data products composed of synchronous measurements from multiple instruments, combined to employ optimal qualities of the separate sensors. used a discrete wavelet transform to merge [*[CLUSTER]{}*]{} fluxgate and search-coil data to study waves downstream of a shock [@ClusterMAG; @ClusterSTAFF]. use a similar method to study the variance anisotropy of kinetic turbulence in the sub-proton range with [*[CLUSTER]{}*]{}. Additionally, perform measurements of scaling functions of the transition from the MHD to kinetic turbulence through continuous analysis of a combination of CLUSTER fluxgate and search-coil magnetometer observations. These previous efforts to combine fluxgate and search-coil data, though made in-flight and without specific optimization towards instrument design and operation, have proved the utility and applicability of multi-rate data fusion methods in space plasma physics [@Hall1992]. More recently, programmatic efforts to perform quantitative end-to-end testing on the Multiscale Magnetospheric Mission (MMS) search-coil and digital and analog fluxgate magnetometer have enabled the development of an optimized merged data set with automated calibration pipeline[@LeContel2016; @Russell2016; @Torbert2016; @Fischer2016].
The FIELDS SCM and MAG share a master clock and were designed with partially overlapping bandwidths, enabling the combination of individual sensors into a single merged dataset. This manuscript outlines the process used to produce survey data product using merged SCM and MAG (SCaM) measurements with a spectral composition that retains an optimal signal to noise ratio. Section \[sec:cal\] provides an overview of preflight ground testing and instrument characterization, as well as inflight calibration routines. Section \[sec:merging\] presents the algorithm used to combine the FIELDS magnetometer data into a merged product with optimal signal to noise characteristics. Section \[sec:inflight\] provides a summary of our in-flight verification of the calibration and a quantitative analysis of the performance of the merged survey data, and discusses merging survey rate data with DFB burst data [@Malaspina2016].
FIELDS Calibration {#sec:cal}
==================
Fluxgate Magnetometers (MAG)
----------------------------
The two fluxgate magnetometers, designed and fabricated at NASA/GSFC, measure DC and low frequency fluctuating magnetic fields. They are placed 1.9 and 2.7 meters from the spacecraft and are respectively referred to as the inboard (MAGi) and outboard (MAGo) sensors. The heritage of the PSP/FIELDS MAGs dates to the 1960’s NASA [*[Explorer 33]{}*]{} mission [@Ness1971; @Acuna1974; @Acuna2002]. Many iterations of the instrument currently operate on both NASA heliophysics and planetary science missions [@Lepping1995; @Acuna2002; @Acuna2008; @Kletzing2013; @Connerney2015; @Connerney2017]. The PSP fluxgate magnetometers have a maximum survey mode sample rate of $f_{svy}^{max}=292.969$ Sa/s. The MAG data is typically downsampled by factors of two with anti-aliasing performed with a Bartlett filter. Generally, MAGi is run at a lower sample rate in order to meet telemetry constraints imposed by the spacecraft. The lower cadence measurements still allow for diagnosis of magnetic noise associated with spacecraft generated magnetic fields. The primary science instrument for the DC magnetic fields is MAGo, which is less sensitive to spacecraft generated fields due to its positioning on the spacecraft boom.
The complex transfer functions associated with the MAGs, shown in Figure \[merging\_resp\](a), are dominated by a single pole low-pass Butterworth filter used for anti-aliasing purposes tuned to -3 dB at the max sample rate Nyquist frequency ($f_{svy}^{Ny}\approx$146.5Hz) [@Acuna2008; @Connerney2015; @Connerney2017]. Due to the low-pass characteristics of the Butterworth transfer function, the MAGs are sensitive to the DC magnetic fields associated with the spacecraft [@Ness1970; @Ness1971; @Belcher1973]. Typically, the minimization of such fields is performed through magnetic control programs [@Ness1970; @Musmann1988]. For PSP a strict magnetic cleanliness program was followed during design and development. Once in space, driven spacecraft maneuvers are used to establish the magnetometer zero offsets relative to the ambient field e.g. [@Acuna2002; @Connerney2015]. However, similar results can be accomplished without controlled maneuvers through statistical analysis of non-compressive Alfvénic rotations in the solar wind [@Belcher1973; @Leinweber2008]. Several multi-sensor techniques to determine sensor zeros have been developed using gradiometric principles, e.g. [@Ness1971] and comparison with scalar magnitude instruments, e.g. [@Olsen2003; @Primdahl2006]. Additionally solar wind electron beams, sensitive to the mean field direction, can be used in calibrating fluxgate offsets [@Plaschke2014; @Connerney2015].
For PSP, the attitude and pointing requirements of the spacecraft preclude the use of controlled maneuvers during perihelion, spacecraft rolls (both sun-pointed and conical rotations) are thus performed before and after each perihelion encounter to establish zero levels of the spacecraft magnetic field. In between such controlled rotations, measurements of Alfvénic rotations of the solar wind magnetic field have been implemented to track variations in the spacecraft field, allowing for an estimate of zero levels during each perihelion encounter, when controlled maneuvers cannot be performed [@Leinweber2008]. Figure \[fig:rolls\] shows the offsets for MAGo over the first encounter computed using both spacecraft rolls, and higher rate estimations from Alfvénic rotations. Spacecraft housekeeping and engineering data are currently under analysis in order to model the effects of variations in the solar panel array, which change orientation over the course of the orbit, on the measured DC offsets. Gradiometric techniques are in development to further verify and monitor offset drifts during each perihelion encounter. In addition to the removal of spacecraft offsets, the vector axes of each MAG are orthogonalized using an alignment matrix determined during pre-flight testing. The alignment matrix, is determined through the process documented in and and verified by methods outlined in .
In addition to the removal of spacecraft offsets and sensor orthogonalization, the merged data set corrects gain and phase shifts associated with the analog Butterworth filter using a convolution of the MAG output with the linear time-invariant inverse filter response. When appropriate, the digital Bartlett filters used to downsample MAG measurements are additionally inverted. Though the phase shifts associated with these low pass filters are quite small (e.g. Figure \[merging\_resp\]) over the range of frequencies observed by both the MAG and SCM, the mis-alignment of the relative phase between the MAG and SCM leads to undesirable artifacts in the merged data product. Correction for complex instrumental response (i.e. filter gain and phase) is not implemented in the un-merged calibrated data.
Search-Coil Magnetometer (SCM)
------------------------------
![(a) MAGo frequency response is dominated by single pole Butterworth filter response tuned to -3 dB at the survey mode Nyquist frequency ($f_{svy}^{max}/2$). (b) SCM frequency response determined from a spectral analyzer. A 4-pole, 2-zero fit analytical fit is performed to the empirically determined function.[]{data-label="merging_resp"}](fig1)
![Magnetic field offsets for FIELDS/MAGo $x$,$y$,$ and z$ axes (panels a, b, and c) over first two encounters of PSP (respectively offset by 219.1, 355.1, -667.4 sensor counts for demonstrative purposes). Data points (+) correspond to estimates from Alfvénic rotations of the solar wind; a seven point median filter used in processing the FIELDS data (orange). Offsets determined from spacecraft rotations are shown in red (sun pointed rolls around the spacecraft $z$-axis) and blue (conical rolls performed off sensor $z$-axis). The blue + corresponds to the MAGo $x$ offset determined from Alfvenic at the time of the conical roles.[]{data-label="fig:rolls"}](fig2.jpg){width="4.5in"}
The PSP/FIELDS search-coil magnetometer was designed at Laboratoire de Physique et Chimie de l’Environnement et de l’Espace (LPC2E) in Orléans, France, and is nearly identical to the search-coils which have been manufactured and delivered for the Solar Orbiter and TARANIS missions [@SeranFergeau2005; @Maksimovic2019]. The SCM, consisting of three mutually orthogonal inductive coils mounted on the end of the spacecraft magnetometer boom, is tailored to study the magnetic fields of the inner heliosphere [@Bale2016] from 10-20 kHz. Additionally, one of the coils (sensor $x$-axis) has a secondary winding with bandwidth of 10 kHz-1 MHz.
The survey mode waveform data, typically captured at a rate of $f_{svy}^{max}=292.969$ Sa/s is sampled and processed by the FIELDS DFB [@Malaspina2016]. In addition to continuous survey mode waveform data, the DFB provides a highly configurable set of operational modes which can be modified in flight to generate a diverse set of burst waveform and spectral data products. In addition to sampling the SCM output, the DFB is designed to inject a programmable calibration signal into the SCM. The response of the SCM to the injected stimulus is captured by the DFB as well as the TDS and the instrumental transfer function can accordingly be determined in-flight.
Figure \[merging\_resp\](b) shows the gain and phase characteristics of the FIELDS SCM $x$-axis which were determined empirically on ground using a spectral analyzer. A complex rational function with 4-poles and 2-zeros: $$R(i\omega) =\frac{A_0+i\omega A_1}{(B_0-B_2\omega^2)+i\omega(B_1-B_3\omega^2)}$$ is fit to the the response using the least square estimation techniques developed by .
The inductive nature of the SCM leads to strong gain and phase shifts in the instrumental response function which must be compensated to obtain an estimate of the observed magnetic field in physical units. During pre-launch integration and testing different methods to invert the SCM frequency response were explored: convolution kernel methods and a windowed fast Fourier transform (FFT) algorithm similar to techniques used in and . Preflight Monte-Carlo simulations testing on synthetic data suggested that convolution in the time domain generated fewer spectral artifacts in the calibrated time series than a windowed FFT algorithm. Compensation filters are developed using the inverse FFT of the response function on an abscissae of 2048 frequencies, corresponding to a 2048 tap (all zero) linear time invariant (LTI) finite impulse response (FIR) filter [@OppenheimSchafer]. The filters are non-causal, such that real time merging of data (e.g. on the spacecraft) is not possible; the future development of causal FIR filters for on-board merging presents an opportunity to increase scientific returns from telemetry limited missions.
Merging {#sec:merging}
=======
Many algorithms for merging data from multiple sensors, occasionally referred to as data fusion, were initially developed in the context of radio system engineering as a method to optimize signal to noise ratios and correct for signal loss due to stochastic fluctuations impacting transmission, [@Kahn1954; @Brennan1959]. Recent research has demonstrated the applicability of data fusion in merging magnetic field measurements from multiple sensors onboard a single spacecraft: [@Alexandrova2004; @Kiyani2009; @Chen2010]. However, not until MMS was significant effort made to design sensors with synchronized timing with pre-launch end-to-end characterization of sensor performance intended to enable optimal merging of the in-flight magnetic field data [@Torbert2016; @Fischer2016]. The shared clock between the FIELDS SCM and MAGs, as well as their simultaneous continuous survey mode operation, likewise facilitates a merged SCM and MAG (SCaM) data product. In order to produce the merged SCaM data product, accurate representations for the complex frequency responses for the individual sensors are required. In addition to the individual characterization of the instruments, multiple efforts were made to inter-calibrate the sensors; however, no strict end-to-end calibration was performed as in [@Fischer2016]. Original ground testing was performed at the Acuña Test facility using FIELDS engineering model hardware; subsequent testing was performed on flight model hardware during final stages of integration onto the PSP spacecraft, verifying the instrument gain and phase characteristics. In addition to the characterization of frequency response, the merged SCaM data product ideally attains minimal noise characteristics. Accordingly, an accurate description of the individual MAG and SCM sensor noise floors is necessary.
To provide an optimal signal to noise merging coefficients, the noise floors of each instrument are assumed to be incoherent, mean zero, gaussian processes. The spectral composition of the instrumental noise was determined during ground testing. The SCM sensitivity was characterized at the magnetic test facility in Chambon-la-Forêt. In addition to the internal sensor noise, the DFB analog electronics as well as analog to digital conversion (quantization) of the SCM signal contribute to the end-to-end instrumental noise. The end-to-end noise floor of the MAGs, incorporating quantization and analog electronic noise, were determined in laboratory using measurements taken over several hours inside of a $\mu$-metal container.
The FIELDS SCaM merging procedure is designed to maintain an optimized signal to noise ratio. Each sensor observes the environmental field, which is a coherent signal between two sensors, in superposition with incoherent, zero mean noise.
$$\begin{aligned}
{B_1}=B(t) +n_1(t)\\
{B_2}=B(t) +n_2(t)\end{aligned}$$
The merged signal $B_m$ is given as a linear combination of the individual sensors, weighted by coefficients $\alpha_1$ and $\alpha_2$ which maintain an optimal signal to noise ratio.
As instrumental noise from each sensor has different spectral characteristics, we develop frequency dependent merging coefficients through consideration of the spectral representation of the linear combination of signal and noise terms $$\begin{aligned}
\tilde{B}_m(\omega)=\alpha_1\tilde{B}(\omega) +\alpha_2\tilde{B}(\omega) \\
\tilde{N}_m(\omega)=\sqrt{\alpha_1^2 \tilde{n_1}^2 +\alpha_2^2 \tilde{n_2}^2}\end{aligned}$$
where the merged noise $\tilde{N}_m(\omega)$ corresponds to the error of each signal, weighted and added in quadrature [@Kahn1954; @Brennan1959]. The condition $\alpha_1 +\alpha_2 =1$ is required such that the merged signal is equal to the coherent environmental field observed by each sensor.
Because signal amplitudes are ideally equal in either sensor, optimizing the ratio $B_m/N_m$ leads to frequency dependent solutions for $\alpha_1(\omega)$ and $\alpha_2(\omega)$ which are independent of the environmental signal, and determined by the spectral composition of the noise floors: $$\begin{aligned}
\label{eq:alpha}
\alpha_1(\omega)=\frac{\tilde{n_2}^2}{ \tilde{n_1}^2 +\tilde{n_2}^2}\\
\alpha_2(\omega)=\frac{\tilde{n_1}^2}{ \tilde{n_1}^2 +\tilde{n_2}^2}\end{aligned}$$ where $\tilde{n_1}^2$ and $\tilde{n_2}^2$ are computed as the spectral densities of the instrument noise. For FIELDS, the coefficients $\alpha_{MAG}$ and $\alpha_{SCM}$ correspond to an effective weighting in instrumental gain which preserves an optimized signal to noise ratio for the merged SCaM data product. The SCM sensor coordinate system is not initially aligned with the MAG sensor axes, accordingly a rotation matrix ${{\ensuremath{\bf{R}}}}$ is applied to bring the SCM measurements in sensor coordinates, ${{\ensuremath{\bf{B}}}}_{SCM}'$, into alignment with the MAG coordinate system, $$\begin{aligned}
{{\ensuremath{\bf{R}}}} =\left(\begin{array}{ccc}
0.8165 & -0.4082 &-0.4082\\
0.0000 & -0.7071 & 0.7071\\
-0.577 & -0.577& -0.577
\end{array}\right)
\end{aligned}$$
$${{\ensuremath{\bf{B}}}}_{SCM}={{\ensuremath{\bf{R}}}}\cdot{{\ensuremath{\bf{B}}}}_{SCM}'.$$
Adhering to an optimal signal to noise merger, spectral composition of the noise of the rotated SCM vector time series in MAG sensor coordinates is then taken as the quadrature weighted error of the SCM sensor axis noise, assuming independence in each sensor channel: i.e. $$n_{SCMx}^2(\omega)= R^2_{xx'}n^2_{SCMx'} +R^2_{xy'}n^2_{SCMy'}+R^2_{xz'}n^2_{SCMz'}.$$ The MAG orthogonalization matrix and rotation from sensor to spacecraft coordinates is approximately equal to the identity matrix such that the measured noise floor for each sensor axis is used without contribution from the other axes.
![Empirically determined noise floors of PSP FIELDS MAGo $x,y,z$ axes (blue, green, black) and the SCM low frequency $x, y, z$ axes with the DFB sampling in a high gain state (red, orange, purple) rotated into MAGo sensor coordinates.[]{data-label="merging_noise"}](fig3)
Figure \[merging\_noise\] shows empirically determined noise-floors for both the SCM and MAGo associated with mean-zero stochastic fluctuations limiting each sensors sensitivity. Merging coefficients are obtained using Equation \[eq:alpha\]. Since the empirically determined noise spectrum is continuous, a smooth weighting of the merged signals is obtained by approximating the MAG merging coefficient using a real-valued rational function of the form $$\label{eq:ratfit}\hat{\alpha}_{MAG}(f)=\frac{N(f)}{D(f)} = \frac{ \sum_0^n A_n f^n }{1 +\sum_1^m B_mf^m}.$$ Below $f_0=2$ Hz the sensitivity of the SCM drops significantly and the full MAG signal is used, i.e. $\alpha_{MAG} =1$ for $f\leq 2$ Hz. Above 2 Hz Equation \[eq:ratfit\] is applied. The coefficients and order of the fit rational function are determined using non-linear least squares fitting [@Markwardt]. Constraints are imposed on $\hat{\alpha}_{MAG}(f)$ to ensure continuity such $\hat{\alpha}_{MAG}(f_0)=1$ and $\hat{\alpha}_{MAG}'(f_0) =0$, where $f_0$=2 Hz. Figure \[merging\_fit\] shows the $n=1, m=3$ (one zero, three pole) fit for $ \hat{\alpha}_{MAG}$.
Fitting $\alpha_{MAG}(f)$ with boundary conditions at $f_0$=2 Hz decreases available degrees of two freedom such that rational functions with three or more fit parameters are required for a reasonable approximation. Figure \[merging\_fit\] shows the best fit rational function (Equation \[eq:ratfit\]) with three poles and one zero (e.g. $m=3$ and $n=1$), to the MAG merging coefficient, $\alpha_{MAG}$ with applied boundary conditions at 2 Hz. Ensuring a piecewise continuous merging coefficient requires constraining the function at 2 Hz to unity gain and a local extremum. For $f< 2$ Hz, the coefficient $\alpha_{MAG}$ is explicitly set to unity, e.g. Figure \[merging\_alpha\], while for $f > 2$ Hz, $\alpha_{MAG}=\hat{\alpha}_{MAG}$. The SCM coefficients are determined from ${{\alpha}}_{SCM} =1-{\alpha}_{MAG}$. The optimal merging coefficients $\alpha_{MAG}$ and $\alpha_{SCM}$ are shown in Figure \[merging\_alpha\](a). The the weighted instrumental noise floors, and the optimal merged SCaM noise floors are shown in Figure \[merging\_alpha\](b).
![Nonlinear least square fit of three pole, one zero rational function to $\alpha_{MAG}$ for $f>f_0$ with $f_0=2$ Hz. At 2 Hz, $\alpha_{MAG}$ is set to unity, the fit function is constrained to maintain a continuous value and first derivative, such that an extremum is obtained.[]{data-label="merging_fit"}](fig4)
![(a) Analytical approximations of MAG and SCM merging coefficients ($y$-axis shown). (b) Noise floors of MAG (blue) and SCM (red), the coefficient weighted noise floors are shown for MAG (purple) and orange (SCM). The optimal noise floor is plotted in green.[]{data-label="merging_alpha"}](fig5)
Calibration and Merger of In-Flight Survey Data {#sec:inflight}
===============================================
![Power spectra densities of observed magnetic field in the spacecraft coordinate $y$ direction from $\approx$ 1 hour interval (2018-11-05/00:00:-01:00) calculated with MAG (blue), SCM (red), and merged SCM and MAG (SCaM, orange), time series. Sensor noise floors are shown for the MAG (teal) and SCM (green).[]{data-label="merging_spect"}](fig6)
![Eight second bandpass filtered waveform near the merging crossover frequency (2-12 Hz) from PSP/FIELDS survey magnetic field data in S/C coordinates. Good phase and gain agreement is observed in between the MAG (blue) SCM (red) and merged SCaM (green) time series.[]{data-label="merging_waves"}](fig7)
During the PSP perihelion encounters, survey mode data are acquired at different cadences, typically varying with solar distance, in order to balance science objectives with telemetry constraints. When possible the FIELDS team intends to operate the instruments with a single high data rate (292.969 Sa/s) over the entire perihelion encounter period. To date, the lowest cadence survey rate during the perihelion encounter is 73.24 Sa/s.
The calibration kernel, corresponding to the inverse response of the MAG instrument response is weighted by the appropriate merging gain coefficients, $\alpha_{MAG}$ to construct the contribution from MAGo to the merged time series. This weighted time-series, subsequently undergoes calibration processes associated with orthogonalization and spacecraft field removal used in generation of public un-merged data. The SCM is similarly calibrated using the instrumental response function with gain weighted by the merge coefficients $\alpha_{SCM}$; the gain and gain phase shifts associated with digitization by the FIELDS DFB are additionally corrected for [@Malaspina2016]. Once convolved with calibration kernels, the SCM is rotated into the MAGo coordinate system. The weighted MAG time series is then interpolated onto the SCM time abscissae. Interpolation onto the SCM time tags are used to preserve the high frequency component of the SCM without introducing artifacts associated with interpolation. The time series are directly summed to generate the merged data set with optimal signal to noise ratio. The merged SCaM data is considered a level 3 (L3) data product.
Power spectra from an approximate hour long interval ($2^{20}$ samples at 292.969 Sa/s) starting 11/05/2018T00:00 during the first PSP perihelion is highlighted in Figure \[merging\_spect\] to demonstrate the results of our calibration and merging algorithm. Power spectra for each of the MAG, SCM, and SCaM time series are computed as an ensemble average of eight power spectra of $2^{17}$ samples. Figure \[merging\_spect\] additionally shows noise-floors associated with the MAG and SCM instruments. Good agreement is observed between the merged data and spectra from either individual instrument. The in-flight observed noise-floor of the MAGo is consistent with on-ground measurements. The SCM noise floor performs similarly to preflight measurements; a slight increase in the sensitivity is observed relative to ground testing which is attributable to a decrease in thermal noise in the instrument. Broadband spectral features near the crossover frequency, corresponding to coherent wave features at several Hz, are captured by both the MAG and SCM and are thus useful in analyzing the performance of the SCaM merging algorithm [@Bale2019; @Bowen2019]. Digital filters are applied to bandpass the MAG, SCM, and merged SCaM time series to between 2 and 12 Hz in order to directly compare the time series in the crossover bandwidth, without contribution from low or high frequency signals. Figure \[merging\_waves\] shows excellent qualitative agreement in phase between the three different axes.
However, in order to ensure quality of the merged SCaM data product, the calibrated, but un-merged, MAG and SCM observations must be analyzed to verify the conditions necessary for the weighted-gain merging algorithm developed in Section \[sec:merging\]: i.e. the MAG and SCM cross calibration, including time synchronization and gain matching, must be verified. Careful inspection of Figure \[merging\_waves\] shows a small gain discrepancy between the MAG and SCM amplitudes. Analysis of the gain discrepancy is required to ensure an artifact-free merged data product. Quantitative determination of the relative MAG and SCM gain calibrations is performed by separating the full day of encounter data from Nov 05, 2019 into 22 non-overlapping intervals of $2^{20}$ samples (a one hour interval where the SCM was in a low-gain state was omitted). The vector spectral density for each interval is estimated for both MAG and SCM sensors by ensemble averaging the power spectrum of 1024 non-overlapping sub intervals computed via FFT e.g.: $$S_{MAG(f)}=\langle\mathcal{F}\{B_{MAG}(t)\}\mathcal{F}^{\dagger}\{B_{MAG}(t)\}\rangle,$$ where $\mathcal{F}\{...\}$ is the Fourier transform and $\langle ...\rangle$ denotes ensemble averaging; a Blackman-Harris window is used to prevent spectral leakage. The frequency dependent gain is then obtained as $$G(f)=10\mathrm{Log}_{10} \frac{S_{SCM}}{S_{MAG}}.$$ Figure \[gain\_error\] shows the measured distribution of $G(f)$ for each vector component as well as the mean at each frequency, and the median gain error computed between 3 and 10 Hz. Systematic gain differences are measured in the $x$, $y$, and $z$ directions of -2.67, -2.50, and -2.58 dB. These values indicate that the typical SCM amplitude is approximately $75\%$ of the measured MAG signal.
Due to the relatively stable gain discrepancy in frequency and time, the SCM may be gain-matched to the MAG through multiplication of scaling factors (1.36,1.33, and 1.34) for the respective $x$, $y$,and $z$ axes. This correction is required in-order to remove artifacts associated with merging signals with un-equal amplitudes. The difference between on-ground and in-flight gain measurements are likely due to differences in the SCM operating temperature and small discrepancies caused by the matching between SCM and DFB, which unfortunately were not quantified due to lack of end-to-end calibration. Continued efforts to quantify and monitor variations in the gain-matching coefficients will be performed throughout the mission. Additionally, both gain-matched and nominal calibrations will be available for public use; though the authors stress that use of non-gain matched data may lead to artifacts in the transition between the MAG and SCM sensor ranges.
![Gain difference between SCM and MAG measured over their shared observations range ($x$, $y$, $z$ axes shown in black, blue, red). The measured distribution of gain differences is plotted as a set of points. The mean at each frequency is plotted as a sold line, while the median gain difference in each axis from 3-10 Hz is plotted as a dashed line. The average gain difference is roughly constant over this range. The feature at 7 Hz corresponds to a reaction wheel (which has a lower signal in the SCM due to the relative positioning of the sensors).[]{data-label="gain_error"}](fig8)
A quantitative determination of the accuracy of the timing between the MAG and SCM data is performed by computing the short time Fourier transform cross-spectra of each of the three combinations of signals. The short time cross spectra is defined as $$S_{12}=\langle\mathcal{F}\{B_1(t)\}\mathcal{F}^{\dagger}\{B_2(t)\}\rangle$$
![Distribution of measured phase delays between the MAG and SCM in spacecraft coordinates ($x$, $y$, $z$) as a function of frequency shown respectively in panels (a ,b, c). The solid black line shows the mean phase error at each frequency. The dashed black lines show linear phase error associated with one and two sample periods ($\Delta t \sim 340 \mu$s).[]{data-label="merging_calphase"}](fig9)
The argument of the cross spectra gives the phase delay between the two signals at a given frequency ${\textrm{arg}}(S_{12})=\textrm{tan}^{-1}\left(\frac{\textrm{Im}\{S_{12})\}}{\textrm{Re}\{S_{12}\}}\right)$. As each sensor observes the same time series, zero-phase difference should exist at each frequency between the sensors. Each of the 22 intervals on 11/05/2018 used in gain calibration are separated into 1024 sub-intervals. The sub-division allows for the calculation of 22528 individual cross-spectra. The distribution of phase delay as a function of frequency is then calculated between the MAG and SCM using the ensemble of cross-spectra.
Figure \[merging\_calphase\] shows the measured distributions of phase difference between the MAG and SCM obtained via cross spectra. The MAG and SCM are shown to be in good agreement in the cross over frequencies: at 4 Hz the mean time-delay between the MAG and SCM measurements is 190 $\mu$s, 84 $\mu$s, and 100 $\mu$s for the respective $x$,$y$, and $z$ axes; the standard deviations are 76 $\mu$s, 59 $\mu$s, and 89$\mu$s respectively. For each vector component, approximately two standard deviations of the measured ensemble fall within $1 \Delta t$ ($\sim 340 \mu$s). These results show that the phase alignment of the MAG and SCM is accurate to within a small phase error in the frequencies surrounding the cross-over point.
Analysis of the relative phase between the SCM and MAG observations verifies timing accuracy to within a single sample period. Additionally, quantification of the relative gain between the instruments allows for the empirical matching of the SCM signal to MAG levels such that a smooth transition over the sensor cross-over range is obtained. Establishing L3 calibrations for the MAG and SCM provide time-synchronized and gain matched signals such that the direct sum of the signals, with gains weighted by $\alpha_{MAG}$ and $\alpha_{SCM}$, results in an optimal signal-to-noise merged SCaM data product .
Merging DFB Burst Data
----------------------
![(a) Burst waveform from FIELDS DFB on 2018-11-05/06:33:58 (blue) with simultaneous SCaM (red) waveform data. (b) Power spectral density from burst interval on 2018-11-05/06:33:58 dashed lines show the corresponding merging coefficients for the DFB burst (blue) and SCaM (red). (c) Merged DFB and SCaM waveform. (d) Merged power spectral density. (e) Waveform data from DFB burst on 2018-11-05/06:33:58 with SCaM data. (f) Power spectral density from 2018-11-05/06:33:58 interval, spectral flattening in the survey waveform can occur either when the DFB quantization limit is reached, or when narrowband spectral features are present above survey Nyquist frequency. []{data-label="BurstCal"}](fig10)
In addition to survey waveform data, the FIELDS DFB produces high resolution burst data from SCM measurements at a maximum sample rate of $f_{brst}=150$ kSa/s. For the three low frequency SCM windings, this is significantly higher than the instrumental 3 dB roll off at $\sim 17$ kHz. The burst buffer is taken as a $N_{brst}=2^{19}$ sample waveform lasting $\sim 3.5$ seconds. Data from the SCaM product is combined with the DFB burst measurements to provide the low frequency spectral composition to contextualize the DFB bursts. The SCM transfer function is applied to the burst data using a finite impulse response calibration kernel of $M_{brst}=$16384 filter coefficients (taps); a cutoff is applied at the high frequency SCM 3 dB roll-off ( $\sim$ 17 kHz) to prevent amplification of high frequency noise. Frequencies $f< f_{brst}/M_{brst}$ (e.g. $\sim 10$ Hz for $M_{brst}=16384$ and $f_{brst}=150$ kSa/s) cannot be captured using a convolution kernel; however, the merged SCaM data is optimized to provide high signal to noise measurements in this frequency range.
Figure \[BurstCal\](a) demonstrates a comparison of the SCaM and DFB burst waveforms from a burst on 2018-11-05/06:33:58. Figure \[BurstCal\](b) shows power spectral densities for the interval. To combine low frequency spectral components with the calibrated burst data, the SCaM data is interpolated onto the DFB burst time-tags, which is rotated into the S/C coordinate system. At frequencies above $\sim 10$ Hz the SCaM data is predominantly derived from the SCM and the intrinsic noise of the DFB burst and SCaM data are thus identical; however, correcting the attenuation from the DFB anti-aliasing filters during the SCaM calibration results in the amplification of noise in the high frequency end of the survey wave-form data. The difference in noise level between the burst and SCaM data corresponds precisely to the DFB anti-aliasing filter, which is taken as the weighting coefficients, shown Figure \[BurstCal\](b,d,e) to merge the SCaM and burst data. Figure \[BurstCal\](c) shows the merged SCaM data with DFB burst waveform data; the corresponding merged power spectral density is depicted in Figure \[BurstCal\](d).
A second burst interval from 2018-11-05/18:30:50 is shown in Figure \[BurstCal\](e-f). On occasion, spectral flattening is observed in the high frequency component of the SCM and SCaM data, e.g. Figure \[BurstCal\](f). Using DFB burst waveform data, it has been determined that this effect likely occurs due to the presence of relatively large amplitude narrowband spectral noise located immediately above the survey waveform Nyquist frequency. Such flattening is also evident with the noise floor of the SCM is reached. Ongoing efforts are made to characterize narrowband spectral noise and its effect on magnetic field measurements made by FIELDS [@Bowen2019].
In-Flight Issues with SCMx Axis
-------------------------------
Since March 2019, the low-frequency $B_x$ channel of the SCM has deviated from nominal operation whenever the sensor is shaded by the TPS, e.g. during perihelion encounters. The main symptoms of this anomaly are a much higher sensitivity to periodic current surges in the SCM heater and a drop in sensitivity in the low frequency wave-form channels. This drop is equivalent to the response of an additional 1st order high-pass filter with a cutoff at 1 kHz. This sudden change in sensitivity mostly impacts measurements of $B_x$ below approximately 600 Hz. The three components of the SCM are rotated in the frame of the MAG before merging. Accordingly the anomaly in one single channel will affect all three vector components in the S/C spacecraft coordinates, making it impossible to properly merge the signals from the two instruments. By rotating the MAG measurements into the SCM frame, it is possible to merge two components. For PSP’s first encounter the full vector merged product in spacecraft and RTN coordinates will be produced. For later events, a 2D merged product will be distributed using the SCM y and z axes. Consequently, the MAG is the only remaining 3-axis measurement below approximately 150 Hz; while the SCM provides vector measurements above approximately 600 Hz. The degraded phase and amplitude response of the SCM maintains remarkable stability, which suggests that it remains possible to deliver properly calibrated data outside of the intermediate frequency range.
Conclusion
==========
This manuscript reports on the development, implementation, and performance of an algorithm to merge magnetic field observations from the PSP/FIELDS fluxgate (MAG) and search-coil (SCM) to create a merged SCM and MAG (SCaM) data product. The techniques used for PSP/FIELDS are similar to efforts made by to combing magnetic field measurements from instrumentation on MMS, which attempt to maintain optimal signal to noise characteristics. These merging algorithms have heritage from techniques developed in radio-systems engineering for linear diversity combining [@Kahn1954; @Brennan1959]. The optimal merging methods takes into account sensor design and operation of the MAG and SCM instrumentation, in order to construct a merged data product with spectral composition which smoothly transitions between the MAG at low frequencies and the SCM at high frequencies with a cross over between $\sim3-10$ Hz. In the cross-over range of frequencies, both sensors contribute significantly to the merged SCaM data product. Using in flight analysis of the calibrated FIELDS observations, we demonstrate that the MAG and SCM sensors are in systematic agreement to within a small fraction of a sample period ( $<340\mu$ s), enabling a smooth transition between dominant signal in the cross over range without phase distortion of the measured waveforms. A small deviation in gain ($\sim$ 2 dB), likely due to temperature effects, between the sensors is measured, which impacts the merging procedure and requires ongoing analysis and correction.
Additionally, the merging algorithm presented is used to combine burst data from the SCM, acquired by the FIELDS DFB at a 150 kSa/s sample rate, with survey rate data from the MAG and SCM at lower frequencies. The merged DFB burst data allows for analysis of magnetic field signals from DC to the SCM LF cutoff (17 kHz) within a single dataset. The successful merging of SCM survey rate data with DFB burst data is promising for ongoing efforts to merge burst measurements from the FIELDS TDS at 1.92 MSa/s with these lower frequency data products. Additionally, the algorithm outlined to merge magnetic field measurements serves as a starting point to merge survey and burst measurements of waveforms made by the FIELDS electric fields antennas as well as other timeseries.
In order to maintain the level three SCaM data product for continued public use, the FIELDS team intends to regularly update the merging algorithm using measured onboard frequency responses, temperatures, noise floors, MAG offsets etc.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Discretizations of the mean curvature and extrinsic curvature components are constructed on piecewise flat simplicial manifolds, giving approximations for smooth curvature values in a mostly mesh-independent way. These constructions are given in combinatoric form in terms of the extrinsic hinge angles, the intrinsic structure of the piecewise flat manifold and a choice of dual tessellation, and can be viewed as the average of $n$-volume integrals. The constructions are also independent of the manifold dimension.
[Keywords: Discrete differential geometry, Regge calculus, piecewise linear, second fundamental form, shape operator.]{}
author:
- Rory Conboye
bibliography:
- 'Ref.bib'
title: Piecewise Flat Extrinsic Curvature
---
Introduction
============
The extrinsic curvature describes the shape of a manifold due to its’ embedding in some higher dimensional manifold. For example the curvature of a smooth curve in Euclidean $2$-space ${\mathbb{E}^{2}}$ is defined as the inverse of the radius of the circle that best fits the curve at each point. Piecewise flat simplicial manifolds are generalizations to higher dimensions of piecewise linear curves, and are formed by joining Euclidean $n$-simplices (line-segments, triangles, tetrahedra) along their $(n-1)$-dimensional faces. Due to this piecewise flat nature the curvature cannot be given in the same way as smooth manifolds. Instead, the shape resulting from an embedding is determined by *hinge-angles* defined at the faces separating each pair of $n$-simplices. This is easily visualised for a piecewise linear curve in ${\mathbb{E}^{2}}$, with an obvious angle between each pair of line-segments.
This piecewise flat interpretation of the extrinsic curvature dates back to Steiner in the 1840’s [@Steiner], where the total mean extrinsic curvature of a tetrahedron is given as the sum for each edge, of $\pi$ minus the exterior dihedral angle, times the length of that edge. This approach is similar to the deficit angles around co-dimension-$2$ simplices which give the intrinsic curvature of a piecewise flat manifold, with the related total sum giving the Regge action [@Regge].
Piecewise flat manifolds can provide a discrete approximation for smooth manifolds in a number of ways, and since they are geometric objects themselves, they provide many advantages over other approximation techniques. However, while hinge angles are appropriate for piecewise flat manifolds, they do not measure curvature in the same way as smooth extrinsic curvature. What they *do* measure, is equivalent to the path integral of certain curvature components across each hinge. Since the choice of path can be ambiguous, dividing these integrals by a path length does not lead to a consistent definition.
In this paper the curvature is instead averaged over a specific collection of $n$-volumes. These $n$-volumes are given by a choice of dual tessellation and defined so as to intersect a collection of hinges appropriate to either the mean curvature at a vertex, or the curvature component orthogonal to a hinge. The $n$-volume integral is constructed using path integrals across each hinge, given in terms of the hinge angles, and divided by the $n$-volume to give an average value. This is an extension of an approach developed for *intrinsic* piecewise flat curvature [@PLCurv], which has lead to a piecewise flat Ricci flow in three dimensions.
Previous approaches for approximating smooth extrinsic curvatures have mostly been for $2$-dimesnional surfaces in ${\mathbb{E}^{3}}$. One of the more prominent of these makes use of the *cotan* formula of Pinkall and Polthier [@PP93], giving the integrated mean curvature over areas dual to vertices as a weighted sum of the edge vectors in ${\mathbb{E}^{3}}$, see for example Meyer et al. [@MeyDebSchBar03] and a discussion about convergence by Wardetzky [@WarDDG]. However as noted in [@BS07], non-Delaunay triangulations lead to negative weightings, which can cause certain issues to arise. While methods exist to adapt a given triangulation to Delaunay for $2$-surfaces, this is more difficult for higher dimensions where Delaunay triangulations also become more restrictive [@BDGintrins]. Other approaches have used weighted sums of the hinge angles bounding individual triangles [@Grin06], and area variations for families of parallel polyhedral meshes [@BobCurv10], while methods related to Regge calculus in the physics literature mostly concentrate on single hinges [@HS81; @Brewin; @KLM89ec; @TraceK].
The main results of the paper are highlighted in theorem \[thm:1.1\] below, giving the piecewise flat integrated curvature along a geodesic segment, the average mean curvature at a vertex, and the average extrinsic curvature component orthogonal to each hinge. While the choice of dual tessellation is left open for the remainder of the paper, for simplicity it is restricted to tessellations which intersect each edge at its’ midpoint below. Such tessellations include Voronoi, barycentric or the mixture area used in [@MeyDebSchBar03].
\[thm:1.1\] Take an $n$-dimensional piecewise flat manifold $S^n$, embedded in ${\mathbb{E}^{n+1}}$.
1. For a geodesic segment $\gamma$ in $S^n$, intersecting a single hinge $h$ at an angle $\theta$ to the vector orthogonal to $h$ in $S^n$, the integral of the second fundamental form $\alpha$ along $\gamma$ is $$\int_\gamma \alpha (\hat \gamma, \hat \gamma) \, \mathrm{d} s
= \cos \theta \ \epsilon_h + O(\epsilon_h^3) ,$$ for small hinge angle $\epsilon_h$, with $\hat \gamma$ giving the unit tangent vector field to $\gamma$.
2. The average mean curvature over the dual $n$-volume $V_v$ at a vertex $v$ is $$H_v
:= \widetilde H_{V_v}
= \frac{1}{|V_v|}
\sum_{h \subset \mathrm{star}(v)} \frac{1}{2}|h| \epsilon_h ,
\label{1.H}$$ with $|V_v|$ and $|h|$ representing the $n$-volume measures of $V_v$ and $h$ respectively.
3. The average curvature orthogonal to $h$ over an $n$-volume $V_h$ (formed by the $n$-volumes dual to the vertices in the closure of $h$, intersecting lines orthogonal to $h$ in $S^n$), is $$\alpha_h
:= \widetilde{\alpha}(\hat \gamma, \hat \gamma)_{V_h}
= \frac{1}{|V_h|} \left(
|h| \epsilon_h
+ \sum_{i} \frac{1}{2} |h_i| \cos^2 \theta_i \, \epsilon_i
\right) ,
\label{1.a}$$ for hinges $h_i$ intersecting $V_h$, making an angle $\theta_i$ with $h$, with hinge angles $\epsilon_i$.
The paper begins with some preliminaries in section \[sec:Prelim\], giving definitions for piecewise flat manifolds, hinge angles and triangulations of smooth embedded manifolds. The integral of the extrinsic curvature over hinge-intersecting paths is then found in section \[sec:Int\]. The following two sections motivate the choice of volumes over which to compute the mean curvature and hinge-orthogonal components, and prove expressions (\[1.H\]) and (\[1.a\]) for these volumes. The curvature values for triangulations of a circle in ${\mathbb{E}^{2}}$ and a $2$-sphere and $2$-cylinder in ${\mathbb{E}^{3}}$ are computed as examples throughout the paper, showing the constructions to match closely with their smooth curvature values. Generalizations to embeddings in higher dimensions is discussed briefly in section \[sec:GenEmb\].
Piecewise Flat Manifolds and Submanifolds {#sec:Prelim}
=========================================
Piecewise flat manifolds
------------------------
A piecewise flat manifold $S^n$ is a differential manifold which can be decomposed into Euclidean segments, with the most simple given by $n$-simplices (line-segments, triangles, tetrahedra). The geometry of a Euclidean $n$-simplex is entirely determined by the lengths of its edges, and so a simplicial piecewise flat manifold is also entirely determined by its simplicial graph and set of edge-lengths.
\[def:PL\]
1. A Euclidean $k$-simplex is an open subspace of a $k$-dimensional Euclidean space ${\mathbb{E}^{k}}$, given by the interior of the convex hull formed by $k + 1$ non-colinear points. An arbitrary $k$-simplex is denoted $\sigma^k$, with its closure denote $\bar \sigma^k$.
2. A homogeneous simplicial $n$-complex is a connected graph of simplices up to dimension $n$, where each simplex is either an $n$-simplex or a face of an $n$-simplex.
3. A piecewise flat manifold $S^n$ is a homogeneous simplicial $n$-complex formed by Euclidean $n$-simplices, such that the metric for each $k$-simplex $\sigma^k$ is consistent with the metrics of all $n$-simplices containing $\sigma^k$ in their closure.
4. The $\mathrm{star}$ of a $k$-simplex $\sigma^k$ is the subspace of a piecewise flat manifold $S^n$ formed by the set of simplices $\sigma_i^m$ containing $\sigma^k$ in their closures, i.e. $\mathrm{star}(\sigma^k) = \{ \sigma_i^m | \bar \sigma_i^m \supset \sigma^k \}$. In two dimensions this is comonly referred to as a $1$-ring.
The development from one $n$-simplex to an adjacent $n$-simplex is still Euclidean, with deviations from a Euclidean manifold arising for closed paths around co-dimension-two simplices. The parallel transport of a vector around such a path deviates by an *intrinsic* deficit angle associated with the co-dimension-two simplex, in the plane orthogonal to it. The intrinsic curvature of a piecewise flat manifold is characterised by these intrinsic deficit angles, which can be derived from the simplicial graph and set of edge-lengths, much like smooth curvature can be given in terms of the metric.
Hinge angles
------------
Any pair of $n$-simplices joined along a common $(n-1)$-face can be linearly embedded in ${\mathbb{E}^{n+1}}$ so that each $n$-simplex is flat. The common $(n-1)$-simplex is known as a *hinge*, denoted $h$, and the relative directions of the two $n$-simplices is determined by a *hinge angle* $\epsilon_h$ in the plane orthogonal to $h$ in ${\mathbb{E}^{n+1}}$. There are many, mostly equivalent, methods for defining the hinge angle, see for example [@SulDDG; @KLM89ec]. One such method uses the angle $\phi$ between the positive normal vector ${}{n}$ to one of the $n$-simplices, from a given orientation, and the inward-directed vector ${}{v}$ tangent to the other $n$-simplex and orthogonal to $h$. The hinge angle $\epsilon_h := \pi/2 - \phi$, and is positive when the $n$-simplices are concave from the given orientation, and negative when convex.
![Pair of triangles linearly embedded in ${\mathbb{E}^{3}}$, showing the hinge $h$ and hinge angle $\epsilon_h$.[]{data-label="fig:hinge"}](fig1)
A piecewise flat manifold $S^n$ is considered as a piecewise flat *submanifold* of ${\mathbb{E}^{n+1}}$ if it can be embedded piecewise-linearly so that each simplex is flat in ${\mathbb{E}^{n+1}}$. The hinge angles $\epsilon_h$ associated with each $(n-1)$-simplex $h$ are determined by the specific embedding of $S^n$ in ${\mathbb{E}^{n+1}}$, and the embedding itself is determined by these angles, up to Euclidean transformations.
Note that the notation for the hinges and hinge angles in this paper coincides with the notation for the co-dimension-two simplices and deficit angles respectively in [@PLCurv]. This is done to emphasize the similarity between the piecewise flat extrinsic curvature constructions developed here, and the *intrinsic* curvature constructions in [@PLCurv].
Approximating submanifolds
--------------------------
A smooth submanifold $M^n \subset {\mathbb{E}^{n+1}}$ can be approximated by a piecewise flat submanifold $S^n \subset {\mathbb{E}^{n+1}}$ in a number of ways. Generally, a homogeneous simplicial complex is first defined on $M^n$, with $S^n$ then defined using the same simplicial complex with either the vertices coinciding with those on $M^n$ or the $n$-simplices tangent to $M^n$ at some point in their interior. The piecewise flat submanifold $S^n$ may then be globally rescaled in order to have an equivalent $n$-volume to $M^n$.
In order for $S^n$ to be a good approximation for $M^n$, both the deficit angles and hinge angles should be uniformly small. This can be achieved by defining the simplicial complex on $M^n$ so that the $n$-simplices are small with respect to both the intrinsic and extrinsic curvature of the manifold. This gives a high resolution where either of the curvatures are high, and a lower resolution in regions where $M^n$ is almost flat.
\[eg:triang\]
1. A regular triangulation of a circle in ${\mathbb{E}^{2}}$ can be given by first subdividing the circle into $k$ equal length arcs. A regular $k$-sided polygon can then be defined in ${\mathbb{E}^{2}}$, with edge-lengths of size $|\ell| = 2 \pi r / k$ where $r$ is the radius of the circle. The hinge angles are based at the vertices of the polygon and are of size $\epsilon_h = - 2 \pi/k$.
2. A regular triangulation of a $2$-sphere in ${\mathbb{E}^{3}}$ can be defined similarly, using an icosahedron formed by 20 equilateral triangles. For a total area equal to that of a $2$-sphere of radius $r$, the length of each edge is $|\ell| = \sqrt{\frac{4 \pi}{5 \sqrt{3}}} \, r$. The hinge angles are based at the edges with an angle of $\epsilon_h = - \arccos [\sqrt{5}/3] \simeq - 0.23228 \pi$.
3. A cylinder in ${\mathbb{E}^{3}}$ can be approximated by starting with regular $k$-sided polygons for the cross-sectional circles, set at regular intervals of length $p$ apart, with edges denoted $\ell_a$. The corresponding vertices for neighbouring polygons can then be joined by new line-segments denoted $\ell_b$, of length $p$, forming a surface of flat rectangles. A simplicial manifold is formed by introducing diagonals to each rectangle, denoted $\ell_c$. The length and hinge angle associated with each type of edge is given in table \[tab:CylApprox\] below.
Edge $|\ell|$ $\epsilon_\ell$
---------- ---------------------------------- -----------------
$\ell_a$ $2 \pi r / k$ $0$
$\ell_b$ $p$ $- 2 \pi / k$
$\ell_c$ $\sqrt{|\ell_a|^2 + |\ell_b|^2}$ 0
: Edge-lengths and hinge angles for a piecewise flat approximation of a cylinder in ${\mathbb{E}^{3}}$.[]{data-label="tab:CylApprox"}
![Triangulations of a circle in ${\mathbb{E}^{2}}$, and of a sphere and cylinder in ${\mathbb{E}^{3}}$.[]{data-label="fig:EgApprox"}](fig2)
Smooth Extrinsic Curvature and Path Integrals {#sec:Int}
=============================================
Smooth extrinsic curvature
--------------------------
For a smooth submanifold $M^n \subset {\mathbb{E}^{n+1}}$, the extrinsic curvature at each point $x \in M^n$ gives a measure of how the manifold near $x$ deviates from the tangent space at $x$. This curvature is given by the second fundamental form, a symmetric bilinear form, see for example [@KobayNomizuII]. For any pair of vector fields ${}{u}$ and ${}{v}$ tangent to $M^n$ in a neighbourhood $U$ of $x$, the second fundamental form is given by the normal component of the change in ${}{u}$ in the direction of ${}{v}$: $$\alpha ({}{u}, {}{v})
:= \left< \nabla_{{}{v}} {}{u}, \hat n \right> ,
\label{alpha}$$ where $\hat n$ is the unit normal vector field to $M^n$ over $U$, $\nabla$ is the covariant derivative associated with the flat connection on ${\mathbb{E}^{n+1}}$ and $\left< \cdot , \cdot \right>$ is the inner product on ${\mathbb{E}^{n+1}}$.
Since the normal vector field $\hat n$ is everywhere orthogonal to the vector fields ${}{u}$ and ${}{v}$ in $U \subset M^n$, the second fundamental form can also be viewed as the change in $\hat n$ along ${}{v}$, taking the ${}{u}$ component of the result: $$\alpha ({}{u}, {}{v})
\equiv - \left< \nabla_{{}{v}} \hat n, {}{u} \right>
=: \left< \mathrm{Q}({}{v}), {}{u} \right> ,$$ with $\mathrm{Q}({}{v})$ known as the shape operator. The covariant derivative can also be given in terms of the infinitesimal parallel transport of vectors, which is trivial in Euclidean space. For an integral curve $\xi$ of the vector field ${}{v}$, $$\nabla_{{}{v}} {}{u} |_x
= \lim_{\Delta x \rightarrow 0}
\frac{T^\xi_{(x+\Delta x) \rightarrow x} {}{u} (x + \Delta x)
- {}{u} (x)}{\Delta x} ,$$ with $T^\xi_{x \rightarrow y} {}{u}$ denoting the parallel transport of the vector ${}{u}$ at $x$, along the curve $\xi$ to $y$.
All of the second fundamental form values at a point $x \in U$ can be found from those given by a set of $n$ linearly independent vector fields in $U$. These consist of $n$ values for a repeated argument, and $(n-1)/2$ for pairs of different vector fields. Over the neighbourhood $U$ these give a set of $(n+1)/2$ field components, forming a symmetric tensor field commonly referred to as the extrinsic curvature tensor. In the physics literature this tensor is denoted by ${\mathrm}{K}$, with ${\mathrm}{K}_{a b} \, {}{u}^a {}{v}^b := \alpha ({}{u}, {}{v})$, with Einstein summation assumed over repeated indices.
The second fundamental form can also be seen as a quadratic form, acting on a single vector field at a time, $$\alpha({}{u})
:= \alpha({}{u}, {}{u})
\equiv {\mathrm}{K}_{a b} \, {}{u}^a {}{u}^b .$$ Values of the bilinear form with mixed arguments can also be given in terms of the quadratic form, using the bilinearity and symmetry of the former. For two vector fields ${}{u}, {}{v}$ tangent to $M^n$ in $U \subset M^n$, the action of the bilinear form on the difference vector field ${}{u} - {}{v}$ can be expanded, $$\begin{aligned}
&&\alpha({}{u} - {}{v}, {}{u} - {}{v})
= \alpha({}{u}, {}{u})
- 2 \alpha({}{u}, {}{v})
+ \alpha({}{v}, {}{v})
\nonumber \\ \Leftrightarrow \qquad
&&\alpha({}{u}, {}{v})
= \frac{1}{2}\left(
\alpha({}{u})
+ \alpha({}{v})
- \alpha({}{u} - {}{v})
\right) .
\label{SFFuv}\end{aligned}$$ The complete second fundamental form at each point $x \in M^n$ can therefore also be given by the quadratic form values for a set of $n$ linearly independent vector fields in $U$, and for the $(n-1)/2$ vector fields given by the difference between each pair of these.
Path integrals of curvature along geodesic segments
---------------------------------------------------
Due to the nature of piecewise flat manifolds, infinitesimal application of the second fundamental form will give zero curvature within each $n$-simplex, and an infinite value at the hinges. However, since the hinge angles should be seen as integrated quantities, the path integral of the second fundamental form along smooth manifold geodesics is first investigated below.
\[lem:PathIntM\] Take a geodesic segment $\gamma (s) : [0, 1] \rightarrow M^n \subset {\mathbb{E}^{n+1}}$, with unit tangent vectors $\hat \gamma_s \in T_{\gamma(s)} M^n$. The integral along $\gamma$ of the second fundamental form $\alpha (\hat \gamma)$, is equal to the angle between $\hat \gamma_0$ and $\hat \gamma_1$ in ${\mathbb{E}^{n+1}}$, $$a_\gamma :=
\int_\gamma \alpha(\hat \gamma) \, \mathrm{d} s
= \psi \left(
\hat \gamma_0, \, T^\gamma_{1 \rightarrow 0} \hat \gamma_1
\right) .
\label{PathIntM}$$ The term $\psi$ represents the angle between its two arguments, and is deemed positive if the normal component of the vector $T^\gamma_{1 \rightarrow 0} \hat \gamma_1 - \hat \gamma_0$ is in the positive $\hat n$ direction, and negative otherwise.
From the definition of the second fundamental form in (\[alpha\]), $$\int_\gamma \alpha(\hat \gamma) \, \mathrm{d} s
= \int_\gamma \left<
\nabla_{\hat \gamma} \hat \gamma, \, \hat n \right> \, \mathrm{d} s .$$ The vector field $\nabla_{\hat \gamma} \hat \gamma$ must be parallel to $\hat n$ at each point of $\gamma$, since it is a geodesic curve in $M^n$ with $\nabla^{M}_{\hat \gamma} \hat \gamma = 0$, where $\nabla^{M}$ is the restriction of $\nabla$ to $M^n$. This derivative can also be given in terms of the parallel transport along $\gamma$, with $$\nabla_{\hat \gamma} \hat \gamma \, |_{\gamma(s)}
= \lim_{\Delta s \rightarrow 0} \frac{
T^\gamma_{(s+\Delta s) \rightarrow s} \hat \gamma_{(s + \Delta s)}
- \hat \gamma_s
}{\Delta s} .$$ In the limit above, the vector in the numerator can be denoted $d \psi \, \hat n$, where $d \psi$ is the infinitesimal angle of rotation between $\hat \gamma_s$ and the parallel transport of $\hat \gamma_{(s+d s)}$ back to the point $\gamma(s)$. This infinitesimal angle will be positive if the difference is in the positive $\hat n$ direction, and negative otherwise. The integral of the curvature along $\gamma$ therefore becomes $$\int_\gamma \alpha(\hat \gamma) \, \mathrm{d} s
= \int_\gamma \left<
\nabla_{\hat \gamma} \hat \gamma, \, \hat n \right> \, \mathrm{d} s
= \int_\gamma \left<
\frac{d \psi}{d s} \, \hat n, \, \hat n \right> \, \mathrm{d} s
= \int_\gamma \mathrm{d} \psi
= \psi \left(
\hat \gamma_0, \, T^\gamma_{1 \rightarrow 0} \hat \gamma_1
\right) ,$$ with the finite angle $\psi$ being positive if the normal component of $T^\gamma_{1 \rightarrow 0} \hat \gamma_1 - \hat \gamma_0$ is in the positive $\hat n$ direction, and negative otherwise.
Applying the result of the lemma above to a hinge $h$ in a piecewise flat manifold $S^n \subset {\mathbb{E}^{n+1}}$ can easily be seen to give the hinge angle $\epsilon_h$ as the integratal of the curvature along a path intersecting $h$ orthogonally. The right hand side of (\[PathIntM\]) can even be used to define the hinge angles. More interestingly, lemma \[lem:PathIntM\] can also be used to define the integrated curvature along paths which are *not* orthogonal to a hinge $h$ in terms of the hinge angle $\epsilon_h$.
\[thm:PathIntS\] For a geodesic segment $\gamma(s): [0,1] \rightarrow S^n \subset {\mathbb{E}^{n+1}}$, intersecting a single hinge $h$ at an angle $\theta$ to the vector orthogonal to $h$, the integrated curvature along $\gamma$ is $$a_h^\theta :=
a_\gamma
= \cos \theta \ \epsilon_h + O(\epsilon_h^3) ,$$ for a small hinge angle $\epsilon_h$, with the higher order terms vanishing for $\theta \in \{0, \pm \pi/2\}$.
A geodesic path in $S^n$ will be a straight line in each $n$-simplex that it intersects. The parallel transport of the vector $\hat \gamma_1$, tangent to $\gamma$ at the point $\gamma (1)$, to the point $\gamma (0)$ will have its component orthogonal to $h$ rotated by the hinge angle $\epsilon_h$, with the remaining components unchanged. From figure \[fig:PFpath\] it is clear that $$\sin \frac{1}{2} \psi(\hat \gamma_0, \, T^\gamma_{1 \rightarrow 0} \hat \gamma_1)
= \cos \theta \, \sin \frac{1}{2} \epsilon_h .$$ For $\cos \theta \in \{0, 1\}$ the $\sin$’s can be removed from both sides of the equation, otherwise for small hinge angle $\epsilon_h$ the equation reduces to $\psi(\hat \gamma_0, \, T^\gamma_{1 \rightarrow 0} \hat \gamma_1)
= \cos \theta \, \epsilon_h + O(\epsilon_h^3)$.
![The star of a hinge $h$ is shown on the left, with a geodesic curve $\gamma$ and the parallel transport of $\hat \gamma_1$ to $\gamma(0)$. The vectors $\hat \gamma_0$ and $T^\gamma_{1 \rightarrow 0} \hat \gamma_1$ are compared on the right.[]{data-label="fig:PFpath"}](fig3)
Since $\lim_{|\gamma| \rightarrow 0} a_\gamma$ gives the infinitesimal curvature $\alpha(\hat \gamma)$ in a smooth manifold, the piecewise flat integrated curvature $a_h^\theta$ can be seen as a discretization with the length of $\gamma$ becoming small with respect to the lattice rather than vanishing. In this sense, small corresponds to intersecting one and only one hinge, though this will be shown to be *too* small in the following sections. Instead these geodesic integrals will be used to give volume integrals over regions intersecting specific collections of hinges.
Mean Curvature {#sec:H}
==============
The mean curvature of a smooth manifold $M^n \subset {\mathbb{E}^{n+1}}$ is a scalar field on $M^n$, given by the trace of the second fundamental form. For an orthonormal set of vector fields $\{ {}{e_1}, ... , {}{e_n} \}$ in a neighbourhood $U \subset M^n$, the mean curvature field over $U$ is $$H
:= \sum_i^n \alpha ({}{e_i}, {}{e_i})
= \sum_i^n \alpha ({}{e_i}).
\label{Hsmooth}$$ This can be used with theorem \[thm:PathIntS\] to compute the integral of the mean curvature over certain $n$-dimensional subregions of a piecewise flat manifold $S^n$.
\[lem:IntKh\] Over any $n$-dimensional region $D \subset \mathrm{star}(h) \subset S^n$, the integral of the mean curvature is $$\int_D H \, \mathrm{d} V^n
= |h_D| \epsilon_h ,$$ where $|h_D|$ represents the $(n-1)$-volume of $h \cap D$.
By the definition of the mean curvature, the integral above can be given in terms of the second fundamental form for an orthonormal set of vector fields $\{ {}{e_1}, ... , {}{e_n} \}$ in $D$, $$\int_D H \, \mathrm{d} V^n
= \int_D \left[ \sum_i^n \alpha({}{e_i}) \right]
\mathrm{d} V^n
$$ Since the star of each hinge is intrinsically Euclidean, these vector fields can be chosen to represent a Cartesian coordinate frame, with their integral curves given as geodesic lines within $D \subset S^n$. The integral and sum therefore commute and can be swapped, with the integrals for each value of $i$ above then given in terms of the geodesic curvature integrals, $$\int_D \left[ \sum_i^n \alpha({}{e_i}) \right]
\mathrm{d} V^n
= \sum_i^n \left[ \int_D \alpha({}{e_i}) \, \mathrm{d} V^n \right]
= \sum_i^n \left[
\int_{h_D} a^{\theta_i}_h \, \mathrm{d} V^{n-1} \right] ,
\label{Proof:IntKh2}$$ with only the lines intersecting $h_D$ having non-zero integrated curvature. Choosing the vector field ${}{e_1}$ to be orthogonal to the hinge $h$, the integral curves of the remainder of the vector fields will not intersect $h$. The sum in (\[Proof:IntKh2\]) can therefore be reduced to a single term, with $$\int_D H \, \mathrm{d} V^n
= \int_{h_D} a^{\theta_1 = 0}_h \, \mathrm{d} V^{n-1}
= \epsilon_h \int_{h_D} \mathrm{d} V^{n-1}
= |h_D| \epsilon_h ,$$ with the hinge angle $\epsilon_h$ invariant over $h$.
This result matches some earlier definitions in the literature [@HS81; @TraceK], associating an integrated mean curvature of $|h| \epsilon_h$ with each hinge. Such a definition also leads naturally to Steiner’s total mean curvature $\sum_h |h| \epsilon_h$ from [@Steiner]. However, unlike the smooth case, hinges have a directionality associated with them for $n > 1$, specifically the direction orthogonal to each hinge in $S^n$. Taking the cylinder in example \[eg:triang\], each type of hinge will have a different value of $|h| \epsilon_h$, while a smooth cylinder has the same mean curvature value everywhere.
The mean curvature at each point of a smooth manifold $M^n \subset {\mathbb{E}^{n+1}}$ can be seen as an average of the second fundamental form over *all* directions in $M^n$. Since regions enclosing single vertices give the most general collection of hinge orientations in a piecewise flat manifold $S^n$, the piecewise flat mean curvature will be given here by integrating over the $n$-dimensional regions of a dual tessellation of $S^n$, similar to [@Taub95; @MeyDebSchBar03]
\[def:V\_v\] A decomposition of a piecewise flat manifold $S^n$, into $n$-dimensional regions $V_v$ dual to each vertex $v$, is defined so that:
1. The vertex $v \in V_v$, but no other vertices are contained within $V_v$.
2. The regions $V_v$ form a complete tessellation of $S^n$, $$|S^n| = \sum_{v \in S^n} |V_v|, \qquad
V_{v_i} \cap V_{v_j} = \emptyset \quad
\forall \ i \neq j .$$ with $|S^n|$ and $|V_v|$ representing the $n$-volumes of $S^n$ and $V_v$ respectively.
3. Only hinges $h_v$ in the star of $v$ intersect $V_v$.
This gives the most general definition of a dual tessellation for the constructions that follow. The third property is not strictly necessary but is deemed a reasonable requirement, with Voronoi tessellations satisfying this condition only where $S^n$ gives a Delaunay triangulation for example. It also seems reasonable to require an additional property which distributes the total $n$-volume of $S^n$ in some consistent manner over the subregions $V_v$. However there are many different methods for doing this, most of which are incompatible with one another, such as the Voronoi and barycentric tessellations. Such a property will therefore not be imposed here.
\[thm:H\_v\] The average mean curvature over a dual $n$-volume $V_n$ is $$\label{H_v}
H_v
:= \widetilde H_{V_v}
= \frac{1}{|V_v|}
\sum_{h \subset \mathrm{star}(v)} |h_{|V_v}| \epsilon_h ,$$ with $h_{|V_v} = h \cap V_v$, and $|V_v|$ representing the $n$-volume of $V_v$.
The region $V_v$ can be decomposed into subregions $D_h$, each enclosing the intersection of the hinge $h$ with $V_v$, and no part of any other hinge. The integral of the mean curvature over each subregion $D_h$ is given by lemma \[lem:IntKh\], with the total integral over $V_v$ then given by the sum of these, $$\int_{V_v} H \ \mathrm{d} V^n
= \sum_{h \subset \mathrm{star}(v)}
\int_{D_h} H \ \mathrm{d} V^n
= \sum_{h \subset \mathrm{star}(v)} |h_{|V_v}| \epsilon_h .$$ The average curvature is then found by dividing the integral by the volume of $V_v$.
Instead of the hinge angles, in [@Taub95] and [@MeyDebSchBar03] the mean curvature comes from the curvature along the edges, essentially given by comparing the tangent vector of the edge with a normal vector constructed at the vertex. These curvatures can be distributed over triangular segments of a Voronoi region $V_v$ associated with each edge. The segments, known as *hybrid cells* in [@MMR; @TraceK], are formed by the given vertex and the circumcenters of the triangles on either side of each edge. Integrating over this distribution of curvatures gives precisely the *cotan* formula of [@PP93; @MeyDebSchBar03; @WarDDG; @BS07]. This can also be extended to higher dimensions, where the curvature associated with each edge can be computed in the same way, and the volume of these circumcentric hybrid cells is also easily found.
Unlike other vertex-based mean curvatures, the piecewise flat mean curvature given in theorem \[thm:H\_v\] above still gives the same total mean curvature expression as that of Steiner [@Steiner].
\[cor:IntH\] The total mean curvature over a piecewise flat manifold $S^n$ is $$\int_{S^n} H_v \ \mathrm{d} V^n
= \sum_{h \subset S^n} |h| \epsilon_h .
\label{TotalH}$$
The integral of $H_v$ over $S^n$ is equal to the sum of the integrals for each vertex volume $V_v$. Each part of a given hinge $h$ must be contained in a single volume $V_v$, since the volumes $V_v$ form a tessellation of $S^n$, and since the deficit angles $\epsilon_h$ are fixed over $h$, the contribution from each hinge will be $|h| \epsilon_h$. The total integral is then given by the sum of these terms for all hinges $h \subset S^n$.
\[eg:mean\]
1. For a piecewise linear curve in ${\mathbb{E}^{2}}$ there is an unambiguous vertex region given by half of each edge in the star of each vertex. For the $k$-sided polygon approximation of a circle, the mean curvature at each vertex is $$H_v^{\mathrm{poly}}
= \frac{\epsilon_v}{|\ell|}
= \frac{- 2 \pi / k}{2 \pi r / k}
= - \frac{1}{r} ,$$ which is exactly equal to the smooth curvature of a circle in ${\mathbb{E}^{2}}$.
2. Due to the high level of symmetry in the icosahedron approximation of a $2$-sphere in ${\mathbb{E}^{3}}$, there is also an unambiguous choice of vertex region $V_v$. This region is formed by the convex hull of the symmetric centres of the equilateral triangles in the star of the vertex. With 12 vertices in an icosahedron, the area of each vertex region $|V_v| = 4 \pi r^2 / 12 = \pi r^2 / 3$, with the mean curvature then given by $$H_v^{\mathrm{icos}}
= \frac{5(\frac{1}{2} \, |\ell| \, \epsilon_\ell)}{|V_v|}
= \frac{15}{2 \pi r} \, \sqrt{\frac{4 \pi}{5 \sqrt{3}}} \,
\left(- \arccos\left[\frac{\sqrt{5}}{3}\right] \right)
\simeq - \frac{2.09851}{r} ,$$ which gives a good approximation for $H = - 2/r$, the mean curvature of a smooth $2$-sphere in ${\mathbb{E}^{3}}$.
3. The most regular vertex regions in the piecewise flat approximation for the cylinder are given by intrinsically rectangular regions with the vertex at the centre, and sides parallel with the edges $\ell_a$ and $\ell_b$. These give the Voronoi regions around each vertex, and an area of $|V_v| = |\ell_a| \times |\ell_b| = 2 \pi r p / k$. For a vertex region intersecting $q$ diagonal edges $\ell_c$ (which can range from $0$ to $4$), the mean curvature is $$H_v^{\mathrm{cyl}}
= \frac{2 \, \frac{1}{2} |\ell_a| \epsilon_a
+ 2 \, \frac{1}{2} |\ell_b| \epsilon_b
+ q \, \frac{1}{2} |\ell_c| \epsilon_c}{|V_v|}
= \frac{- 2 \pi p / k + 0 + 0}{2 \pi r p / k}
= - \frac{1}{r} .$$ This value is constant for all vertices, is invariant to the number of diagonal edges $\ell_c$ intersected, and gives exactly the mean curvature of a smooth cylinder in ${\mathbb{E}^{3}}$.
Hinge-Orthogonal Component {#sec:Kh}
==========================
Due to the structure of the hinge angles, the most natural values of the second fundamental form are those given by vector fields orthogonal to each hinge. However regions enclosing single hinges can be seen to be insufficient, similar to the mean curvature, which can be shown from both the sphere and cylinder triangulations in example \[eg:triang\].
The regular shape of the icosahedron approximation of a $2$-sphere means it can be tessellated regularly into regions $D_h$ associated with each hinge. Using part of the arguments from the proof of lemma \[lem:IntKh\], the average value of the component orthogonal to a hinge $h$ can then be given by the expression $|h| \epsilon_h / D_h$. However the resulting value differs from the smooth curvature by a factor of about two. Even more problematic are the diagonal edges $\ell_c$ in the cylinder example. These have zero hinge angles, and so the expression $|\ell_c| \epsilon_c / D_c$ vanishes, which is clearly not the case for the corresponding smooth curvature component. The issue with the sphere can be attributed to the fact that there are $n$ linearly independent vectors at each point, while using tessellations assigns each point to a single curvature component. The cylinder problem implies that a region intersecting a larger collection of hinges is required.
A plausible solution is to follow the mean curvature approach, but to use a union of the vertex regions $V_v$ for the vertices $v$ in the closure of each hinge. Unfortunately the vectors orthogonal to a given hinge cannot be defined over the regions $V_v$ in an unambiguous way. A new set of $n$-dimensional regions is therefore defined as a subspace of this union, where vectors can be parallel transported unambiguously, and where the integrated curvature components are stable to small deviations of hinges across the boundaries internal to each vertex region.
\[def:Vh\] The $n$-dimensional regions $V_h$ associated with each hinge $h$ are defined as the union of the regions $V_v$ dual to the vertices $v$ in the closure of $h$, intersected with the geodesic extensions $\gamma^\perp_h$ of the lines orthogonal to $h$, $$V_h
:= \left(\cup_v \, V_v\right)
\cap \int_h \gamma^\perp_h \, \mathrm{d} V^{n-1} ,
\quad \text{such that} \quad
v \in \bar h .$$
Each region $V_h$ contains all of the hinge $h$, and parts of other hinges, and will overlap with regions of these other hinges. The average curvature orthogonal to each hinge $h$ is now given by integrating over these hinge regions.
\[thm:alpha\_h\] In a piecewise flat manifold $S^n \subset {\mathbb{E}^{n+1}}$, the average curvature component orthogonal to a hinge $h$ over the region $V_h$ is $$\label{alpha_h}
\alpha_h
:= \widetilde{\alpha}(\hat \gamma^\perp_h)_{V_h}
= \frac{1}{|V_h|} \left(
|h| \epsilon_h
+ \sum_{i} |h_i \cap V_h| \, \cos^2 \theta_i \, \epsilon_i
+ O(\epsilon_i^3)
\right) ,$$ with the sum taken over the other hinges $h_i$ intersecting $V_h$, with $\epsilon_i$ representing the hinge angle at $h_i$, and $\theta_i$ the angle between $\gamma^\perp_h$ and the lines orthogonal to each hinge $h_i$.
![Hinge region for an icosahedron in ${\mathbb{E}^{3}}$.[]{data-label="fig:IIh"}](fig4)
For each geodesic segment $\gamma := \gamma^\perp_h \ \cap \ V_h$, the path integral of the second fundamental form $\alpha(\hat \gamma)$ along $\gamma$ is given by the sum of the contributions from each hinge it intersects, $$ a_{\gamma}
= a_h + \sum_i a^{\theta_i}_{h_i}
= \epsilon_h
+ \sum_i \cos \theta_i \, \epsilon_i
+ O(\epsilon_i^3) .$$
The integral of $\alpha(\hat \gamma)$ over the entire region $V_h$ is then given by integrating these path integrals over all of $h$, which can also be given by a sum of contributions from each hinge, $$\int_{V_h} \alpha(\hat \gamma) \, \mathrm{d} V^n
= \int_h
a_{\gamma} \, \mathrm{d} V^{n-1}
= |h| \epsilon_h
+ \sum_i \int_h a^{\theta_i}_{h_i} \, \mathrm{d} V^{n-1} .
\label{alpha_h_2}$$ For each hinge $h_i$, the cross-sectional $(n-1)$-volume of the geodesic segments $\gamma$ which intersect $h_i$ is $|h_i \cap V_h| \cos \theta_i$, see figure \[fig:IIh\] for example. Since $\theta_i$ and $\epsilon_i$ are constant over each hinge $h_i$, the final expression for the integral of $\alpha(\hat \gamma)$ over $V_h$ is $$\int_{V_h} \alpha(\hat \gamma) \mathrm{d} V^n
= |h| \epsilon_h
+ \sum_i |h_i \cap V_h| \cos^2 \theta_i \, \epsilon_i
+ O(\epsilon_i^3) .$$ The average value is then given by dividing by the $n$-volume of the region $V_h$.
For a two dimensional piecewise flat manifold $S^2 \subset {\mathbb{E}^{3}}$, the second fundamental form can be completely determined within each triangle, using the curvature orthogonal to each of the edges (hinges) $\ell \subset S^2$. For any given triangle, with edges denoted $\ell_1$, $\ell_2$ and $\ell_3$, $$|\ell_1| \, \hat \ell^\perp_1
+ |\ell_2| \, \hat \ell^\perp_2
+ |\ell_3| \, \hat \ell^\perp_3
= 0 .$$ Taking the normal vectors to two of the sides within $S^2$ as basis vectors, the relation above can be used instead of the difference vector in equation (\[SFFuv\]) to give the second fundamental biliear form for mixed arguments as $$\begin{aligned}
\alpha(\hat \ell^\perp_1, \hat \ell^\perp_2)
&=& \frac{1}{|\ell_1| \, |\ell_2|}\left(
- \alpha(|\ell_1| \, \hat \ell^\perp_1)
- \alpha(|\ell_2| \, \hat \ell^\perp_2)
+ \alpha(|\ell_1| \, \hat \ell^\perp_1 + |\ell_2| \, \hat \ell^\perp_2)
\right) \nonumber \\
&=& \frac{1}{|\ell_1| \, |\ell_2|}\left(
- |\ell_1|^2 \, \alpha_{\ell_1}
- |\ell_2|^2 \, \alpha_{\ell_2}
+ |\ell_3|^2 \, \alpha_{\ell_3}
\right) .\end{aligned}$$ This gives a complete extrinsic curvature tensor within each triangle, using only the hinge-orthogonal curvatures.
\
\[eg:mean\]
1. For a piecewise linear curve in ${\mathbb{E}^{2}}$ the curvature component $\alpha_h \equiv H_v$. In the case of the $k$-sided regular polygon, this gives a curvature component of $\alpha = - 1/r$ which is equal to the smooth curvature of a circle in ${\mathbb{E}^{2}}$.
2. The hinge region $V_h$ for the icosahedron consists of two thirds of the area of each of the equilateral triangles on either side of the hinge (edge) $h$, and a sixth of the four triangles bounding these, see figure \[fig:IIhEg\], so that $|V_h| = 2 \times 4 \pi r^2 / 20 = 2 \pi r^2 / 5$. There are four other edges with half of their lengths intersecting this volume, each making an angle of $\theta = \pi/3$ with $h$. The curvature component orthogonal to each hinge is therefore $$\alpha_h
= \frac{|\ell| \epsilon
+ 4\left(\frac{1}{2} |\ell| \cos^2 \frac{\pi}{3} \, \epsilon\right)
}{|V_h|}
= \frac{15}{4 \pi r} \, \sqrt{\frac{4 \pi}{5 \sqrt{3}}} \,
\left( - \arccos\left[\frac{\sqrt{5}}{3}\right] \right)
\simeq - \frac{1.04926}{r} ,$$ which approximates the curvature components of a smooth $2$-sphere, $\alpha({}{\hat u}) = - 1/r$, to the same level of accuracy as the mean curvature.
3. Since the piecewise flat approximation of the cylinder is intrinsically flat, with zero deficit angles around each vertex, the hinge regions will consist of half of the areas of each of the two vertex regions, so $|V_a| = |V_b| = |V_c| = 2 \pi r p / k$. For the hinge regions $V_a$ and $V_b$, there are edges $\ell_b$ and $\ell_a$ on the boundaries, with $q$ representing the number of diagonal edges $\ell_c$ (from zero to four).
![Hinge region for cylinder triangulation edge $\ell_c$.[]{data-label="fig:IIhEg"}](fig5)
The angles between each pair of edges are $\theta_{ab} = \pi/2$, $\theta_{ac} = \arctan\left[|\ell_b| / |\ell_a|\right]$ and $\theta_{bc} =\arctan\left[|\ell_a| / |\ell_b|\right]$, and the curvature components orthogonal to each type of hinge are $$\begin{aligned}
\alpha_a
&= \frac{|\ell_a| \epsilon_a
+ 2 \frac{1}{2} |\ell_b| \cos^2 \frac{\pi}{2} \, \epsilon_b
+ q \frac{1}{2} |\ell_c| \cos^2 \theta_{a c} \, \epsilon_c
}{|V_a|}
= 0 ,
\\
\alpha_b
&= \frac{|\ell_b| \epsilon_b
+ 2 \frac{1}{2} |\ell_a| \cos^2 \frac{\pi}{2} \, \epsilon_a
+ q \frac{1}{2} |\ell_c| \cos^2 \theta_{b c} \, \epsilon_c
}{|V_b|}
= - \frac{1}{r} ,
\\
\alpha_c
&= \frac{|\ell_c| \epsilon_c
+ 2 \frac{1}{2} |\ell_a| \cos^2 \theta_{a c} \, \epsilon_a
+ 2 \frac{1}{2} |\ell_b| \cos^2 \theta_{b c} \, \epsilon_b
}{|V_a|}
= - \frac{\cos^2 \theta_{b c}}{r} .\end{aligned}$$ These exactly match the smooth values, where the curvature component for a unit vector ${}{\hat u}$ tangent to a cylinder, making an angle $\theta$ with a circular cross-section in ${\mathbb{E}^{3}}$, is $\alpha({}{\hat u}) = - \cos^2 \theta / r$.
Generalized Embeddings {#sec:GenEmb}
======================
For a smooth manifold $M^n \subset {\mathbb{E}^{n+m}}$ the second fundamental form is valued in the normal tangent bundle to $M^n$ in ${\mathbb{E}^{n+m}}$. For $m = 1$, $M^n$ is a hypermanifold and the normal bundle is one dimensional, so the scalar values of the magnitude is all that is required to determine the extrinsic curvature. For $m > 1$, the second fundamental form is instead defined as $$\alpha ({}{u}, {}{v})
:= \left< \nabla_{{}{v}} {}{u}, \hat n_{\alpha} \right> \hat n_\alpha ,$$ where $\hat n_\alpha$ is the unit normal vector to $M^n$ which gives the projection of $\nabla_{{}{v}} {}{u}$ into the normal bundle. As can be seen by the Serret-Frenet frame for a curve in ${\mathbb{E}^{3}}$, see for example [@SulDDG], the curvature is no longer sufficient to describe the embedding of $M^n$ into ${\mathbb{E}^{n+m}}$ completely, with information about the variation in $\hat n_\alpha$ over $M^n$ also required.
Not all piecewise flat manifolds can be embedded in a Euclidean space of only one extra dimension, but can be embedded if the dimension of the Euclidean space is increased. For an embedding of $S^n$ into ${\mathbb{E}^{n+m}}$, with $m > 1$, the normal vector to each $n$-simplex is no longer unique. However for each hinge $h$ there is an unambiguous subspace ${\mathbb{E}^{n+1}} \subset {\mathbb{E}^{n+m}}$ containing the two $n$-simplices on either side of $h$, where the angle $\epsilon_h$ can be defined. The hinge angle can then be multiplied by the unit vector $\hat n_h$, normal to $h$ in ${\mathbb{E}^{n+1}}$, which makes equal angles with the normal vectors to each $n$-simplex in the star of $h$. Since these $(n+1)$-dimensional spaces will not be consistent across all hinges, the relationship between them is also required for a complete description of the embedding, as with the smooth case.
The integral of the *magnitude* of the second fundamental form, tangent to a geodesic curve $\gamma \subset M^n \subset {\mathbb{E}^{n+m}}$, will still be given by $a_\gamma$ from lemma \[lem:PathIntM\]. As a result, the integral of the magnitude of the piecewise flat curvature across a hinge $h \subset S^n \subset {\mathbb{E}^{n+m}}$ will still be given by $a_h^\theta$ from theorem \[thm:PathIntS\]. Equations (\[H\_v\]) and (\[alpha\_h\]) then give the average magnitudes of the mean and hinge-orthogonal curvatures. Orientations for these curvatures could then be determined by a weighted average of the unit normal vectors at each hinge. However the details should depend on a complete piecewise flat version of the smooth variation of the unit normal vectors $\hat n_\alpha$.
Conclusion {#sec:Con}
==========
Expressions have been given for the piecewise flat mean curvature at each vertex (\[H\_v\]) and the extrinsic curvature components orthogonal to the hinges (\[alpha\_h\]) of a piecewise flat manifold $S^n \subset {\mathbb{E}^{n+1}}$, and shown to give good approximations for triangulations of a circle in ${\mathbb{E}^{2}}$ and a sphere and cylinder in ${\mathbb{E}^{3}}$. Since these expressions depend on a collection of hinges, they should be stable to different triangulations of the same smooth manifold, as long as the triangulations give small deficit and hinge angles everywhere. The definitions are also not dependent on any individual choice of dual tessellation, as long as the properties of definition \[def:V\_v\] are satisfied, and can be used in any dimension.
Acknowledgements {#acknowledgements .unnumbered}
================
I’d like to thank Warner A. Miller, Maximilian Hanush and Christopher Duston for many helpful discussions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper we study metastable behaviour at low temperature of Glauber spin-flip dynamics on random graphs. We fix a large number of vertices and randomly allocate edges according to the Configuration Model with a prescribed degree distribution. Each vertex carries a spin that can point either up or down. Each spin interacts with a positive magnetic field, while spins at vertices that are connected by edges also interact with each other via a ferromagnetic pair potential. We start from the configuration where all spins point down, and allow spins to flip up or down according to a Metropolis dynamics at positive temperature. We are interested in the time it takes the system to reach the configuration where all spins point up. In order to achieve this transition, the system needs to create a sufficiently large droplet of up-spins, called critical droplet, which triggers the crossover.
In the limit as the temperature tends to zero, and subject to a certain *key hypothesis* implying metastable behaviour, the average crossover time follows the classical *Arrhenius law*, with an exponent and a prefactor that are controlled by the *energy* and the *entropy* of the critical droplet. The crossover time divided by its average is exponentially distributed. We study the scaling behaviour of the exponent as the number of vertices tends to infinity, deriving upper and lower bounds. We also identify a regime for the magnetic field and the pair potential in which the key hypothesis is satisfied. The critical droplets, representing the saddle points for the crossover, have a size that is of the order of the number of vertices. This is because the random graphs generated by the Configuration Model are expander graphs.
address:
- 'Department of Mathematics, Ruhr University Bochum, Universitätsstraße 150, 44780 Bochum, Germany'
- 'Mathematical Institute, Leiden University, P.O. Box 9512, 2300 RA Leiden, The Netherlands.'
- 'Mathematical Institute, Leiden University, P.O. Box 9512, 2300 RA Leiden, The Netherlands.'
- 'Eindhoven University of Technology, Department of Mathematics and Computer Science, P.O. Box 513, 5600 MB Eindhoven, The Netherlands.'
author:
- 'S. Dommers'
- 'F. den Hollander'
- 'O. Jovanovski'
- 'F.R. Nardi'
title: Metastability for Glauber dynamics on random graphs
---
[^1]
Introduction and main theorems
==============================
A physical system is in a *metastable state* when it remains locked for a very long time in a phase that is different from the one corresponding to thermodynamic equilibrium. The latter is referred to as the *stable state*. Classical examples are supersaturated vapours, supercooled liquids, and ferromagnets in the hysteresis loop. The main three objects of interest for metastability are the transition time from the metastable state to the stable state, the gate of configurations the system has to cross in order to achieve the transition, and the tube of typical trajectories the system follows prior to and after the transition.
Metastability for interacting particle systems on *lattices* has been studied intensively in the past three decades. Various different approaches have been proposed. After initial work by Cassandro, Galves, Olivieri and Vares [@CGOV84], Neves and Schonmann [@NS91], [@NS92], a powerful method – known as the *pathwise approach* to metastability based on large deviation theory – was developed in Olivieri and Scoppola [@OS95], [@OS96], Catoni and Cerf [@CC97], Manzo, Nardi, Olivieri and Scoppola [@MNOS04], Cirillo and Nardi [@CN13], Cirillo, Nardi and Sohier [@CNSo15]. This was successfully applied to low-temperature Ising and Blume-Capel models subject to Glauber spin-flip dynamics (in two and three dimensions, with isotropic, anisotropic and staggered interactions) in Kotecký and Olivieri [@KO92], [@KO93], [@KO94], Cirillo and Olivieri [@CO96], Ben Arous and Cerf [@BAC96], Nardi and Olivieri [@NO96]. Later, another powerful method – known as the *potential-theoretic approach* to metastability based on the analogy between Markov processes and electric networks – was developed in Bovier, Eckhoff, Gayrard and Klein [@BEGK00], [@BEGK01], [@BEGK02], [@BEGK04]. This was shown in Bovier and Manzo [@BM02], Bovier, den Hollander and Spitoni [@BdHS10] to lead to a considerable sharpening of earlier results. For other approaches to metastability, as well as further examples of metastable stochastic dynamics and relevant literature, we refer the reader to the monographs by Olivieri and Vares [@OV05], Bovier and den Hollander [@BdH15].
Recently, there has been interest in the Ising model on *random graphs* (Dembo and Montanari [@DM10], Dommers, Giardinà and van der Hofstad [@DGvdH10], Mossel and Sly [@MS13]). The only results known to date about metastability subject to Glauber spin-flip dynamics are valid for $r$-regular random graphs (Dommers [@Dpr]). In the present paper we investigate what can be said for more general degree distributions. Metastability is much more challenging on random graphs than on lattices. Moreover, we need to capture the metastable behaviour for a *generic realisation* of the random graph.
In Section \[IsingGlauber\] we define the Ising model on a random multigraph subject to Glauber spin-flip dynamics. We start from the configuration where all spins point down, and allow spins to flip up or down according to a Metropolis dynamics at positive temperature. We are interested in the time it takes the system to reach the configuration where all spins point up. In Section \[metastability\] we introduce certain geometric quantities that play a central role in the description of the metastable behaviour of the system, and state three general theorems that are valid under a certain key hypothesis. These theorems concern the average transition time, the distribution of the transition time, and the gate of saddle point configurations for the crossover, all in the limit of low temperature. They involve certain key quantities associated with the random graph. Our goal is to study the scaling behaviour of these quantities as the size of the graph tends to infinity.
In Section \[literature\] we describe four examples to which the three general theorems apply: three refer to regular lattices, while one refers to the Erdős-Rényi random graph. In Section \[configurationmodel\] we recall the definition of the Configuration Model, which is an example of a random graph with a non-trivial geometric structure. In Section \[maintheorems\] we state our main metastability results for the latter. In Section \[discussion\] we place these results in their proper context and give an outline of the remainder of the paper.
Ising model and Glauber dynamics {#IsingGlauber}
--------------------------------
Given a finite connected non-oriented multigraph $G=(V,E)$, let $\Omega=\{-1,+1\}^V$ be the set of configurations $\xi=\{\xi(v)\colon\,v \in V\}$ that assign to each vertex $v\in V$ a spin-value $\xi(v)\in\{-1,+1\}$. Two configurations that will be of particular interest to us are those where all spins point up, respectively, down: $$\boxplus \equiv +1, \qquad \boxminus \equiv -1.$$ For $\beta \geq 0$, playing the role of *inverse temperature*, we define the Gibbs measure $$\mu_\beta(\xi) = \frac{1}{Z_\beta}\,e^{-\beta {\mathcal{H}}(\xi)},
\qquad \xi\in\Omega,
\label{eq:Gibbs}$$ where ${\mathcal{H}}\colon\,\Omega\to{\mathbb{R}}$ is the *Hamiltonian* that assigns an energy to each configuration given by $${\mathcal{H}}(\xi) = -\frac{J}{2} \sum_{(v,w) \in E} \xi(v)\xi(w)
-\frac{h}{2} \sum_{v \in V} \xi(v), \qquad \xi\in\Omega,
\label{eq:hamiltonian}$$ with $J>0$ the *ferromagnetic pair potential* and $h>0$ the *magnetic field*. The first sum in the right-hand side of runs over all non-oriented edges in $E$. Hence, if $v,w \in V$ have $k \in {\mathbb{N}}_0$ edges between them, then their joint contribution to the energy is $-k\,\frac{J}{2}\,\xi(v)\xi(w)$.
We write $\xi\sim\zeta$ if and only if $\xi$ and $\zeta$ agree at all but one vertex. A transition from $\xi$ to $\zeta$ corresponds to a flip of a single spin, and is referred to as an *allowed move*. Glauber spin-flip dynamics on $\Omega$ is the continuous-time Markov process $(\xi_t)_{t \geq 0}$ defined by the transition rates $$c_\beta(\xi,\zeta) = \begin{cases}
e^{-\beta[{\mathcal{H}}(\zeta)-{\mathcal{H}}(\xi)]_+}, &\xi\sim\zeta,\\
0, &\mbox{otherwise}.
\end{cases}$$ The Gibbs measure in is the reversible equilibrium of this dynamics. We write $P^{G,\beta}_\xi$ to denote the law of $(\xi_t)_{t \geq 0}$ given $\xi_0=\xi$, ${\mathcal{L}}^{G,\beta}$ to denote the associated generator, and $\lambda^{G,\beta}$ to denote the principal eigenvalue of ${\mathcal{L}}^{G,\beta}$. The upper indices $G,\beta$ exhibit the dependence on the underlying graph $G$ and the interaction strength $\beta$ between neighbouring spins. For $A\subseteq\Omega$, we write $$\tau_A = \inf\big\{t>0\colon\,\xi_t \in A,\,\exists\,0<s<t\colon\,\xi_s \neq \xi_0\big\}$$ to denote the first hitting time of the set $A$ after the starting configuration is left.
Metastability
-------------
To describe the metastable behaviour of our dynamics we need the following geometric definitions.
\(a) The communication height between two distinct configurations $\xi,\zeta\in\Omega$ is $$\Phi(\xi,\zeta) = \min_{\gamma\colon\,\xi\to\zeta} \max_{\sigma\in\gamma}
{\mathcal{H}}(\sigma),$$ where the minimum is taken over all paths $\gamma\colon\,\xi\to\zeta$ consisting of allowed moves only. The communication height between two non-empty disjoint sets $A,B\subset\Omega$ is $$\Phi(A,B) = \min_{\xi\in A,\zeta\in B} \Phi(\xi,\zeta).$$ (b) The stability level of $\xi\in\Omega$ is $$V_\xi = \min_{ {\zeta\in\Omega:} \atop {{\mathcal{H}}(\zeta)<{\mathcal{H}}(\xi)} } \Phi(\xi,\zeta)-{\mathcal{H}}(\xi).$$ (c) The set of stable configurations is $$\Omega_{\mathrm{stab}}= \left\{\xi\in\Omega\colon\,{\mathcal{H}}(\xi)
= \min_{\zeta\in\Omega} {\mathcal{H}}(\zeta)\right\}.$$ (d) The set of metastable configurations is $$\Omega_{\mathrm{meta}}= \left\{\xi\in\Omega\backslash\Omega_{\mathrm{stab}}\colon\,
V_\xi = \max_{\zeta\in\Omega\backslash\Omega_{\mathrm{stab}}} V_\zeta\right\}.$$
It is easy to check that $\Omega_{\mathrm{stab}}= \{\boxplus\}$ for all $G$ because $J,h>0$. For general $G$, however, $\Omega_{\mathrm{meta}}$ is not a singleton, but we will be interested in those $G$ for which the following *hypothesis* is satisfied:
- $\Omega_{\mathrm{meta}}= \{\boxminus\}$. \[itm:Hhyp\]
The energy barrier between $\boxminus$ and $\boxplus$ is $$\Gamma^\star = \Phi(\boxminus,\boxplus)-{\mathcal{H}}(\boxminus).$$
Let $({\mathcal{P}}^\star,{\mathcal{C}}^\star)$ be the unique maximal subset of $\Omega\times\Omega$ with the following properties (see Fig. [\[fig-protocrcr\]]{}):
1. $\forall\,\xi\in{\mathcal{P}}^\star\,\exists\,\xi\prime\in{\mathcal{C}}^\star\colon\,\xi\sim\xi\prime$,\
$\forall\,\xi\prime\in{\mathcal{C}}^\star\,\exists\,\xi\in{\mathcal{P}}^\star\colon\,\xi\prime\sim\xi$.
2. $\forall\,\xi\in{\mathcal{P}}^\star\colon\,\Phi(\xi,\boxminus)<\Phi(\xi,\boxplus)$.
3. $\forall\xi\,\in{\mathcal{C}}^\star\,\exists\,\gamma\colon\,\xi\to\boxplus\colon\,\\
{\rm (i)} \max_{\zeta\in\gamma} {\mathcal{H}}(\zeta)-{\mathcal{H}}(\boxminus) \leq \Gamma^\star$.\
[(ii)]{} $\gamma \cap\{\zeta\in\Omega\colon\,\Phi(\zeta,\boxminus)
<\Phi(\zeta,\boxplus)\} = \emptyset$.
(15,15)(0,-1) [(0,0)(0,5)(0,10) (7,0)(7,5)(7,10) ]{} (-2,11)[${\mathcal{P}}^\star$]{} (5,11)[${\mathcal{C}}^\star$]{} (0.4,5.8)[$\xi$]{} (5.8,7.8)[$\xi\prime$]{} (-5,1.5)[$\boxminus$]{} (10.6,-2)[$\boxplus$]{} (-8.6,4.8)[$<\Gamma^\star+{\mathcal{H}}(\boxminus)$]{} (8.5,6)[$\leq \Gamma^\star+{\mathcal{H}}(\boxminus)$]{} (0,5) (7,7.2) (-4.5,3) (11,-.5) (0,5)[(3,1)[6.8]{}]{} (7,7.2)[(1,-2)[3.5]{}]{} (0,5)[(-2,-1)[3.5]{}]{}
Think of ${\mathcal{P}}^\star$ as the set of configurations where the dynamics, on its way from $\boxminus$ to $\boxplus$, is ‘almost at the top’, and of ${\mathcal{C}}^\star$ as the set of configurations where it is ‘at the top and capable over crossing over’. We refer to ${\mathcal{P}}^{\star}$ as the *protocritical set* and to ${\mathcal{C}}^\star$ as the *critical set*. Uniqueness follows from the observation that if $({\mathcal{P}}_1^\star,{\mathcal{C}}_1^\star)$ and $({\mathcal{P}}_2^\star,{\mathcal{C}}_2^\star)$ both satisfy conditions (1)–(3), then so does $({\mathcal{P}}_1^\star\cup{\mathcal{P}}_2^\star,{\mathcal{C}}_1^\star
\cup{\mathcal{C}}_2^\star)$. Note that $$\begin{array}{lll}
&{\mathcal{H}}(\xi)<\Gamma^\star+{\mathcal{H}}(\boxminus) &\forall\,\xi \in {\mathcal{P}}^\star,\\[0.1cm]
&{\mathcal{H}}(\xi)=\Gamma^\star+{\mathcal{H}}(\boxminus) &\forall\,\xi \in {\mathcal{C}}^\star.
\end{array}$$
It is shown in Bovier and den Hollander [@BdH15 Chapter 16] that *subject to hypothesis* (H) the following three theorems hold.
\[thm:critdrop\] $\lim_{\beta\to\infty} P^{G,\beta}_\boxminus(\tau_{{\mathcal{C}}^\star}<\tau_\boxplus
\mid \tau_\boxplus < \tau_\boxminus)=1$.
\[thm:nucltime\] There exists a $K^\star \in (0,\infty)$ such that $$\lim_{\beta\to\infty}
e^{-\beta\Gamma^\star}\,E^{G,\beta}_\boxminus(\tau_\boxplus) = K^\star.$$
\[thm:explaw\] (a) $\lim_{\beta\to\infty} \lambda^{G,\beta}_\beta\,E^{G,\beta}_\boxminus(\tau_\boxplus)=1$.\
(b) $\lim_{\beta\to\infty} P^{G,\beta}_\boxminus(\tau_\boxplus/
E^{G,\beta}_\boxminus(\tau_\boxplus)>t) = e^{-t}$ for all $t \geq 0$.
The proofs of Theorems \[thm:critdrop\]–\[thm:explaw\] in [@BdH15] do not rely on the details of the graph $G$, provided it is finite, connected and non-oriented (i.e., allowed moves are possible in both directions). For concrete choices of $G$, the task is to verify hypothesis (H) and to identify the triple (see Fig. \[fig-abstractdoublewell\]) $$\label{eq:triple}
\big({\mathcal{C}}^\star,\Gamma^\star,K^\star\big).$$ For lattice graphs this task has been carried out successfully (even for several classes of dynamics: see [@BdH15 Chapters 17–18]). For random graphs, however, the triplet in is random, and describing it represents a *very serious challenge*. In what follows we focus on a particular class of random graphs called the *Configuration Model*. But before doing so, we first summarise what is known in the literature.
(8,6)(0,0) (0,0)[(11,0)[11]{}]{} (0,0)[(0,9)[9]{}]{} (3.3,3.8)(3.3,1.9)(3.3,0) (4.8,5.6)(4.8,3)(4.8,0) (6.8,3.0)(6.8,1.5)(6.8,0) (0,5.8)(2.5,5.8)(4.8,5.8) [(2,8)(3,2)(4,5) (4,5)(5,7)(6,4) (6,4)(7,2)(8,5) ]{} (2.2,-1.3)[$\Omega_\mathrm{meta}$]{} (4.55,-.9)[${\mathcal{C}}^\star$]{} (5.8,-1.3)[$\Omega_\mathrm{stab}$]{} (11.5,-.3)[$\xi$]{} (-1.3,9.5)[${\mathcal{H}}(\xi)$]{} (-1.4,5.6)[$\Gamma^\star$]{} (3.3,4) (4.8,5.75) (6.8,3.20)
Examples of applications {#literature}
------------------------
### Torus
If the underlying graph is a torus, then the computations needed to identify the critical set ${\mathcal{C}}^\star$ and the prefactor $K^\star$ simplify considerably. As shown in Bovier and den Hollander [@BdH15 Chapter 17], for Glauber dynamics on a finite box $\Lambda\subset{\mathbb{Z}}^2$ (wrapped around to form a torus), the set ${\mathcal{C}}^\star$ consists of all $\ell_c \times (\ell_c -1)$ quasi-squares (located anywhere in $\Lambda$ in any of the two orientations) with an extra vertex attached to one of its longest sides, where $\ell_c = \lceil \tfrac{2J}{h}\rceil$ (the upper integer part of $\tfrac{2J}{h}$). Hypothesis (H) has been verified, and the exponent and the prefactor equal $$\Gamma^\star = J(4\ell_c)-h(\ell_c(\ell_c-1)+1), \qquad
K^\star = \frac{1}{|\Lambda|}\,\frac{1}{\frac{4}{3}(2\ell_c-1)}.$$ Metastable behaviour occurs if and only if $\ell_c \in (1,\infty)$, and for reasons of parity it is assumed that $\tfrac{2J}{h} \notin {\mathbb{N}}$. Similar results apply for a torus in ${\mathbb{Z}}^3$.
### Hypercube
For Glauber dynamics on the $n$-dimensional hypercube, Jovanovski [@Jpr] gives a complete description of the set ${\mathcal{C}}^\star$ (under the assumption that $\tfrac{h}{J}
\neq\frac{a}{b}$ for some $a\in \mathbb{N}$ and $b\in \left\lbrace 1,2,\ldots,2^{n}\right\rbrace$) and shows that $$\Gamma_n^\star = \tfrac{1}{3}\left(1-\tfrac{h}{J}+\left\lceil \tfrac{h}{J}\right\rceil \right)
\left(2^{\left\lceil n-\tfrac{h}{J}\right\rceil }-4+2\epsilon\right)-\epsilon, \qquad
K_n^\star = \frac{\left\lceil \tfrac{h}{J}\right\rceil !}{n!\,2^{n-4}\left(3-\epsilon\right)},$$ with $\epsilon=\left\lceil n-h\right\rceil \mbox{mod }2$. Hypothesis (H) has been verified.
### Complete graph
For Glauber dynamics on the complete graph $K_n$, it is easy to see that any monotone path from $\boxminus$ to $\boxplus$ is an optimal path. It is straightforward to show that ${\mathcal{C}}^\star=\{U\subseteq V\colon\,|U|=n^\star\}$ with $n^\star = \lceil\frac{1}{2}
(n-1-\tfrac{h}{J})\rceil$, whenever $\frac{h}{J}$ is not an integer, and to compute $$\label{GKexp}
\Gamma_n^\star=n^\star(J(n-n^\star)-h), \qquad K_n^\star=\frac{1}{|{\mathcal{C}}^\star|}
\frac{n}{n-n^\star}.$$ Metastable behaviour occurs for any value of $h$ and $J$, provided $n$ is large enough. Hypothesis (H) is also easy to confirm by observing that every configuration lies on some optimal path. Like the hypercube, $K_n$ is an expander graph and consequently the communication height $\Gamma^\star$ grows at least linearly with the number of vertices (quadratically for $K_n$).
We can reduce the quadratic growth by introducing an interaction parameter that is inversely proportional to the size of the graph: e.g. $J=\tfrac{J'}{n}$ for some constant $J'>0$, with $h>0$ fixed. It follows that $$\Gamma_{n}^\star=n^\star\left(J'\left(\frac{n-n^\star}{n}-h\right)\right),$$ where this time $n^\star=\lceil\frac{n}{2}(1-\tfrac{h}{J'})-\frac{1}{2}\rceil$, and $K_n^\star$ is the same as in . Metastable behaviour occurs if and only if $\tfrac{h}{J'}
<1-\frac{1}{n}$.
### Erdős-Rényi random graph
Sharp results of the above type become infeasible when the graph is random. The Erdős-Rényi random graph is the result of performing bond percolation on the complete graph, and is a toy model of a graph with a random geometry. Let $\mathrm{ER}_n(p)$ denotes the resulting random graph on $n$ vertices with percolation parameter $p=f(n)/n$ for some $f(n)$ satisfying $\lim_{n\to\infty} f(n)
=\infty$, the so-called *dense* Erdős-Rényi random graph. Then, as shown in the appendix, metastable behaviour occurs for any $h,J>0$, and $$\lim_{n\to\infty} \frac{\Gamma_n^\star}{\tfrac14 Jnf(n)} = 1 \quad
\text{ in distribution under the law of } \mathrm{ER}_n(f(n)/n),
\label{eq:er1}$$ which is accurate up to leading order. The computation of ${\mathcal{C}}_n^\star$ and $K_n^\star$, however, is a *formidable task*. The reason for this is that, while allows for a small error in the energy, the set ${\mathcal{C}}_n^\star$ is made up of configurations that have *exactly* the critical energy $\Gamma_n^\star$.
When $f(n)=\lambda$ for some constant $\lambda>1$, the *sparse case*, an analysis similar to the one carried out in this paper can be used to obtain lower and upper bounds on the communication height. However, we have been unable to prove a convergence of the form in .
Configuration Model {#configurationmodel}
-------------------
In this section we recall the construction of the random *multi-graph* known as the *Configuration Model* (illustrated in Fig. \[fig:CM\]). We refer to van der Hofstad [@vdHpr Chapter 7] for further details.
Fix $n\in{\mathbb{N}}$, and let $V=\{v_1,\ldots,v_n\}$. With each vertex $v_i$ we associate a *random degree* $D_i$, in such a way that $D_1,\ldots,D_n \in {\mathbb{N}}$ are i.i.d.with marginal probability distribution $f$ conditional on the event $\{\sum_{i=1}^n D_i
= \mbox{even}\}$. Consider a uniform matching of the elements in the set of *stubs* (also called half-edges), written $$\{x_{i,j}\}_{1 \leq i \leq n,1 \leq j \leq D_i}.$$ By erasing the second label of the stubs, we can associate with it a *multi-graph* ${\mathrm{CM}}_n$ satisfying the requirement that the degree of $v_i$ is $D_i$ for $1 \leq i \leq n$. The total number of edges is $\tfrac12\sum_{i=1}^n D_i$.
Throughout the sequel we use the symbol ${\mathbb{P}}_n$ to denote the law of the random multi-graph ${\mathrm{CM}}_n$ on $n$ vertices generated by the Configuration Model. To avoid degeneracies we assume that $${d_\mathrm{min}}= \min\{k\in{\mathbb{N}}\colon\,f(k)>0\} \geq 3,
\qquad {d_\mathrm{ave}}= \sum_{k\in{\mathbb{N}}} kf(k) < \infty,$$ i.e., all degrees are at least three and the average degree is finite. In this case the graph is connected *with high probability* (w.h.p.), i.e., with a probability tending to $1$ as $n\to\infty$ (see van der Hofstad [@vdHpr]).
Main theorems {#maintheorems}
-------------
We are interested in proving hypothesis (H) and identifying the key quantities in for $G = {\mathrm{CM}}_n$, which we henceforth denote by $({\mathcal{C}}^\star_n,\Gamma^\star_n,K^\star_n)$, in the limit as $n\to\infty$.
Our first main theorem settles hypothesis (H) for small magnetic field.
\[thm:hyp\] Suppose that the inequality in equation holds. Then $$\lim_{n\to\infty} {\mathbb{P}}_n\big({\mathrm{CM}}_n \text{ satisfies {\rm (H)}}\big) = 1.$$
Our second and third main theorem provide upper and lower bounds on $\Gamma^\star_n$. Label the vertices of the graph so that their degrees satisfy $d_{1}\leq\ldots\leq d_{n}$. Let $\gamma\colon\,\boxminus\to\boxplus$ be the path that successively flips the vertices $v_{1},\ldots,v_{n}$ (in that order), and let $\ell_{m}=\sum_{i=1}^{m}d_{i}$.
\[thm:Ubound\] Define $$\label{eq-defmbar}
\bar{m} = \min\left\{1 \leq m \leq n\colon\,\ell_m\left(1-\frac{\ell_m}{\ell_n}\right)
\geq \ell_{m+1}\left(1-\frac{\ell_{m+1}}{\ell_n}\right)-\frac{h}{J}\right\} < \frac{n}{2}.$$ Then, w.h.p., $$\Gamma_n^{\star} \leq \Gamma_n^+,
\qquad \Gamma_n^+ = J\ell_{\bar{m}}\Big(1-\frac{\ell_{\bar{m}}}{\ell_n}\Big)
-h\bar{m} \pm O\big(\ell_n^{3/4}\big).$$
For $0<x\leq\tfrac12$ and $\delta>1$, define (see Fig. \[fig:functionI\]) $$\begin{aligned}
I_{\delta}\left(x\right) &= \inf\Big\{0<y\leq x\colon\\
&1<x{}^{x\left(1-1/\delta\right)}
\left(1-x\right)^{\left(1-x\right)\left(1-1/\delta\right)}
\left(1-x-y\right)^{-\left(1-x-y\right)/2}\left(x-y\right)^{-\left(x-y\right)/2}y^{-y}\Big\}.
\end{aligned}
\label{eq:functionI}$$
(0,0) – (5,0) node\[anchor=north\] [$x$]{}; (0,0) – (0,4) node\[anchor=east\] [$I_{\delta}$]{}; (0,0) node\[anchor=north\] [0]{} (2,0) node\[anchor=north\] [$\tfrac{1}{4}$]{} (4,0) node\[anchor=north\] [$\tfrac{1}{2}$]{}; (0,1.4) node\[anchor=east\] [$\tfrac{11}{250}$]{} (0,2.8) node\[anchor=east\] [$\tfrac{11}{125}$]{}; plot \[smooth\] coordinates [(0,0) (1/5,30\*0.0111) (2/5,30\*0.0205) (3/5,30\*0.0286) (4/5,30\*0.0361) (5/5,30\*0.0429) (6/5,30\*0.0490) (7/5,30\*0.0546) (8/5,30\*0.0596) (9/5,30\*0.0641) (10/5,30\*0.0683) (11/5,30\*0.0721) (12/5,30\*0.0753) (13/5,30\*0.0780) (14/5,30\*0.0805) (15/5,30\*0.0825) (16/5,30\*0.0844) (17/5,30\*0.0854) (18/5,30\*0.0864) (19/5,30\*0.0869) (20/5,30\*0.0870)]{};
\[thm:Lbound\] Define $$\label{eq-defmtilde}
\tilde{m} = \min\left\lbrace 1 \leq m \leq n\colon\, \ell_m \geq \tfrac12\ell_n\right\rbrace.$$ Then, w.h.p., $$\Gamma_n^\star \geq \Gamma_n^-, \qquad
\Gamma_n^- = J\,d_\mathrm{ave}\,I_{d_\mathrm{ave}}
\left(\tfrac12\right) n - h\tilde{m} - o(n).
\label{eq:lboundGamma}$$
Under hypothesis [(H)]{} (or the weaker version of [(H)]{} introduced in Section \[althyp\]), Theorems [\[thm:Ubound\]–\[thm:Lbound\]]{} yield the following bounds on the crossover time (see Dommers [@Dpr Proposition 2.4]): $$\label{timesandwich}
\lim_{\beta\to\infty} P_{\boxminus}^{G,\beta}\left(e^{\Gamma_n^- -\varepsilon}
\leq \tau_\boxplus \leq e^{\Gamma_n^+ +\varepsilon}\right) =1.$$
In Corollary \[cor:mvalues\] we compute $\bar{m}$, $\ell_{\bar{m}}$, $\tilde{m}$ for two degree distributions: Dirac distributions and power-law distributions. It is clear that $\tilde{m}=\lceil \tfrac12 n\rceil$ for Dirac distributions.
The bounds we have found in Theorems \[thm:Ubound\]–\[thm:Lbound\] are tight in the limit of large degrees. Indeed, by the law of large numbers we have that $$\ell_{n}\frac{\ell_{\bar{m}}}{\ell_{n}}
\left(1-\frac{\ell_{\bar{m}}}{\ell_{n}}\right)
\leq\tfrac{1}{4}\ell_{n}=\tfrac{1}{4} d_{\mathrm{ave}}\,n\left[1+o(1)\right].$$ Hence $$\frac{\Gamma_n^+}{\Gamma_n^-}
= \frac{\frac{1}{4}d_{\mathrm{ave}}\left[1+o\left(1\right)\right]
-\frac{h}{J}\frac{\bar{m}}{n}+o(1)}{d_{\mathrm{ave}}
I_{d_{\mathrm{ave}}}\left(\frac{1}{2}\right)-\frac{h}{J}\frac{\tilde{m}}{n}
-o(1)}.
\label{eq:ratioofbounds}$$ In the limit as $d_{\mathrm{ave}}\to\infty$ we have $I_{d_{\mathrm{ave}}}
\left(\frac{1}{2}\right)\to\frac{1}{4}$, in which case tends to 1.
Discussion
----------
We close this introduction by discussing our main results.
[**1.**]{} We believe that Theorem \[thm:hyp\] holds as soon as $$0 < h < (d_\mathrm{min}-1)J,$$ i.e., we believe that in the limit as $\beta\to\infty$ followed by $n\to\infty$ this choice of parameters corresponds to the *metastable regime* of our dynamics, i.e., the regime where $(\boxminus,\boxplus)$ is a *metastable pair* in the sense of [@BdH15 Chapter 8].
[**2.**]{} The scaling behaviour of $\Gamma_n^\star$ as $n\to\infty$, as well as the geometry of ${\mathcal{C}}_n^\star$ are hard to capture. We can only offer some conjectures.
\[conj:scalnucltime\] There exists a $\gamma^\star \in (0,\infty)$ such that $$\lim_{n\to\infty} {\mathbb{P}}_n\Big( \big| n^{-1} \Gamma^\star_n - \gamma^\star\big|
> \delta\Big) = 0 \qquad \forall\,\delta>0.$$
\[conj:scalcritdrop\] There exists a $c^\star \in (0,1)$ such that $$\lim_{n\to\infty} {\mathbb{P}}_n\Big( \big| n^{-1} \log|{\mathcal{C}}^\star_n| - c^\star\big|
> \delta\Big) = 0 \qquad \forall\,\delta>0.$$
\[conj:scalprefac\] There exists a $\kappa^\star \in (1,\infty)$ such that $$\lim_{n\to\infty} {\mathbb{P}}_n\Big( \big| |{\mathcal{C}}^\star_n|\,K^\star_n - \kappa^\star\big|
> \delta\Big) = 0 \qquad \forall\,\delta>0.$$
As is clear from the results mentioned in Section \[literature\], all three conjectures are true for the torus, the hypercube and the complete graph. This supports our belief that they should be true for a large class of random graphs as well.
[**3.**]{} In Section \[CMdynamic\] we will give a *dynamical construction* of ${\mathrm{CM}}_n$ in which vertices are added one at a time and edges are relocated. This leads to a *random graph process* $({\mathrm{CM}}_n)_{n\in{\mathbb{N}}}$ whose marginals respect the law of the Configuration Model. In Section \[tailprop\] we will show that this process is *tail trivial*, i.e., all events in the tail sigma-algebra $${\mathcal{T}}= \cap_{N \in {\mathbb{N}}}\, \sigma\left(\cup_{n \geq N} {\mathrm{CM}}_n\right)$$ have probablity 0 or 1. Consequently, the associated communcation height process $(\Gamma^\star_n)_{n\in{\mathbb{N}}}$ with $\Gamma^\star_n = \Gamma^\star({\mathrm{CM}}_n)$ is tail trivial as well. In particular, both $\gamma^*_- = \liminf_{n\to\infty} n^{-1}
\Gamma^\star_n$ and $\gamma^*_+ = \limsup_{n\to\infty} n^{-1} \Gamma^\star_n$ exists and are constant a.s. Theorems \[thm:Ubound\]–\[thm:Lbound\] show that $0<\gamma^*_- \leq \gamma^*_+<\infty$. Settling Conjecture \[conj:scalnucltime\] amounts to showing that $\gamma^*_-=\gamma^*_+$.
[**4.**]{} It was shown by Dommers [@Dpr] that for the Configuration Model with $f=\delta_r$, $r \in {\mathbb{N}}\backslash\{1,2\}$, i.e., for a random regular graph with degree $r$, there exist constants $0<\gamma_-^\star(r)<\gamma_+^\star(r)<\infty$ such that $$\label{eq:sandwich}
\lim_{n\to\infty} \lim_{\beta\to\infty}
{\mathbb{E}}_n\left( P^{{\mathrm{CM}}_n}_\boxminus\left( e^{\beta n\gamma_-^\star(r)}
\leq \tau_\boxplus \leq e^{\beta n\gamma_+^\star(r)}\right) \right) = 1,$$ provided $\frac{h}{J} \in (0,C_0\sqrt{r})$ for some constant $C_0 \in (0,\infty)$ that is small enough. Moreover, there exist constants $C_1 \in (0,\tfrac14\sqrt{3})$ and $C_2 \in
(0,\infty)$ (depending on $C_0$) such that $$\gamma_-^\star(r) \geq \tfrac14 Jr - C_1J\sqrt{r}, \qquad
\gamma_+^\star(r) \leq \tfrac14 Jr + C_2J\sqrt{r}, \qquad r \in {\mathbb{N}}\backslash\{1,2\}.$$ The result in is derived without hypothesis (H), but it is shown that hypothesis (H) holds as soon as $r \geq 6$.
[**Outline.**]{} The rest of the paper is organised as follows. In Section \[metstate\] we prove that hypothesis (H) holds under certain constraints on the magnetic field $h$ and the minimal degree of the graph $d_\mathrm{min}$. Section \[althyp\] gives an alternative to hypothesis (H), which holds for a broader range parameters, yet still permits us to claim our bounds on the crossover time. In Section \[thmproofs\] we prove our upper and lower bounds on $\Gamma_n^\star$. Part of this proof depends on a *dynamical construction* of ${\mathrm{CM}}_n$. In Section \[tailprop\] we derive certain properties of this construction.
Proof of Theorem \[thm:hyp\] {#metstate}
============================
This section gives a proof of hypothesis (H). We start with the following remark about the configurations in $\Omega$.
[A natural isomorphism between configurations and subsets of vertices of the underlying graph $G=(V,E)$ comes from identifying $\xi\in\Omega$ with the set $\{v \in V\colon\,\xi(v)=+1\}$. With this in mind, we denote by $\overline{\xi}$ the configuration corresponding to the complement of this set: $\{v\in V\colon\,\xi(v)=-1\}$. Furthermore, for $\zeta,\sigma\in\Omega$ we denote by $E(\zeta,\sigma) \subseteq E$ the set of all unoriented edges $\{(v,w)\in E\colon\,\zeta(v)=\sigma(w)=+1\}$. The main use of the last definition will be for $\sigma=\overline{\zeta}$: $E(\zeta,\overline{\zeta})$ is the edge boundary of the set $\{v\in V\colon\,\zeta(v)=+1\}$.]{}
We next give two lemmas that will be useful later on.
\[lem:Ideltabound\] For all $\delta \geq 2$ and $0<x\leq\tfrac12$, $I_{\delta}\left(x\right) \leq \left(1-x\right)
-\left(1-x\right)^{2\left(1-1/\delta\right)}$.
The claim can be verified numerically. For $w \in \left(0,\frac{1}{2}\right]$, let $\tilde{y}
=\left(1-x\right)-\left(1-x\right)^{2\left(1-w\right)}$. Fig. \[fig:cplot\] gives a contour plot of the function $$\tilde{I}\left(x,w\right) = x^{x\left(1-w\right)}\left(1-x\right)^{\left(1-x\right)
\left(1-w\right)}\left(1-x-\tilde{y}\right)^{-\left(1-x-\tilde{y}\right)/2}
\left(x-\tilde{y}\right)^{-\left(x-\tilde{y}\right)/2}\tilde{y}^{-\tilde{y}}.$$
\[htbp\]
Note that $\tilde{I}\left(x,w\right)\geq 1$, which immediately implies Lemma \[lem:Ideltabound\] when we take $w=1/\delta$. It is easy to verify that the boundary values corresponding to $x \downarrow 0$ and $x \uparrow 1$ result in $\tilde{I}\left(x,w\right) \downarrow 1$.
\[lem:Ideltamonotone\] The function $x\rightarrow\frac{I_{\delta}\left(x\right)}{x}$ is non-increasing on $\left(0,\frac{1}{2}\right]$.
By definition of $I_{\delta}\left(x\right)$, the function $$\hat{I}\left(x,z\right) = x^{x\left(1-1/\delta\right)}\left(1-x\right)^{\left(1-x\right)
\left(1-1/\delta\right)}\left(1-x-xz\right)^{-\left(1-x-xz\right)/2}
\left(x-xz\right)^{-\left(x-xz\right)/2}\left(xz\right)^{-xz}$$ satisfies $\hat{I}(x,\frac{I_{\delta}\left(x\right)}{x})=1$ for all $x\in\left(0,\frac{1}{2}\right]$. It will therefore suffice to show that $$\label{Itask}
\frac{\partial}{\partial x}\vert_{z=\frac{I_{\delta}\left(x\right)}{x}} \hat{I}\left(x,z\right) \geq 0,$$ since this implies that, for $\epsilon$ sufficiently small, $\hat{I}(x+\epsilon,\frac{I_{\delta}
\left(x\right)}{x})\geq1$, and hence that $$\inf\left\{w\colon\,\hat{I}\left(x+\epsilon,w\right)\geq1\right\} \leq\frac{I_{\delta}\left(x\right)}{x},$$ and thus $\frac{I_{\delta}(x+\epsilon)}{x+\epsilon} \leq \frac{I_{\delta}\left(x\right)}{x}$. Observe that $$\begin{aligned}
&\frac{\partial}{\partial x}\hat{I}\left(x,z\right)\\
&= \hat{I}\left(x,z\right)
\left\{ \log\left(\left(\frac{x}{1-x}\right)^{\left(1-1/\delta\right)}
\left(1-x-xz\right)^{\left(1+z\right)/2}\left(x-xz\right)^{-\left(1-z\right)/2}
\left(xz\right)^{-z}\right)\right\}.
\end{aligned}$$ For $z=\frac{I_{\delta}\left(x\right)}{x}$, $\hat{I}\left(x,z\right)=1$ implies $$\left(x^{x\left(1-1/\delta\right)}\left(1-x\right)^{\left(1-x\right)
\left(1-1/\delta\right)}\right)^{\frac{1}{x}}=\left(1-x-xz\right)^{\frac{1}{2x}
-\left(1+z\right)/2}\left(x-xz\right)^{\left(1-z\right)/2}\left(xz\right)^{z}$$ and hence $$\begin{aligned}
&\frac{\partial}{\partial x}\hat{I}\left(x,z\right)\\
&=\hat{I}\left(x,z\right)\left\{ \log\left(\left(\frac{x}{1-x}\right)^{\left(1-1/\delta\right)}
\left(x^{x\left(1-1/\delta\right)}\left(1-x\right)^{\left(1-x\right)
\left(1-1/\delta\right)}\right)^{-\frac{1}{x}}\left(1-x-xz\right)^{\frac{1}{2x}}\right)\right\}.
\label{eq:derivIhat}
\end{aligned}$$ The term inside the logarithm in (\[eq:derivIhat\]) simplifies to $\left(1-x\right)^{-\frac{1}{x}
\left(1-1/\delta\right)}\left(1-x-xz\right)^{\frac{1}{2x}}$, which satisfies $\left(1-x\right)^{-\frac{1}{x}\left(1-1/\delta\right)}\left(1-x-xz\right)^{\frac{1}{2x}}\geq1$ whenever $1-x-\left(1-x\right)^{2\left(1-1/\delta\right)} \geq xz=I_{\delta}\left(x\right)$. By Lemma \[lem:Ideltabound\], this is true for all $x\in\left(0,\frac{1}{2}\right]$, and so follows.
We can now proceed with the proof of hypothesis (H). Let $\sigma\in\Omega$ be any configuration that satisfies $x=\ell_{\sigma}/\ell_n\leq\tfrac12$, where $\ell_\sigma
=\sum_{i\in\sigma} d_i$. We will construct a path from $\sigma$ to some $\sigma\prime
\in\Omega$ satisfying $\mathcal{H}\left(\sigma\prime\right)<\mathcal{H}\left(\sigma\right)$ by removing one vertex at a time, obtaining a path $\sigma=\sigma_{0},\ldots,\sigma_{m}
=\sigma\prime$. In particular, at step $t$ we remove any vertex $v_{t}\in\sigma_{t-1}$ that minimises the quantity $\left|E\left(v_{t},\sigma_{t-1}\backslash v_{t}\right)\right|
-\left|E\left(v_{t},\overline{\sigma_{t-1}}\right)\right|$. It will follow that for every $\sigma_{i}$ in this path, we have $\left|\mathcal{H}\left(\sigma_{i}\right)-\mathcal{H}\left(\sigma_{0}
\right)\right|<\Gamma^{\star}$, which proves the claim of the theorem.
The probability that some configuration $\sigma$, chosen uniformly from all configurations in $\Omega$ with $\ell_\sigma=L$, has a boundary of size $\left|E\left(\sigma,
\overline{\sigma}\right)\right|=K$ equals $$\begin{aligned}
&{L \choose K}K!\left(L-K-1\right)!!{\ell_{n}-L \choose K}
\left(\ell_{n}-L-K-1\right)!!/\left(\ell_{n}-1\right)!! \\
&\approx \frac{\left(L\right)!}{\left(L-K\right)!!}
\frac{\left(\ell_{n}-L\right)!}{K!\left(\ell_{n}-L-K\right)!!}\frac{1}{\ell_{n}!!} \\
&\approx L^L\left(\ell_{n}-L\right)^{\left(\ell_{n}-L\right)}
\left(\ell_{n}-L-K\right)^{-\left(\ell_{n}-L-K\right)/2}
\left(L-K\right)^{-\left(L-K\right)/2}K^{-K}\ell_{n}^{-\ell_{n}/2},
\end{aligned}$$ where the symbol $\approx$ stands for equality up to polynomial terms (here of order $O(n^{2})$). Let $x=L/\ell_{n}$ and $y=K/\ell_{n}$, so that the above expression becomes $$\exp\left[\ell_{n}\log\left(x{}^{x}\left(1-x\right)^{\left(1-x\right)}
\left(1-x-y\right)^{-\left(1-x-y\right)/2}\left(x-y\right)^{-\left(x-y\right)/2}y^{-y}\right)\right].$$ Furthermore, if we define $\eta\left(x\right)$ by $$\label{eq:defeta}
\exp\left[\ell_{n}\log\eta\left(x\right)\right]
=\left|\left\{ U\subseteq V\colon\,\ell_U=\ell_{n}x\right\} \right|,$$ then the probability of there being any configuration of total degree $L$ having a boundary size $K$ is bounded from above by $$\label{sd-expbound}
\exp\left[\ell_{n}\log\left(\eta\left(x\right)x{}^{x}\left(1-x\right)^{\left(1-x\right)}
\left(1-x-y\right)^{-\left(1-x-y\right)/2}\left(x-y\right)^{-\left(x-y\right)/2}y^{-y}\right)\right].$$
It is easy to see that, by using $\delta=d_{\min}$, the cardinality in the right-hand side of is bounded from above by ${\ell_{n}/\delta \choose x\ell_{n}/\delta}$. Using Stirling’s approximation for this term, and substituting in , we get $$\label{eq:probineq}
\begin{aligned}
&\mathbb{P}\left[\exists\, A\subseteq V\colon\,\ell_{A}=x\ell_{n}
\mbox{ and }\left|E\left(A,\overline{A}\right)\right|=y\ell_{n}\right]\\
&\leq \exp\left[\ell_{n}\log\left(x{}^{x\left(1-1/\delta\right)}
\left(1-x\right)^{\left(1-x\right)\left(1-1/\delta\right)}
\left(1-x-y\right)^{-\left(1-x-y\right)/2}\left(x-y\right)^{-\left(x-y\right)/2}y^{-y}\right)\right].
\end{aligned}$$ Recall the definition of $I_{\delta}$ from and note that is exponentially small for $y<I_{\delta}\left(x\right)$, and by a union bound it is exponentially small for all such $y$.
Suppose that after $s$ vertices have been removed, we reach a configuration $\sigma_{s}$ with $\mathcal{H}\left(\sigma_{s}\right) < \mathcal{H}\left(\sigma\right)$, such that for every vertex $v\in\sigma_{s}$ we have $$\left|E\left(v,\sigma_{s}\backslash v\right)\right|+\frac{h}{J}
>\left|E\left(v,\overline{\sigma_{s}}\right)\right|.
\label{eq:afters}$$ In other words, equation (\[eq:afters\]) states that after removing $s$ vertices we are at a configuration of lower energy, and removing any additional vertex leads to a configuration of higher energy. Note that if no such $s$ exists, then we keep on removing vertices until $\boxminus$ has been reached. By the assumption that $h$ is sufficiently small (by , it would suffice if $h<\tfrac12\,
J\,d_\mathrm{ave}\,I_{d_\mathrm{ave}} \left(\tfrac12\right) \frac{n}{\tilde{m}}$, where $\tilde{m}$ was also defined in the aforementioned equation), w.h.p., every configuration $\sigma$ of total degree $\ell_{\sigma}\leq\tfrac12 \ell_n$ satisfies $\mathcal{H}
\left(\sigma\right)>\mathcal{H}\left(\boxminus\right)$. If $v\in\sigma_{s}$ has no self-loops, then we have $$\left|E\left(v,\sigma_{s}\backslash v\right)\right|
=d_{v}-\left|E\left(v,\overline{\sigma_{s}}\right)\right|$$ and thus the condition in is satisfied when for all $v\in\sigma_{s}$, $$\label{eq:}
\tfrac12\left(d_{v}+\frac{h}{J} \right) > \left|E\left(v,\overline{\sigma_{s}}\right)\right|.$$ The total number of vertices with self-loops is w.h.p. of order $o(n)$, and so it will be evident from the bounds below that this assumption is immaterial. The second inequality in $$\left|E\left(\sigma_{s},\overline{\sigma_{s}}\right)\right|
< \tfrac12 \sum_{v\in\sigma_{s}} \left( d_{v}+\frac{h}{J} \right)
= \tfrac12 \bigg(x\ell_{n}-\sum_{i=1}^{s}d_{i}+\frac{h\left(\left|\sigma\right| -s\right)}{J} \bigg)
\leq\left|E\left(\sigma,\overline{\sigma}\right)\right|$$ holds whenever $$x\ell_{n}-2\left|E\left(\sigma,\overline{\sigma}\right)\right|+\frac{h\left(\left|\sigma\right| -s\right)}{J}
\leq \sum_{i=1}^{s} d_{i},$$ which in particular is true when we take the smallest $s$ such that $$\sum_{i=1}^{s} d_{i} \geq x\ell_{n}-2I_{\delta}\left(x\right)\ell_{n}
+\frac{h\left(\left|\sigma\right| -s\right)}{J}.
\label{eq:condits}$$ Furthermore, by removing $s$ vertices, the change in the size of the boundary at step $t$ is given by $$\begin{aligned}
&\left|E\left(\sigma_{t},\overline{\sigma_{t}}\right)\right|
-\left|E\left(\sigma,\overline{\sigma}\right)\right|
= \sum_{i=1}^{t}\big(\left|E\left(v_{i},\sigma_{i-1}\backslash v_{i}\right)\right|
-\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\big)\\
&\qquad \leq \sum_{i=1}^{t}\left(d_{i}-2\left\lceil d_{i}\,
\frac{I_{\delta}\left(x\right)}{x}\right\rceil \right)
\leq \left(1-2\,\frac{I_{\delta}\left(x\right)}{x}\right)\sum_{i=1}^{t} d_{i}.
\label{eq:stepssum}
\end{aligned}$$ The first inequality in follows from the following observation: note that w.h.p. $|E(\sigma,\overline{\sigma})|$ $\geq I_{\delta}(x)\ell_n$, and hence the “proportion” of the total degree of $\sigma$ that is paired with vertices in $\overline{\sigma}$ is at least $I_{\delta}(x)/x$. This implies that there must be some vertex $v_i$ with a proportion of at least $I_{\delta}(x)/x$ of its degree connected with vertices in $\overline{\sigma}$. In other words, $v_i$ shares at least $\lceil d_i \frac{I_{\delta}(x)}{x}\rceil$ edges with $\overline{\sigma}$.
By the definition of $s$, we that $|E(\sigma,\overline{\sigma})|=|E(\sigma_s,\overline{\sigma_s})|
+o(n)$ (when $d_s=o(n)$), and hence dropping the $o(n)$-term is of no consequence in the following computations. This implies that if $t$ is such that $|E(\sigma_t,\overline{\sigma_t})|
-|E(\sigma,\overline{\sigma})|$ is maximised, then we get (again, possibly after dropping a term of order $o(n)$) $$\begin{aligned}
\sum_{i=1}^t \big(\left|E\left(v_{i},\sigma_{i-1}\right)\right|
-\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\big)
& = & \sum_{i=t+1}^s \big(\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|
-\left|E\left(v_{i},\sigma_{i-1}\right)\right|\big).
\label{eq:peakequality}
\end{aligned}$$ Let $m_{t}$ denote the left-hand side of $\eqref{eq:peakequality}$, so that $$\begin{aligned}
m_{t} &= \sum_{i=1}^{t}\left(d_{i}-2\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\right)
= \sum_{i=t+1}^{s}\left(d_{i}-2\left|E\left(v_{i},\sigma_{i-1}\right)\right|\right)\\
&= \sum_{i=1}^{s}\left(d_{i}-2\left|E\left(v_{i},\sigma_{i-1}\right)\right|\right)
-\sum_{i=1}^{t}\left(d_{i}-2\left|E\left(v_{i},\sigma_{i-1}\right)\right|\right).
\end{aligned}$$ Hence $$\sum_{i=1}^{t}d_{i}\left(\frac{\sum_{i=1}^{t}\left(d_{i}
-2\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\right)}{\sum_{i=1}^{t}d_{i}}+1\right)
=\sum_{i=1}^{s}d_{i}-\sum_{i=t+1}^{s}2\left|E\left(v_{i},\sigma_{i-1}\right)\right|,$$ and thus $$\begin{aligned}
m_{t} &= \left(\frac{\sum_{i=1}^{t}\left(d_{i}
-2\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\right)}
{\sum_{i=1}^{t}d_{i}}\right)\left(\frac{\sum_{i=1}^{t}\left(d_{i}
-2\left|E\left(v_{i},\overline{\sigma_{i-1}}\right)\right|\right)}
{\sum_{i=1}^{t}d_{i}}+1\right)^{-1}\\
&\qquad\qquad\qquad \times \left(\sum_{i=1}^{s}d_{i}
-\sum_{i=t+1}^{s}2\left|E\left(v_{i},\sigma_{i-1}\right)\right|\right)\\
&\leq \frac{1}{2}\left(1-2\frac{I_{\delta}\left(x\right)}{x}\right)
\left(1-\frac{I_{\delta}\left(x\right)}{x}\right)^{-1}\left(x\ell_{n}
-2I_{\delta}\left(x\right)\ell_{n}+\frac{h\left(\left|\sigma\right|-s\right)}{J}\right),
\end{aligned}$$ where for the last inequality we use – and the monotonicity of $y\rightarrow y(y+1)^{-1}$. From , using the fact that $nd_{\mathrm{ave}}=\ell_{n}+o\left(n\right)$, we get that $ \mathcal{H}\left(\sigma_{t}\right) - \mathcal{H}\left(\sigma\right) < \Gamma^{\star}$ whenever $$\frac{h}{J\ell_{n}}\left(2\tilde{m}+t+\left(\left|\sigma\right|
-s\right)\left(\frac{x-2I_{\delta}\left(x\right)}{x-I_{\delta}\left(x\right)}\right)\right)
<2I_{d_{\mathrm{ave}}}\left(\tfrac{1}{2}\right)
-\left(x-2I_{\delta}\left(x\right)\right)^{2}\left(x-I_{\delta}\left(x\right)\right)^{-1}.
\label{eq:ihalfbound}$$ Note that if $x<2I_{\delta}\left(x\right)$, then for sufficiently small $h$ we can find a monotone downhill path to $\boxminus$. More precisely, $x<2I_{\delta}\left(x\right)$ implies that the terms in the right-hand side of become negative, and hence for $\frac{h}{J}<d_{\min}(\frac{2I_{\delta}\left(x\right)}{x}-1)$ every step in our path is a downhill step. For $x\geq2I_{\delta}\left(x\right)$, observe first that since the function $u\rightarrow\frac{\left(1-2u\right)^{2}}{\left(1-u\right)}$ is non-increasing for $u\leq\frac{1}{2}$, by Lemma \[lem:Ideltamonotone\] $$\frac{\left(x-2I_{\delta}\left(x\right)\right)^{2}}{\left(x-I_{\delta}\left(x\right)\right)}
= x\frac{(1-2\frac{I_{\delta}\left(x\right)}{x})^{2}}{(1-\frac{I_{\delta}\left(x\right)}{x})}
\leq \tfrac{1}{2}\frac{(1-4I_{\delta}(\tfrac{1}{2}))^{2}}
{(1-2I_{\delta}\left(\tfrac{1}{2}\right))},$$ and thus a sufficient condition for to hold is $$\tfrac{h}{J}\left(\tfrac{1}{d_{\mathrm{ave}}} + \tfrac{1}{2} \right)
< 2I_{d_{\mathrm{ave}}}\left(\tfrac{1}{2}\right)
-\tfrac{1}{2}\left(1-4I_{d_{\mathrm{min}}}\left(\tfrac{1}{2}\right)\right)^{2}
\left(1-2I_{d_{\mathrm{min}}}\left(\tfrac{1}{2}\right)\right)^{-1}.
\label{eq:Hcond1}$$ Hence we have a path $\sigma \to \sigma_{s}$ (or, when such an $s$ satisfying does not exist, a path $\sigma \to \boxminus$) with $\mathcal{H}\left(\sigma_{s}\right)
< \mathcal{H}\left(\sigma\right)$ that never exceeds $\mathcal{H}\left(\sigma\right)$ by $\Gamma^{\star}$ or more, whenever $h$ is sufficiently small and holds. This proves the claim of the theorem for all configurations $\sigma$ with $\ell_{\sigma}
\leq \tfrac12\ell_n$.
Note also that, for $\ell_{\sigma} > \tfrac12\ell_n$, the same argument can be repeated by adding a vertex at each step, which will also come at a lower cost since at each step the magnetisation changes by $-h$.
An alternative to hypothesis (H) {#althyp}
================================
In this section gives a weaker version of hypothesis (H), which nonetheless suffices as a prerequisite for Theorem \[thm:nucltime\]. This weaker version can be verified for a parameter range that is larger than the one needed in Section \[metstate\].
We can repeat the arguments given in Section \[metstate\]. But, instead of insisting that $\mathscr{V}_{\sigma}<\Gamma^{\star}$ for every configuration $\sigma \in \Omega$, we require that $\mathscr{V}_{\sigma}$ is bounded from above by our upper bound on $\Gamma^{\star}$, since this guarantees that our upper bound on the crossover time is still valid and still holds (see Dommers [@Dpr Lemma 5.3]). Thus, it follows from the arguments leading to that we only need the condition $$\frac{h}{J\ell_{n}}\left(\bar{m}+t+\left|\sigma\right|
-s\left(\frac{x-2I_{\delta}\left(x\right)}{x-I_{\delta}\left(x\right)}\right)\right)
\leq 2\frac{\ell_{\bar{m}}}{\ell_n}\Big(1-\frac{\ell_{\bar{m}}}{\ell_n}\Big)
-\left(x-2I_{\delta}\left(x\right)\right)^{2}\left(x-I_{\delta}\left(x\right)\right)^{-1}.
\label{eq:ihalfbound2}$$ For $h$ sufficiently small, the ratio $\frac{\ell_{\bar{m}}}{\ell_n}$ can be made arbitrarily close to $\tfrac12$, in which case the right-hand side of becomes strictly positive. This implies that the inequality in holds for any $\delta \geq 3$ whenever $h$ is sufficiently small.
Proof of Theorems \[thm:Ubound\] and \[thm:Lbound\] {#thmproofs}
===================================================
A dynamic construction of the configuration model {#CMdynamic}
-------------------------------------------------
Prior to giving the proof of Theorems \[thm:Lbound\] and \[thm:Ubound\], we introduce a *dynamical construction* of the CM graph. This will be used to obtain the upper bound in Theorem \[thm:Ubound\].
Let $V=\{v_{i}\} _{i=1}^{n}$ be a sequence of vertices with degrees $\{d_{i}\} _{i=1}^{n}$. In this section we construct a graph $G=(V,E)$ with the same distribution as a graph generated through the Configuration Model algorithm, but in a *dynamical way*, as follows.
Suppose that $\xi_{m}$ is a uniform random matching of the integers $\{1,\ldots,2m\} $, denoted by $\xi_{m}=\{(x_{1},x_{2}),\ldots,(x_{2m-1},
x_{2m}\}$, where the pairs are listed in the order they were created (which is not an important issue, so long as we agree on some labeling). Next, let $u$ be uniform on $\{1,\ldots,2m,2m+1\} $ and set $\xi_{m+1}
=\xi_{m}\cup\{(2m+2,u)\}$ if $u=2m+1$. Else if $u \neq 2m+1$, then w.l.o.g.$u=x_{2i-1}$ for some $i\leq m$, and we set $\xi_{m+1}=\{\xi_{m}\backslash
\{(x_{2i-1},x_{2i})\}\}\cup\{(2m+2,x_{2i-1}),(2m+1,x_{2i})\}$. Then $\xi_{m+1}$ is a uniform matching of the points $\{1,\ldots,2m,2m+2\}$. It is now obvious how the construction of $G$ follows from the given scheme.
Energy estimates
----------------
Label the vertices of the graph so that their degrees satisfy $d_{1}\leq\ldots\leq
d_{n}$. Let $\gamma\colon\,\boxminus\to\boxplus$ be the path that successively flips the vertices $v_{1},\ldots,v_{n}$ (in that order), and let $\ell_{m}=\sum_{i=1}^{m}
d_{i}$. We show that, w.h.p., for every $1\leq m\leq n$, $$\mathcal{H}\left(\gamma_{m}\right)-\mathcal{H}\left(\boxminus\right)
=J\ell_{m}\Big(1-\frac{\ell_{m}}{\ell_{n}}\Big)-mh \pm O\big(\ell_{n}^{3/4}\big).
\label{eq:pathbound}$$
We are particularly interested in the maximum of over all $1\leq m \leq n$. To this avail, observe that the function defined by $$g\left(x\right) = Jx\left(1-x\right)-h\left(x\right)
\label{eq:functiong}$$ has at most one maximum for $x\in[0,1]$ if $x \mapsto h(x)$ is non-decreasing. Thus, taking $x=\frac{\ell_{m}}{\ell_{n}}$ and $h(x)=hx\frac{m}{\ell_{m}}$, we see that our definition of $\bar{m}$ in Theorem \[thm:Ubound\] is justified. Furthermore, note the equivalent conditions $$\ell_{m}\left(1-\frac{\ell_{m}}{\ell_{n}}\right)
\geq \ell_{m+1}\left(1-\frac{\ell_{m+1}}{\ell_{n}}\right)-\frac{h}{J}
\iff \frac{h}{J} \geq d_{m+1}\left(1-\frac{2\ell_{m}}{\ell_{n}}\right)
-O\left(\frac{d_{m}^{2}}{\ell_{n}}\right),
\label{eq:equivalentconditions}$$ with the last term in disappearing whenever $d_{m+1}
=o\left(\sqrt{\ell_{n}}\right)$. Note that gives us an alternative formulation for $\bar{m}$ in the statement of Theorem \[thm:Ubound\], which we will use to compute $\bar{m}$ below.
Two examples
------------
Two commonly studied degree distributions for the Configuration Model are the Dirac distribution $$q_r(k) = \delta_r(k), \qquad k \in {\mathbb{N}}_0,
\label{eq:Diracdist}$$ for some $r \in {\mathbb{N}}$ (i.e., the $r$-regular graph), and the power-law distribution $$q_{\tau,\delta}(k) = \mathbb{P}\left[d_{i}=\delta+k\right]
= \frac{\left(\delta+k\right)^{-\tau}}{\sum_{i\in{\mathbb{N}}_0}
\left(\delta+i\right)^{-\tau}}, \qquad k \in {\mathbb{N}}_0,
\label{eq:powerlawdist}$$ for some exponent $\tau \in (2,\infty)$ and shift $\delta \in {\mathbb{N}}$.
For these degree distributions we get the following corollary of Theorems \[thm:Ubound\]–\[thm:Lbound\]:
\[cor:mvalues\] (a) For the Dirac-distribution in , $$Jr I_r\left(\tfrac12\right)n -\frac{hn}{2}-o(n) \leq \Gamma_{n}^{\star} \leq \frac{Jr}{4}
n \left(1-\left(\frac{h}{Jr}\right)^2\right) \pm O\big(n^{3/4}\big).$$ (b) For the power-law distribution distribution in , $\bar{m}$ and $\ell_{\bar{m}}$ are given by and .\
(c) For the power-law distribution distribution in , $\tilde{m}$ is given by .
\(a) Straightforward.\
(b) $\{d_{i}\} _{i=1}^{n}$ are i.i.d. with degree distribution $q_{\tau,\delta}$. Let $s_{\tau,\delta,k} = \sum_{i=1}^{n} \mathbf{1}\left\{d_{i}\leq \delta + k\right\}$, and note that $$\mathbb{E}\left[s_{\tau,\delta,k}\right]
=n\left(1-\frac{\xi_{\tau}\left(\delta+k+1\right)}{\xi_{\tau}\left(\delta\right)}\right)$$ with $\xi_{\tau}\left(a\right)=\sum_{i=a}^{\infty}i^{-\tau}$ for $a\geq0$. We claim that, for $k$ sufficiently small, $s_{\tau,\delta,k}$ is concentrated around its mean. Indeed, define $a_{\delta,k}=\sum_{i}\mathbf{1}\left\{ d_{i}=\delta+k\right\}$, $k\in{\mathbb{N}}_0$, and note that for any i.i.d. sequence we have $a_{\delta,k}\stackrel{d}{=}\mathrm{Bin}\left(n,
p_{\delta,k}\right)$, where $p_{\delta,k}=\mathbb{P}\left[d_{i}=\delta+k\right]$. From Hoeffding’s inequality we get that $$\mathbb{P}\left[\left|a_{\delta,k}-np_{\delta,k}\right|>n^{\frac{1}{2}+\frac{1}{6}}\right]
\leq\exp\left(-2n^{\frac{1}{3}}\right).$$ Hence, for any $k=O(n^{1/6})$, $$\begin{aligned}
\mathbb{P}\left[\left|s_{\tau,\delta,k}-\mathbb{E}
\left[s_{\tau,\delta,k}\right]\right|>n^{\frac{1}{2}+\frac{1}{3}}\right]
&\leq \mathbb{P}\left[\bigcup_{m=0}^{k}
\left|a_{\delta,m}-np_{\delta,m}\right|>n^{\frac{1}{2}+\frac{1}{6}}\right]
\leq n^{\frac{1}{6}}\exp\left(-2n^{\frac{1}{3}}\right).
\label{eq:bnd-s}
\end{aligned}$$ Note that if $p_{\delta,k}=q_{\tau,\delta}(k)$, then $\mathbb{E}[a_{\delta,k}]
=n\,\frac{(\delta+k)^{-\tau}}{\xi_{\tau}(\delta)}$, and w.h.p. $\ell_{n}=n
\frac{\xi_{\tau-1}(\delta)}{\xi_{\tau}(\delta)}+o(n)$. Hence we define $$\kappa = \min\left\{k\in{\mathbb{N}}\colon\,\frac{h}{J}\geq\left(\delta+k-1\right)
\left(1-\left(\frac{\xi_{\tau-1}\left(\delta+k\right)}
{\xi_{\tau-1}\left(\delta\right)}\right)\right)\right\} -1,
\label{eq:defkappa}$$ which certainly satisfies $\kappa= o\left(n^{1/6}\right)$. From (\[eq:equivalentconditions\]) and the monotonicity of (\[eq:functiong\]) it follows that w.h.p. $$\begin{aligned}
\bar{m}
&=n\Bigg[\left(1-\frac{\xi_{\tau}\left(\delta+\kappa\right)}{\xi_{\tau}\left(\delta\right)}\right)\\
&+\min\left\{ y\in\left[0,1\right]\colon\,\frac{h}{J}\geq\left(\delta+\kappa\right)
\left(1-\left(\frac{\xi_{\tau-1}\left(\delta+\kappa\right)}{\xi_{\tau-1}\left(\delta\right)}
+\frac{y\kappa\xi_{\tau}\left(\delta\right)}{\xi_{\tau-1}
\left(\delta\right)}\right)\right)\right\} \Bigg]+o\left(n\right)
\label{eq:mprimeval}
\end{aligned}$$ and $$\frac{\ell_{\bar{m}}}{\ell_{n}}
=\left(\frac{\xi_{\tau-1}\left(\delta+\kappa\right)}{\xi_{\tau-1}\left(\delta\right)}
+\frac{y\kappa\xi_{\tau}\left(\delta\right)}{\xi_{\tau-1}\left(\delta\right)}\right)
+o\left(1\right),
\label{eq:lm/ln}$$ where $y$ is the taken as the argument of the minimum in (\[eq:mprimeval\]). Since we know $\ell_{n}$ up to $o(n)$, this also gives the value of $\ell_{\bar{m}}$.\
(c) Note that $\tilde{m} = \sum_{i=0}^{\kappa} a_{\delta,i}$ where $\kappa$ is the least integer such that $$\delta a_{\delta,0} + (\delta + 1)a_{\delta,0} + \ldots
+ (\delta + \kappa)a_{\delta,\kappa} \geq \tfrac12\ell_n.$$ By the concentration results given above, we see that w.h.p. $$\kappa = \min\left\lbrace m\in{\mathbb{N}}\colon\,
\frac{\xi_{\tau -1}(\delta) +\xi_{\tau -1}(\delta + m+1) }{\xi_{\tau}(\delta)}
\geq \tfrac12 d_{\mathrm{ave}} \right\rbrace$$ and $$\tilde{m} = \frac{n}{\xi_{\tau}(\delta)} \sum_{i=0}^{\kappa} (\delta + i)^{\tau} \: + o(n).
\label{eq:tildem}$$
Proof of Theorem \[thm:Ubound\]
-------------------------------
Consider a sequence of matchings $\left\{\xi_{1},\xi_{2},\ldots,\xi_{M/2}\right\}$ constructed in a dynamical way as outlined above, where $M$ is some even integer. Let $0\leq x\leq M$ be even, and define $z_{x,0}=x$ and $z_{x,t}=\sum_{i=1}^{x}
\sum_{m=1}^{x}\mathbf{1}\left\{\left(i,m\right)\in\xi_{x/2+t}\right\}$. Then $$z_{x,t+1}=z_{x,t}-2\mathbf{1}\left\{ Y_{t+1}\right\},
\label{eq:ziplus1}$$ where $Y_{t}$ is the event that $x+2t-1$ and $x+2t$ are both paired with terms in $\left[x\right]$. Note that $$\mathbb{P}\left[Y_{t+1}\left|\mathscr{G}{}_{x+t}\right.\right]=\frac{z_{x,t}}{x+2+1},
\label{eq:Ytplus1}$$ where $\mathscr{G}_{x+t}$ is the $\sigma$-algebra generated by $\left\{\xi_{1},\ldots,
\xi_{x/2+t}\right\}$. Therefore $$\mathbb{E}\left[z_{x,t+1}\left|\mathscr{G}{}_{x+t}\right.\right]
= z_{x,t}-\frac{2z_{x,t}}{x+2t+1}
= z_{x,t}\left(\frac{x+2t-1}{x+2t+1}\right)$$ and so $$\mathbb{E}\left[z_{x,t+1}\right]
= z_{0}\prod_{j=0}^{t}\left(\frac{x+2j-1}{x+2j+1}\right)
= x\left(\frac{x-1}{x+2t+1}\right).
\label{eq:Expztplus1}$$ To compute the second moment, observe that $$z_{x,t+1}^{2}=z_{x,t}^{2}-4z_{x,t}\mathbf{1}\left\{ Y_{t+1}\right\}
+4\mathbf{1}\left\{ Y_{t+1}\right\},
\label{eq:zmoment2}$$ and so $$\mathbb{E}\left[z_{x,t+1}^{2}\left|\mathscr{G}{}_{x+t}\right.\right]
= z_{x,t}^{2}-\frac{4z_{x,t}^{2}}{x+2t+1}+\frac{4z_{x,t}}{x+2t+1}
= z_{x,t}^{2}\left(\frac{x+2t-3}{x+2t+1}\right)+\frac{4z_{x,t}}{x+2t+1}.$$ Then $$\begin{aligned}
&\mathbb{E}\left[z_{x,t+1}^{2}\right]\\
& = \mathbb{E}\left[z_{x,t}^{2}\right]\left(\frac{x+2t-3}{x+2t+1}\right)
+\frac{4\mathbb{E}\left[z_{x,t}\right]}{x+2t+1}\\
&= x^{2}\prod_{i=0}^{t}\left(\frac{x+2i-3}{x+2i+1}\right)+\sum_{i=0}^{t}
\frac{4\mathbb{E}\left[z_{x,i}\right]}{x+2i+1}
\prod_{j=i+1}^{t}\left(\frac{x+2j-3}{x+2j+1}\right)\\
&= \frac{x^{2}\left(x-3\right)\left(x-1\right)}{\left(x+2t+1\right)\left(x+2t-1\right)}
+\sum_{i=0}^{t}\frac{4x\left(x-1\right)}{\left(x+2i+1\right)\left(x+2i-1\right)}
\frac{\left(x+2i-1\right)\left(x+2i+1\right)}{\left(x+2t+1\right)\left(x+2t-1\right)} \\
&= \frac{x^{2}\left(x-3\right)\left(x-1\right)}{\left(x+2t+1\right)\left(x+2t-1\right)}
+\frac{4x\left(x-1\right)\left(t+1\right)}{\left(x+2t+1\right)\left(x+2t-1\right)} \\
&= \frac{x\left(x-1\right)}{\left(x+2t+1\right)\left(x+2t-1\right)}\left(x\left(x-3\right)
+4\left(t+1\right)\right),
\label{eq:expzmoment2}
\end{aligned}$$ while $$\left(\mathbb{E}\left[z_{x,t+1}\right]\right)^{2}=x^{2}\left(\frac{x-1}{x+2t+1}\right)^{2}$$ and so $$\mathbb{E}\left[z_{x,t+1}^{2}\right]-\left(\mathbb{E}\left[z_{x,t+1}\right]\right)^{2}
=\frac{x\left(x-1\right)}{x+2t+1}\left(\frac{x\left(x-3\right)
+4\left(t+1\right)}{x+2t-1}-\frac{x\left(x-1\right)}{x+2t+1}\right).$$ It follows that if we let $w_{x,t}=\frac{z_{x,t}}{x+2t}$, $t\geq1$, then $$\begin{aligned}
&\mathbb{E}\left[w_{x,t+1}^{2}\right]-\left(\mathbb{E}\left[w_{x,t+1}\right]\right)^{2}\\
& = \frac{\frac{x}{x+2\left(t+1\right)}\frac{\left(x-1\right)}{x+2\left(t+1\right)}}
{x+2t+1}\left(\frac{4\left(t+1\right)}{x+2t-1}+\frac{x-3}{1+2\frac{t}{x}-\frac{1}{x}}
-\frac{x-1}{1+2\frac{t}{x}+\frac{1}{x}}\right)\\
&= \frac{\frac{x}{x+2\left(t+1\right)}\frac{\left(x-1\right)}{x+2\left(t+1\right)}}
{x+2t+1}\left(\frac{4\left(t+1\right)}{x+2t-1}+\frac{-4\frac{t}{x}
-4\frac{1}{x}}{\left(1+2\frac{t}{x}-\frac{1}{x}\right)\left(1+2\frac{t}{x}
+\frac{1}{x}\right)}\right)\\
&= \frac{4\frac{x}{x+2\left(t+1\right)}\frac{\left(x-1\right)}{x+2\left(t+1\right)}}
{x+2t+1}\frac{t+1}{x+2t-1}\left(1-\frac{1}{1+2\frac{t}{x}+\frac{1}{x}}\right).
\label{eq:wvar}
\end{aligned}$$
Observe also that for $$\bar{z}_{x,t} = \sum_{i=1}^{x}\sum_{m=x+1}^{M}
\mathbf{1}\left\{ \left(i,m\right)\in\xi_{x/2+t}\right\} =x-z_{x,t}
\label{eq:zbar}$$ and $$\bar{w}_{x,t} = \frac{\bar{z}_{x,t}}{x+2t}=\frac{x}{x+2t}-w_{x,t}$$ the same variance calculations follow, so that $\mathbb{E}\left[\bar{w}_{x,t+1}^{2}\right]
-\left(\mathbb{E}\left[\bar{w}_{x,t+1}\right]\right)^{2}$ is also given by (\[eq:wvar\]). For $\alpha\in\left(0,1\right)$ and $1\leq i\leq k\left(\alpha\right)$ with $k\left(\alpha\right)
=M/\left\lfloor M^{\alpha}\right\rfloor $, let $x_{i}=i\left\lfloor M^{\alpha}\right\rfloor $ and note that $$\begin{aligned}
&\sum_{i=1}^{k\left(\alpha\right)} \left(\mathbb{E}\left[\bar{w}_{x,t+1}^{2}\right]
-\left(\mathbb{E}\left[\bar{w}_{x,t+1}\right]\right)^{2}\right)\\
&=\frac{1}{2}M^{-2}\left(M-1\right)^{-1}\left(M-3\right)^{-1}\sum_{i=1}^{k\left(\alpha\right)}
x_{i}\left(x_{i}-1\right)\left(M-x_{i}\right)\left(1-\frac{x_{i}}{M-1}\right)
=O\left(M^{-\alpha}\right).
\label{eq:2ndmomentbndforpath}
\end{aligned}$$ From Markov’s inequality we have that $$\begin{aligned}
&\mathbb{P}\left[\exists\,i\mbox{ such that }\frac{\left|\bar{z}_{x_{i},\left(M-x_{i}\right)/2}
-\mathbb{E}\left[\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right]\right|}{M}
>M^{-\frac{\alpha}{3}}\right]\\
&= \mathbb{P}\left[\exists\, i\mbox{ such that }\frac{\left|\bar{z}_{x_{i},\left(M-x_{i}\right)/2}
-\mathbb{E}\left[\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right]\right|^{2}}{M^{2}}
>M^{-\frac{2\alpha}{3}}\right]\\
&= \mathbb{P}\left[\exists\, i\mbox{ such that } \left|\bar{w}_{x_{i},\left(M-x_{i}\right)/2}
-\mathbb{E}\left[\bar{w}_{x_{i},\left(M-x_{i}\right)/2}\right]\right|^{2}
>M^{-2\alpha/3}\right]= O\left(M^{-\alpha/3}\right)
\end{aligned}$$ and thus we have that w.h.p. for every $1\leq i\leq k\left(\alpha\right)$, $$\left|\bar{z}_{x_{i},\left(M-x_{i}\right)/2}-\mathbb{E}\left[\bar{z}_{x_{i},
\left(M-x_{i}\right)/2}\right]\right|=O\left(M^{1-\alpha/3}\right).
\label{eq:bndziconcentration}$$ Now suppose that $x_{i}\leq x\leq x_{i+1}$. Then, clearly, $\bar{z}_{x_{i},
\left(M-x_{i}\right)/2}-M^{\alpha}\leq\bar{z}_{x,\left(M-x\right)/2}\leq
\bar{z}_{x_{i},\left(M-x_{i}\right)/2}+M^{\alpha}$, and via (\[eq:Expztplus1\]) and (\[eq:zbar\]) we conclude that w.h.p. for every $1\leq x\leq M$ we have $$\label{eq:boundzbarvariance}
\begin{aligned}
&\left|\bar{z}_{x,\left(M-x\right)/2}-\mathbb{E}\left[\bar{z}_{x,\left(M-x\right)/2}\right]\right|\\
& \leq\left|\bar{z}_{x,\left(M-x\right)/2}-\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right|
+\left|\bar{z}_{x_{i},\left(M-x_{i}\right)/2}
-\mathbb{E}\left[\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right]\right|\\
&\qquad +\left|\mathbb{E}\left[\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right]
-\mathbb{E}\left[\bar{z}_{x,\left(M-x\right)/2}\right]\right|\\
& =\left|\bar{z}_{x,\left(M-x\right)/2}-\bar{z}_{x_{i},\left(M-x_{i}\right)/2}\right|
+\left|\bar{z}_{x_{i},\left(M-x_{i}\right)/2}-x_{i}\left(\frac{M-x_{i}}{M-1}\right)\right|\\
&\qquad +\left|x_{i}\left(\frac{M-x_{i}}{M-1}\right)-x\left(\frac{M-x}{M-1}\right)\right|\\
&\leq M^{\alpha}+O\left(M^{1-\alpha/3}\right)+M^{\alpha}.
\end{aligned}$$ Now let $\gamma_{s}$ be any configuration on the path $\gamma:\boxminus\rightarrow\boxplus$ defined above. Then w.h.p. $$\begin{aligned}
\mathcal{H}\left(\gamma_{s}\right)-\mathcal{H}\left(\boxminus\right)
&= J\left|E\left(\gamma_{s},\overline{\gamma_{s}}\right)\right|-hs\\
&= J\bar{z}_{\ell_{\gamma_{s}},\left(\ell_{n}-\ell_{\gamma_{s}}\right)/2}-hs
=J\ell_{\gamma_{s}}\left(1-\frac{\ell_{\gamma_{s}}}{\ell_{n}}\right)-hs+O\left(\ell_{n}^{3/4}\right),
\end{aligned}$$ where the last line follows from with $x=\ell_{m}$, $M=\ell_{n}$ and $\alpha=\frac{3}{4}$, and uses the fact that $\mathbb{E}\left[\bar{z}_{x,t}\right]
=x\left(1-\frac{x-1}{x+2t+1}\right)$. By definition, this quantity is maximised when $\ell_{\gamma_{s}}$ is replaced by $\ell_{\bar{m}}$, from which the statement of the theorem follows.
Proof of Theorem \[thm:Lbound\]
-------------------------------
Setting $x=\tfrac12$ in , we get that the probability of there being any configuration of total degree $\ell_n/2$ having a boundary size $y\ell_n$ is bounded from above by $$\exp\left[\ell_{n}\log\left(\tfrac12\eta\left(\tfrac12\right)
\left(\tfrac12-y\right)^{-\left(\tfrac12-y\right)}y^{-y}\right)\right].
\label{eq:expbound}$$ By the law of large numbers, w.h.p. we have that $n=\ell_{n}/d_{\mathrm{ave}}+
o\left(n\right)$. Combining this with $\left|\left\{ U\subseteq V\colon\,\ell_U
=\ell_{n}x\right\} \right|\leq2^n$ we get that $\eta\left(\tfrac12\right) \leq 2^{1/d_{\mathrm{ave}}}$. It follows that if $y< I_{d_{\mathrm{ave}}}\left(\tfrac12\right)$, then (\[eq:expbound\]) decays exponentially. Hence, all configurations $\sigma$ with $\ell_\sigma=\ell_n/2$ have, w.h.p., an energy at least $${\mathcal{H}}(\sigma) \geq J I_{d_{\mathrm{ave}}}\left(\tfrac12\right) \ell_n - h |\sigma| + {\mathcal{H}}(\boxminus),$$ and the lower bound on $\Gamma^\star_n$ in follows.
Tail properties of the dynamically constructed ${\mathrm{CM}}_n$ {#tailprop}
================================================================
In this section we explore some properties of the dynamical construction of ${\mathrm{CM}}_n$ introduced in Section \[CMdynamic\].
Trivial tail $\sigma$-algebra
-----------------------------
Let $V_{n}=\left(v_{1},\ldots,v_{n}\right)$ be the vertices with corresponding degree sequence $$\vec{d}_{n}=\left(d_{1},\ldots,d_{n}\right),$$ and let $G_{n}=\left(V_{n},E_{n}\right)$ and $G_{n}\prime=\left(V_{n},E_{n}\prime
\right)$ be two independent Configuration Models with the same degree sequence $\vec{d}_{n}$. We will extend $G_{n}$ and $G_{n}\prime$ to larger graphs, $G_{n+t}
=\left(V_{n+t},E_{n+t}\right)$ and $G_{n+t}\prime=\left(V_{n+t},E_{n+t}\prime\right)$, respectively, with degree sequence $$\vec{d}_{n}=\left(d_{1},\ldots,d_{n},d_{n+1},\ldots d_{n+t}\right),$$ by utilising a pairing scheme similar to the one introduced in Section \[CMdynamic\]:
- If $\xi_{m}$ is a uniform random matching of the integers $\left\{ 1,\ldots,2m\right\}$, denoted by $\xi_{m}=\left\{ \left(x_{1},x_{2}\right),\ldots,\left(x_{2m-1},x_{2m}\right)
\right\} $, then take $u_{1}$ to be uniform on $\left\{ 1,\ldots,2m\right\} $ and $u_{2}$ to be uniform on $\left\{ 1,\ldots,2m,2m+1\right\} $. If $u_{2}=2m+1$, then set $\xi_{m+1}=\xi_{m}\cup\left\{ \left(2m+1,2m+2\right)\right\} $. Otherwise add to $\xi_{m}$ the pairs $\left(2m+1,u_{1}\right)$ and $\left(2m+2,u_{2}\right)$ when $u_{1}\neq u_{2}$, and when $u_{1}=u_{2}$, only add to $\xi_{m}$ the pair $\left(2m+2,u_{2}\right)$. In either case, if there are two remaining terms that are unpaired, then pair them to each other and add this pair to $\xi_{m}$. Needless to say, we also remove from $\xi_{m}$ old pairs that were undone by the introduction of $2m+1$ and $2m+2$. Again, this construction leads to $\xi_{m+1}$, a uniform matching of the points $\left\{1,\ldots,2m+2\right\}$.
Now construct the coupled graphs $\left(G_{n+t},G_{n+t}\prime\right)$ by starting with $\left(G_{n},G_{n}\prime\right)$ and using the same uniform choice $$\left\{u_{i}\right\} _{i=1}^{|\vec{d}_{n+t}|-|\vec{d}_{n}|}$$ to determine new edges in both graphs. Note that, under this scheme, every term (half-edge) $s>|\vec{d}_{n}|$ is paired with the same term in $G_{n+t}$ as in $G_{n+t}\prime$. In other words, for all $1\leq j\leq|\vec{d}_{n+t}|$ we have $\left(s,j\right)\in E_{n+t}$ if and only if $\left(s,j\right)\in E_{n+t}\prime$. For $s\leq|\vec{d}_{n}|$ and $1\leq j\leq|\vec{d}_{n+t}|$, we have $$\begin{aligned}
&\mathbb{P}\left[\mathbf{1}_{\left\{ \left(s,j\right)\in E_{n+t}\right\} }
\neq \mathbf{1}_{\left\{ \left(s,j\right)\in E_{n+t}\prime\right\} }\right]\\
&\leq \mathbb{P}\left[\bigcap_{i=1}^{|\vec{d}_{n+t}|
-|\vec{d}_{n}|}\left\{ u_{i}\neq s\right\} \right]
= \prod_{i=|\vec{d}_{n}|}^{|\vec{d}_{n+t}|}\left(1-\frac{1}{i-1}\right)
=\frac{|\vec{d}_{n}|-2}{|\vec{d}_{n+t}|-1}.
\end{aligned}$$ Hence $$\mathbb{P}\left[\bigcup_{s=1}^{\left|\vec{d}_{n}\right|}
\left\{ \mathbf{1}_{\left\{ \left(s,j\right)\in E_{n+t}\right\} }
\neq \mathbf{1}_{\left\{ \left(s,j\right)\in E_{n+t}\prime\right\} }\right\} \right]
\leq\frac{|\vec{d}_{n}|(|\vec{d}_{n}|-2)}
{|\vec{d}_{n+t}|-1}.$$ Thus, we conclude that $\mathbb{P}\left[G_{n+t}\neq G_{n+t}\prime\right] =
O\left(\frac{1}{t}\right)$.
We can now make the following standard argument to show that the process above has a trivial tail-sigma-algebra. Let $\mathscr{F}_{t}=\sigma\left(\xi_{1},
\ldots.\xi_{t}\right)$ and $\mathscr{F}_{t}^{+}=\sigma\left(\xi_{t+1},\ldots\right)$. The tail sigma-algebra is given by $\mathscr{T}=\cap_{n\in\mathbb{N}}
\mathscr{F}_{t}^{+}$. For any $A\in\mathscr{T}$, there is a sequence of events $A_{1},A_{2},\ldots$ such that $$\lim_{t\to\infty}\mathbb{P}\left[A_{t}\triangle A\right]=0,$$ and hence also $$\lim_{t\to\infty}\mathbb{P}\left[A_{t}\cap A\right] =\mathbb{P}\left[A\right],
\qquad \lim_{t\to\infty}\mathbb{P}\left[A_{t}\right]=\mathbb{P}\left[A\right].$$ But, since $A\in\mathscr{F}_{t}^{+}$ for all $t$, it follows that $\mathbb{P}\left[A_{t}
\cap A\right]=\mathbb{P}\left[A_{t}\right]\mathbb{P}\left[A\right]$, and hence $\mathbb{P}\left[A\right]=\mathbb{P}\left[A\right]^{2}$. This shows that $\mathscr{T}$ is a trivial sigma-algebra. Therefore, given $\left\{ d_{i}\right\} _{i\in{\mathbb{N}}}$ (but also by the law of large numbers for i.i.d. sequences), $$\limsup_{n\to\infty}\frac{\Gamma_{n}^{\star}}{n}= \gamma^*_+,
\qquad \liminf_{n\to\infty}\frac{\Gamma_{n}^{\star}}{n}= \gamma^*_-,$$ for some $\gamma^*_+,\gamma^*_- \in{\mathbb{R}}$ with $\gamma^*_+ \geq \gamma^*_-$.
Oscillation bounds
------------------
It is possible to obtain bounds on the possible oscillations of $n\mapsto
\Gamma_{n}^{\star}/n$.
\[lem: gammadifference\] Let $G=(V,E)$ and $\tilde{G}=(\tilde{V},\tilde{E})$ be two connected graphs. Suppose that $\left|E\nabla\tilde{E}\right|\leq k$ under some labelling of the vertices in $V$ and $\tilde{V}$ (i.e., a one-to-one map from $V$ to $\tilde{V}$). Then $$|\Gamma^{\star}-\tilde{\Gamma}^{\star}| \leq Jk+h\big| |\tilde{V}|-|V| \big|.$$
W.l.o.g. assume that $|\tilde{V}|\geq|V\|$. Given the labelling of the vertices that satisfies the above condition, let $\gamma\colon\,\boxminus\to\boxplus$, denoted by $\gamma=\left(\gamma_{1},\ldots,\gamma_{m}\right)$, be an optimal path for the Glauber dynamics on $G$. Now let $\tilde{\gamma}\colon\,
\tilde{\boxminus}\to\tilde{\boxplus}$ be the Glauber path of configurations on $\tilde{G}$, denoted by $\tilde{\gamma}=(\tilde{\gamma}_{1},\ldots,\tilde{\gamma}_{m},
\tilde{\gamma}_{m+1},\ldots,\tilde{\gamma}_{m+|\tilde{V}|-|V|})$, and defined by the following rule: whichever vertex $v\in V$ is flipped at step $i$ in the path $\gamma$, flip the corresponding vertex $\tilde{v}\in\tilde{V}$ also at step $i$ in $\tilde{\gamma}$. For steps $m+1,\ldots,m+|\tilde{V}|-|V|$, flip the remaining $-1$ valued vertices in any arbitrary order. Then it follows that, for $1\leq i\leq m$, $$\tilde{\mathcal{H}}\left(\tilde{\gamma}_{i}\right)
-\tilde{\mathcal{H}}\left(\tilde{\boxminus}\right)
=J|\tilde{E}\left(\tilde{\gamma}_{i},\overline{\tilde{\gamma}_{i}}\right)|
-h\left|\tilde{\gamma}_{i}\right|
\leq J\left(|E\left(\gamma_{i},\overline{\gamma_{i}}\right)|+k\right)
-h\left|\tilde{\gamma}_{i}\right|.$$ Similarly, for $m\leq i\leq m+|\tilde{V}|-|V|$, we have $$\tilde{\mathcal{H}}\left(\tilde{\gamma}_{i}\right)
-\tilde{\mathcal{H}}\left(\tilde{\boxminus}\right)
\leq Jk-h\left(\left|\gamma_{i}\right|-\left|\boxplus\right|\right).$$ It follows that $\tilde{\Gamma}^{\star}\leq\Gamma^{\star}+Jk-h(|\tilde{V}|-|V|)$. A similar argument gives $\Gamma^{\star}\leq\tilde{\Gamma}^{\star}+Jk
+h||\tilde{V}|-|V||$.
Now let $G=\left(V,E\right)$ and $\tilde{G}=(\tilde{V},\tilde{E})$ be two Configuration Models and suppose w.l.o.g. that the total degree of the vertices in $V$ is $\ell_{V}$ and the total degree of vertices in $\tilde{V}$ is $\ell_{\tilde{V}}\geq\ell_{V}$. Let $G_{t}$ and $\tilde{G}_{t}$ be the extension of each these two graphs, obtained by adding vertices $\left\{ v_{1},\ldots,v_{t}\right\} $ and $\left\{ \tilde{v}_{1},\ldots,\tilde{v}_{t}\right\}$, both with the same degree sequence $\left\{d_{1},\ldots,d_{t}\right\}$. We will couple the construction leading to the two graphs $G_{t}$ and $\tilde{G}_{t}$ in the following manner: for $1\leq i\leq\sum_{k=1}^{t}d_{k}$, choose $u_{i}$ uniformly as described above, and pair $i$ with $u_{i}$ in $G_{t}$. Let $\delta_{i}=\left(\ell_{\tilde{V}}-\ell_{V}\right)/\left[\left(\ell_{\tilde{V}}
+i-1\right)\left(\ell_{V}+i-1\right)\right]$ and set $\tilde{u}_{i}=u_{i}$ with probability $1-\delta_{i}$, and with probability $\delta_{i}$ independently and uniformly pick one of the remaining $\left(\ell_{\tilde{V}}-\ell_{V}\right)$ points. Then $$\mathbb{E}[|E_{t}\nabla\tilde{E_{t}}|]
\leq |E\nabla\tilde{E}|+\sum_{i}\delta_{i} \leq |E\nabla\tilde{E}|
+2\left(\ell_{\tilde{V}}-\ell_{V}\right).$$ Hence, from Markov’s inequality and from Lemma \[lem: gammadifference\], it follows that, w.h.p. and for any function $f(t)$ such that $\lim_{t\to\infty}
f(t)=\infty$, $$\left|\tilde{\Gamma^{\star}}_{t}-\Gamma^{\star}_{t}\right|
\leq J\left(|E\nabla\tilde{E}|+2(\ell_{\tilde{V}}-\ell_{V})+f(t)\right)
-h\left(\big||\tilde{V}|-|V|\big|\right).$$ Hence, by this pairing scheme, we have that $\tilde{\Gamma^{\star}_{t}}/\Gamma^{\star}_{t}
\to 1$ as $t\to \infty$.
Appendix A {#appendix-a .unnumbered}
==========
Note that any configuration $\sigma$ chosen uniformly from all configurations of size $|\sigma|$ satisfies ($\mathbb{E}$ denotes expectation w.r.t. bond percolation) $$\mathbb{E}\left(\left|E\left(\sigma,\overline{\sigma}\right)\right|\right)
= \mu_{\left|\sigma\right|} \quad \mbox{ with } \quad
\mu_{\left|\sigma\right|} = p\left|\sigma\right|\left(n-\left|\sigma\right|\right).$$ Using Chernoff’s inequality and a union bound, we can show that if $|\sigma|=\Theta(n)$, then $$\mathcal{H}\left(\sigma\right)-\mathcal{H}\left(\boxminus\right)
= J\mu_{\left|\sigma\right|}[1\pm o(1)]-h\left|\sigma\right|.$$ Furthermore, any $\sigma$ of size $|\sigma|\leq[1-o(1)]\frac{n}{2}$ has (modulo small fluctuations) a downhill path to $\boxminus$ (e.g. by flipping $+1$ spins in any arbitrary order), while every $|\sigma|\geq[1+o(1)]\frac{n}{2}$ has a downhill path to $\boxplus$ (e.g. by flipping $-1$ spins in any arbitrary order). This proves hypothesis (H), and the claim in .
[99]{}
G. Ben Arous and R. Cerf, Metastability of the three-dimensional Ising model on a torus at very low temperature, Electron. J. Probab. 1, paper 10, 1996.
A. Bovier, M. Eckhoff, V. Gayrard and M. Klein, Metastability and small eigenvalues in Markov chains, J. Phys. A 33, L447–L451, 2000.
A. Bovier, M. Eckhoff, V. Gayrard and M. Klein, Metastability in stochastic dynamics of disordered mean-field models, Probab. Theory Relat. Fields 119, 99–161, 2001.
A. Bovier, M. Eckhoff, V. Gayrard and M. Klein, Metastability and low lying spectra in reversible Markov chains, Comm. Math. Phys. 228, 219–255, 2002.
A. Bovier, M. Eckhoff, V. Gayrard and M. Klein, Metastability in reversible diffusion processes. I. Sharp asymptotics for capacities and exit times, J. Eur. Math. Soc. 6, 399–424, 2004.
A. Bovier and F. den Hollander, *Metastability – A Potential-Theoretic Approach*, Grundlehren der mathematischen Wissenschaften 351, Springer, 2015.
A. Bovier, F. den Hollander and C. Spitoni, Homogeneous nucleation for Glauber and Kawasaki dynamics in large volumes and low temperature, Ann. Probab. 38, 661–713, 2010.
A. Bovier and F. Manzo, Metastability in Glauber dynamics in the low temperature limit: beyond exponential asymptotics, J. Stat. Phys. 107, 757–779, 2002.
M. Cassandro, A. Galves, E. Olivieri and M.E. Vares, Metastable behavior of stochastic dynamics: A pathwise approach, J. Stat. Phys. 35, 603–634, 1984.
O. Catoni and R. Cerf, The exit path of a Markov chain with rare transitions, ESAIM Probab. Statist. 1, 95–144, 1997.
E.N.M. Cirillo and F.R. Nardi, Relaxation height in energy landscapes: an application to multiple metastable states, J. Stat. Phys. 150, 1080–1114, 2013.
E.N.M. Cirillo, F.R. Nardi and J. Sohier, Metastability for general dynamics with rare transitions: escape time and critical configurations, J. Stat. Phys. 161, 365–403, 2015.
E.N.M. Cirillo and E. Olivieri, Metastability and nucleation for the Blume-Capel model: different mechanisms of transition, J. Stat. Phys. 83, 473–554, 1996.
A. Dembo and A. Montanari, Ising models on locally tree-like graphs, Ann. Appl. Probab. 20, 565–592, 2010.
S. Dommers, Metastability of the Ising model on random regular graphs at zero temperature, to appear in Probab. Theory Relat. Fields, DOI: 10.1007/s00440-015-0682-0, 2015.
S. Dommers, C. Giardinà and R. van der Hofstad, Ising models on power-law random graphs, J. Stat. Phys. 141, 638–660, 2010.
R. van der Hofstad, *Random Graphs and Complex Networks, Volume I*, to appear with Cambridge University Press. (File can be downloaded from <http://www.win.tue.nl/~rhofstad/>.)
O. Jovanovski, Metastability for the Ising Model on the hypercube, \[arXiv:1508.07769\], 2015.
R. Kotecký and E. Olivieri, Stochastic models for nucleation and crystal growth, Probabilistic Methods in Mathematical Physics (Siena 1991), riverside NY: World Scientific 264–275, 1992.
R. Kotecký and E. Olivieri, Droplet Dynamics for asymmetric Ising model, J. Stat. Phys. 70, 1121–1148, 1993.
R. Kotecký and E. Olivieri, Shapes of growing droplets - a model of escape from the metastable phase, J. Stat. Phys. 75, 409–506, 1994.
F. Manzo, F.R. Nardi, E. Olivieri and E. Scoppola, On the essential features of metastability: tunnelling time and critical configurations, J. Stat. Phys. 115, 591–642, 2004.
E. Mossel and A. Sly, Exact thresholds for Ising-Gibbs samplers on general graphs, Ann. Appl. Probab. 41, 294–328, 2013.
F.R. Nardi and E. Olivieri, Low temperature stochastic dynamics for an Ising model with alternating field, Markov Proc. Relat. Fields. 2, 117–166, 1996.
E.J. Neves and R.H. Schonmann, Critical droplets and metastability for a Glauber dynamics at very low temperature, Comm. Math. Phys. 137, 209–230, 1991.
E.J. Neves and R.H. Schonmann, Behavior of droplets for a class of Glauber dynamics at very low temperature, Probab. Theory Relat. Fields. 91, 331–354, 1992.
E. Olivieri and E. Scoppola, Markov chains with exponentially small transition probabilities: First exit problem from general domain I. The reversible case, J. Stat. Phys. 79, 613–647, 1995.
E. Olivieri and E. Scoppola, Markov chains with exponentially small transition probabilities: First exit problem from general domain II. The general case, J. Stat. Phys. 84, 987–1041, 1996.
E. Olivieri and M.E. Vares, *Large Deviations and Metastability*, Encyclopedia of Mathematics and its Applications 100, Cambridge University Press, Cambridge, 2005.
[^1]: The research in this paper was supported by NWO Gravitation Grant 024.002.003-NETWORKS. The work of SD was supported also by DFG Research Training Group 2131.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The polarization dependence of photoconductivity response at cyclotron-resonance harmonics in a nondegenerate two-dimensional (2D) electron system formed on the surface of liquid helium is studied using a setup in which a circular polarization of opposite directions can be produced. Contrary to the results of similar investigations reported for semiconductor 2D electron systems, for electrons on liquid helium, a strong dependence of the amplitude of magnetoconductivity oscillations on the direction of circular polarization is observed. This observation is in accordance with theoretical models based on photon-assisted scattering and, therefore, it solves a critical issue in the dispute over the origin of microwave-induced conductivity oscillations.'
author:
- 'A. A. Zadorozhko'
- 'Yu. P. Monarkha'
- 'D. Konstantinov'
title: 'Circular-Polarization-Dependent Study of Microwave-Induced Conductivity Oscillations in a Two-Dimensional Electron Gas on Liquid Helium'
---
Studies of microwave (MW) photoconductivity in a two-dimensional (2D) electron gas of semiconductor heterostructures subjected to a perpendicular magnetic field $B$ have revealed remarkable magnetotransport phenomena: giant microwave-induced resistance oscillations (MIRO) [@ZudDuSim-2001; @YeEngTsu-2001] and associated zero resistance states (ZRS) [@ManSmeKlit-2002; @ZudDuPfe-2003]. These discoveries have opened a prominent research area and triggered a large body of theoretical works. The universality of the effect of MIRO was proved by similar observations in hole systems [@ZudMirEbn-2014] and in a nondegenerate 2D electron system formed on the free surface of liquid helium [@YamAbdBad-2015].
It is very surprising that by now there is a great body of different theoretical mechanisms explaining MIRO which use quantum and classical effects (for a review, see [@DmiMirPol-2012]), but the origin of these oscillations is still under debate. Among these mechanisms there is a large group of models whose description is based on the concept of the photon-assisted scattering off disorder which overcomes the selection rules existing for direct photon-induced transitions; direct transitions can be only between adjacent Landau levels: $n^{\prime }-n=\pm 1$ . The photon-assisted scattering leads to two important effects. Firstly, it gives a separate contribution to magnetoconductivity $\sigma _{xx}$ where the displacement of the electron orbit center $X^{\prime }-X$ caused by energy conservation changes its sign when the ratio $\omega /\omega _{c}$ passes an integer $m=n^{\prime }-n=1,2,...$ (here $\omega $ is the MW frequency and $%
\omega _{c}$ is the cyclotron frequency) [@Ryz-1969; @DurSachRea-2003]. This mechanism is called the displacement model (DM).
Secondly, electron scattering to higher Landau levels ($n^{\prime }-n=2,3...$) changes the distribution function of electrons at these levels $f\left( \varepsilon \right) $: it acquires an oscillatory form with maxima, and, therefore, a sort of population inversion occurs [@DmiMirPol-2003; @DmiVavAle-2005]. This mechanism is called the inelastic model (IM) because the inelastic thermalization rate is an important quantity for the description of the effect. The maxima of $f\left( \varepsilon \right) $ affect the contribution to $\sigma _{xx}$ caused by usual scattering processes and also lead to a sign-changing correction to $%
\sigma _{xx}$.
Both theoretical models (DM and IM) give satisfactory descriptions of MIRO in semiconductor electron systems. Still, there is a critical unresolved issue which concerns the dependence of MIRO on the direction of circular polarization: the results of the DM and IM are very sensitive to the direction of circular polarization, while the MIRO observed in semiconductor heterostructures are notably immune to the sense of circular polarization [@SmetGorJia-2005], or have a very weak dependence on the direction of circular polarizations in the THz range [@HerDmiKoz-2016] which is at odds with existing theories of MIRO.
The theoretical analysis [@Mon-2017] of the DM and IM performed for a nondegenerate 2D electron gas on liquid helium also indicates that the both models result in practically the same strong dependence of the amplitude of MIRO on the direction of circular polarization if the number $m=2,3,...$ is not large. Since the MIRO observed for electrons on liquid helium [@YamAbdBad-2015] are in good accordance with the IM, and the recent observation of MW-induced oscillations in magnetocapacitance of a semiconductor system [@DorKapUma] also supports the concept of a nonequilibrium distribution function oscillating with energy, there is a strong need to investigate the dependence of the amplitude of magnetoconductivity oscillations on the direction of circular polarization for the 2D electron system on liquid helium.
Here, we report the first observation of a strong dependence of the amplitude of microwave-induced magnetoconductivity oscillations on the direction of circular polarization in the electrons-on-helium system. The analysis of data given here shows that this observation is in accordance with theoretical models based on photon-assisted scattering. Thus, contrary to the mysterious contradiction between experiment and theory existed in semiconductor heterostructures, the circular-polarization-dependent study of MIRO in a nondegenerate 2D electron gas on liquid helium provides strong support for the photon-assisted scattering as the origin of the magnetoconductivity oscillations.
The experiments are done in a 2D electron system formed on the free surface of liquid $^3$He, which is contained in a closed cylindrical cell and cooled to $T=0.2$ K. At this temperature the scattering of electrons is dominated by capillary-wave excitations of the liquid helium surface (ripplons) and is very well understood [@Andrei_book; @Monarkha_book]. The magnetic field $B$ is applied perpendicular to the liquid surface, and the longitudinal conductivity of electrons $\sigma_{xx}$ is measured by the capacitive-coupling method using a pair of gold-plated concentric circular electrodes (Corbino disk) placed beneath and parallel to the liquid surface. The electrodes have outer diameters of 14 and 19.8 mm and are separated by a gap of width $0.005$ mm. Similar to our previous experiment [@YamAbdBad-2015], conductivity oscillations are excited by the electric field component of the fundamental TEM$_{002}$ mode in a semiconfocal Fabry-Perot resonator [@Kog-1966]. The resonator is formed by the Corbino disk acting as a flat reflecting mirror and a copper concave (radius of curvature 30 mm) mirror of diameter 35.2 mm placed above and parallel to the Corbino disk at a distance about 13 mm. At liquid helium temperatures, the TEM$_{002}$ mode has a frequency $\omega_r/2\pi\approx 35.21$ GHz and the quality factor is about 10$^4$.
The MW excitation is supplied from a room-temperature source followed by a linear-to-circular polarization converter, transmitted into the cryostat via a circular (inner diameter 6.25 mm) 1.5 meter long waveguide, and then coupled to the resonator via a circular aperture (diameter 1.8 mm) drilled in the center of the concave mirror. The axial symmetry and alignment of this setup are very important to ensure the circular polarization of the resonant mode field excited in the resonator. To check the latter, we observed the dependence of the cyclotron resonance (CR) excited in 2D electron system by the circular-polarized MWs on the direction of the applied magnetic field. For this, the electrons were placed in the maximum of the MW $E$-field in the resonator by adjusting the height of liquid helium to be about 2.1 mm above the Corbino disk, and the value of the magnetic field $B$ was adjusted such that the cyclotron frequency of electrons, $\omega_c=eB/m_e c$, was close to the frequency of the resonant mode.
To detect the photoconductivity response of electrons we applied a 20 mV excitation voltage at the frequency of 1.117 kHz to the inner electrode of the Corbino disk and measured the current induced in the outer electrode by the electron motion. This current is plotted in Fig. \[fig:1\] as a function of the magnetic field $B$ and frequency of MW excitation $\omega$. Two panels correspond to two opposite directions of the magnetic field $B$. For a given direction of circular polarization, strong CR absorption should occur only for the proper direction of the magnetic field and should be strongly suppressed for the opposite direction. As was demonstrated earlier [@AbdYamBad-2016], for a resonator of sufficiently high quality factor the strong coupling of cyclotron motion of electrons to the resonator mode leads to the appearance of two polaritonic branches of coupled electron-mode motion. This is shown in Fig. \[fig:1\](a) where two polaritonic branches are revealed in the conductivity response of electrons due to their strong heating by the CR absorption. For the opposite direction of the magnetic field, see Fig. \[fig:1\](b), the polaritonic branches are barely visible, pointing out that the CR absorption is strongly suppressed. The origin of a strong response appearing around the crossover point $\omega_r=\omega_c$, which corresponds to $B\approx 1.26$ T, is not clear. For the opposite direction of circular polarization, the polaritonic branches are clearly observed for negative $B$ and are strongly suppressed for positive $B$. Finally, for linear polarized MWs polaritonic branches of nearly equal intensity are observed for both directions of the field $B$.
![(color online) Photocurrent response of the 2D electron system at $T=0.2$ K and electron density $n_s= 8.2\times 10^7$ cm$^{-2}$ to a circular-polarized microwave excitation plotted as a function of MW frequency $\omega$ and perpendicular magnetic field $B$. Panels (a) and (b) are for two opposite directions of the field distinguished by the sign of $B$.[]{data-label="fig:1"}](fig1.eps){width="8.5cm"}
![(color online) (a) Photocurrent response of the 2D electron system at $T=0.2$ K and electron density $n_s=5.2\times 10^6$ cm$^{-2}$ to circular polarized MWs plotted as a function of MW frequency $\omega$ and the perpendicular magnetic field $B$. (b) Photocurrent response taken at $\omega=\omega_r$ (indicated by a dashed line in panel (a)).[]{data-label="fig:2"}](fig2.eps){width="8.5cm"}
Next, we consider the photoconductivity response of electrons at the harmonics of the CR by varying the value of $B$ in a wide range for both directions of the field. Figure \[fig:2\](a) shows the current signal measured at $T=0.2$ K for MW excitation which is circularly polarized in the same direction as for the data shown in Fig. \[fig:1\]. Strong photocurrent response is observed near values of $B$ which satisfy the condition $\omega=m\omega_c$, where $m=2,3,...$ . In addition, the photocurrent response is observed only when the frequency of microwave excitation $\omega$ is in the vicinity of the resonant frequency $\omega_r$. In particular, Figure \[fig:2\](b) displays the current signal measured at $\omega=\omega_r$. As shown previously [@YamAbdBad-2015], the photoconductivity oscillations can be observed only when the amplitude of the microwave $E$-field is sufficiently large, which occurs near the cavity resonance. For the input MW power used in Fig. \[fig:2\] we crudely estimate a maximum amplitude of the $E$-field in the resonant TEM$_{002}$ mode of about 5 V/cm.
The most important feature of plots shown in Figs. \[fig:2\](a,b) is a significant asymmetry in the amplitude of the photocurrent response with respect to directions of $B$. In particular, the large current oscillations are observed in the direction of the $B$-field corresponding to strong CR absorption, while the oscillations are strongly reduced for the direction of the field corresponding to suppressed CR absorption, c.f. Fig. \[fig:1\](a) and \[fig:1\](b). For the opposite direction of circular polarization (data not shown), the situation is reversed with respect to the direction of $B$, which is consistent with the behavior of CR absorption described above. Finally, for linear polarized microwaves, the amplitude of oscillations was found to be the same for both directions of $B$.
Figure \[fig:3\] shows values of $\sigma_{xx}$ extracted from the measured current signals and plotted as a function of the magnetic field $B$ for two opposite directions of circular polarization. The strong dependence of the amplitude of oscillations on the direction of circular polarization indicated in Fig. \[fig:3\] is the central result of this work and is at least in qualitative agreement with predictions of the theories based on photon-assisted scattering. For the sake of a quantitative comparison, we consider theoretical predictions with some more details below.
![(color online) Longitudinal conductivity $\sigma_{xx}$ of 2D electrons at $T=0.2$ K and $n_s=5.1\times 10^6$ cm$^{-2}$ versus magnetic field $B$ for two directions of circular polarization of MWs at frequency $\omega/2\pi=35.213$ GHz. A black line is for the circular polarization direction ($p=+ $) the same as for data shown in Figs. \[fig:1\] and \[fig:2\], while a red line is for the opposite direction ($p=-$).[]{data-label="fig:3"}](fig3.eps){width="9.75cm"}
To obtain probabilities of photon-assisted scattering there is a nonperturbative treatment resulting in the Landau-Floquet states which includes the MW field in an exact form (for a recent examples, see [Par-2004,TorKun-2005,Mon-2017]{}). The wave function of these states has a time-dependent shift of the center position $\xi \left( t\right) $ and a special phase factor. Considering two components of the MW field $E^{\left( x\right)
}=a_{p}E_{\mathrm{mw}}\cos \omega t$, and $E^{\left( y\right) }=b_{p}E_{%
\mathrm{mw}}\sin \omega t$ (here the numbers $a_{p}$ and $b_{p}$ describe the MW polarization, and $p$ is the polarization index) one have to introduce two time-dependent shifts $\xi $ (along $x-$axis) and $\zeta $ (along $y-$axis), which satisfy classical equations of motion [@Mon-2017]. The important consequence of this treatment is that the new expression for the matrix elements $%
\left\langle n^{\prime },X^{\prime }\right\vert e^{-i\mathbf{q}\cdot \mathbf{%
r}}\left\vert n,X\right\rangle _{L-F}$ describing electron scattering off disorder acquires an additional exponential factor$$\exp \left[ -i\beta _{p,\mathbf{q}}\sin \left( \omega t+\gamma _{p}\right) %
\right] , \label{e1}$$as compared to the conventional form obtained for the Landau basis. Here$$\beta _{p,\mathbf{q}}=\lambda \omega _{c}l_{B}\frac{\sqrt{q_{y}^{2}\left(
a_{p}\omega _{c}+b_{p}\omega \right) ^{2}+q_{x}^{2}\left( a_{p}\omega
+b_{p}\omega _{c}\right) ^{2}}}{\left( \omega ^{2}-\omega _{c}^{2}\right) },
\label{e2}$$$\lambda =eE_{\mathrm{mw}}l_{B}/\hbar \omega $ is the parameter describing the strength of the MW field, $l_{B}^{2}=\hbar c/eB$, and the exact form of $%
\gamma _{p}$ is not important for the following discussion. Calculation of scattering probabilities is reduced to the usual treatment by means of the Jacobi-Anger expansion $e^{iz\sin \phi
}=\sum_{k}J_{k}\left( z\right) e^{ik\phi }$ \[here $J_{k}\left( z\right) $ is the Bessel function\]. Thus, the probability of one-photon assisted scattering is proportional to $J_{1}^{2}\left( \beta _{p,\mathbf{q}%
}\right) \simeq \beta _{p,\mathbf{q}}^{2}/4$, if the parameter $\beta _{p,%
\mathbf{q}}$ is small. Therefore, the ratio of MW-induced corrections to the dc dissipative conductivity obtained for different directions of circular polarization ($p=+$ and $p=-$; $a_{\pm }=1$, $b_{\pm }=\pm 1$) is described by the simple relationship $$\frac{\Delta \sigma _{xx}^{\left( +\right) }}{\Delta \sigma _{xx}^{\left(
-\right) }}=\frac{\left( \omega /\omega _{c} +1\right) ^{2}}{\left( \omega /\omega _{c}
-1\right) ^{2}}. \label{e3}$$This equation is valid for the both DM and IM because $\beta _{p,\mathbf{q}}$ is now independent of the direction of the wave-vector $\mathbf{q}$. The ratio of Eq. (\[e3\]) is large for $m=2$ (it equals 9) and $m=3$ (it equals 4), but it approaches unity if $m$ increases. Calculations based on Eq. (\[e2\]) indicate that deviations from circularity (elliptic polarization) can reduce the ratio $\Delta \sigma _{xx}^{\left( +\right) }/\Delta \sigma _{xx}^{\left(
-\right) }$.
![(color online) The ratio $\Delta \sigma _{xx}^{\left( +\right) }/\Delta \sigma _{xx}^{\left(
-\right) }$ versus $\omega / \omega _{c}$: results obtained from experimental data for conductivity maxima (filled circles and squares) and minima (open circles and squares), the ratio of peak-to-peak amplitudes (triangles and crosses), the ratio of amplitudes obtained for positive and negative $B$ using the same curve of Fig. \[fig:3\] are marked in the legend by R-L, theory \[Eq. (\[e3\])\] calculated for the circular polarization (CP, solid curve), theory using the elliptic polarization (EP) with $a_{\pm }=0.7$ and $b_{\pm }=\pm 1.3$ calculated for minima (open stars).[]{data-label="fig:4"}](fig4.eps){width="9.5cm"}
For the both DM and IM, the shape of MIRO was shown to be similar to experimental observations [@YamAbdBad-2015; @Mon-2017]. Therefore, here we concentrate on the comparison of polarization dependencies of MIRO. The ratio of the amplitudes of conductivity oscillations obtained here for two opposite directions of circular polarization is illustrated in Fig. \[fig:4\]. The experimental results were plotted separately for maxima (filled circles and squares) and minima (open circles and squares) because the ratio of Eq. (\[e3\]) shown in this figure by the solid curve depends strongly on $B$. The ratio of peak-to-peak amplitudes shown by triangles agrees with these data. To ensure that a possible power difference didn’t affect our results here we plotted (squares and crosses) the ratio of respective amplitudes obtained at positive and negative $B$ from the same curve of Fig. \[fig:3\].
Fig. \[fig:4\] indicates that the ratio $\Delta \sigma _{xx}^{\left( +\right) }/\Delta \sigma _{xx}^{\left(
-\right) }$ found in the experiment increases with lowering $\omega / \omega _{c}$ in accordance with the theory, still it at an average is lower than theoretical values by a factor of the order of unity (about $1.4$). As a possible reason for such a reduction we considered a deviation from circularity. By way of illustration, calculations employing the DM and Eq. (\[e2\]) with $a_{\pm }=0.7$ and $b_{\pm }=\pm 1.3$ are shown in Fig. \[fig:4\] by star symbols with lines. Actually, the chosen parameters of ellipticity give even stronger reduction of the ratio $\Delta \sigma _{xx}^{\left( +\right) }/\Delta \sigma _{xx}^{\left(
-\right) }$ than that observed in the experiment. It should be noted that chosen ellipticity still leads to a strong suppression (about 10 times) of the photocurrent response at CR conditions for MWs with $p=-$, which is consistent with observations shown in Fig. \[fig:1\]. For inverted parameters ($a_{\pm }=1.3$, $b_{\pm }=\pm 0.7$), the reduction of the ratio of amplitudes is less than 5% because the integrand of the conductivity equation contains the symmetry-breaking factor $q_{y}^{2}$. Since the Corbino experiment gives $\sigma _{xx}$ data averaged over all directions of the driving electric field, the actual reduction of the ratio $\Delta \sigma _{xx}^{\left( +\right) }/\Delta \sigma _{xx}^{\left(
-\right) }$ for $a_{\pm }=0.7$ and $b_{\pm }=\pm 1.3$ can be noticeably less than that shown in Fig. \[fig:4\] by star symbols.
It should be noted also that under the conditions of the experiment the average Coulomb interaction energy per electron is more than 30 times larger than the average kinetic energy (temperature). Additionally, one cannot completely exclude heating of surface electrons by MWs. Therefore, the agreement between the experiment and the theories based on photon-assisted scattering can be considered as satisfactory.
At $E_{\mathrm{mw}}=5\,\mathrm{V/cm}$, the both DM and IM are estimated to provide sufficient amplitudes of MIRO under the conditions of this experiment, but the circular-polarization study cannot determine the contribution of which model dominates. For liquid $^{3}\mathrm{He}$, there is also an uncertainty in the definition of the inelastic thermalization rate because short wavelength ripplons responsible for inter-level scattering are heavily damped at low temperatures. Assuming that we can still rely on the usual ripplon spectrum, the IM yields somewhat stronger amplitudes of MIRO than the DM. Still, for a strict conclusion, an experimental setup with different linear polarizations and a fixed direction of the dc field is required.
In summary, by irradiating a nondegenerate 2D electron gas formed on liquid $^{3}\mathrm{He}$ with MWs of different polarizations we discovered a strong dependence of MIRO on the MW circular polarization direction. This allowed us to report the first observation of the effect of radiation helicity which provides a crucial information for understanding the origin of MW-induced magnetoconductivity oscillations in a 2D electron gas. In particular, our experiments unambiguously support theoretical mechanisms of MIRO based on photon-assisted scattering off disorder.
This work is supported by an internal grant from Okinawa Institute of Science and Technology (OIST) Graduate University. We are grateful to V. P. Dvornichenko for providing technical support.
[99]{}
M.A. Zudov, R.R. Du, J. A. Simmons, and J.L. Reno, Phys. Rev. B **64**, 201311 (2001).
P.D. Ye, L.W. Engel, D.C. Tsui, J.A. Simmons, J.R. Wendt, G.A. Vawter, and J.L. Reno, Appl. Phys. Lett. **79**, 2193 (2001).
R. Mani, J.H. Smet, K. von Klitzing, V. Narayanamurti, W.B. Johnson, and V. Umansky, Nature **420**, 646 (2002).
M.A. Zudov, R.R. Du, L.N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **90**, 046807 (2003).
M. A. Zudov, O. A. Mironov, Q. A. Ebner, P. D. Martin, Q. Shi, and D. R. Leadley, Phys. Rev. B **89**, 125401 (2014).
R. Yamashiro, L.V. Abdurakhimov, A.O. Badrutdinov, Yu.P. Monarkha, and D. Konstantinov, Phys. Rev. Lett. **115**, 256802 (2015). I. A. Dmitriev, A. D. Mirlin, D. G. Polyakov, and M. A. Zudov, Rev. Mod. Phys. **84**, 1709 (2012). V. I. Ryzhii, Fiz. Tverd. Tela (Leningrad) **11**, 2577 (1969) \[Sov. Phys. Solid State **11**, 2078 (1970)\].
A.C. Durst, S. Sachdev, N. Read, and S.M. Girvin, Phys. Rev. Lett. **91**, 086803 (2003). I.A. Dmitriev, A.D. Mirlin, and D.G. Polyakov, Phys. Rev. Lett. **91**, 226802 (2003). I.A. Dmitriev, M.G. Vavilov, I.L. Aleiner, A.D. Mirlin, and D.G. Polyakov, Phys. Rev. B **71**, 115316 (2005). J. H. Smet, B. Gorshunov, C. Jiang, L. Pfeiffer, K. West, V. Umansky, M. Dressel, R. Meisels, F. Kuchar, and K. von Klitzing, Phys. Rev. Lett. **95**, 116804 (2005). T. Herrmann, I. A. Dmitriev, D. A. Kozlov, M. Schneider, B. Jentzsch, Z. D. Kvon, P. Olbrich, V. V. Bel’kov, A. Bayer, D. Schuh, D. Bougeard, T. Kuczmik, M. Oltscher, D.Weiss, and S. D. Ganichev, Phys. Rev. B **94**, 081301(R) (2016). Yu.P. Monarkha, Fiz. Nizk. Temp. **43**, 819 (2017) \[Low Temp. Phys. **43**, 650 (2017)\] S.I. Dorozhkin, A.A. Kapustin, V. Umansky, K. von Klitzing, and J.H. Smet, Phys. Rev. Lett., **117**, 176801 (2016). , edited by E. Y. Andrei (Kluwer Academic Publishers, Dordrecht, 1997).
Yu. P. Monarkha and K. Kono, [*Two-Dimensional Coulomb Liquids and Solids*]{} (Springer-Verlag, Berlin, 2004).
H. Kogelnik and T. Li, Appl. Opt. **5**, 1550 (1966).
L.V. Abdurakhimov, R. Yamashiro, A.O. Badrutdinov, and D. Konstantinov, Phys. Rev. Lett. **117**, 056803 (2016). K. Park, Phys. Rev. B **69**, 201301(R) (2004). M. Torres and A. Kunold, Phys. Rev. B **71**, 115313 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
James T. Wheeler\
[*Utah State University, Logan, UT 84322*]{}
title: ' $SU(3) \times SU(2) \times U(1):$ The residual symmetry of extended conformal gravity.'
---
Motivation
==========
There are two points crucial to the understanding of gravity and its relationship to the standard model of the remaining forces [@Stdmodel]. These observations motivate the subsequent search for an additional symmetry and ultimately lead to a natural extension of the conformal algebra.
1. Conformal symmetry must underlie our mathematical description of nature.
2. Naive conformal gravity fails.
The first claim is true because we can measure only relative magnitudes. When the size of a table is given in meters it is being compared to a certain number of wavelengths of light; the mass of a chair is ultimately in comparison to the masses of particles or nuclei. The situation is no different at the subatomic level, where we maintain our use of standard units.
Nor is it necessary for our standard units to be chosen in the same way everywhere. The choice is made in accordance with the purpose at hand. A physicist generally chooses a uniform scale because with such a scale motion has a simple description. But an artist represents objects using a scale having distant objects smaller than the same types of object nearby because this is consistent with our visual process. It is certainly possible and consistent to translate our physical descriptions into such units.
While the Poincaré group does not generate scale changes, the conformal group respects this freedom of choosing scale by preserving only relative lengths. The preservation of one scale choice by the Poincaré group is not physical – scale need not be preserved. Thus, we reach the first conclusion: the local spacetime symmetry which does not claim more than we can know is the conformal group, not the Poincaré group.
To understand the second conclusion requires some understanding of the conformal group. We have, of course, the usual Poincaré symmetries – Lorentz transformations and translations. Lorentz transformations $(o(3,1))$ or $sl(2,C))$ include both rotations and boosts, and are generated by $4 \times 4$ matrices, $M_{ab}$. The translations, with generators $P_{a}$, give simple displacements. In addition to these are two new transformations:
1. Scale changes, generated by the dilation operator, $D$ directly produce the allowed choices of units by rescaling lengths locally.
2. Conformal translations. The four generators, $K_{a}$, translate infinitesimal vectors in inverse coordinates. Such a vector is inverted through a unit sphere, translated, then re-inverted. The transformation clearly involves an arbitrary displacement vector. The thing to notice – or rather not to notice – is that the vector is also rotated and rescaled. Since both rotating and stretching are already independent transformations, the group closes.
It is central to what follows to understand that with the scale symmetry available, one mode of translation is just as good as the other. Since we only compare relative magnitude, translating with $K_{a}$ or $P_{a}$ always gives the same physical result: things are moved from here to there and have the same relative magnitude. We now return to the second motivational observation. Seeking a conformal theory of gravity, we introduce gauge fields for each of the generators of the algebra:
$$M_{ab}, P_{a}, K_{a}, D \longrightarrow \omega_{\alpha}^{ab},
u_{\alpha}^a, v_{\alpha}^a, W_{\alpha}$$
When we gauge the Poincaré group, the first of these, $\omega_{\alpha}^{ab}$, is the spin connection, while the second, the gauge field of translations, $u_{\alpha}^a,$ is identified with the vierbein, $e_{\alpha}^a$. The vierbein (the soldering form of the fibre bundle) is related to the metric tensor, $g_{\alpha\beta}$, via the relation $$g_{\alpha\beta} = e_{\alpha}^a e_{\beta}^b \eta_{ab}$$ where $\eta_{ab}$ is the Minkowski metric.
If we compute the conformal curvature tensor we find that it has certain components corresponding to a generalization of the usual Riemannian curvature, and additional components related to the $u_{\alpha}^a,$ $v_{\alpha}^a,$ and $W_{\alpha}$ fields. A problem arises as soon as we write down a scale-invariant action from this curvature. It has been shown [@Wh1; @Kaku] that for any such action we write, the $v_{\alpha}^a$ field is auxiliary. What happens is that, when the action (which contains 9 independent terms) is varied with respect to $v_{\alpha}^a,$ the resulting field equation for $v_{\alpha}^a$ can be solved and substituted back into the action. The resulting effective action is independent of $v_{\alpha}^a$. A second problem arises as soon as we consider the quantization of a scale-invariant theory. The action of a dilation on the generator of translations is $$e^{-\lambda D} P_{a} e^{\lambda D} = e^{\lambda} P_{a}$$ which simply says that the action of a dilation on an infinitesimal translation is to change the magnitude, but not the direction, of the displacement. But since $P^{2}$ is identified with the square of the mass, this implies the relation $$e^{-\lambda D} P^{2} e^{\lambda D} = e^{2\lambda} P^{2}.$$ As a consequence, masses may be changed by an arbitrary factor, $e^{2\lambda}$. The mass spectrum becomes continuous instead of discrete, in contradiction with experiment. For this reason, it is generally assumed that scale-invariance is not a good symmetry at the quantum level [@Wess; @Carruthers].
Thus, naive implementation of the conformal group leads to the conclusion that the gauge field of the conformal translations may be eliminated and the dilation symmetry must be broken. We are left with the Poincaré symmetry from which we started.
The metric structure of conformal gauge theory
==============================================
The conclusions above concerning the elimination of the gauge field, $v_{\alpha}^a$, hold because we chose to identify the gauge field of translation, $u_{\alpha}^a$, with the vierbein, and hence with the metric. Had we chosen $v_{\alpha}^a$ as the vierbein instead, it would have been $u_{\alpha}^a$ that was auxiliary. The only reason for picking $u_{\alpha}^a$ is that that is what is done when gauging the Poincaré group. With the conformal group, there is an option. This observation is central. It means that there is an additional symmetry, implicit in the conformal group, which did not need to be broken. The only thing that breaks the symmetry between the two translations, $P_{a}$ and $K_{a},$ is our arbitrary choice. If we can rewrite or alter the conformal group in such a way as to make this symmetry explicit, then we may find some new content to conformal gauge theory after all.
This reworking of the group is dependent on the introduction of a metric on the underlying manifold. What we’re going to do is to find all metrics which can be constructed from the gauge fields $u_{\alpha}^a$ and $v_{\alpha}^a$, and rebuild the group in a way that guarantees that we can independently pick any of the possibilities as a gauge choice.
It is easy to establish that there are precisely three rank-2, symmetric tensor fields constructible from the gauge fields $u_{\alpha}^a$ and $v_{\alpha}^a$. They are:
$$g^{1}_{\alpha\beta} = u_{\alpha}^a u_{\beta}^b \eta_{ab}$$ $$g^{2}_{\alpha\beta} = u_{(\alpha}^a v_{\beta)}^b \eta_{ab}$$ $$g^{3}_{\alpha\beta} = v_{\alpha}^a v_{\beta}^b \eta_{ab}$$
While there is no guarantee that any of these metrics is invertible or torsion-free, the same is true of the gauge theory of the Poincaré group. Invertibility and vanishing torsion are assumptions which must be made to reproduce general relativity from the gauge theory, and we make the same assumptions here. The status of these assumptions is a subject of debate. Certainly, invertibility holds generically. As for the vanishing of the torsion, it is still an open question whether torsion does vanish. While the macroscopic limits are quite stringent, there is always the possibility of consistently interpreting some physical field as torsion.
The next step is to write the symmetry so that $g^{1}_{\alpha\beta}$, $g^{2}_{\alpha\beta}$, and $g^{3}_{\alpha\beta}$ are possible gauge choices. This is achieved by introducing a vierbein, $e^{i}_{\alpha}{}^{a} (i = 1, 2,
3)$, for each possible metric, and introducing a translation generator, $T^{i}_{a}$, as the operator for which $e^{i}_{\alpha}{}^{a}$ is the gauge field. Clearly, we can let $T^{1}_{a} = P_{a}$ and $T^{3}_{a} = K_{a}$, but $T^{2}_{a}$ is new. One might think that $e^{2}_{\alpha}{}^{a}$ could be gotten by taking some combination $\lambda_{1}P_{a} + \lambda_{2}K_{a}$, but such combinations always introduce some measure of $g^{1}_{\alpha\beta}$ and $g^{3}_{\alpha\beta}$ in addition, so the middle metric would not be independent.
The presence of the new generator $T^{2}_{a}$ changes the conformal algebra, and we are faced with a choice of several ways to close the new algebra. We could simply let the new translation commute with all of the other generators, but this means that it rotates as four scalars instead of as a vector. We can extend the group until it can contain the generator we want, or we can contract the group. Both of these latter ways work. For example, $O(5,3)$ contains two translations and gle of 4-spinor. Because the norm is preserved, transformations of $\psi$ produce rotations of $g^{A}.$
When a particular metric is singled out, the $SU(4)$ gauge is partially fixed. There remains an $SU(3)$ subgroup which leaves $\psi$ invariant. In addition, there additional symmetries present. These symmetries may depend on the particular $\psi$ chosen. We prove the following theorem:
Fixing the metric reduces the $SU(4)$ symmetry to $SU(3) \times C(1) \times K,$ where $C(1)$ is a bounded, one-parameter group and $K$ is a discrete group. When the spacetime metric has definite scaling weight, $C(1) = U(1)$ and $K$ includes the integers.
Proof: We first demonstrate that the subgroup which leaves a fixed spinor invariant is $SU(3).$ Let $\psi$ be fixed, let $U$ be a unitary transformation and let $\bar{U}$ be that transformation that maps $\psi$ to $$\bar{U}\psi = \psi_{0} = \left( \begin{array}{c}
\alpha \\ \alpha* \\ 0 \\ 0
\end{array} \right)$$ Then for every transformation $U$ that leaves $\psi$ invariant, $$U\psi = \psi,$$ we can construct another one, $$U_{0} = \bar{U}U\bar{U}^{\dagger},$$ that leaves $\psi_{0}$ invariant, and vice-versa: $$U_{0}\psi_{0} = (\bar{U}U\bar{U}^{\dagger}) (\bar{U}\psi) = \bar{U}U\psi
= \bar{U}\psi = \psi_{0}$$ $$U\psi = (\bar{U}^{\dagger}U_{0}\bar{U}) (\bar{U}^{\dagger}\psi_{0}) =
\bar{U}^{\dagger}U_{0}\psi_{0} = \bar{U}^{\dagger}\psi_{0} = \psi$$
Therefore, the group that leaves $\psi$ invariant is the same as the group that leaves $\psi_{0}$ invariant. But this group was shown by Wheeler [@Wh2] to be $SU(3),$ by a direct construction of the infinitesimal Hermitian generators. These generators are required to satisfy $H\psi = 0.$
The existence of a one-parameter, bounded symmetry, $C(1),$ follows immediately from the expression for $[g]$ in terms of $\psi$. Each component of $\psi$ is bounded by the norm of $[g].$ Furthermore, $\psi$ has four complex components, constrained by the constancy of the norm, $\psi^{\dagger}\psi$. Therefore, seven degrees of freedom in $\psi$ parameterize the six independent components of $[g]$, leaving a one-parameter family of solutions to the algebraic equations for $\psi_{\alpha}([g]).$ Since the equations to be solved for $\psi([g])$ are of quadratic or higher order, there will be more than a single root, providing a discrete symmetry, $K.$ Additional discrete symmetry may also be provided by the phase transformations of the components.
The spacetime metric has definite Weyl weight if and only if $[g]$ has only one nonvanishing component. In these cases $\psi$ has exactly two nonvanishing components. For example, we may have $$\psi = \left( \begin{array}{c} \alpha \\
\beta \\ 0 \\ 0 \end{array} \right)$$ If the phase of $\alpha$ is shifted by $\delta$ and the phase of $\beta$ by $-\delta$, $[g]_{12} = Re(\alpha\beta)$ remains invariant. This is a $U(1)$ symmetry. Note that it does not commute with the $SU(3)$ symmetry, but for any change of phase, $\delta$, it is trivial to write down the new generators of $SU(3).$
Finally, when the metric is of definite Weyl weight, there is an Hermitian transformation, $A$, which commutes with $SU(3)$ satisfying $A\psi = \psi.$ When exponentiated, the net effect is a phase change $$U\psi = e^{i\delta A} \psi = e^{i\delta} \psi$$ In general this does not leave $[g]$ invariant. However, when $\delta = n\pi,
[g]$ is unchanged. The transformation remains nontrivial when acting on spinors other than $\psi$, and is distinct for different values of $n.$ The symmetry group, $K,$ therefore contains the integers.
The $SU(2)$ symmetry
====================
There is a remaining symmetry of the metric choice, which is most naturally thought of as arising because the metric is symmetric in the gauge fields $e^{i}_{\alpha}{}^{a}$. It is natural to ask about the character of the antisymmetric combination $$F^{i}_{\alpha\beta} = \epsilon^{i}_{jk} e^{j}_{\alpha}{}^{a}
e^{k}_{\beta}{}^{b} \eta_{ab} = - F^{i}_{\beta\alpha}$$ where $\epsilon_{ijk}$ is the Levi-Civita symbol. $F^{i}_{\alpha\beta}$ has the index structure of a Yang-Mills field. Applying the same criteria used to arrive at $SU(4),$ we have the space $R^{3} - \{0\}$ covered by $R \times
O(3),$ with compact part $O(3)$ and covering group $SU(2).$ This is precisely the additional group required to give the standard model.
For $F^{i}_{\alpha\beta}$ to be a gauge field, it must arise from a gauge potential. Interestingly, the necessary condition depends on the vanishing of part of the torsion. Normally, in the gauging of Poincaré symmetry, we impose the condition $$D_{[ \alpha}e_{\beta ]}^{a} = T^{a}_{\alpha\beta} = 0$$ on the vierbein. It is sufficient to demand that each metric in our class be torsion-free. We therefore require the same condition of each of the three vierbein components, $$D_{[ \alpha}e^{k}_{\beta ]}{}^{a} = 0$$ Contracting with $\eta_{ab} \epsilon^{i}_{jk} e^{j}_{\mu}{}^{b}$ and antisymmetrizing yields the Bianchi identity for $F^{i}_{\alpha\beta}:$ $$D_{[ \alpha}F^{i}_{\mu\beta ]} = 0$$ $F^{i}_{\alpha\beta}$ therefore arises from an $SU(2)$ gauge potential, providing the final symmetry required for the standard model.
Gauging
=======
We now have shown the existence of $SU(3) \times SU(2) \times U(1) \times Z$ symmetry as the residual gauge group following any choice of metric of definite scaling weight. The full symmetry is therefore the standard model, together with the Poincaré group and a discrete symmetry. We assumed that the dilation symmetry is broken even though it does not directly give a scaling of the mass as assumed by Wess [@Wess]. Even if dilations were allowed, the standard model symmetries would still remain.
It is important to note that the new unitary symmetries are independent of the Poincaré symmetry. The Poincaré symmetry is the remnant of the original conformal symmetry. The unitary symmetry was introduced to classify the metrics allowed by the conformal gauge fields, but has no direct relationship to the translation, rotation or boost symmetries. The gauging of the group may proceed along the usual lines, with the exception that the product of the electromagnetic and strong symmetries is semi-direct and not direct.
Still more interesting is the possibility of investigating what happens if the entire $SU(4)$ symmetry is maintained. The spacetime metric may be regarded as an $SU(4)$-valued tensor field, $g^{A}_{\alpha\beta}$, and the curvature for the full symmetry derived. It remains to be seen whether an appropriate action emerges naturally in this approach.
Finally, we conjecture that the correct way to introduce matter fields is through supersymmetrization of the model.
[9]{} C. N. Yang and R. L. Mills, Phys. Rev. [**96**]{} (1954) 191.\
R. Utiyama, Phys. Rev. [**101**]{} (1956) 1597.\
S. L. Glashow, Nucl. Phys. [**22**]{} (1961) 579.\
S. Weinberg, Phys. Rev. Lett. [**19**]{} (1967) 1264.\
A. Salam, In [*Elementary Particle Theory,*]{} ed. N. Svartholm, Stockholm, Almquist Forlag AB, (1968) 367.\
C. N. Yang, Ann. N. Y. Acad. Sci. [**294**]{} (1977) 86. J.T.Wheeler, to be published in Phys.Rev.D. M. Kaku, P.K. Townsend and P. Van Nieuwenhuizen, Phys. Lett. [**69B**]{} (1977) 304. J. Wess, Nuovo Cim.[**18**]{} (1960) 1086. P. Carruthers, Phys. Rep. 1, no.1 (1971) 1. J.T. Wheeler, USU preprint FTG-105-USU, June 1991 (submitted for publication).
| {
"pile_set_name": "ArXiv"
} |
‘@=11 \#1[@underline\#1 $\@@underline{\hbox{#1}}$]{} ‘@=12
==c===ł=Ł=ø=Ø==u c i ł ø u Ł Ø ¶ §
\#1 \#1
=cmex10 scaled 1200
\#1[\^[\#1]{}]{} \#1[\_[\#1]{}]{} \#1[/\#1]{} /\#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1[\#1]{} \#1 \#1 \#1 \#1[\#1|]{} \#1[| \#1]{} \#1[\#1]{} \#1[| \#1|]{} \#1 \#1[[ ]{}]{} \#1 \#1 \#1 \#1\#2 \#1\#2[[1 1]{}]{} \#1\#2[[1 1]{}]{} \#1\#2[[1\#2]{}]{} \#1\#2[[\#1\#2]{}]{} \#1\#2[[\#1\#2]{}]{} \#1\#2\#3[[\^2 \#1\#2 \#3]{}]{} \#1\#2[\^[\#1]{}]{} \#1 \#1 $${\lfloor{\hskip 0.35pt}\!\!\!\lceil}
\def$$ \#1\#2[\_[\#1]{}\^[\#2]{}]{} \#1\#2[\^[\#1]{}\_[\#2]{}]{} \#1\#2\#3[\_[\#1]{}\^[\#2]{}\_[\#3]{}]{} \#1\#2\#3[\^[\#1]{}\_[\#2]{}\^[\#3]{}]{}
\#1\#2 \#1
\
[\[\#1\]]{}
\#1\#2[[[\#1]{}]{}]{} \#1[[^1]]{} \#1[[-3pt]{}]{}
\#1 \#1 \#1 \#1
\#1[0.015in]{} \#1
plus 1000pt minus 1000pt \#1 \#1[= to]{} \#1[= to]{}
\#1[$\sp{#1)}$]{}
\#1\#2\#3[Phys. Lett. [**[\#1]{}B**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Nucl. Phys. [**B[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Phys. Rev. Lett. [**\#1**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Phys. Rev. [**D[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Class. and Quantum Grav. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Commun. Math. Phys. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[J. Math. Phys. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Ann. of Phys. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Phys. Rep. [**[\#1]{}C**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Progr. Theor. Phys. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Int. J. Mod. Phys. [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Mod. Phys. Lett. [**A[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[Nuovo Cim. [**[\#1]{}**]{} (19[\#2]{}) \#3]{} \#1\#2\#3[[*ibid.*]{} [**[\#1]{}**]{} (19[\#2]{}) \#3]{}
=3ex =4ex = =1em plus.3em minus.5em =1em plus.3em minus.5em =.5em plus.2em minus.4em =.5em plus.2em minus.4em \#1[[**[\#1]{}**]{} ]{} \#1[\#1]{}
1.3cm
[**The String Measure and Spectral Flow\
of Critical $~N=2~$ Strings**]{} [^2]\
.3in
Sergei V. Ketov [^3] and Olaf Lechtenfeld .1in [*Institut für Theoretische Physik, Universität Hannover*]{}\
[*Appelstraße 2, 30167 Hannover, Germany*]{}\
[*ketov, lechtenf @itp.uni-hannover.de*]{}
.2in
[**Abstract**]{}
The general structure of $N{=}2$ moduli space at arbitrary genus [*and*]{} instanton number is investigated. The $N{=}2$ NSR string measure is calculated, yielding picture- and $U(1)$ ghost number-changing operator insertions. An explicit formula for the spectral flow operator acting on vertex operators is given, and its effect on $N{=}2$ string amplitudes is discussed.
[**1**]{} [*Introduction*]{}. The recent interest [@bv; @b; @o] in computing loop scattering amplitudes of critical $N{=}2$ strings in $2{+}2$ dimensions is partially motivated by a rich structure of $N{=}2$ moduli space, which leads to some novel features of string perturbation theory when compared to the $N{=}1$ and $N{=}0$ cases. In refs. [@bv; @b; @o], an $N{=}4$ (twisted) supersymmetrical topological description of critical $N{=}2$ strings was used to bypass problems with the (bosonized) $N{=}2$ superconformal ghosts and the location of the $N{=}2$ picture-changing operators. Still, it is of interest to pursue the non-topological approach of integrating correlation functions of BRST-invariant vertex operators over all $N{=}2$ supergravity moduli. We shall show, in particular, how the integration over $U(1)$ moduli generates $U(1)$ ghost number-changing operators and implements the spectral flow for string amplitudes.
In refs. [@klp; @bkl; @l], we have investigated the critical $N{=}2$ string in the BRST formalism, but confined ourselves to calculating the BRST cohomology and tree amplitudes without $U(1)$ instantons. In this letter, we are going to describe the structure of the $N{=}2$ moduli space at arbitrary genus [*and*]{} $U(1)$ instanton number, and calculate the string measure resulting from integrating out fermionic as well as $U(1)$ moduli.
The gauge-invariant $N{=}2$ string world-sheet action is given by the minimal coupling of $N{=}2$ matter represented by two $N{=}2$ scalar multiplets $(X^\m,\j^\m)$ [^4] to $N{=}2$ supergravity comprising a zweibein $e^a_{\a}$ or a metric $g_{\a\b}$, a real $U(1)$ gauge field $A_{\a}$, and two real gravitini $\c^{\pm}_{\a}$, [^5] all living on the two-dimensional (2d) world-sheet. The local symmetries are given by 2d reparametrizations, local Lorentz and Weyl invariances, $N{=}2$ supersymmetry and $N{=}2$ super-Weyl symmetry, phase $U_{\rm V}(1)$ and chiral $U_{\rm A}(1)$ invariances. [^6] The world-sheet $\S$ is supposed to arise from a closed orientable $N{=}2$ super-Riemann surface, whose topology is characterized by two integers, its Euler number $\c$ (related to the genus $h$) and its first Chern class (or $U(1)$ instanton number [^7] ) $c\,$, $$\c=\fracmm{1}{2\p}\int_{\S}R~=2-2h~,\quad h\in {\bf N}~;\qquad
c=\fracmm{1}{2\p}\int_{\S}F~=k~,\quad k\in {\bf Z}~.
\eqno(1)$$ Here, the curvature two-form $R$ and the $U(1)$ field strength two-form $F=dA$ have been introduced. Finally, correlation functions require the introduction of a third integer, $n\in{\bf N}$, which counts the number of punctures on $\S_{h,n}$. .1in
[**2**]{} [*The U(1) bundles*]{}. The $U(1)$ gauge field $A_{\a}=(A_z,A_{\bar{z}})$ is the vertical connection on a principal $U(1)$ bundle over $\S_{h,n}$. It can always be split as $A=A^{\rm inst} + A^0$, where the instantonic part $A^{\rm inst}$ saturates the instanton number, $c(A^{\rm inst})=k$, and $A^0$ is globally defined on $\S_{h,n}$ since $c(A^0)=0$. The ambiguity in choosing a representative $A^{\rm inst}$ is contained in the space of $A^0$’s. Since we are going to integrate over the moduli space of all $\,U(1)$ connections, we need to decompose $A^0$ into gauge and moduli parts. The Hodge decomposition on $\S_{h,n}$ reads $$A^0 \ =\ d\f +\d \o + A^{\rm Teich}~,
\eqno(2)$$ where the function $\f$ and two-form $\o$ represent the $U_{\rm V}(1)$ and $U_{\rm A}(1)$ gauge degrees of freedom, respectively, and $A^{\rm Teich}$ is a harmonic one-form on $\S_{h,n}$ representing the $U(1)$ Teichmüller degrees of freedom. [^8] In the presence of punctures, however, this decomposition is not yet unique. There exist harmonic one-forms which are also exact or co-exact but diverge at the punctures. In locally flat holomorphic coordinates around a puncture, the co-exact prototype is $$\tilde A\ =\ {1\over2i}\left({dz\over z}-{d\bar z\over\bar z}\right)
\ =\ d\tilde\f\ =\ \d\tilde\o \quad,\qquad
\tilde\f\ =\ {1\over2i}\ln {z\over\bar z} \quad,\quad
\tilde\o\ =\ {i\over2}\ln|z|\,dz\wedge d\bar z
\eqno(3)$$ which is harmonic because $$\tilde F\ =\ d\tilde A\ =\ 2\p i\,\d^{(2)}(z)\,dz\wedge d\bar z
\eqno(4)$$ vanishes away from the puncture, implying $\D\tilde A=0$ at least locally on $\S_{h,n}$. $\tilde A$ is not exact because $\tilde\f$ is multi-valued around $z{=}0$. Clearly, $\l\tilde A$ represents a Dirac monopole with magnetic charge $\l$. [^9] Similarly, one may utilize the Hodge dual of $\tilde A$, $$*\tilde A\ =\ -{1\over2}\left({dz\over z}+{d\bar z\over\bar z}\right)
\ =\ \d *\tilde\f\ =\ -d *\tilde\o \quad,
\eqno(5)$$ to manufacture an exact (but not co-exact) harmonic one-form which does not contribute to the curvature $F$ at all. We shall render the decomposition (2) unique by putting such gauge fields into the $A^{\rm Teich}$ part, since they do not really represent $U_{\rm V}(1)$ or $U_{\rm A}(1)$ gauge degrees of freedom but do something non-trivial to the punctures. With $\f$ and $\o$ being smooth at the punctures, the $U_{\rm A}(1)$ gauge symmetry changes $F$ but not $c$. Likewise, the $\,U(1)$ Teichmüller variations must respect $c(A^0)=0$, $$0\ =\ \fracmm{1}{2\p}\int_{\S_{h,n}}dA^0\ =\
-\fracmm{1}{2\p}\sum_{\ell=1}^n \oint_{z_\ell} A^{\rm Teich} \quad.
\eqno(6)$$ A second condition arises from the fact that the Lorentz gauge condition, $\d A=0$, can be chosen globally, $$0\ =\ \fracmm{1}{2\p}\int_{\S_{h,n}}*\d A^0\ =\
-\fracmm{1}{2\p}\sum_{\ell=1}^n \oint_{z_\ell} *A^{\rm Teich} \quad.
\eqno(7)$$
The complex structure of $\S$ naturally suggests the complex linear combinations $$A^\pm\ =\ A^0\pm i*A^0 \quad,\qquad d^\pm\ =\ d\pm i*d \quad,\qquad
\f^\pm\ =\ \f\mp i*\o \quad
\eqno(8)$$ so that $$A^\pm\ =\ d^\pm\f^\pm+A_{\rm Teich}^\pm \quad,\qquad
F^\pm\ =\ dA^\pm\ =\ F^0\pm i*\d A^0 \quad,
\eqno(9)$$ and the singular one-forms of eqs. (3) and (5) combine to $$\tilde A^+\ =\ {dz\over z}\quad,\qquad\tilde A^-\ =\ {d\bar z\over\bar z}\quad.
\eqno(10)$$
Independent of $c$, it is easy to count the dimension of the $U(1)$ Teichmüller space. On $\S_{h,n}$, one has $2h$ real abelian one-forms $\a_i$ and $\b_i$, $i=1,\dots,h$, which may be chosen to be dual to the basis $\{a_i,b_i\}$ of homology cycles, $$\oint_{a_i}\a_j=\d_{ij}\quad,\qquad \oint_{b_i}\a_j=0\quad,\qquad
\oint_{a_i}\b_j=0\quad,\qquad \oint_{b_i}\b_j=\d_{ij}\quad,
\eqno(11)$$ as well as $2n$ real one-forms of the type given in eqs. (3) and (5), namely two for each puncture. However, eqs. (6) and (7) put two real constraints on the coefficients multiplying the latter, so that the total real dimension equals $2h+2n-2$ for $n{>}0$.
Not all $U(1)$ Teichmüller variations are moduli, however. Whenever the Wilson loops $$W[A;\g]\ =\ \exp \Bigl\{ i\oint_\g A \Bigr\} \qquad {\rm and} \qquad
*\!W[A;\g]\ =\ \exp \Bigl\{ i\oint_\g *A \Bigr\}
\eqno(12)$$ become trivial ($=1$) for some $A^{\rm Teich}$ and all cycles $\g$, we have encountered a ‘big’ $U(1)$ gauge transformation. Invariably, this will happen if the coefficients multiplying the harmonic one-forms get large enough, as for $\l=1$ in $$1 = \exp\Bigl\{i\oint_0\l\tilde A\Bigr\} = e^{2\p i\l} \qquad
\Longleftrightarrow\qquad
i\tilde A = g^{-1} dg \quad{\rm with}\quad g(z)=e^{i\tilde\f(z)}={z\over|z|}~,
\eqno(13)$$ or for $\vf=1$ in $$1 = \exp\Bigl\{i\oint_{a_1} 2\p\vf\a_1\Bigr\} = e^{2\p i\vf} \qquad
\Longleftrightarrow\qquad
2\p i\a_1 = g^{-1} dg \quad{\rm with}\quad g(z)=e^{2\p i\int^z \a_1}~.
\eqno(14)$$ Therefore, the $U(1)$ moduli space of $\S_{h,n}$ is parametrized by $2h+2n-2$ real phases (twists), $$\oint_{a_i}A = 2\p\vf_i \quad,\qquad
\oint_{b_i}A = 2\p\vq_i \quad,\qquad
\oint_{z_\ell}A = 2\p\l_\ell \quad,\qquad
\oint_{z_\ell}*A= 2\p\m_\ell \quad,
\eqno(15)$$ subject to $\sum_\ell \l_\ell=0=\sum_\ell \m_\ell$. All phases range from $0$ to $1$. It is important to note, however, that one may change the value of $c$ simply by shifting the $\l_\ell$ so that the total phase shift is integral. In summary, one obtains (for $n{>}0$) a real torus $$\cm^{U(1)}_{h,n}\ =\ {{\bf R}^{2(h+n-1)}\over{\bf Z}^{2(h+n-1)}}~.
\eqno(16)$$ This picture does not depend on the value $k$ of the instanton number $c$. For a thorough discussion of such matters in the context of $N{=}0$ and $N{=}1$ strings, see ref. [@dp].
From now on we shall employ the $U_{\rm V}(1)$ Lorentz gauge, $\d A^0=0$, and its $U_{\rm A}(1)$ counterpart, $F^0=0$, on $\S_{h,n}$, effectively setting $\f=\o=0$ in eq. (2). The $U(1)$ moduli space can then be seen as the moduli space of flat connections on $\S_{h,n}$, which is a product of two factors. One factor is the Jacobian variety of flat $U(1)$ connections on $\S_{h,0}$, $$J(\S_{h,0})\ =\ \fracmm{{\bf C}^h}{{\bf Z}^h+\O{\bf Z}^h}
\ \sim\ \fracmm{{\bf R}^{2h}}{{\bf Z}^{2h}}\ =\ Pic(h,0) \quad,
\eqno(17)$$ where $\O$ is the period matrix of $\S$. It is diffeomorphic to the torus $Pic(h,0)$ parametrizing all flat holomorphic line bundles over $\S_{h,0}$, with twists $\vf_i$ and $\vq_i$ of eq. (15) on the homology cycles. The other factor is the torus $$Pic(0,n)\ =\ \fracmm{{\bf R}^{2(n-1)}}{{\bf Z}^{2(n-1)}}
\eqno(18)$$ encoding the $2n-2$ independent twists $\l_\ell$ and $\m_\ell$ around the punctures on the Riemann sphere $\S_{0,n}$. The corresponding flat connections are given by linear combinations of the singular one-forms given in eq. (10), for each puncture.
To construct the $N{=}2$ string measure, we need to know the instanton solution $A^{\rm inst}$ and the flat connection $A^{\rm Teich}$ explicitly, in terms of moduli. As far as $A^{\rm inst}$ is concerned, it can be chosen to satisfy the Laplace-Beltrami equation (of motion) and the gauge condition $\d A^{\rm inst}=0$ on the punctureless Riemann surface $\S\equiv\S_{h,0}$. Taken together with the conformal gauge for the 2d metric $g$, they lead to the simple equations $$\pa_{\bar{z}}g^{z\bar{z}}\pa_zA^{\rm inst}_{\bar{z}}\ =\ 0\qquad {\rm and}
\qquad \pa_zA^{\rm inst}_{\bar{z}}+\pa_{\bar{z}}A^{\rm inst}_z\ =\ 0\quad,
\eqno(19)$$ respectively. Since $\S$ is compact, orientable and without boundary, eq. (19) implies that $\pa_zA^{\rm inst}_{\bar{z}}=-\pa_{\bar{z}}A^{\rm inst}_z\sim g_{z\bar{z}}$. The coefficient is easily fixed by the instanton number constraint, $c(A^{\rm inst})=k$, $$\pa_zA^{\rm inst}_{\bar{z}}\ =\ -\pa_{\bar{z}}A^{\rm inst}_z\ =\
\fracmm{\p k}{A}g_{z\bar{z}}\quad,
\eqno(20)$$ where the total area $A=\int_\S d^2z\,g_{z\bar{z}}$ of $\S$ has been introduced. The solution to eq. (20) is ([*cf*]{}. ref. [@f]) $$\eqalign{
A^{\rm inst}_z(z,\bar{z})\ =&\ -\fracmm{\p k}{A}\int_{\S}d^2w\,\pa_z K(z,w)
g_{w\bar{w}} +\p k \sum^h_{i,j=1}\o_i(z)\left|{\rm Im}\O_{ij}\right|^{-1}
\int_{\bar{z}_0}^{\bar{z}}\bar{\o}_j(\bar{w})d\bar{w}\quad,\cr
A^{\rm inst}_{\bar{z}}(z,\bar{z})\ =&\ +\fracmm{\p k}{A}\int_{\S}d^2w\,
\pa_{\bar{z}} K(z,w) g_{w\bar{w}} - \p k \sum^h_{i,j=1}\bar{\o}_i(\bar{z})
\left|{\rm Im}\O_{ij}\right|^{-1} \int_{z_0}^z \o_j(w)dw\quad,\cr}
\eqno(21)$$ where $K(z,w)$ is the Green function to the scalar Laplacian, $$\pa_z\pa_{\bar{z}}K(z,w)\ =\ \d^{(2)}_{z\bar{z}}(z,w) +
\o_i(z)\left|{\rm Im}\O_{ij}\right|^{-1}\bar{\o}_j(\bar{z})\quad~,
\eqno(22)$$ and $\o_i=\a_i+\O_{ij}\b_j$, $i=1,\ldots,h$, are the holomorphic abelian differentials with normalization $$\oint_{a_i}\o_j\ =\ \d_{ij}\qquad,\qquad \oint_{b_i}\o_j\ =\ \O_{ij}\quad.
\eqno(23)$$ The solution to eq. (22) can be expressed in terms of the prime form $E(z,w)$ [^10] as $$K(z,w)\ =\ \ln\left|E(z,w)\right|^2 + \sum^h_{i,j=1}\left[ {\rm Im}
\int^z_w \o_i(u)du\right] \left|{\rm Im}\O_{ij}\right|^{-1}
\left[ {\rm Im} \int^z_w \o_j(u)du\right]~.
\eqno(24)$$ The fields $A^{\rm inst}_{z,\bar{z}}$ are neither holomorphic nor single-valued around the homology cycles $a_i$ or $b_j$ but change by a gauge transformation (as long as $k$ is integer).
Turning to $A^{\rm Teich}$, a convenient parametrization is given by $$A^{\rm Teich}\ =\ 2\p \sum_{i=1}^h (\vf_i\a_i + \vq_i\b_i) +
\sum_{\ell=1}^n (\l^+_\ell \tilde A^+_\ell + \l^-_\ell \tilde A^-_\ell)\quad,
\eqno(25)$$ where, in locally flat complex coordinates, $$A^+_\ell(z)\ =\ {dz\over z-z_\ell} \qquad{\rm and}\qquad
A^-_\ell(z)\ =\ {d\bar z\over \bar z-\bar z_\ell} \quad.
\eqno(26)$$ .1in
[**3**]{} [*The gravitini bundles*]{}. The 2d gravitini $\c^{\pm}_{\a}$ transform inhomogeneously under the $N{=}2$ local supersymmetry and $N{=}2$ super-Weyl (fermionic) gauge symmetry as $$\d_{\rm S}\c^{\pm}_{\a}\ =\ \hat{D}_{\a}\ve^{\pm}\quad,\qquad
\d_{\rm W}\c^{\pm}_{\a}\ =\ \g_{\a}\z^{\pm}\quad,
\eqno(27)$$ where $\hat{D}_{\a}(\hat{\o},A)$ is the $N{=}2$ supergravitational covariant derivative containing the spin connection $\hat{\o}_{\a}$ and the $U(1)$ gauge connection $A_{\a}$. The $\c^{\pm}_{\a}$ transform homogeneously under all other local symmetries, with a $U(1)$ charge of $\pm1$. Hence, the gravitini are sections of some complex spinor bundle associated to the principal $U(1)$ bundle over $\S_{h,n}$.
The local symmetries of eq. (27) allow one to gauge away all 8 real Grassmann degrees of freedom of $\c^{\pm}_{\a}$, except for those in the kernel of $\hat{D}^{\pm}$, $$\hat{D}^{\pm} \c^\pm_z \ \equiv\
\left( \bar{\pa} \mp iA_{\bar{z}} \right) \c^\pm_z \ =\ 0 \quad,
\eqno(28)$$ and similarly for $\c^{\pm}_{\bar{z}}$ (signs are correlated). We have used the superconformal gauge, in which $\hat{\o}_{\bar z}=0$ and all $\c$’s are $\g$-traceless. The solutions of eq. (28) are the fermionic moduli on $\S_{h,n}$. Their number depends on $h$ and $n$ as well as on the instanton number $k$ and is dictated by the Riemann-Roch theorem, $${\rm ind} \hat{D}^{\pm}\ \equiv\
{\rm dim~ker}\,\hat{D}^{\pm} - {\rm dim~ker}\,\hat{D}^{\pm\dg}\
=\ 2(h-1)\pm k +n ~\quad.
\eqno(29)$$ For $h>1$, the contributions to a positive index generically come from the first term, so that on reads off $2(h-1)+n+k$ positively charged and $2(h-1)+n-k$ negatively charged fermionic moduli. When the index becomes negative, $\det\hat{D}^{\pm\dg}$ develops zero modes, which implies the vanishing of the corresponding $n$-point correlator. This restricts the range of the sum over $k$ to $$\abs{k}\ \leq\ 2(h-1)+n \quad.
\eqno(30)$$
The issue of a complex structure for the gravitini bundles with $k\ne0$ is a subtle one since $A_{\bar{z}}$ in $\hat{D}^{\pm}$ contains the non-holomorphic $A_{\bar{z}}^{\rm inst}$ of eq. (21). Even without punctures ($n{=}0$), the gravitini bundles cannot in general be [*holomorphic*]{} line bundles. The latter are (twisted) integral or half-integral powers of the canonical line bundle and as such always yield integral multiples of $h{-}1$ for ${\rm ind}\hat{D}$. The r.h.s. of eq. (29), however, is of this form only when $k$ itself is a multiple of $h{-}1$, in which case the gauge connection may be absorbed in the spin connection, effectively shifting the conformal weights of the gravitini from $\fracmm32$ to $\fracmm32$, to $\fracmm32{\pm}\fracmm12$, or to $\fracmm32{\pm}1$. .1in
[**4**]{} [*The string measure*]{}. Our starting point is the formal expression for scattering amplitudes, $$A_n\ =\ \sum_{h,k}\fracmm{1}{\cn}\int \! D(X\j g\c A)\,
e^{-S_{\rm m}}\,V_1\ldots V_n \quad,
\eqno(31)$$ where $S_{\rm m}$ is the gauge-invariant $N{=}2$ string (matter) action, $V_\ell$ represent vertex operators for particles, and $\cn$ denotes the volume of the gauge group. We fix the $N{=}2$ superconformal gauge and use the BRST method [@bkl]. A careful treatment of the Faddeev-Popov determinant, naively [^11] $$\int \! D(bc\b\g\tilde b\tilde c)\, e^{-S_{\rm gh}} \quad,
\eqno(32)$$ yields anti-ghost zero mode insertions for each moduli direction. As we already know from $N{=}0$ and $N{=}1$ string theory [@dp], these anti-ghost insertions come paired with the corresponding Beltrami differentials which are the tangents to the moduli slice. Let us take $h{>}1$ for simplicity; the cases of the sphere and torus require obvious minor modifications due to isometries. We get $$\abs{\prod^{3(h-1)+n}_{m=1}\int_{\S}\m_m b\,}^2\ =\
\abs{\prod^{3(h-1)+n}_{m=1}\oint_{C_m}\! b\,}^2
\eqno(33)$$ for the metric moduli, where the Beltrami differentials $(\m_m)^z_{\bar{z}}=\de_{\bar{z}}(v_m)^z$ have been represented in terms of quasiconformal vector fields $v_m$ with a unit jump across closed contours $C_m\,$. The commuting superconformal ghosts yield $$\abs{\prod_{a^+=1}^{2(h-1)+k+n}\d\Bigl(\b^+(z_{a^+})\Bigr)\quad
\prod_{a^-=1}^{2(h-1)-k+n}\d\Bigl(\b^-(z_{a^-})\Bigr)\,}^2
\eqno(34)$$ for the $N{=}2$ fermionic moduli, with a delta-function choice for the fermionic Beltrami differentials. [^12] Finally, as a novel feature we obtain $$\prod_{i=1}^{h}\;\biggl[\oint_{a_i}\tilde{b}\;\oint_{b_i}\tilde{b}\biggr]
\quad \abs{ \prod_{\ell=1}^{n-1} \oint_{z_\ell}\tilde{b} \,}^2
\eqno(35)$$ for the $U(1)$ moduli, after taking the real abelian one-forms for the $U(1)$ Beltrami differentials. It must be added that this counting presumes that the vertex operators $V_\ell$ are taken from the natural picture- and ghost-number sector, namely $(\p^+,\p^-)=(-1,-1)$ and of $\tilde c c$ type. [^13]
The $N{=}2$ supergravity fields enter the full (BRST-invariant) $N{=}2$ string action $S_{\rm tot}=S_{\rm m}+S_{\rm gh}$ as Lagrange multipliers [@klp; @bkl]. Since the action $S_{\rm tot}$ is [*linear*]{} in the fermionic and $U(1)$ moduli, we may integrate those out and arrive at an additional insertion of $$\abs{
\prod_{a^+=1}^{2(h-1)+k+n} \! G^+_{\rm tot}(z_{a^+}) \,
\prod_{a^-=1}^{2(h-1)-k+n} \! G^-_{\rm tot}(z_{a^-}) \quad
\prod_{\ell=1}^{n-1} \d\Bigl( \oint_{z_\ell}\! J_{\rm tot} \Bigr)\,}^2 \quad
\prod_{i=1}^{h} \biggl[
\d\Bigl( \oint_{a_i}\! J_{\rm tot} \Bigr) \,
\d\Bigl( \oint_{b_i}\! J_{\rm tot} \Bigr) \biggr]
\eqno(36)$$ where $G^{\pm}_{\rm tot}$ and $J_{\rm tot}$ are the full (BRST-invariant) supercurrents and $U(1)$ current of the $N{=}2$ string, respectively [@klp; @bkl].
Combining eqs. (33)–(36), we find a product of $N{=}2$ picture-changing operators ([*cf.*]{} ref. [@pco]), $$\abs{
\prod_{m=1}^{3(h-1)+n}\!\oint_{C_m}\!b \;
\prod_{a^+=1}^{2(h-1)+k+n}\! Z^+(z_{a^+}) \,
\prod_{a^-=1}^{2(h-1)-k+n}\! Z^-(z_{a^-}) \;
\prod_{\ell=1}^{n-1} Z^0(z_\ell) \,}^2 \;
\prod_{i=1}^{h} \Bigl[ Z^0(a_i)\,Z^0(b_i) \Bigr] \,
\eqno(37)$$ where $$Z^{\pm}\ :=\ \d(\b^{\pm})\,G^{\pm}_{\rm tot}\ =\ \{Q_{\rm BRST},\x^{\pm}\}
\eqno(38)$$ and $$Z^0(\g)\ :=\ \Bigl( \oint_{\g} \tilde{b} \Bigr) \;
\d \Bigl( \oint_{\g} J_{\rm tot} \Bigr)
\eqno(39)$$ for a closed contour $\g$ (being a homology cycle or encircling a puncture). Up to $n$ of the $b$ insertions can be used to convert $c$-type vertex operators to integrated ones, and any number of $Z^\pm$ may be taken to upwardly change their $(-1,-1)$ pictures, keeping the total picture at $(2(h{-}1){+}k,2(h{-}1){-}k)$. The novel $U(1)$ picture-changing operators $Z^0$ serve two purposes: First, a $Z^0$ associated with a homology cycle enforces a projection onto charge-neutral excitations propagating across the loop and ensures factorization on neutral states when pinching the cycle. Second, a $Z^0$ attached to a puncture can be absorbed by the $\tilde c$-type vertex operator, changing its $U(1)$ ghost number by removing the $\tilde c$ factor. In this way, only a single $\tilde c$-type vertex operator remains in the end, plus the $2h$ real $Z^0$ insertions associated with the homology cycles. Thus not only $Z^\pm$ but also $Z^0$ provides a map between BRST cohomology classes. [^14]
Eq. (37) is formally BRST-invariant, which is important for the consistency and BRST-invariance of $N{=}2$ string amplitudes. The integration over the $N{=}2$ matter fields does not present a principal problem, but it has to be done in the presence of a background instanton field $A^{\rm inst}$ minimally coupled to the fermions. No sum over fermionic spin structures appears since it has already been carried out as part of the $U(1)$ moduli integration. The integration over the metric moduli follows the lines of the familiar $N{=}0$ and $N{=}1$ cases.
[**5**]{} [*The spectral flow*]{}. The $N{=}2$ NSR string fermionic matter fields $\j^{\pm\m}$ are sections of a complex twisted spinor bundle, just like the gravitini. For vanishing instanton number, $k{=}0$, this becomes a twisted holomorphic spinor bundle. The associated spin structures parametrizing the NS/R sectors or monodromies of $\j$ in an $n$-point function are labeled by the half-points $(\fracmm12{\bf Z}/{\bf Z})^{2(h+n-1)}$ in the Picard variety of eq. (16). Now observe that by a unitary transformation via $$U(z)\ =\ \exp\Bigl\{ i\int^z_{z_0} \hat{A} \Bigr\}
\eqno(40)$$ of a given spinor bundle we can always change the monodromies with a suitable $\hat{A}\in\{A^{\rm Teich}\}$, because the $\j$ carry $U(1)$ charge [@amv]. We can in fact reach any point in the Picard variety and, in particular, move to any other spin structure. Since the unitary transformation of eq. (40) is equivalent to a shift in the integration variables $A^{\rm Teich}\to A^{\rm Teich}+\hat{A}$ and the $U(1)$ moduli space has no boundary, we conclude that the sum over the NSR spin structures (and, in fact, over all intermediate monodromies) is automatically contained in the integration over the $U(1)$ moduli. This is the so-called [*spectral flow*]{} of Ooguri and Vafa [@ov].
As a consequence, the distinction between NS and R sectors is $U(1)$ moduli-dependent and thus cannot be physical. This feature is not restricted to $h\geq 1$, but appears just as well for the $n$-punctured sphere, i.e. for tree-level $N{=}2$ string amplitudes. More precisely, any pair of R-type punctures can be turned into a pair of NS-type, since there are $n{-}1$ independent cycles and the sum of all twists has to vanish. Hence, the R-type and NS-type states of the $N{=}2$ string cannot be physically distinguished and their correlators must coincide, which is consistent with our previous explicit calculations [@bkl]. Note, however, that the correlation functions should depend on the value $k$ of $c$, which may be changed by allowing for a non-zero but integral total twist of the external states.
A shift in the $U(1)$ puncture moduli, $\l^+_\ell\to\l^+_\ell+\q_\ell$, which shifts the puncture monodromies by $\q_\ell$, also modifies the vertex operators $V_\ell(z_\ell)$ present in the string path integral, $$V_\ell(z_\ell)\ \longrightarrow\
V_\ell^{\q_\ell}(z_\ell)\ \equiv\ \sfo(\q_\ell,z_\ell)\,V_\ell(z_\ell)\quad,
\eqno(41)$$ where, remarkably, the spectral-flow operator $\sfo(\q,z)$ can be written in the explicit form [^15] $$\sfo(\q,z)\ =\ \exp\Bigl\{ \q\int^z \! J_{\rm tot}(z')dz' \Bigr\}\
=\ \exp\bigl\{ \q\,(\f^+ -\f^- -\vf^+ +\vf^- +\tilde{b}c)(z) \bigr\}~.
\eqno(42)$$ Here, we have bosonized [@bkl] $$J_{\rm tot}\ =\ \{Q_{\rm BRST},\tilde{b}\}\
=\ -\fracmm12\j^+{\cdot}\j^- +\pa(\tilde{b}c)+\fracmm12(\b^+\g^- {-}\b^-\g^+)
\eqno(43)$$ in the holomorphic basis via ($\e{=}\pm1$) $$\j^{\pm,\e} = 2e^{\e\f^{\pm\e}} \quad,\qquad
\g^\pm = \h^{\pm}e^{+\vf^{\pm}} \quad,\qquad
\b^\pm = e^{-\vf^{\mp}}\pa\x^{\pm} \quad,
\eqno(44)$$ and obtained a [*local*]{} operator. It is easy to check that $\sfo$ is BRST invariant but not BRST trivial. However, like for the $Z^\pm$, the derivative $\pa\sfo$ is BRST trivial, so we may move spectral-flow operators around at will on the world-sheet. Therefore, an insertion of $\prod_\ell \sfo(\q_\ell,z_\ell)$ is BRST-equivalent to unity as long as $\sum_\ell \q_\ell=0$, implying again the invariance of string amplitudes under shifts of the puncture monodromies, i.e. $$\Bigl\langle V_1 V_2\ldots V_n \Bigr\rangle\ =\
\VEV{V_1^{\q_1} V_2^{\q_2}\ldots V_n^{\q_n}} \qquad{\rm for}\qquad
\sum_\ell \q_\ell =0 \quad.
\eqno(45)$$
The $N{=}2$ superconformal algebra generated by $(T_{\rm tot}, G^\pm_{\rm tot}, J_{\rm tot})$ is extended to the [*small*]{} $N{=}4$ superconformal algebra by adding the $SU(2)$ ladder operators $\sfo(\pm1)$ and closing the algebra.
The NS$\leftrightarrow$R exchange is accomplished at, $\q=\pm\fracmm12$, which may be symbolically written as $\sfo^\pm{\rm NS}={\rm R}^\pm$ where $\sfo^\pm\equiv\sfo(\pm\frac12)$ and ${\rm NS(R)}\equiv V_{\rm NS(R)}$. The index on R indicates that we can flow to two different Ramond vertex operators. The $N{=}2$ string vertex operators are known to exist in different (holomorphic) pictures $(\p^+,\p^-)$, which are connected by the process of picture-changing $(\p^+,\p^-)\to(\p^+{+}1,\p^-)$ or $(\p^+,\p^-)\to(\p^+,\p^-{+}1)$ via the picture-changing operators $Z^+$ or $Z^-$ of eq. (38), respectively. The spectral flow operators $\sfo^{\pm}$ can actually be interpreted as yet additional picture-changing operators: $$\sfo^\pm:\qquad (\p^+,\p^-)\ \to\ (\p^+{\pm}\frac12,\p^-{\mp}\frac12)\quad.
\eqno(46)$$ This leads, for example, to the following identifications among tree-level two-, three- and four-point functions at instanton number $k{=}0$: $$\eqalign{
&\Bigl\langle{\rm NS}\;{\rm NS}\Bigr\rangle\ =\
\VEV{{\rm R}^+\,{\rm R}^-}\quad,\quad
\Bigl\langle{\rm NS}\;{\rm NS}\;{\rm NS}\Bigr\rangle\ =\
\VEV{{\rm NS}\;{\rm R}^+\,{\rm R}^-} \quad,\cr
&\Bigl\langle{\rm NS}\;{\rm NS}\;{\rm NS}\;{\rm NS}\Bigr\rangle\ =\
\VEV{{\rm NS}\;{\rm NS}\;{\rm R}^+\,{\rm R}^-}\ =\
\VEV{{\rm R}^+\,{\rm R}^-\,{\rm R}^+\,{\rm R}^-}\ =\ 0
\quad,\cr}
\eqno(47)$$ which were all verified by explicit calculations in ref. [@bkl]. At tree-level, non-vanishing correlators require $\abs{k}\leq n{-}2$, so that the complete three-point amplitude, for example, also has $k{=}\pm1$ contributions. These can be generated from $k{=}0$ by inserting $\sfo(\pm1)=\sfo^\pm\sfo^\pm$ into $\VEV{{\rm NS}\,{\rm NS}\,{\rm NS}}$, resulting in $$\VEV{{\rm NS}\;{\rm R}^+\,{\rm R}^+}\ +\ \VEV{{\rm NS}\;{\rm R}^-\,{\rm R}^-}
\quad.
\eqno(48)$$ Noting that $V_{\rm R}^-=f(p)V_{\rm R}^+$ with a momentum-dependent factor $f(p)$ [@bkl], we relate eq. (48) to the $k{=}0$ case and find that the two terms in eq. (48) cancel each other on-shell. This leaves us with the standard $U(1,1)$-invariant result for the tree-level three-point function and removes the $U(1,1)$ non-invariant $k{\neq}0$ terms (see also ref. [@lp]). It would be interesting to check whether such a mechanism ensures global $U(1,1)$ symmetry in general.
Finally, it should be noticed that the discussion of spectral flow heavily relied on the use of the [*holomorphic*]{} basis for bosonization. In contrast, in the [*real*]{} basis (discussed also at length in refs. [@bkl; @lp]) the spectral flow is obscured and the operator defined by eq. (43) is non-local.
With pleasure we thank the participants of the conference “Strings’95” at the University of Southern California in March 13–18 and, in particular, Nathan Berkovits, Hong Lü, Hiroshi Ooguri and Chris Pope, for fruitful discussions. Our special gratitude to the crew of the British Airways flight \# 268, from Los Angeles to London, where this work was completed. .2in
[99]{}
N. Berkovits and C. Vafa, . N. Berkovits, [*Vanishing Theorems for the Self-Dual N=2 String*]{},\
San Paulo preprint IFUSP–P–1134, December 1994, hep-th/9412179. H. Ooguri, [*Loop Amplitudes of N=2 Strings*]{}, Talk given at the Intern. Conference “Strings ’95”, Univ. of Southern California, Los Angeles, 13–18 March 1995,\
to appear in the Proceedings (World Scientific). S. Ketov, O. Lechtenfeld and A. Parkes, [*Twisting the N=2 String*]{},\
DESY and Hannover preprint, DESY 93–191, ITP–UH–24/93, December 1993,\
hep-th/9312150; to appear in Phys. Rev. [**D**]{} (1995). J. Bischoff, S. Ketov and O. Lechtenfeld, . O. Lechtenfeld, [*From N=2 Fermionic Strings to Superstrings?*]{},\
Talk given at the 28th Intern. Symposium on the Theory of Elementary Particles,\
Wendisch-Rietz, Germany, 30 August – 03 September 1994, hep-th/9412242. E. D’Hoker and D. Phong, Rev. Mod. Phys. [**60**]{} (1988) 1012. F. Ferrari, Helv. Phys. Acta [**67**]{} (1994) 702. D. Friedan, E. Martinec and S. Shenker, ;\
E. Martinec, ;\
E. Verlinde and H. Verlinde, ; . L. Alvarez-Gaumé, G. Moore and C. Vafa, . H. Ooguri and C. Vafa, ; . N. Berkovits and C. Vafa, . N. Berkovits, [*Super-Poincaré Invariant Superstrings*]{},\
San Paulo preprint IFUSP–P–1143, March 1995, hep-th/9503099. H. Lü and C. N. Pope, [*BRST Quantization of the N=2 String*]{},\
Texas and Trieste preprint, CTP–TAMU–62/94, SISSA–175/94/EP,\
November 1994, hep-th/9411101.
[^1]: =6.5in [\#1]{}
[^2]: Supported in part by the ‘Deutsche Forschungsgemeinschaft’
[^3]: On leave of absence from: High Current Electronics Institute of the Russian Academy of Sciences, ${~~~~~}$ Siberian Branch, Akademichesky 4, Tomsk 634055, Russia
[^4]: The fields $X^\m$ and $\j^\m$ are complex-valued, and $\m=0,1$ is a vector index w.r.t. the spacetime ‘Lorentz group’ $U(1,1)$.
[^5]: The superscript $\pm$ of a field denotes its $U(1)$ charge $\pm 1$, and $\a,\b=0,1$ are 2d world-sheet indices.
[^6]: Note that there is only a single gauge field for both $U(1)$ invariances.
[^7]: After Wick-rotating the world-sheet to the Euclidean regime, it is better to think of $c$ as a $U(1)$ monopole charge.
[^8]: We use the notation $\d=*d*$, where the star means Hodge conjugation, so that $\D=\d d+d\d$ and $d^2=0=\d^2$. The connection $A^{\rm Teich}$ is both closed and co-closed, $dA^{\rm Teich}=0=\d A^{\rm Teich}$.
[^9]: Globally, the condition $c(A^0)=0$ forces the sum of all magnetic charges to vanish.
[^10]: The prime form $E(z,w)$ is a holomorphic $(-\fracmm12,0)$ form in $z$ and $w$, with a single zero at $z=w$.
[^11]: Our notation is as follows (see refs. [@klp; @bkl] for more details): $(b,c)$ stand for conformal ghosts, $(\b^{\pm},\g^{\pm})$ for $N{=}2$ superconformal ghosts, and $(\tilde{b},\tilde{c})$ for $U(1)$ ghosts. The ghosts for the Weyl and super-Weyl symmetries are ignored since they do not propagate. $S_{\rm gh}$ is the ghost action.
[^12]: Like in the $N{=}1$ string, this raises issues of globality and boundary terms in moduli space. We have no comment here.
[^13]: This allows only NS states; R states will be generated below.
[^14]: Neither $Z^\pm$ nor $Z^0$ have a local inverse [@bkl], while $Z^0$ is nilpotent.
[^15]: Normal ordering is implicit in our formulae. $\sfo(\q)$ generalizes the instanton number-changing operators $\sfo(\pm1)$ of refs. [@bv2; @b2].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'BellQMA protocols are a subclass of multi-prover quantum Merlin-Arthur protocols in which the verifier is restricted to perform nonadaptive, unentangled measurements on the quantum states received from each Merlin. In this paper, we prove that $m$-clause 3-SAT instances have BellQMA proofs of satisfiability with constant soundness gap, in which $\tilde{O}(\sqrt{m})$ Merlins each send $O(\log m)$ qubits to Arthur. Our result answers a question of Aaronson et al., who gave a protocol with similar parameters that used entangled measurements; the analysis of our protocol is significantly simpler than that of Aaronson et al. Our result also complements recent work of Brandao, Christandl, and Yard, who showed upper bounds on the power of multi-prover quantum proofs with unentangled but adaptive (LOCC) measurements.'
author:
- 'Jing Chen [^1]'
- 'Andrew Drucker [^2]'
title: |
Short Multi-Prover Quantum Proofs for SAT\
without Entangled Measurements
---
Introduction
============
In quantum Merlin-Arthur (QMA) proof systems, a computationally unbounded but untrusted prover Merlin tries to convince a polynomial-time quantum verifier Arthur that a given statement is true, by sending to Arthur a quantum state as a “proof”. We desire that the protocol have two properties. The first is “completeness”: if the statement is true, then there should exist a proof which makes Arthur accept with at least some high probability $c$. The second is “soundness": if the statement is false, then for any proof received, Arthur should accept with at most some lower probability $s < c$. In general, the complexity class ${\mathsf{QMA}}_{\ell, c, s}$ consists of all languages whose membership can be proved by a quantum Merlin-Arthur proof system using $\ell$-qubit proofs, with completeness $c$ and soundness $s$. The complexity class ${\mathsf{QMA}}$ is defined to be ${\mathsf{QMA}}_{\mbox{\scriptsize{poly}}(n), 2/3, 1/3}$ where $n$ is the input length.
The generalized multi-prover version of ${\mathsf{QMA}}_{\ell, c, s}$, denoted ${\mathsf{QMA}}(k)_{\ell, c, s}$, was introduced by Kobayashi, Matsumoto, and Yamakami in [@KMY'03]. In such a proof system, $k$ Merlins are trying to convince a single Arthur that a given statement is true, by each sending Arthur a quantum state with $\ell$ qubits, and these $k$ states are assumed to be unentangled with each other. The class ${\mathsf{QMA}}(k)$ is defined to be ${\mathsf{QMA}}(k)_{\mbox{\scriptsize{poly}}(n), 2/3, 1/3}$.
One piece of evidence for the power of multiple quantum provers was given by Blier and Tapp [@BT'09], who showed that every language in ${\mathsf{NP}}$ has a 2-prover proof system with extremely short proofs, of $\ell = O(\log n)$ qubits each. Unfortunately, the soundness gap in their proof system (i.e., the quantity $c - s$) is very small: their protocol has $c = 1, s = 1 - 1/\operatorname{poly}(n)$. A related but incomparable result was shown by Aaronson et al. [@ABDFS'09]: they showed that $m$-clause 3-SAT instances can be proved satisfiable by a proof system using $\tilde{O}(\sqrt{m}))$ Merlins, each sending $O(\log m)$ bits, and with an improved soundness gap $c = 1, s = 1 - \Omega(1)$. This still gives an almost-quadratic improvement in total proof length compared to known classical proofs, at least in the regime where the number $n$ of variables satisfies $n = \Theta(m)$.
A recent paper by Harrow and Montanaro [@HM'10] answers several important questions about ${\mathsf{QMA}}(k)$. They prove that soundness amplification for ${\mathsf{QMA}}(k)$ is possible and that ${\mathsf{QMA}}(k) = {\mathsf{QMA}}(2)$, for any $k = O(\operatorname{poly}(n))$. Building on the result of [@ABDFS'09], Harrow and Montanaro also show that there exists a 2-prover proof system with proof length $\tilde{O}(\sqrt{m})$ for $m$-clause 3-SAT instances.
In all results mentioned above, Arthur uses the so-called *swap test* measurement [@BCWdW'01] as an important step in the protocol. This is an efficient method to test whether two unentangled states are approximately equal. A natural question thus arises: how crucial is the swap test to the power of multi-prover quantum proof systems? The swap test is an example of an *entangled measurement*, in which the states may become entangled by the measurement process; so more generally, how crucial are entangled measurements to these proof systems? To make such questions formal, [@ABDFS'09] defined the complexity classes ${\mathsf{LOCCQMA}}(k)$ and ${\mathsf{BellQMA}}(k)$. The class ${\mathsf{LOCCQMA}}(k)$ consists of all languages whose membership can be proved by a $k$-prover proof system where Arthur is constrained to make unentangled measurements on the states provided by the Merlins, but is allowed to make these measurements adaptively based on the outcome of previous measurements. ${\mathsf{BellQMA}}(k)$ is the subclass of ${\mathsf{LOCCQMA}}(k)$ in which we additionally require that no choice of measurement depends on the outcomes of other measurements. (For more precise definitions of LOCCQMA and BellQMA protocols, see Sec. \[belldefsec\].) Brandao [@B'08] showed that ${\mathsf{BellQMA}}(k) = {\mathsf{QMA}}$ for constant $k$. Quite recently Brandao, Christandl, and Yard [@BCY'10] made a breakthrough in the study of entanglement, and used their techniques to show that ${\mathsf{LOCCQMA}}(k) = {\mathsf{QMA}}$ for constant $k$. The situation for growing values of $k$ remains unclear.[^3]
#### Our contribution.
In this paper, we exhibit a BellQMA proof system for 3-SAT, which essentially matches the parameters of the earlier protocol of [@ABDFS'09]. Formally, we prove the following theorem:
\[mainthm\] There is a BellQMA proof system which, given a 3-SAT instance with $m$ clauses, uses $\tilde{O}(\sqrt{m})$ Merlins, each of which sends $O(\log m)$ qubits. The proof system has completeness $1 - \exp\{-\Omega(\sqrt{m})\}$ and soundness $1 - \Omega(1)$.
Our result shows that entangled measurement is not necessary for short proofs of membership in 3-SAT, and thus answers a question raised in [@ABDFS'09 Sec. 6.3]. Our system (just barely) loses perfect completeness as achieved by the protocol of [@ABDFS'09], but retains the constant soundness gap of that protocol. The analysis of our protocol is also significantly simpler than that of [@ABDFS'09], which may be viewed as another contribution of this work.
Our protocol also complements a negative result from [@BCY'10] (although our work was independent of theirs). Corollary 5 of [@BCY'10] implies that if there exists a 2-prover LOCCQMA protocol for 3-SAT with proof length $o(\sqrt{m})$ and with constant soundness gap, then there exists a deterministic algorithm solving 3-SAT in subexponential time.[^4] This result seems to pose a significant barrier to achieving shorter proof length using unentangled measurements. Our positive result nearly reaches this barrier, except for the fact that we use more than 2 provers. If our protocol could be converted to a 2-prover BellQMA or LOCCQMA protocol with similar proof length, then (under the plausible assumption that 3-SAT requires exponential time) we would obtain a nearly tight understanding of the power of these restricted quantum proof systems for 3-SAT (and of many other ${\mathsf{NP}}$ languages, via standard reductions).
#### Our Techniques.
The construction of our proof system, which we sketch next, adapts techniques used by Blier and Tapp in their proof system for 3-colorability from [@BT'09], and combines them with sampling and PCP ideas similar to those used by Aaronson et al. [@ABDFS'09]. In the Blier-Tapp protocol, Arthur receives two states ${|\Psi^1\rangle}, {|\Psi^2\rangle}$ of form $${|\Psi^i\rangle} = \sum_{v, c} \alpha^i_{v, c} {|v\rangle}{|c\rangle} , \quad{} i = 1, 2,$$ where $v \in \{0, 1, \ldots, n - 1\}$ indexes a vertex in a graph $G$ to be properly 3-colored and $c \in \{0, 1, 2\}$ is a color for $v$. The protocol randomly performs one of three tests on ${|\Psi^1\rangle}, {|\Psi^2\rangle}$:
- An “Equality Test" uses the swap test to check that ${|\Psi^1\rangle}, {|\Psi^2\rangle}$ are nearly equal. This is the only entangled measurement.
- A “Uniformity Test" uses the quantum Fourier transform to check that each state has amplitudes which are almost uniformly spread over the $n$ vertices.
- A “Consistency Test" directly measures the vertex and color registers on each proof, rejecting if it sees two adjacent vertices with the same color or two differently-colored copies of the same vertex.
This proof is extremely succinct—only two states of $\log n$ qubits each—but as mentioned, its soundness gap is only inverse-polynomial in $n$. Intuitively[^5] this is because, if the Merlins send proofs uniformly spread over the vertices, each equipped with a coloring violating only one edge constraint, then the Consistency test can only succeed if the two vertices sampled come from this edge, which happens with probability $2/n^2$.
We modify this protocol as follows. First, we ask for $O(\sqrt{n})$ proofs instead of 2. The “birthday paradox” then ensures that the Consistency Test will turn up pairs of equal vertices. These will cause rejection unless almost all vertices are nearly-unanimous in their colorings across the supplied proofs. With this added assurance, we simply omit the Equality Test. Our modified Uniformity Test ensures that there are enough states in which the amplitudes are almost uniformly spread over the vertices, although to improve the soundness of this test, we are forced to sacrifice perfect completeness.
In the Consistency Test, we now also expect to sample pairs of vertices adjacent in $G$. However, this will only lead to rejection with noticeable probability if the constraint problem is “highly unsatisfiable", in the sense that every coloring violates an $\Omega(1)$ fraction of the edge constraints. To ensure this, we apply the size-efficient PCP reduction of Dinur [@D'07] to our original 3-coloring problem (or 3-SAT instance), which incurs only a polylogarithmic blowup in the instance size. This completes the sketch of our protocol and the basic ideas of the analysis; the full proof of correctness is slightly more involved.
#### Open Problems.
Some questions raise from our result and those mentioned above. The most immediate one is whether the number of provers in our system can be further reduced without expanding the total proof length much. In particular, is there a 2-prover LOCCQMA proof for 3-SAT with length $\tilde{O}(\sqrt{m})$? As we discussed earlier, if 3-SAT requires exponential time, this would be a nearly tight result in terms of proof length. Whether general 2-prover protocols can achieve even shorter proofs of satisfiability remains an interesting question.
It also seems promising to see whether the entanglement theory ideas of [@BCY'10] can be extended to give a fuller understanding of entanglement between more than 2 quantum states. As just one benefit, this could yield new information about the power of ${\mathsf{LOCCQMA}}(k)$ and ${\mathsf{BellQMA}}(k)$ for superconstant $k$.
Preliminaries
=============
We assume familiarity with (uniform) polynomial-time quantum algorithms. Such algorithms are describable by a polynomial-size quantum circuit with polynomially many auxiliary qubits; the circuit is required to be constructible by a classical logarithmic-space algorithm.
BellQMA and LOCCQMA protocols {#belldefsec}
-----------------------------
We now more formally define the restricted multi-prover proof systems called BellQMA and LOCCQMA protocols. The complexity classes ${\mathsf{BellQMA}}(k), {\mathsf{LOCCQMA}}(k)$ are defined in perfect analogy with ${\mathsf{QMA}}(k)$, using these restricted protocols.[^6] In BellQMA protocols, Arthur performs a so-called “Bell test" upon the quantum proofs; in LOCCQMA protocols, Arthur performs a test involving only “local operations and classical communication" (LOCC) upon the proofs. This motivates the terminology.
Our definition of BellQMA protocols is slightly broader than that given in [@B'08], and we discuss the difference below. The more general class of LOCCQMA protocols will not be important in this paper, but we choose to provide a definition since previous discussions presume familiarity with the framework of LOCC tests (see [@Ben'96; @Nie'99]).
In a $k$-prover QMA protocol, the verifier (Arthur) receives a classical input $x \in \{0, 1\}^n$, as well as $k$ “proof" states ${|\Psi_1\rangle}, \ldots, {|\Psi_k\rangle}$ from $k$ provers (Merlins). The $k$ proofs are required to be unentangled. Arthur performs some quantum-polynomial time test on the proofs, after which he either accepts or rejects. We say that a QMA protocol for a language $L \subseteq \{0, 1\}^*$ possesses *completeness* $c$ and *soundness* $s < c$ if:
1. If $x \in L$, some collection ${|\Psi_1\rangle}, \ldots, {|\Psi_k\rangle}$ causes Arthur to accept with probability at least $c$;
2. If $x \notin L$, any collection ${|\Psi_1\rangle}, \ldots, {|\Psi_k\rangle}$ causes Arthur to accept with probability at most $s$.
In a $k$-prover BellQMA protocol, we restrict the form of Arthur’s test as follows: Arthur first performs a polynomial-time quantum computation on $x$ alone. The workspace is then measured fully, yielding a description of measurements $M_1, \ldots, M_k$ described by polynomial-size quantum circuits; the $i$-th measurement, which may output more than one bit, is required to act locally on the $i$-th proof. The measurements are then performed, and we let $y_i$ denote the output of the $i$-th measurement. Finally, Arthur performs a quantum polynomial-time computation on $(x, y_1, \ldots, y_k)$ to decide whether to accept or reject.
Since the measurements $M_1, \ldots, M_k$ are fully determined by the intermediate measurement and act separately on the $k$ unentangled proof states, the proof states remain unentangled after the $M_i$ are performed. A second observation about BellQMA protocols is that the identities of the measurements $M_1, \ldots, M_k$ can be random variables, and these random variables need not be independent.[^7]
In [@ABDFS'09], the definition of BellQMA protocols was informal and slightly ambiguous. In Brandao’s thesis [@B'08], the definition of BellQMA protocols required the final computation on $(x, y_1, \ldots, y_k)$ to be performed by a *classical* polynomial-time algorithm. We feel that, since Arthur is allowed to use arbitrary polynomial-time quantum measurements $M_i$ on the $k$ proofs, it is natural to allow polynomial-time quantum computations in the final stage. Indeed, Brandao’s proof in [@B'08] that ${\mathsf{BellQMA}}(k) = {\mathsf{BellQMA}}$ works equally well if this final computation is allowed to be quantum. The BellQMA protocol that we give in this paper actually obeys Brandao’s more restrictive definition.
In LOCCQMA protocols, Arthur is allowed to repeatedly and adaptively choose measurements to perform on the proofs. However, these measurements are required to act locally on a single proof state, and they must be performed when Arthur’s workspace is in a computational basis state. This forces the proofs to remain unentangled throughout the computation.
Formally, $k$-prover LOCCQMA protocols can be defined as follows. Arthur’s verification algorithm consists of a polynomial number $p(n)$ of stages. Each stage $t \leq p(n)$ has the following form:
1. Arthur first performs a polynomial-time quantum computation acting on his workspace qubits alone. Arthur’s full workspace is then measured, yielding a tuple $(i_t, M_t, z_t)$. Here $M_t$ describes a polynomial-time quantum measurement to be performed locally on the $i_t$-th proof, and $z_t$ is an auxiliary memory string.
2. $M_t$ is then performed, yielding an outcome $y_t$ of one or more bits. Arthur then begins the $(t+1)$-st stage with his workspace initialized to the computational basis state ${|y_t, z_t\rangle}$.
Finally, Arthur accepts or rejects based upon the first bit of $y_{p(n)}$. We remark that Arthur is allowed to measure individual proof states more than once.
Note that BellQMA protocols can be defined as LOCCQMA protocols in which all measurements to be performed on the $k$ proof states are determined in the first computation phase and described by the string $z_1$, then nonadaptively performed in the following phases.
Dinur’s PCP reduction {#dinursec}
---------------------
The recent version of the PCP Theorem given by Dinur [@D'07] is a reduction from the Boolean Satisfiability problem to a so-called *constraint graph* problem, or 2-CSP. A constraint graph is an undirected graph (possibly with self-loops) along with a set $\Sigma$ of “colors". For each edge $e = (u, v) \in E$ the constraint graph has an associated constraint $R_e: \Sigma \times \Sigma \rightarrow \{0, 1\}$. A coloring $\tau: V\rightarrow \Sigma$ *satisfies* the constraint $R_e$ if $R_e(\tau(u), \tau(v)) = 1$. We say that $G$ is satisfiable if there exists a mapping $\tau$ that satisfies all constraints. We say that $G$ is $(1 - \eta)$-unsatisfiable if for all mappings $\tau: V\rightarrow \Sigma$, the fraction of constraints satisfied by $\tau$ is at most $(1 - \eta)$.
[[@D'07 Thm. 8.1 and its proof]]{} There exists a reduction $T$ from 3-SAT instances to 2-CSP instances, with the following properties:
1. Completeness: If $\varphi$ is a satisfiable formula, $T(\varphi)$ is a satisfiable 2-CSP instance;
2. Soundness: There exists an absolute constant $\eta > 0$ such that if $\varphi$ is unatisfiable, $G = T(\varphi)$ is $(1 - \eta)$-unsatisfiable;
3. Size-Efficiency: If $\varphi$ has $m$ clauses, then $|V(G)| = n = O(m\cdot \operatorname{polylog}m)$ and also $|E(G)| = O(m\cdot \operatorname{polylog}m)$;
4. Alphabet Size: $|\Sigma| = K = O(1)$;
5. Regularity: $G$ is $d$-regular (with self-loops), where $d = O(1)$.
The last point is not quite explicit in the main statement of Dinur’s result, but can be readily extracted from her proof: simply apply the “preprocessing" transformation of [@D'07 Lemma 1.9] to the graph output by her main reduction. Also, Dinur’s main reduction takes as input a constraint graph, not a formula, but we can simply begin by transforming any 3-SAT instance of $m$ clauses into an equivalent instance of an ${\mathsf{NP}}$-hard 2-CSP such as 3-Colorability, yielding a constraint graph whose number of edges is $O(m)$.
In our protocol, Arthur first performs the above reduction, yielding a 2-CSP $G$ on $n = \tilde{O}(m)$ vertices that is either satisfiable or $(1 - \eta)$-unsatisfiable. We now describe our BellQMA protocol for the problem, starting directly from the constraint graph $G$.
The BellQMA protocol
====================
Given a constraint graph $G$, let the proof states Arthur receives be denoted ${|\Psi_1\rangle},\dots,{|\Psi_{C\sqrt{n}}\rangle}$, with $C$ a constant (to be determined later). Each ${|\Psi_i\rangle}$ consists of a “vertex” register with base states ${|0\rangle},\dots,{|n-1\rangle}$ (describable by $\lceil \log n \rceil $ qubits) and a “color” register with base states ${|0\rangle},\dots,{|K-1\rangle}$ (describable with $\lceil \log K \rceil = O(1)$ qubits). Let $\mu \triangleq C\sqrt{n}/K$. The verifier’s protocol is given below.
[**Verifier ${{\cal V}}$:**]{}
- Flip a fair coin. If Heads, do the Uniformity Test; if Tails, do the Consistency Test.
- **Uniformity Test:**
- For each ${|\Psi_i\rangle}$, perform a Fourier transform $F_K$ on the color register and measure it.\
Let $Z=\{i: \mbox{the color register of } {|\Psi_i\rangle}\ \mbox{is measured 0}\}$. If $|Z|< \frac{99\mu}{100}$, *reject*; otherwise continue.
- For each ${|\Psi_i\rangle}$ such that the measurement in Step 1 gets 0, perform a Fourier transform $F_n$ on the vertex register and measure it. If there exits a ${|\Psi_i\rangle}$ such that the measurement doesn’t get 0, *reject*; otherwise *accept*.
- **Consistency Test:**
- For each ${|\Psi_i\rangle}$, measure it and denote the value in the two registers as $(v_i, c_i)$.
- If there exists two states ${|\Psi_i\rangle}$ and ${|\Psi_j\rangle}$ such that $e=(v_i,v_j)\in E$ but $R_e(c_i,c_j)=0$, *reject*. Also *reject* if $v_i = v_j$ but $c_i \neq c_j$. Otherwise, *accept*.
Note that, since $n = \tilde{O}(m)$, we have $ \tilde{O}(m)$ proofs, each consisting of $\log n + O(1) = O(\log m)$ qubits, as needed. The verifier is clearly polynomial-time and performs only nonadaptive, unentangled measurements, so it defines a valid BellQMA protocol.
Completeness of our protocol
----------------------------
In the rest of the paper, we use $\hat{i}$ to denote the square root of $-1$, and reserve the symbol $i$ as an index of states sent by the provers. We first consider the case where the 3-SAT instance $\varphi$ is satisfiable, so that the constraint graph $G$ is also satisfiable.
If $G$ is satisfiable, then there exist (unentangled) states ${|\Psi_1\rangle},\dots,{|\Psi_{C\sqrt{n}}\rangle}$ such that ${{\cal V}}$ accepts with probability at least $1-\exp\left(-\mu/(2\cdot 10^4)\right) = 1 - \exp\left(-\Omega(\sqrt{m})\right)$.
Let ${|\Psi_i\rangle} = {|\Psi\rangle}\triangleq \frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}{|v\rangle}{|\tau(v)\rangle}$ for all $i\leq C\sqrt{n}$, where $\tau$ is a coloring satisfying the constraint graph $G$. Since $\tau$ is satisfying, the Consistency Test will accept with probability 1. Below we analyze the probability that the Uniformity Test will accept if that test is performed.
Observe that a Fourier transform on the color register changes ${|\Psi\rangle}$ into $$\label{eq1}
(I_n\otimes F_K)\frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}{|v\rangle}{|\tau(v)\rangle} = \frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}{|v\rangle}\frac{1}{\sqrt{K}}\sum_{k=0}^{K-1}\exp \left( \frac{2\pi \hat{i} \tau(v) k}{K} \right){|k\rangle}.$$ Therefore for each ${|\Psi_i\rangle}$, the measurement in Step 1 of the Uniformity Test will see 0 with probability $n(1/\sqrt{n})^2(1/\sqrt{K})^2 = 1/K$. Accordingly, ${\mathbb{E}}[|Z|] = C\sqrt{n}/K = \mu$. Since the ${|\Psi_i\rangle}$’s are unentangled, their measurement outcomes are independent. By a Chernoff bound, the probability that the Uniformity Test passes Step 1 is $$\begin{aligned}
1 - \Pr\left[|Z|< \frac{99\mu}{100} \right] > 1-\exp\left(-\frac{\mu}{2\cdot 10^4}\right).\end{aligned}$$ Further notice that according to Eq. \[eq1\], conditioned on the color register measuring to 0 in Step 1 of the Uniformity Test, the state in the vertex register of ${|\Psi\rangle}$ becomes $\frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}{|v\rangle}$, and a Fourier transform $F_n$ will change this state into $$F_n \frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}{|v\rangle} = \frac{1}{\sqrt{n}}\sum_{v=0}^{n-1}\frac{1}{\sqrt{n}}\sum_{u=0}^{n-1}\exp \left(\frac{2\pi \hat{i} vu}{n} \right){|u\rangle} = {|0\rangle}.$$ Thus for each ${|\Psi_i\rangle}$ which is measured 0 in Step 1 of the Uniformity Test, Step 2 of this test will measure 0 with probability 1. Accordingly, if the Uniformity Test passes Step 1, it will accept in Step 2 with probability 1.
Thus the probability that ${{\cal V}}$ accepts is at least $$\frac{1}{2}\cdot \left(1-2\exp\left(-\frac{\mu}{2\cdot 10^4}\right) \right)\cdot 1 + \frac{1}{2}\cdot 1 = 1-\exp\left(-\frac{\mu}{2\cdot 10^4}\right).$$
Soundness of our protocol
=========================
Now we consider the case where the 3-SAT instance $\varphi$ is unsatisfiable, so that the constraint graph $G$ is $(1 - \eta)$-unsatisfiable. We show that for any sequence of proof states ${|\Psi_1\rangle}, \dots,{|\Psi_{C\sqrt{n}}\rangle}$, ${{\cal V}}$ will reject with probability $\Omega(1)$. The proof depends on three lemmas, corresponding to three cases that cover all possible sequences of states sent by the Merlins.
First, we can assume without loss of generality that the states Arthur receives are pure states, since by convexity some sequence of pure states maximizes Arthur’s acceptance probability. For each $i\in [C\sqrt{n}]$, we can express ${|\Psi_i\rangle}$ as $${|\Psi_i\rangle}=\sum_{v=0}^{n-1} \alpha^i_v {|v\rangle} \sum_{j=0}^{K-1}\beta^i_{v,j}{|j\rangle},$$ where $\sum_{v=0}^{n-1}|\alpha^i_v|^2 = 1$ for each $i$, and $\sum_{j=0}^{K-1}|\beta^i_{v,j}|^2 = 1$ for each $i, v$.
Let $p^i_0$ be the probability that the color register of ${|\Psi_i\rangle}$ is measured 0 after the Fourier transform in Step 1 of the Uniformity Test (conditioned on our performing that test). Let $$Z' \triangleq \left\{i: p^i_0\geq \frac{1}{4K}\right\}.$$ We claim:
\[lem:fourier1\] If $|Z'|\leq \frac{\mu}{2}$, then Step 1 of the Uniformity Test rejects with probability $\Omega(1)$.
Let $Z_1= Z\cap Z'$ and $Z_2= Z\setminus Z'$. We have $|Z_1|\leq |Z'|\leq \mu /2$, and $\Pr[i\in Z_2] < 1/(4K)$ independently for every $i\in [C\sqrt{n}]$ . Let $W$ be a random subset of $[C\sqrt{n}]$ such that $\Pr[i\in W] = 1/(4K)$ independently for every $i$. Then $|W|$ stochastically dominates $|Z_2|$ and we have ${\mathbb{E}}[|W|]= C\sqrt{n}/(4K) = \mu/4$. The probability that Step 1 of the Uniformity Test accepts is $$\begin{aligned}
\Pr\left[|Z|\geq \frac{99\mu}{100}\right]
&= \Pr\left[|Z_1|+|Z_2|\geq \frac{\mu}{2}+\frac{\mu}{4}+\frac{24\mu}{100}\right] \\
&\leq \Pr\left[|Z_2|\geq \frac{\mu}{4}+\frac{24\mu}{100}\right] \\
&\leq \Pr\left[|W|\geq \frac{\mu}{4}+\frac{24\mu}{100}\right] \\
&\leq \exp\left( -\frac{24^2}{25^2\cdot 2}\cdot \frac{\mu}{4}\right) = o(1),\end{aligned}$$ where we used a Chernoff bound.
Let ${\varepsilon}<\eta/20$ be a constant (recall that $\eta$ is the soundness constant in Dinur’s PCP reduction), and for each $i\in [C\sqrt{n}]$, let $$R_i \triangleq \{v: v\in V, |\alpha^i_v|^2<1/(8Kn)\}.$$ The next lemma considers the case where one of the sets $R_i$ is noticeably large.
\[lem:fourier2\] Suppose there exists $i\in Z'$ such that $|R_i|\geq {\varepsilon}n$. Then the Uniformity Test rejects with probability $\Omega(1)$.
We focus on any such index $i$. After the Fourier transform on the color register in Step 1 of the Uniformity Test, ${|\Psi_i\rangle}$ becomes the state ${|\Phi_i\rangle}$ defined by $$\begin{aligned}
\label{eq2}
{|\Phi_i\rangle}& \triangleq & (I_n\otimes F_K)\sum_{v=0}^{n-1} \alpha^i_v {|v\rangle} \sum_{j=0}^{K-1}\beta^i_{v,j}{|j\rangle}
=\sum_{v=0}^{n-1}\alpha^i_v{|v\rangle}\sum_{j=0}^{K-1}\beta^i_{v,j}\frac{1}{\sqrt{K}}\sum_{k=0}^{K-1}\exp\left( \frac{2\pi \hat{i} jk}{K}\right){|k\rangle}\nonumber \\
&=& \frac{1}{\sqrt{K}}\sum_{k=0}^{K-1}\left(\sum_{v=0}^{n-1}\alpha^i_v\left(\sum_{j=0}^{K-1}\beta^i_{v,j}\exp\left( \frac{2\pi\hat{i}jk}{K}\right)\right){|v\rangle}\right){|k\rangle}.\end{aligned}$$ Let ${| \gamma \rangle} = \sum_{v=0}^{n-1}\gamma^i_v{|v\rangle}$ with $\sum_{v=0}^{n-1}|\gamma^i_v|^2 = 1$ be the state left in the vertex register of ${|\Phi_i\rangle}$, after conditioning on the color register of ${|\Phi_i\rangle}$ measuring to 0, which occurs with probability $p^i_0$ by definition. For each $v\in \{0, 1, \ldots, n-1\}$, let $P^i_{0, v}$ be the probability that the color register of ${|\Phi_i\rangle}$ is measured $0$ [*and*]{} the vertex register of ${|\Phi_i\rangle}$ is measured $v$. We have that $$P^i_{0, v} = p^i_0 \cdot |\gamma^i_v|^2.$$ On the other hand, by Eq. \[eq2\] we have $$P^i_{0, v} = \left|\frac{\alpha^i_v}{\sqrt{K}} \sum_{j=0}^{K-1}\beta^i_{v,j}\right|^2 = \frac{|\alpha^i_v|^2}{K}\left|\sum_{j=0}^{K-1} \beta^i_{v,j}\right|^2 \leq \frac{|\alpha^i_v|^2}{K} \cdot K\sum_{j=0}^{K-1}|\beta^i_{v,j}|^2 = |\alpha^i_v|^2,$$ where we used the Cauchy-Schwarz inequality and the fact that $\sum_{j=0}^{K-1}|\beta^i_{v,j}|^2=1$. Combining the above two equations, we find $p^i_0 \cdot |\gamma^i_v|^2 \leq |\alpha^i_v|^2$. Because $i\in Z'$, we have $p^i_0 \geq 1/(4K)$. Thus $|\gamma^i_v|^2 \leq 4K|\alpha^i_v|^2$ for each $v$. Accordingly, for each $v\in R_i$, $$|\gamma^i_v|^2\leq \frac{4K}{8Kn} = \frac{1}{2n}.$$ Define $${|\psi\rangle} \triangleq F_n^{-1}{|0\rangle} = \frac{1}{\sqrt{n}}\sum_{v = 0}^{n - 1}{|v\rangle}.$$ For each $v \in R_i$, $| \gamma^i_v - 1/\sqrt{n} |^2 \geq |1/\sqrt{2n}- 1/\sqrt{n}|^2 = (1 - 1/\sqrt{2})^2/n$. Then, using unitarity of $F_n$, we have $$|| F_n {|\gamma\rangle} - {|0\rangle} ||_2^2 = || {|\gamma\rangle} - {|\psi\rangle} ||_2^2 \geq \sum_{v \in R_i} (1 - 1/\sqrt{2})^2 /n \geq {\varepsilon}n \cdot (1 - 1/\sqrt{2})^2 / n = \Omega(1).$$ Since $||F_n {|\gamma\rangle}||_2 = 1$, it follows that the amplitude of ${|0\rangle}$ in $F_n {|\gamma\rangle}$ is of norm $\leq 1 - \Omega(1)$. Thus if the color register of ${|\Phi^i\rangle}$ measures 0 in the Uniformity Test (as happens with probability $p^i_0 \geq 1/(4K)$ since $i \in Z'$), the vertex register measures to some $v \neq 0$ with probability $\Omega(1)$. The Uniformity Test’s rejection probability is therefore $\Omega(1/K) = \Omega(1)$ as claimed.
In light of Lemmas \[lem:fourier1\] and \[lem:fourier2\], we need only to address the case when $|Z'|>\mu/2$ and $|R_i|<{\varepsilon}n$ for all $i\in Z'$. We show that in this case, the Consistency Test rejects with probability $\Omega(1)$.
Consider an arbitrary state index $i \in Z'$. Let $D_{i}$ denote the distribution on vertex/color pairs when ${|\Psi_{i}\rangle}$ is measured by the Consistency Test. We can equivalently generate each $D_i$ as $D_i = g_{i}(U_{i})$, where each $U_{i}$ is a uniform, independent value from $[0, 1]$, and $g_{i}: [0, 1] \rightarrow V(G) \times \Sigma$ is a function such that each preimage $g_{i}^{-1}((v, c))$ is an interval of length equal to $\Pr[D_{i} = (v, c)]$. Then for each $v \notin R_i$, the set $g_{i}^{-1}(v, \star)$ is of measure $|\alpha^i_v|^2\geq 1/(8Kn)$. Select $J_{i, v} \subseteq g_{i}^{-1}(v, \star)$ of measure *exactly* $1/(8Kn)$ for each such $v$, and let $J_i = \bigcup_{v \notin R_i} J_{i, v}$. Observe the following: first, $J_i$ has measure greater than $(1 - {\varepsilon})/(8K)$. Second, conditioned on $U_i \in J_i$, the posterior distribution of the vertex $v_i$ that $g_i$ outputs is now *uniform* over $S_i\triangleq \{0, 1, \ldots, n-1\}\setminus R_i$.
So let us consider the Consistency Test applied to a sequence of states satisfying $|Z'|>\mu/2$ and $|R_i|<{\varepsilon}n$ for all $i\in Z'$. Letting the measurement outcomes be generated as described above, define the random set $$Z'' \triangleq \{i: i\in Z', U_i\in J_i\} .$$ Notice that $Z''$ is itself a random variable determined by the $U_i$’s. Notice also that for each $i\in Z'$, the probability that $i\in Z''$ is at least $(1-\epsilon)/(8K)$, and these events are independent from each other. Therefore we have that $${\mathbb{E}}[|Z''|] \geq \frac{(1-\epsilon)|Z'|}{8K} > \frac{(1 - {\varepsilon})\mu}{16K} = \frac{(1 - {\varepsilon})C\sqrt{n}}{16K^2}.$$ Since $|Z''|$ never exceeds $C\sqrt{n}$, the total number of proof states, we find that with probability $\Omega(1)$, $$\label{eq3}
|Z''| \geq \frac{C\sqrt{n}}{32K^2} .$$ The following lemma tells us that if $C$ is chosen as a suitably large constant, then conditioned on Eq. \[eq3\] holding, the Consistency Test rejects with $\Omega(1)$ probability.
\[lem:fourier3\] Let $(G, \{R_{e}\})$ be an $n$-vertex, $d$-regular constraint graph (possibly with self-loops, and $d > 1$) with alphabet $K$, such that $G$ is $(1 - \eta)$-unsatisfiable. Let $D_1, \ldots D_{m'}$ be independent distributions on $V(G) \times \Sigma$, with $(v_i, c_i)$ denoting the output of $D_i$. Suppose for each $i \leq m'$ there exists an $S_i \subseteq V(G)$ of size at least $(1 - {\varepsilon})n$, such that $v_i$ is uniformly distributed over $S_i$, where ${\varepsilon}< \eta/20$.
Then we can set $m' = O(\sqrt{n})$ large enough so that with probability at least $.99$ there exists $i < j \leq m'$ such that: *either* $e = (v_i, v_j)$ is an edge of $G$ and $R_e(c_i, c_j) = 0$; *or* $v_i = v_j, c_i \neq c_j$. (The constant in the $O()$ notation depends on $d$ and $\eta$, but not $K$.)
To apply Lemma \[lem:fourier3\] to our Consistency Test when $|Z'|>\mu/2$ and $|R_i|<{\varepsilon}n$ for all $i\in Z'$, choose $C = O(1)$ such that $C\sqrt{n}/(32K^2) \geq m'$. Then conditioned on $|Z''|\geq m'$, which occurs with probability $\Omega(1)$, we can select $D_1,\dots, D_{m'}$ from the distributions of ${|\Psi_i\rangle}$’s (when measured in the Consistent Test) such that $i\in Z''$. By definition of $Z''$, these $D_1,\dots, D_{m'}$ satisfy the hypothesis of Lemma \[lem:fourier3\], so the Consistency Test rejects with probability .99. Thus in this case ${{\cal V}}$ also rejects with probability $\Omega(1)$. This completes the proof that ${{\cal V}}$ possesses soundness $1-\Omega(1)$, proving Theorem \[mainthm\].
For $i < j \leq m'$, let $V_{i,j}$ be the indicator random variable for the event that *either* $e = (v_i, v_j) \in E(G)$ and $R_e(c_i, c_j) = 0$, *or* $v_i = v_j$ and $c_i \neq c_j$. Let $V = \sum_{i < j} V_{i, j}$. To prove Lemma \[lem:fourier3\] it is enough to show that Pr$[V = 0] \leq .01$. We show this using the second moment method.
A first observation is that we can generate $D_i$ in the following way: first, randomly select a coloring $\tau_i$ according to some distribution $H_i$; next select $v_i$ uniformly from $S_i$, and set $c_i = \tau_i(v_i)$. To be explicit, each $H_i$ independently chooses colors according to the rule Pr$[\tau_i(v) = c] = $ Pr$[c_i = c| v_i = v]$. It is easily verified that this process yields $D_i$.
Next we lower-bound ${\mathbb{E}}[V] = \sum_{i < j}{\mathbb{E}}[V_{i, j}]$. Fix any pair $i, j$, $1 \leq i < j \leq m'$. Condition on any values of the colorings $\tau_i, \tau_j$; we’ll show that ${\mathbb{E}}[V_{i,j}|\tau_i, \tau_j] \geq {\varepsilon}/ n$. Let $P_{i, j} \subseteq V(G)$ be the subset of vertices $v$ for which $\tau_{i}(v) = \tau_j(v)$. Suppose first that $|P_{i, j}| \leq (1 - 3{\varepsilon}) n$. In this case there are at least ${\varepsilon}n$ vertices contained in $S_i \cap S_j \cap \overline{P_{i, j}}$, and $V_{i, j} = 1$ whenever a vertex in this set is selected as both $v_i$ and $v_j$. Thus in this case ${\mathbb{E}}[V_{i,j}|\tau_i, \tau_j] \geq {\varepsilon}n\cdot |S_i|^{-1} \cdot |S_j|^{-1} \geq {\varepsilon}/n$.
For our second case, suppose $|P_{i, j}| > (1 - 3{\varepsilon}) n$. Consider the induced subgraph $G[S_i \cap S_j \cap P_{i, j}]$, which contains at least $n - 2{\varepsilon}n - 3{\varepsilon}n = (1 - 5{\varepsilon})n$ vertices. Since $G$ has maximum degree $d$, $|E(G)| = dn/2$, and the set $\overline{S_i} \cup \overline{S_j} \cup \overline{P_{i, j}}$ is incident on at most $d(5{\varepsilon}n)$ edges, we have that $|E(G[S_i \cap S_j \cap P_{i, j}])| \geq dn/2 - 5d {\varepsilon}n = (1 - 10{\varepsilon})(dn)/2 = (1 - 10{\varepsilon})|E(G)|$. By $(1 - \eta)$-unsatisfiability of $(G, \{R_e\})$, the coloring $\tau_i$ satisfies at most a $(1 - \eta)/(1 - 10{\varepsilon})$ fraction of the edge constraints in $G[S_i \cap S_j \cap P_{i, j}]$. Thus the fraction of these constraints which are *violated* by $\tau_i$ is at least $1 - \frac{1 - \eta}{1 - 10{\varepsilon}} = \frac{1 - 10{\varepsilon}-( 1- \eta )}{1 - 10{\varepsilon}} > \frac{\eta}{2(1 - 10{\varepsilon})} > \frac{\eta}{2}$, since $\eta > 20{\varepsilon}$.
Now $\tau_j \equiv \tau_i$ on $P_{i, j}$. We can thus lower-bound ${\mathbb{E}}[V_{i, j}|\tau_i,\tau_j]$ by the probability that $v_i, v_j \in S_i \cap S_j \cap P_{i, j}$ and that $v_i, v_j$ form (in either order) an edge violated by the color assignment $(\tau_i(v_i), \tau_j(v_j))$. Note that some edges are self-loops and so may only be chosen in one way. We get $${\mathbb{E}}[V_{i, j}|\tau_i, \tau_j] \geq (1 - 5{\varepsilon})^2\cdot \frac{ \frac{\eta}{2}|E(G[S_i \cap S_j \cap P_{i, j}])|}{n^2}$$ $$\geq (1 - 5{\varepsilon})^2\cdot \frac{ \eta (1 - 10{\varepsilon})|E(G)| }{2n^2} = \frac{\eta(1 - 5{\varepsilon})^2(1 - 10{\varepsilon})d}{4n} .$$ Recall that ${\varepsilon}< \eta/20 < 1/20$, so the quantity above is greater than $\eta\cdot 2^{-3}d/4n > \eta/(20n) > {\varepsilon}/n$, as needed (using $d > 1$). Thus in either of our two cases we conclude ${\mathbb{E}}[V_{i, j}|\tau_i, \tau_j] \geq {\varepsilon}/n$, so ${\mathbb{E}}[V_{i, j}] \geq {\varepsilon}/n$ unconditioned as well. Summing over all $i < j$, we find ${\mathbb{E}}[V] \geq {\varepsilon}{m' \choose 2}/n = \Omega\left((m')^2/n \right)$.
Next we upper-bound ${\mathbb{E}}[V^2] = \sum_{i < j, k < l} {\mathbb{E}}[V_{i, j}V_{k, l}]$. There are ${m' \choose 2}$ terms for which $(i, j) = (k, l)$. For each such term ${\mathbb{E}}[V_{i,j}^2] = {\mathbb{E}}[V_{i, j}]$. Condition on the vertex $v_i$ outputted by $D_i$. Fixing any such choice of $v_i$, the probability that $V_{i, j} = 1$ is of course upper-bounded by the probability that $v_j$ is equal or adjacent to $v_i$ in $G$. Since $v_j$ is uniform on $S_j$ and $v_i$ is of degree $d$, this probability is at most $(d+1)/|S_i| \leq (d + 1)/((1 - {\varepsilon})n)$, so ${\mathbb{E}}[ V_{i, j} | v_i] \leq (d+1)/((1 - {\varepsilon})n)$. As $v_i$ was an arbitrary conditioning, we conclude ${\mathbb{E}}[ V_{i, j}] \leq (d + 1)/((1 - {\varepsilon})n)$. Thus the contribution to ${\mathbb{E}}[V^2]$ from terms where $(i, j) = (k, l)$ is at most ${m' \choose 2}(d + 1)/((1 - {\varepsilon})n) = O\left((m')^2/n\right)$.
If $(i, j), (k, l)$ consists of three distinct indices, assume that $j = l$, the other cases being handled similarly. Condition on any choice of $v_j$. Then $V_{i,j}V_{j, k} = 1$ can only occur if $v_j$ and $v_k$ are each either incident on or equal to $v_i$. These two events are independent after conditioning on $v_j$ since $D_i, D_j, D_k$ are independent. Thus ${\mathbb{E}}[V_{i, j}V_{k, j}] \leq \left[(d + 1)/((1 - {\varepsilon})n)\right]^2$.
For any three distinct indices $a < b < c \leq m'$, there are six tuples $(i < j), (k < l)$ for which $\{i, j, k, l\} = \{a, b, c\}$. Thus the contribution to ${\mathbb{E}}[V^2]$ from these “triplet” terms is at most $6{m' \choose 3} \cdot \left[(d + 1)/((1 - {\varepsilon})n)\right]^2 = O\left( (m')^3/n^2 \right)$.
If $(i, j), (k, l)$ are four distinct elements of $[m']$, then the pair $V_{i, j}, V_{j, k}$ depend on disjoint sets of independent random variables, so that ${\mathbb{E}}[V_{i, j}V_{k, l}] = {\mathbb{E}}[V_{i, j}]{\mathbb{E}}[V_{k, l}]$. Thus the contribution to ${\mathbb{E}}[V^2]$ from these terms is upper-bounded by $\sum_{i < j, k < l} {\mathbb{E}}[V_{i, j}]{\mathbb{E}}[V_{k, l}] = {\mathbb{E}}[V]^2$.
Putting things together, $${\mathbb{E}}[V^2] < O\left( \frac{(m')^2}{n} + \frac{(m')^3}{n^2} \right) + {\mathbb{E}}[V]^2 .$$ With this bound in hand, we apply Chebyshev’s inequality: $$\begin{aligned}
\Pr[V = 0] &\leq \Pr\left[|V - {\mathbb{E}}[V]| \geq {\mathbb{E}}[V] \right]\\ &\leq \frac{{\mathbb{E}}[V^2] - {\mathbb{E}}[V]^2}{{\mathbb{E}}[V]^2} \\
&\leq O \left( \frac{n^2}{(m')^4} \left( \frac{(m')^2}{n} + \frac{(m')^3}{n^2} \right) \right) \\
&= O\left( \frac{n}{(m')^2} + \frac{1}{m'} \right) ,\end{aligned}$$ which is at most $.01$ if we take $m'$ to be a suitably large value in $O(\sqrt{n})$. This proves the lemma.
Acknowledgements
================
We thank Scott Aaronson for helpful comments and suggestions.
[BCWdW01]{}
S. Aaronson, S. Beigi, A. Drucker, B. Fefferman, and P. Shor. The power of unentanglement. , 5(1):1–42, 2009. Earlier version in Complexity’08.
C. Bennett, H. Bernstein, S. Popescu, and B. Schumacher. Concentrating partial entanglement by local operations. , 53(4):2046–2052, Apr 1996.
H. Buhrman, R. Cleve, J. Watrous, and R. de Wolf. Quantum fingerprinting. , 87(16):167902, Sep 2001. arXiv:quant-ph/0102001v1.
F. Brandao, M. Christandl, and J. Yard. Faithful squashed entanglement, 2010. arXiv:1010.1750.
S. Beigi. vs [QMA]{}$_{\log}$(2). , 54(1&2):0141–0151, 2010. arXiv:0810.5109.
F. Brandao. . PhD thesis, Imperial College, London, 2008. arXiv:0810.0026.
H. Blier and A. Tapp. All languages in [NP]{} have very short quantum proofs. In [*Proc. of 3rd International Conference on Quantum, Nano and Micro Technologies*]{}, 2009. arXiv:0709.0738.
I. Dinur. The [PCP]{} theorem by gap amplification. , 54(3):12, 2007. Earlier version in STOC’06.
A. Harrow and A. Montanaro. An efficient test for product states, with applications to quantum [M]{}erlin-[A]{}rthur games, 2010. arXiv:1001.0017v3. To appear in FOCS’10.
H. Kobayashi, K. Matsumoto, and T. Yamakami. Quantum [M]{}erlin-[A]{}rthur proof systems: Are multiple [M]{}erlins more helpful to [A]{}rthur? In Toshihide Ibaraki, Naoki Katoh, and Hirotaka Ono, editors, [ *ISAAC*]{}, volume 2906 of [*Lecture Notes in Computer Science*]{}, pages 189–198. Springer, 2003.
M. Nielsen. Conditions for a class of entanglement transformations. , 83(2):436–439, Jul 1999.
[^1]: MIT. Email: [email protected].
[^2]: MIT. Email: [email protected]. Supported by a DARPA YFA grant. Supported during part of this work by an Akamai Presidential Graduate Fellowship.
[^3]: On the one hand, the ideas of [@HM'10] rely on the swap test and do not apply to ${\mathsf{LOCCQMA}}$ and ${\mathsf{BellQMA}}$. On the other hand, both the proof for ${\mathsf{BellQMA}}(k)={\mathsf{QMA}}$ and that for ${\mathsf{LOCCQMA}}(k) = {\mathsf{QMA}}$ for constant $k$ blow up the total length of the proofs to $n^{\exp(\Omega(k))}$, and thus cannot be used for $k = \omega(1)$. We can at least say that, if ${\mathsf{QMA}}(2)={\mathsf{QMA}}$, then all classes here collapse to ${\mathsf{QMA}}$.
[^4]: This is not quite made explicit in [@BCY'10], so we elaborate. [@BCY'10 Corollary 5] gives an explicit construction of certain mappings called “approximate disentanglers" (for LOCC measurements). Following the use of such mappings as described in [@ABDFS'09] (where they were defined), a 2-prover LOCCQMA proof system for 3-SAT of proof length $o(\sqrt{m})$ would imply a single-prover QMA protocol for 3-SAT of proof length $\ell = o(m)$. The maximum acceptance probability of such a protocol can be approximately determined using semidefinite-programming solvers in time $\operatorname{poly}(2^\ell) = 2^{o(m)}$.
[^5]: The actual soundness gap shown in [@BT'09] is $\Omega(n^{-6})$, even worse than our sketch would suggest; this was improved to $\Omega(n^{-3 - {\varepsilon}})$ by Beigi in [@Bei'10] by a modified protocol that still uses two $O(\log n)$-sized proofs, but loses perfect completeness.
[^6]: In [@BCY'10], the notations ${\mathsf{QMA}}_{\mathsf{LO}}(k) = {\mathsf{BellQMA}}(k)$ and ${\mathsf{QMA}}_{\mathsf{LOCC}}(k) = {\mathsf{LOCCQMA}}(k)$ are used.
[^7]: In our protocol, the measurements will be chosen in a dependent fashion; however, it is not hard to modify our protocol to make these choices independent (for a fixed input $x$), with only a constant-factor increase in the number of provers.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the problem of warehouse multi-robot automation system in discrete-time and discrete-space configuration with focus on the task allocation and conflict-free path planning. We present a system design where a centralized server handles the task allocation and each robot performs local path planning distributively. A genetic-based task allocation algorithm is firstly presented, with modification to enable heuristic learning. A semi-complete potential field based local path planning algorithm is then proposed, named the recursive excitation/relaxation artificial potential field (RERAPF). A mathematical proof is also presented to show the semi-completeness of the RERAPF algorithm. The main contribution of this paper is the modification of conventional artificial potential field (APF) to be semi-complete while computationally efficient, resolving the traditional issue of incompleteness. Simulation results are also presented for performance evaluation of the proposed path planning algorithm and the overall system.'
author:
- 'Kam Fai Elvis Tsang, Yuqing Ni, Cheuk Fung Raphael Wong and Ling Shi[^1][^2]'
bibliography:
- 'IEEEabrv.bib'
- 'bibliography.bib'
title: '**A Novel Warehouse Multi-Robot Automation System with Semi-Complete and Computationally Efficient Path Planning and Adaptive Genetic Task Allocation Algorithms** '
---
Introduction
============
In the recent decades, the rapid advancement of multi-robot systems has been attractive to both academia and industries because of the wide range of potential applications. Some of the promising applications include transportation, industrial plant inspection [@mrta_mrpp_astar_ga] and logistic management [@mrta_hospital]. In particular, conflict-free path planning algorithm [@mrpp_mstar] and task allocation policy [@mrta_hospital] are the main challenges in multi-robot automation system such as warehouse management which will be the focus of this paper. Unfortunately, both of these problems are in general NP-hard [@mrta_mrpp], thus of great incentives to investigate.
Multi-robot task allocation refers to the process of allocating tasks and the execution orders for each robot in the system. Although it can normally be considered as a combinatorial optimization problem [@mrta_cnp_nn], it is difficult to solve efficiently due to the large search space. Yuan et al. [@mrta_cnp_nn] introduced a neural network based solution for effective bidding in auction at the expense that a large training set is necessary. Liu and Kroll [@mrta_mrpp_astar_ga] implemented genetic algorithm (GA) to generate task allocations but the performance greatly depends on the quality of heuristics for the fitness function. Whilst each of these approaches have their merits, their respective flaws have limited their performances to a large extent.
Path planning algorithms can be categorized into global and local planning [@rpp_svm]. Global planning assumes all information are available and plan the robot path accordingly. Despite the promising solutions contributed by global information, it is computationally expensive thus typically infeasible for large scale system to operate at real time. Local planning, on the other hand, mainly considers the immediate or recent ambience of the robot for online planning with low computation load. A widely used computationally efficient path planning algorithm is the artificial potential field (APF) algorithm [@mrpp_apf], albeit being incomplete due to local minima which has been a challenging problem [@rpp_apf_deadlock], the details of which will be discussed in depth in . Numerous attempts have been made to overcome this issue. For example, Tuazon et al. [@mrpp_apf_fuzzy] proposed the integration of fuzzy logic into the APF algorithm to identify and escape from local minimum. Kovács et al. [@mrpp_apf_animal] implemented the BUG algorithm in additional to APF method as a complement. While these approaches showed effectiveness in simulation, seldom did they prove the completeness of the algorithms.
This paper presents a semi-complete and computationally efficient potential-based local path planning algorithm, named the recursive excitation/relaxation artificial potential field algorithm, hereinafter referred to as the RERAPF algorithm. It has completely eradicated the incompleteness flaw of APF algorithm, and the proof of completeness will be presented in later sections. In addition, we also present a genetic-based task allocation algorithm and an adaptive integrated system with learning ability for the fitness function of the genetic algorithm to improve the overall system performance. We show in simulation the significant improvement in computation speed of RERAPF compared with A\* algorithm, and the overall performance of the integrated system.
The remaining of the paper is organized as follows. Some preliminaries and notations are defined in , then the problem setup is introduced in . The main results, including the proposed algorithms and proof of completeness, are presented in and the simulation results in .
Preliminaries and Notations {#section:prelim}
===========================
In this paper, a warehouse is defined as a grid world $\mathcal{W}$, an example layout of which is shown in , with $N$ homogenous autonomous robots, $M$ static obstacles (including storage shelves and walls), $P$ reachable positions and $K$ tasks. Denote $\mathcal{R} = \{r_1, \dots, r_N \}$ as the set of robots, $\mathcal{O} = \{o_1, \dots, o_M\} $ the set of obstacles, $\mathcal{T} = \{t_1, \dots, t_K \}$ the set of tasks and $\mathcal{S} = \{s_1, \dots s_P\} \subset \mathbb{Z}^2$ the set of all reachable positions. Note that $\mathcal{W} = \mathcal{S} \cup \mathcal{O}$. A position $s$ is represented by a positional vector $[x,y]' \in \mathcal{S}$. Each robot $r_i \in \mathcal{R}$ has a position $s_{r_i}(k)$ at time $k$, a goal position $s_g^i$ and an ordered set of $K_i$ tasks $T_i = \{ t_{i,1}, \dots t_{i,K_i} \} \subset \mathcal{T}$. Also each obstacle $o_i \in \mathcal{O}$ and task $t_i \in \mathcal{T}$ has a position $s_{o_i}$ and $s_{t_i}$ respectively. When $T_i \neq \emptyset$, the goal position for $r_i$ is simply the position of first task, i.e., $s_g^i = s_{t_{i,1}}$.
i\[evaluate=ias using i\] in [0,...,]{} [ in [0,...,]{} [ at (,); ]{} ]{}
in [0,...,]{} [ at (0,); at (5.7,); ]{}
i\[evaluate=ias using i\] in [0,...,]{} [ at (,0); at (,6.3); ]{}
i\[evaluate=ias using i\] in [3,4,7,8,11,12,15,16]{} [ in [3,4,5,6,9,10,11,12,15,16,17,18]{} [ at (,); ]{} ]{}
at (3,5.1); at (0.3,2.7); at (5.4,0.6); at (3,5.1) [0]{}; at (0.3,2.7) [1]{}; at (5.4,0.6) [2]{}; at (2.4,1.5) ; at (1.2,4.8) ; at (4.8,2.7) ; at (3.6,2.1) ;
at (6.25,5); at (6.25,4); at (6.25,4.5); at (6.25, 4.5); at (6.25,4.5) [0]{}; at (6.25,4.0) ; at (6.55,5)[Obstacles]{}; at (6.55,4.5)[Robots]{}; at (6.55,4)[Tasks]{};
Consider the metric space $M_p=(\mathcal{S}, d_p)$ with metric $d_p(s_i, s_j) = \vert \vert s_i - s_j \vert \vert_p$, i.e., the $p$-norm. We define the neighborhood of $s$, denoted by $\mathcal{N}(s)$, as the closed unit ball in $M_1$ centered at $s$, i.e., $\mathcal{N}(s) = \overline{B_1(s)} = \{s' \in \mathcal{S} \colon d_1(s, s') \leq 1 \}$, and the adjacent neighborhood of $s$ as $\mathcal{N}_a(s) = \mathcal{N}(s)\backslash\{s\}$. Also, the adjacent neighborhood of a space $U$ is defined as $\mathcal{N}_a(U) = \{s' \in \mathcal{S} \colon \min_{s \in U} d_1(s, s') = 1 \}$. The set of successors for the robot $r_i$ at time $k$ is $\mathcal{N}(s_{r_i}(k))$.
Problem Setup {#section:problem}
=============
= \[draw, fill=white!20, rectangle, minimum height=3em, minimum width=6em\] = \[draw, fill=white!20, circle, node distance=1cm\] = \[coordinate\] = \[coordinate\] = \[pin edge=[to-,thin,black]{}\]
at (1.5, -0.5) ; ;
(task\_scheduler) [Task Scheduler]{}; (task1) ; (task2) ; (task3) ; (task4) ; (estimator) [Estimator]{}; (sensor) [Position Sensor]{};
(controller) [Controller]{}; (system) [System]{}; (path) [Path Planner]{}; (path1) ; (path2) ;
(sensor2) [Proximity Sensor]{};
(input1) – (task1); (estimator) |- (task2); (sensor) – (estimator); (estimator) -| (controller); (system) -| (sensor); (task3) -| (path2); (path1) |- (task4); (path) – (controller); (controller) – (system); (system) |- (sensor2); (sensor2) -| (path);
(5.7,0) rectangle (8.79,-1.5); (0.1,-1.2) rectangle (2.9,-4.3); (3.15,-3.65) rectangle (11.35,-6.55);
at (2.5,-4.75) [$s_{r_i}$]{}; at (7.5,-2.75) [$\hat{s}_{r_i}$]{}; at (4.5,-1.25) [$\hat{s}_{r_i}$]{};
at (7.2, 0.35) [Central Server]{}; at (7.25, -6.85) [Autonomous Robots]{}; at (0.65, -4.575) [Observer]{};
at (1.25, -0.45) [$\mathcal{T}$]{}; at (10.5, -2.75) [$T_i$]{}; at (9.5, -2.75) [$D_i$]{};
We consider a warehouse multi-robot automation system shown in on a discrete time horizon $k=0,1,\ldots \in \mathbb{N}_0$. The overall system is divided into three subsystems, namely the centralized server, observer and autonomous robots.
The central server handles incoming tasks, and perform task allocation algorithm to allocate tasks $T_i$ from $\mathcal{T}$ to each robot $r_i$ in a centralized manner. The central server communicates with the robots via a wireless channel, which is assumed to have no bit error or packet drop for simplicity.
Each robot is equipped with proximity sensors to retrieve local information at each time $k$, with a sensing range $\mathcal{P}(s_{r_i}(k)) \subset \mathcal{S}$. We further extend the notation to $\mathcal{P}(s)$ representing the sensing range for any robot at $s$. When $r_i$ receives the task assignment from the server, they will distributively perform path planning to reach $s_g^i$ based on the ambient information retrieved by proximity sensors. When the robot $r_i$ has reached its goal, it will feedback the distance travelled $D_i$ to the central server to improve the task allocation performance.
The observer consists of position sensor and estimator for estimation of robot positions $s_{r_i}$, denoted by $\hat{s}_{r_i}$. In this paper, we assume a perfect estimator such that $\hat{s}_{r_i} = s_{r_i}$. This subsystem can be an on-board system of the robot such as odometer, or a separate subsystem such as visual camera.
We aim to design the task scheduler and path planner, which are handled in centralized and distributive manners respectively. To evaluate the system performance, we first define a path cost $J_1$ for the proposed path planning algorithm, to be the ratio of the total travel distance $D_i$ to optimal distance $D_i^*$ (determined by A\* search) of each $r_i$. This quantifies the optimality of the path planning algorithm. $$J_1 = \dfrac{\sum_{i=1}^N D_{i}}{\sum_{i=1}^N D_{i}^*}$$ As for the task allocation policy, the paramount concern is the travel distance for each robot $r_i$ with tasks $T_i$. We define an average distance cost function $J_2$ as the average travel distance per robot per task. This is directly correlated to the energy consumption. In addition, the bottleneck for completion time is the longest distance travelled amongst the robots since the system has to wait for the slowest robot to finish its tasks. We further define a $J_3$ as the bottleneck distance per task, as follows: $$\begin{aligned}
J_2 &=& \dfrac{1}{KN}\sum_{i=1}^N D_i \\
J_3 &=& \dfrac{1}{K}\max_{i \in [1,N]} D_i\end{aligned}$$ Finally, we also hope to evaluate the entire system from the perspective of task completion efficiency. Define $J_4$ as the task completion rate, given in $$J_4 = \dfrac{K}{k_{total}}$$ where $k_{total}$ is the total time needed to finish all $K$ tasks. The design problem is to individually minimize $J_1, J_2, J_3$ and maximize $J_4$ to ensure an effective and efficient warehouse automation system.
Main Results {#section:main}
============
We present our main results, i.e., the proposed genetic multi-robot task allocation algorithm and RERAPF algorithm in this section. The generic GA and conventional APF algorithm are first described in each subsection. The proposed solutions are then built upon each of these algorithms.
Genetic Multi-Robot Task Allocation Algorithm
---------------------------------------------
In order to allocate the tasks effectively and efficiently, a meta-heuristic approach is adopted. In this subsection, genetic algorithm is used along with a proposed learning rule to adapt for the warehouse and multi-robot setup. It should be noted that other meta-heuristic algorithm could also be used with the same learning rule.
### Generic Genetic Algorithm
In the generic GA, a chromosome is used to represent the potential solution, which, in this case, is the task allocation policy. A fitness function $F$ is defined to evaluate the quality of the chromosomes where higher fitness means better quality. An outline of the standard and generic GA is presented in Algorithm \[algo:ga\] [@Tomioka2007]. A more detailed description of the GA can be found in [@Tsang2018]. The detailed design of chromosome representation, fitness function, crossover and mutation operations are described in the following subsections.
[latex@errorgobble]{}
\[algo:ga\]
$n \leftarrow 0$\
initialize population $P(n)$\
evaluate the fitnesses of $P(n)$\
### Chromosome Representation
The chromosome representation is used to encode tasks allocation policy for the robots. For a set of $N$ robots and $K$ tasks, the length of chromosome is $N+K-1$ and each element of the chromosome is called a genome. In each chromosome, $K$ genomes represents the tasks index, and $N-1$ genomes represent the delimiters of chromosome which are arbitrary negative integers, hereinafter represented by $-1$ to $-(N-1)$. The set of all possible chromosomes $\mathcal{C}$ is all permutations of $\{ 1, 2, \dots, K, -1, -2, \dots, -(N-1) \}$ and each set of tasks are encoded as consecutive genomes, separated by a delimiter. In general, any $C \in \mathcal{C}$ has the form $C = \{c_1, \dots c_{N+K-1} \} = \{T_1, n_1, T_2, n_2, \dots, n_{N-1}, T_{N}\}$, where $n_i$ is a negative integer as delimiter, represents the task allocation policy for robot $r_i$ to $r_N$. For example, the sets of tasks $T_1 = \{t_3, t_5, t_1\}$, $T_2 = \{t_4, t_6\}$, $T_3 = \{t_2, t_7\}$, $T_4 = \emptyset$ can be encoded as $\{3,5,1,-1,4,6,-2,2,7,-3\}$.
### Fitness Function
The fitness function, based on the cost function in , directly reflects the quality of the chromosome, hence task allocation. A higher value of fitness function means higher quality. The fitness function $F \colon \mathcal{C} \rightarrow \mathbb{R}$ is shown in . $$\label{eq:fitness}
F(C) = \bigg( \dfrac{1}{KN}\sum_{i=1}^N \hat{D}_{i}(T_i) + \displaystyle \dfrac{1}{K} \max_i \hat{D}_{i}(T_i) \bigg)^{-1}$$ where $\hat{D}_{i}(T_i)$ is the heuristic of the travelled distance for robot $r_i$ with $T_i$, and is given by $$\hat{D}_{i}(T_i) = \hat{d}\big(s_{r_i}(k_0), s_{t_{i,1}}\big) + \sum_{j=1}^{K_i-1} \hat{d}(s_{t_{i,j}}, s_{t_{i,j+1}})$$ and $\hat{d}(s_i, s_j) = d_1(s_i, s_j)$ is the initial heuristic of distance between $s_i$ and $s_j$. This fitness function is in similar form as $(J_2 + J_3)^{-1}$ as it aims to reduce the average and bottleneck distance cost simultaneously. If $(J_2 + J_3)^{-1}$ is at global maximum, then $J_2, J_3$ are both at their global minimum.
### Crossover and Mutation
The crossover operation is used to reproduce offsprings $C_{c1}$ and $C_{c2}$ from two parents $C_{p1}$ and $C_{p2}$. shows an example of this operation. Two crossover points $i, j$ are randomly chosen. Then the $i$-th to $j$-th genomes of $C_{p1}$, labeled in blue, are preserved to the child, and the unused genomes $C_{p2}$, labeled in red, will fill the remaining genomes of $C_{c1}$. The second offspring $C_{c2}$ is reproduced with the same procedure with reversed roles of $C_{p1}$ and $C_{p2}$.
---------- --- ---- -- -- -- -- --- ---- --- ----
Parent 1 3 -2 4 -1 7 -3
Parent 2
Child 1
---------- --- ---- -- -- -- -- --- ---- --- ----
\
The mutation operation is relatively simpler. Two random points $m,n$ are first chosen. Then one randomly permute the $m$-th to $n$-th genomes. This is also called the scramble mutation.
### Learning Rule
After the robots have finished the paths from any $s_i$ to $s_j$, the actual travelled distance $D_i(s_i, s_j)$ will be used to update the heuristic $\hat{d}(s_i, s_j)$. To generalize the learning rule, let $\mathbf{\hat{D}} = [\hat{d}_{ij}]$ where $\hat{d}_{ij} = \hat{d}(s_i, s_j)$ and $\mathbf{A} = [a_{ij}]$ where $a_{ij} = \lambda D_i(s_i, s_j) + (1-\lambda) \hat{d}_{ij}$. The $\lambda$ is a binary indicator of whether or not an update is received. We consider a quadratic error cost function $J$, given by $$J = \dfrac{1}{2} \text{Tr}(\mathbf{A} - \mathbf{\hat{D}})'(\mathbf{A} - \mathbf{\hat{D}})$$ and apply gradient descent to update the heuristics: $$\Delta \hat{d}_{ij} = - \eta \dfrac{\partial J}{\partial \hat{d}_{ij}} = \eta (a_{ij} - \hat{d}_{ij})$$ $$\hat{d}_{ij} \leftarrow \hat{d}_{ij} + \Delta \hat{d}_{ij}$$ where $\eta$ is the learning rate.
Recursive Excitation/Relaxation APF Algorithm
---------------------------------------------
In this section, we will present a novel recursive excitation/relaxation APF algorithm that completely eliminates the well-known problem of local minima with proof, which will be discussed in more details.
### Conventional Artificial Potential Field Algorithm
The conventional APF algorithm is based on the potential function in the form of [@mrpp_apf]. $$\label{eq:conventional_APF}
U_i(k,s) = U_i^{att}(s) + U_i^{rep}(s)$$ where $U_i(k,s)$ is the potential at position $s$ and time $k$ for robot $r_i$, $U_i^{att}(s)$ and $U_i^{rep}(s)$ are the attractive and repulsive potentials. Typically, the attractive potential is contributed by the goal position while the repulsive potential by obstacles and other robots to avoid collision [@mrpp_apf_animal]. Also, the potential from other robots is a function of time as robots are dynamic objects. Equation can therefore be rewritten as . $$\label{eq:conventional_APF2}
U_i(k,s) = U_i^g(s) + U_i^o(s) + U_i^r(k,s)$$ where $U_i^g(s), U_i^o(s), U_i^r(k,s)$ are the potential components from goal, obstacle and other robots. The potential component at a point $s$ under influence of $s'$, denoted by $\phi(s, s')$, regardless of types, is generally expressed as a function of $d_p(s, s')$ [@mrpp_apf_guideline]. $$\label{eq:general_potential}
\phi(s, s') = g\big(d_p(s,s')\big)$$ where $g(x)$ is monotonically increasing for attraction and non-increasing for repulsion. At each $k$, the robot selects a neighbor $s' \in \mathcal{N}(s_{r_i}(k))$ as the next step based on discrete gradient descent [@rpp_apf_num], a modified version is shown in . $$\label{eq:robot_pos_update}
s_{r_i}(k+1) = \operatorname*{arg\,min}_{s' \in \mathcal{N}(s_{r_i}(k))} U_i(k, s')$$
### Proposed Artificial Potential Field Algorithm
While the conventional APF algorithm is highly efficient for path finding in various applications, it has a fatal limitation, the existence of local minima [@mrpp_apf].
\[def:local\_min\] A position $s_m$ is a local minimum for robot $r_i$ at time $k$ iff $\forall s' \in \mathcal{N}_a(s_m)$, $U_i(k, s_m) \leq U_i(k, s') $.
When a robot is at a local minimum $s_m$, it will be trapped indefinitely according to . This is considered a key disadvantage of the APF algorithm [@mrpp_apf_regression]. The proposed RERAPF algorithm is an extension of the conventional APF algorithm designed to overcome the local minima problem. In this algorithm, two operations are introduced, namely the Excitation $f_e$ and Relaxation $f_r$ defined in and . $$\begin{aligned}
\label{eq:excitation}
f_e(x(k)) &=& \gamma x(k-1) \\
\label{eq:relaxation}
f_r(x(k)) &=& (1-\alpha)x(k-1) + \alpha x(k_0)\end{aligned}$$ where $k_0$ is an arbitrary starting time, also $\gamma \in (1,\infty)$ and $\alpha \in [0,1)$ are called the excitation factor and relaxation factor. Rewriting to : $$\label{eq:APF2}
U_i(k,s) = U_i^s(s) + U_i^d(k,s)$$ where $U_i^s(s) = U_i^g(s) + U_i^o(s)$ is the potential contributed by static objects and $U_i^d(k, s) = U_i^r(k, s)$ by dynamic objects, hereinafter referred to as the static potential and dynamic potential respectively. Instead of being a time-independent function, the static potential is modified into a time-recursive function. The potential update is shown in . $$\label{eq:potential_update}
U_i^s(k,s) =
\begin{cases}
f_e(U_i^s(k-1,s)), & s_{r_i}(k) = s \\
f_r(U_i^s(k-1,s)), & otherwise
\end{cases}$$ $\forall s \in \mathcal{N}(s_{r_i}(k))$ with initial condition $U_i^s(k_0,s) = U_i^s(s)$. The purpose of excitation is to increase the potential at the position of the robot $r_i$, i.e., $U_i(k, s_{r_i}(k))$, hence attempting to remove any local minimum in case the robot $r_i$ is trapped. For each position $s$, we only calculate the static potential once until the current task is finished. The explored positions are stored in a set $\mathcal{E}_i$ for each robot $r_i$. We now formally present the RERAPF algorithm in Algorithm \[algo:rapf\].
[latex@errorgobble]{}
\[algo:rapf\]
$k \leftarrow 0$\
A local minima $s_m$ is temporary to $r_i$ if $\forall k_0$ $\exists k' \in (k_0, \infty)$ $\exists s' \in \mathcal{N}_a(s_m)$ s.t. $U_i(k',s') < U_i(k',s_m)$.
\[assum:apf\_pos\] The static potential function satisfies $U_i^s(k,s) \geq 0$ $\forall i,k,s$ with unique global minimum at $s=s_g^i$.
\[assum:finite\_dynamic\_potential\] The dynamic potential function satisfies $U_i^d(k,s) \in [0,\infty)$ $\forall i,k,s$, i.e., always non-negative and finite, and $U_i^d(k,s) \gg U_i^s(s')$ if $\exists r \neq r_i$ $s_r(k) = s$ $\forall i,k,s,s'$.
Before we introduce the main theorem in this paper, we first introduce the following lemma about local minimum:
\[lemma:local\_min\_temp\] Any local minimum reached by robot $r_i$ is temporary to $r_i$ in the proposed RERAPF algorithm.
Under assumption \[assum:apf\_pos\] and \[assum:finite\_dynamic\_potential\], consider an arbitrary local minimum $s_m$ in which robot $r_i$ is trapped at time $k_t$, i.e., $s_{r_i}(k_t)=s_m$. Also assume there exists an unoccupied neighbor for $s_m$. Then by the definition \[def:local\_min\], $\forall s' \in \mathcal{N}_a(s_m)$, $U_i(k_t,s_m) \leq U_i(k_t,s')$ and $s_{r_i}(k_t+1)=s_m$. It is also possible that $s_{r_i}(k_t+1) = s'$ if $U_i(k_t,s')=U_i(k_t,s_m)$ which immediately implies that the robot has escaped from the local minimum $s_m$. Therefore we will focus on the case where $s_{r_i}(k_t+1)=s_m$. $$\begin{aligned}
U_i(k_t+1,s_m) &=& f_e(U_i(k_t,s_m)) + U_i^d(k_t+1,s_m) \quad \\
U_i(k_t+1,s') &=& f_r(U_i(k_t,s')) + U_i^d(k_t+1,s')
\end{aligned}$$ If $\forall s' \in \mathcal{N}_a(s_m)$, $U_i(k_t+1,s_m)\leq U_i(k_t+1,s')$ then the above update rules will be repeated until $\exists k'>k_t, s' \in \mathcal{N}_a(s_m)$, $U_i(k',s_m)>U_i(k',s')$. The dynamic potential $U_i^d(k,s)$ is bounded with $\sup U_i^d(k,s)=m$ and $\inf U_i^d(k,s)=n$ where $m$ and $n$ are two non-negative finite numbers that $m - n \geq 0$. Consider the extreme worst case scenario where $U_i^d(k,s_m)=n$ and $U_i^d(k,s')=m$ $\forall s' \in \mathcal{N}_a(s_m)$. For some $k'>k_t$, $$\begin{aligned}
U_i(k',s_m)&=&\gamma^{k'-k_t}U_i^s(k_t,s_m) + n \\
U_i(k',s')&=&(1-\alpha)^{k'-k_t}U_i^s(k_t,s') \nonumber \\
&&{}+[1-(1-\alpha)^{k'-k_t}]U_i^s(s')+m
\end{aligned}$$ Let $s^* = \operatorname*{arg\,min}_{s' \in \mathcal{N}_a(s_m)} U_i(k',s')$. If $r_i$ is to escape from $s_m$, the inequality must be satisfied. $$\begin{aligned}
\label{eq:k_ineq_unsol}
U_i(k',s_m) &>& U_i(k',s^*) \\
\label{eq:k_ineq_unsol2}
\gamma^{\Delta k}U_i^s(k_t,s_m) &>& (1-\alpha)^{\Delta k} (U_i^s(k_t,s^*)-U_i^s(s^*)) \nonumber \\
&& {}+ U_i^s(s^*) + m - n
\end{aligned}$$ where $\Delta k = k'- k_t$. It can be easily verified that $U_i^s(k_t, s^*) \geq (1-\alpha)^{\Delta k} (U_i^s(k_t,s^*) - U_i^s(s^*)) + U_i^s(s^*)$. Then consider the following inequality, $$\label{eq:k_ineq}
\gamma^{\Delta k'} U_i^s(k_t,s_m) > U_i^s(k_t,s^*) + m - n$$ where $\Delta k' \geq \Delta k$. The inequality has a simple analytical solution for the critical point of $\Delta k'$ where $\Delta k' > (\Delta k')_{crit}$ is the region of solutions. $$\label{eq:k_over}
(\Delta k')_{crit} = \log_\gamma \dfrac{U_i^s(k_t,s^*) + m - n}{U_i^s(k_t,s_m)} \geq (\Delta k)_{crit}$$ Similarly, $\Delta k > (\Delta k)_{crit}$ is the region of solutions for . Note that is an overestimation of the solution of for general case (where the dynamic potentials are not at extreme values), thus an upper bound. In other words, the time required to eliminate local minimum $s_m$ does not exceed $(\Delta k')_{crit} + 1$ regardlessly. Since $\exists \Delta k \leq (\Delta k')_{crit}+1 < \infty$ such that $U_i(k',s_m)>U_i(k',s^*)$, the local minimum $s_m$ is temporary to $r_i$.
Lemma \[lemma:local\_min\_temp\] shows that the introduction of excitation can remove local minimum in case a robot is trapped. From (25) of [@Tsang2018], it is observed that the upper bound of $\Delta k$ decreases with increasing $\gamma$, meaning that a larger excitation factor shortens entrapment time. In addition, the introduction of relaxation factor also facilitates the escape of local minimum as it reduces the potentials of the neighbors to the original values over time.
### Semi-Completeness for Proposed RERAPF Algorithm
The proposed RERAPF algorithm is capable of eliminating local minimum as proven in Lemma \[lemma:local\_min\_temp\], which results in semi-completeness shown in the following.
An algorithm is semi-complete if and only if it is guaranteed to find a solution when there exists one but may not return false when there is none.
\[thm:main\] For every $\alpha \in [0,1)$, $\exists \gamma \in (1,\infty)$ such that the proposed RERAPF algorithm is semi-complete.
Assume robot $r_i$ is contained in a fixed region $V \subset \mathcal{S}$ indefinitely where $s_g^i \not \in V$ and $s_{r_i}(k) \in V$ $\forall k$. Also, we apply an additional constraint on $V$ such that all positions in $V$ must be reached by $r_i$ at some $k$, i.e., $\forall s \in V$ $\exists k$ $s_{r_i}(k)=s$ with finite maximum time intervals. Lemma \[lemma:local\_min\_temp\] implies that the additional constraint is always achievable for small region $V$ and $\vert V \vert > 1$. For each $s \in V$, the corresponding potential can be modeled as a recursive Bernoulli random process with $P[s_{r_i}(k)=s]=p$. Because of the constraint on $V$, $p$ is not the true probability but rather a relative frequency, due to the assumption that all states will be reached by $r_i$ with finite interval, and therefore $p \in (0,1)$. $$\begin{aligned}
\mathbb{E}[U_i^s(k,s)] &=& p\mathbb{E}[f_e(U_i^s(k-1,s))] \nonumber \\
& & +{}(1-p)\mathbb{E}[f_r(U_i^s(k-1,s))] \\
&=& \bigg[\beta^{k-k_0} + (1-p)\alpha \sum_{i=0}^{k-k_0-1} \beta^i\bigg]U_i^s(s)
\end{aligned}$$ where $\beta = p\gamma+(1-p)(1-\alpha)$. It can be verified that $\lim_{k\rightarrow \infty} \mathbb{E}[U_i^s(k,s)] \rightarrow \infty$ when $\beta > 1$, and therefore $\gamma > 1 - \alpha + \alpha/p$. This implies that $U_i^s(k,s)$ will tend to increase, albeit not necessarily monotone, with large $k$. Since $p \in (0,1)$, there must exist a finite $\gamma$ satisfying this inequity. Assume the condition $\gamma > 1 - \alpha + \alpha/p$ is satisfied. If there exists a solution, then $\exists k'' \in (k_0, \infty)$ such that $\exists s'' \in \mathcal{N}_a(V)$ with finite non-increasing potential. For some finite $k^* > k''$ when $s_{r_i}(k^*) \in \mathcal{N}_a(s'') \cap V$, then $s'' \in \mathcal{N}_a(s_{r_i}(k^*))$ with $U_i(k^*, s'') < U_i(k^*, s_{r_i}(k^*))$, which implies that $s_{r_i}(k^*+1) \not \in V$ as at least one neighbor of $s_{r_i}(k^*)$ has lower potential. By contradiction, it is impossible for $r_i$ to be contained in a fixed region $V$ indefinitely. In other words, if $r_i$ is contained in a region $V$ indefinitely, $V$ must be expanding until $s_g^i \in V$. By definition of $V$, the robot $r_i$ will reach $s_g^i$ in finite time. However, if there is no possible path to the goal, it will continue move inside $V$ indefinitely without reaching $s_g^i$, thus only semi-complete.
[0.3292]{} {width="\textwidth"}
[0.3292]{} {width="\textwidth"}
[0.3292]{} {width="\textwidth"}
[0.3292]{} {width="\textwidth"}
[0.3292]{} {width="\textwidth"}
Although a closed-form inequality was derived in , it is impractical to design $\alpha$ and $\gamma$ directly based on the condition because $p$ is unknown. The parameters still need to be chosen by trial and error. However, it is clear that $p$ decreases as $\vert V \vert$ increases. In other words, a larger layout should lead to smaller $p$, hence larger $\gamma$ and/or smaller $\alpha$ required.
### Implementation
The main implementation challenge is the design of potential functions. In the following context, we introduce an implementation based on proximity sensors information. We firstly rewrite the general potential function as the weighted sum of the $p^j$-norms, addition to a constant $\varepsilon^j$, to the power of $e^j$. The superscript $j$ denotes the source of the potential ($g$ for goals, $o$ for obstacles, $r$ for robots), as different object may produce a different potential function. $$\phi^j(s,s') = \sum_{i=0}^\infty c_i^j \Big(d_{p_i^j}(s,s') + \varepsilon_i^j\Big)^{e_i^j}$$ where $p_i^j, c_i^j, e_i^j, \varepsilon_i^j$ are design parameters. More specifically, $\varepsilon_i^r > 0$ $\forall i \in \{ x \colon c_x^r \neq 0 \}$ in order to bound the dynamic potential function to be finite. To utilize the local sensors information, we introduce an alternative form of the potential function as follows. $$\varphi^j(s, s') =
\begin{cases}
\phi^j(s,s'), & s' \in \mathcal{Q}(s) \\
0, & otherwise
\end{cases}$$ where $\mathcal{Q}(s) = \bigcap_{s^* \in \mathcal{N}(s)} \mathcal{P}(s^*)$ to ensure a consistent static potential calculated at different location. This can prevent duplicate computation of potential at the same position, and ensure that the excitation/relaxation operations are correctly performed. The potential function is highly dependent on the sensor information for robot $r_i$ at time $k$. $$\begin{aligned}
U_i^s(s) &=& \phi^g(s, s_g^i) + \sum_{o \in \mathcal{O}} \varphi^o(s, s^o) \\
U_i^d(k, s) &=& \sum_{r \in \mathcal{R}\backslash\{r_i\}} \varphi^r(s, s_r(k))\end{aligned}$$
It is clear that this implementation can satisfy all assumptions made previously on the potential functions with proper $p_i^j, c_i^j, e_i^j, \varepsilon_i^j$. The advantage of this approach is the adaptivity to change of the warehouse environment, for example, change of layout or unexpected obstacles.
Simulation Results {#section:sim}
==================
In this section, we present some simulation results for the proposed warehouse automation system based on the cost functions introduced in . The simulation parameters are listed in , whilst the layout used in the simulation is in the same pattern as . For each set of parameters, 500 runs of simulation were conducted to obtain the corresponding results.
Layout Size $81\times 80$
---------------------------- ---------------
Number of Robots $1 - 100$
Number of Tasks $1 - 100$
Excitation Factor $\gamma$ $15$
Relaxation Factor $\alpha$ $0.05$
: Simulation parameters[]{data-label="table:sim_par"}
For the sake of simplicity, the following potential functions are adopted in the simulation. $$\begin{aligned}
U_i^s(s) &= \phi^g(s, s_g^i) + \sum_{o \in \mathcal{O}} \varphi^o(s, s^o) \\
U_i^d(k, s) &= 0.01\sum_{r \in \mathcal{R}\backslash\{r_i\}} \varphi^o(s, s_r(k)) \\
\phi^g(s, s_g^i) &= d_\infty(s,s_g^i) \\
\varphi^o(s, s^o) &=
\begin{cases}
0.1 (d_2(s, s^o) + \varepsilon)^{-2} , & s^o \in \mathcal{Q}(s) \\
0, & otherwise
\end{cases}
$$ where $\varepsilon = 10^{-9}$ is an arbitrarily chosen small number, and $\mathcal{Q}(s) = \bigcap_{s^* \in \mathcal{N}(s)} \mathcal{P}(s^*)$.
In addition to the cost functions mentioned previously, we also evaluate the computational time of the proposed RERAPF algorithm compared with traditional A\* algorithm. The simulation results are shown in . It is shown that the computation time of A\* algorithm is approximately $13$ times higher than the RERAPF algorithm on average. Also, since the proposed algorithm can be executed distributively on each robot, the actual computational load could be much lower for each robot. As for the path cost $J_1$, it is below $1.2$ within the simulation environment, i.e., the difference between the actual path length and the optimal length is less than $20\%$, with a maximum slope of $0.09\%$ per additional robot. This is the tradeoff between the computation time and optimality for the path planning algorithm.
The results for the cost functions $J_2$, $J_3$ and task completion rate $J_4$ with $N=10,20,40,100$ are shown in to . It is seen that all of $J_2$, $J_3$ and $J_4$ gradually converges as $K$ increases albeit with a slower converging rate for larger $K$, with diminishing marginal improvement with respect to increase in $N$. It is interesting to notice that $J_3$ and $J_4$ can be improved with larger $K$, and $J_2$ was slightly worsen as $K$ increases. The intuition behind this is that when the number of tasks is sufficiently large, it is more likely for the workload of robots distributed more uniformly, hence a lower bottleneck cost $J_3$. Also, multiple tasks can be completed by a single robot at once, hence the tasks done more efficiently, which leads to a lower $J_3$ and higher task completion rate $J_4$. On the other hand, when there are significantly more robots than tasks, the probability that there exist a robot close to the tasks is relatively high, therefore resulting in a lower average travel distance $J_2$.
Conclusion and Future Work
==========================
In this paper, we have presented a computationally efficient local path planning algorithm, namely the RERAPF algorithm, and a genetic task allocation algorithm with heuristic learning rule. The semi-completeness of the RERAPF algorithm was proven by showing that it is impossible for any robot to be trapped in any fixed region that does not contain the goal position. The performances of each subsystem, and the overall system were shown in simulation. Future work may include consideration of additional practical constraints, such as battery and robot capacities. Also, an equivalent RERAPF algorithm of the continuous time or continuous space configuration may also be considered to derive a more general algorithm.
[^1]: Kam Fai Elvis Tsang, Yuqing Ni, Cheuk Fung Raphael Wong and Ling Shi are with the Department of Electronic and Computer Engineering, Hong Kong University of Science and Technology, Hong Kong. Emails: [{kftsang, yniac, cfrwong, eesling}@ust.hk]{}
[^2]: The work is supported by a Hong Kong ITC research fund ITS/066/17FP-A.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This tutorial is intended to give an accessible introduction to Hopf algebras. The mathematical context is that of representation theory, and we also illustrate the structures with examples taken from combinatorics and quantum physics, showing that in this latter case the axioms of Hopf algebra arise naturally. The text contains many exercises, some taken from physics, aimed at expanding and exemplifying the concepts introduced.'
address:
- ' $^a$ LIPN - UMR 7030CNRS - Université Paris 13F-93430 Villetaneuse, France '
- |
$^\ddag$ H. Niewodniczański Institute of Nuclear Physics, Polish Academy of Sciences\
ul. Eliasza-Radzikowskiego 152, PL 31342 Kraków, Poland
- |
$^\diamondsuit$ Laboratoire de Physique Théorique des Liquides,\
Université Pierre et Marie Curie, CNRS UMR 7600\
Tour 24 - 2e ét., 4 pl. Jussieu, F 75252 Paris Cedex 05, France
- |
$^{\sharp}$ The Open University, Physics and Astronomy Department\
Milton Keynes MK7 6AA, United Kingdom
author:
- |
G H E Duchamp$^{\dag}$, P Blasiak$^{\ddag}$, A Horzela$^\ddag$, K A Penson$^{\diamondsuit}$\
and A I Solomon$^{\diamondsuit,\sharp}$
title: 'Hopf Algebras in General and in Combinatorial Physics: a practical introduction'
---
Introduction
============
Quantum Theory seen in action is an interplay of mathematical ideas and physical concepts. From a modern perspective its formalism and structure are founded on the theory of Hilbert spaces [@Isham; @Peres]. Using a few basic postulates, the physical notions of a system and apparatus, as well as transformations and measurements, are described in terms of linear operators. In this way the algebra of operators constitutes a proper mathematical framework within which quantum theories may be constructed. The structure of this algebra is determined by two operations, namely - addition and multiplication of operators; and these lie at the root of all fundamental aspects of Quantum Theory.
The formalism of quantum theory represents the physical concepts of states, observables and their transformations as objects in some Hilbert space $\mathcal{H}$ and subsequently provides a scheme for measurement predictions. Briefly, vectors in the Hilbert space describe states of a system, and linear forms in $\mathcal{V}^*$ represent basic observables. Both concepts combine in the measurement process which provides a probabilistic distribution of results and is given by the Born rule. Physical information about the system is gained by transforming the system and/or apparatus in various ways and performing measurements. Sets of transformations usually possess some structure – such as that of a group, semi-group, Lie algebra, etc. – and in general can be handled within the concept of an algebra $\mathcal{A}$. The action of the algebra on the vector space of states $\mathcal{V}$ and observables $\mathcal{V}^*$ is simply its representation. Hence if an algebra is to describe physical transformations it has to have representations in all physically relevant systems. This requirement directly leads to the Hopf algebra structures in physics.
From the mathematical viewpoint the structure of the theory, modulo details, seems to be clear. Physicists, however, need to have some additional properties and constructions to move freely in this arena. Here we will show how the structure of Hopf algebras enters into the game in the context of representations. The first issue at point is the construction of tensor product of vector spaces which is needed for the description of composite systems. Suppose, we know how some transformations act on individual systems, i.e. we know representations of the algebra in each vector space $\mathcal{V}_1$ and $\mathcal{V}_2$, respectively. Hence natural need arises for a canonical construction of an induced representation of this algebra in $\mathcal{V}_1\otimes\mathcal{V}_2$ which would describe its action on the composite system. Such a scheme exists and is provided by the *co-product* in the algebra, i.e. a morphism $\Delta:\mathcal{A}\longrightarrow\mathcal{A}\otimes\mathcal{A}$. The physical plausibility of this construction requires the equivalence of representations built on $(\mathcal{V}_1\otimes\mathcal{V}_2)\otimes\mathcal{V}_3$ and $\mathcal{V}_1\otimes(\mathcal{V}_2\otimes\mathcal{V}_3)$ – since the composition of three systems can not depend on the order in which it is done. This requirement forces the co-product to be *co-associative*. Another point is connected with the fact that from the physical point of view the vector space $\mathbb{C}$ represents a trivial system having only one property – “being itself” – which can not change. Hence one should have a canonical representation of the algebra on a trivial system, denoted by $\epsilon:\mathcal{A}\longrightarrow\mathbb{C}$. Next, since the composition of any system with a trivial one can not introduce new representations, those on $\mathcal{V}$ and $\mathcal{V}\otimes\mathbb{C}$ should be equivalent. This requirement imposes the condition on $\epsilon$ to be a *co-unit* in the algebra. In this way we motivate the need for a *bi-algebra* structure in physics. The concept of an antipode enters in the context of measurement. Measurement in a system is described in in terms of $\mathcal{V}^*\times\mathcal{V}$ and measurement predictions are given through the canonical pairing $c:\mathcal{V}^*\times\mathcal{V}\longrightarrow \mathbb{C}$. Observables, described in the dual space $\mathcal{V}^*$, can also be transformed and representations of appropriate algebras are given with the help of an anti-morphism $\alpha:\mathcal{A}\longrightarrow\mathcal{A}$. Physics requires that transformation preformed on the system and apparatus simultaneously should not change the measurement results, hence the pairing should trivially transform under the action of the bi-algebra. We thus obtain the condition on $\alpha$ to be an *antipode*, which is the last ingredient of a Hopf Algebra.
Many Hopf algebras are motivated by various theories (physical or close to physics) such as renormalization [@BF1; @BK; @B1; @CK; @K2], non-commutative geometry [@CK; @CM], physical chemistry [@PC; @BF2], computer science [@DT], algebraic combinatorics [@B1; @H1; @MR; @CHNT], algebra [@F1; @F2; @F3].
Operators
=========
Generalities
------------
Throughout this text we will consider (linear) operators $\omega:V\longrightarrow V$, where $V$ is a vector space over $k$ ($k$ is a field of scalars which can be thought of as $\mathbb{R}$ or $\mathbb{C}$). The set of all (linear) operators $V\longrightarrow V$ is an algebra (see appendix) which will be denoted by $End_\mathbb{K}(V)$.
What is a representation
------------------------
It is not rare in Physics that we consider, instead of a single operator, a set or a family of operators $(\omega_\alpha)_{\alpha\in A}$ and often the index set itself has a structure. In the old books one finds the family-like notation, where $\rho(\alpha)$ is denoted, say, $\omega_\alpha$. As a family of operators $(\omega_\alpha)_{\alpha\in A}$ is no more than a mapping $\rho : A\mapsto End_k(V)$ (see [@B_Set], Ch. II 3.4 remark), we wprefer to exhibit the mapping by considering it defined as such. This will be precisely the concept of *representation* that we will illustrate by familiar examples. Moreover using arrows allows, as we will see more clearly below, for extension and factorization procedures.
1. [First case: $A$ is a group]{}
In this case, we postulate that the action of the operators be compatible with the laws of a group; that is, for all $\alpha,\beta\in A$,
$$\begin{aligned}
\label{group_rep1}
\left\{\begin{array}{l}
\rho(\alpha.\beta)=\rho(\alpha)\circ\rho(\beta)\\
\rho(\alpha^{-1})=(\rho(\alpha))^{-1}
\end{array}\right.\end{aligned}$$
which is equivalent to
$$\begin{aligned}
\label{group_rep2}
\left\{\begin{array}{l}
\rho(\alpha.\beta)=\rho(\alpha)\circ\rho(\beta)\\
\rho(1_A)=1_{End(V)}
\end{array}\right.\end{aligned}$$
Note that each of these conditions implies that the range of $\rho$ is in $Aut_k(V)$ ($=GL_n(k)$, the linear group), the set of one-to-one elements of $End_k(V)$ (called automorphisms).
1. [Second case: $A$ is a Lie algebra]{}
In this case, one requires that $$\begin{aligned}
\label{Lie_rep}
\rho([\alpha,\beta])=\rho(\alpha)\circ\rho(\beta)-\rho(\beta)\circ\rho(\alpha)=[\rho(\alpha),\rho(\beta)].\end{aligned}$$
We will see that these two types of action (of a group or Lie algebra) can be unified through the concept of the representation of an algebra (or which amounts to the same thing, of a module).
ßIn the first case, one invokes the *group algebra* $k[A]$ (see appendix). In the case of a Lie algebra, one invokes the *enveloping algebra* $\mathcal{U}(A)$ (or $\mathcal{U}_k(A)$ see appendix). In both cases, the original representation $\rho$ is extended to a representation of an *associative algebra with unit* (AAU) as follows:
$$\begin{aligned}
\nonumber
\ \ \ \ \ \ \xymatrix{
G\ar[r]^{\rho}\ar[d]^{can}&End_k(V)\\
k[G]\ar[ur]_{\hat{\rho}} &\\
}
&&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\xymatrix{
\mathfrak{G}\ar[r]^{\rho}\ar[d]^{can}&End_k(V)\\
\mathcal{U}_k(\mathfrak{G})\ar[ur]_{\hat{\rho}} &\\
}\end{aligned}$$
So far we have not defined what a representation of AAU is. Keeping the philosophy of (or ) and , we can state the following definition:
Let $(\A,+,\cdot)$ be an AAU. A collection of operators $\{\rho(\alpha)\}_{\alpha\in \A}$ in a vector space $V$ is said to be a representation of $\A$ iff the mapping $\rho : \A \mapsto End(V)$ is compatible with the operations and units of $A$. This means that, identically (i. e. for all $\alpha,\beta\in \A$ and $\lambda\in\mathbb{K}$). $$\begin{aligned}
\left\{\begin{array}{l}
\rho(\alpha+\beta)=\rho(\alpha)+\rho(\beta),\ \ \ \ \ \
\rho(\lambda\alpha)=\lambda\rho(\alpha),\\
\rho(\alpha\cdot\beta)=\rho(\alpha)\circ\rho(\beta),\\
\rho(1_\A)=Id_V,
\end{array}\right.\end{aligned}$$ where $\circ$ denotes composition of operators.
\(i) This is equivalent to saying that the arrow $\rho : \A \mapsto End(V)$ from $\A$ to $End(V)$ is a morphism of algebras (with units).
\(ii) In this case, it is sometimes convenient to denote by $\alpha.v$ the action of $\rho(\alpha)$ on $v$ (i.e. the element $\rho(\alpha)[v]$) for $\alpha\in \A$ and $v\in V$.
\(iii) It may happen (and this often occurs) that a representation has relations that are not present in the original algebra. In this case the representation is said to be not *faithful*. More rigourously a representation is said to be faithful iff $\rho$ is injective or, which is equivalent, $ker(\rho)=\rho^{-1}(\{0\})=\{0\}$ (for algebras and Lie algebras) and $ker(\rho)=\rho^{-1}(\{1_V\})=\{1_G\}$ (for groups).
: Let $G=\{1,c,c^2\}$ be the cyclic group of order 3 ($c$ is the cycle $1\rightarrow 2\rightarrow 3\rightarrow 1$), $G$ admits the plane representation by $$\begin{aligned}
\hspace{-2.4cm}\rho(c)=\left(\begin{array}{cc}-1/2&-\sqrt{3}/2\\\sqrt{3}/2&-1/2\end{array}\right)\ \text{(it is the matrix corresponding to a rotation of $2\pi/3$)}.\end{aligned}$$ Thus, $$\begin{aligned}
\rho(c^2)=\left(\begin{array}{cc}-1/2&\sqrt{3}/2\\-\sqrt{3}/2&-1/2\end{array}\right) \ \ \ \ \text{and, of course,}\ \ \ \ \rho(1)=\left(\begin{array}{cc}1&0\\0&1\end{array}\right). \end{aligned}$$ The representation $\rho$ is faithful while its extension to the group algebra is not, as seen from: $$\begin{aligned}
\rho(1+c+c^2)=\left(\begin{array}{cc}0&0\\0&0\end{array}\right) \text{whereas}\ 1+c+c^2\neq0\ \text{in}\ \mathbb{C}[G].\end{aligned}$$
Note that the situation is even worse for a Lie algebra, as $\mathcal{U}_k(\mathfrak{G})$ is infinite dimensional iff $\mathfrak{G}$ is not zero.
Operations on representations
=============================
Now, we would like to see the representations of AAU as building blocks to construct new ones. The elementary operations on vector spaces are:
- sums
- tensor products
- duals
Hence, an important problem is:\
Given representations $\rho_i:\A\mapsto V_i$ on the building blocks $V_i\ ;\ i=1,2$ how does one naturally construct representations on $V_1\oplus V_2$, $V_1\otimes V_2$ and $V_i^*$.
Sums will cause no problem as the sum $V_1\oplus V_2$ of two vector spaces $V_1$ and $V_2$ amounts to taking their cartesian product $V_1\oplus V_2\cong V_1\times V_2$. Then, if $\rho_i : \A\mapsto V_i\ ;\ i=1,2$ are two representations of $\A$ then the mapping $\rho_1\oplus\rho_2:\A\mapsto V_1\otimes V_2$ such that $$\rho_1\oplus\rho_2(a)[(v_1,v_2)]=(\rho_1(a)[v_1],\rho_2(a)[v_2])$$ which can be symbolically written $$\rho_1\oplus\rho_2=
\left(
{\begin{array}{cc}
\rho_1 & 0 \\
0 & \rho_2
\end{array}}
\right),$$ is a representation of $\A$ in $V_1\oplus V_2$.
ßDualization will be discussed later and solved by the existence of an *antipode*. Now, we start with the problem of constructing representations on tensor products. This will be solved by means of the notion of “scheme of actions” which is to be formalized, in our case, by the concept of comultiplication (or coproduct).
Arrows and addition or multiplication formulas
----------------------------------------------
Let us give first some examples where comultiplication naturally arises.\
We begin with functions admitting an “addition formula” or “multiplication formula”. This means functions such that for all $x$, $y$ $$\begin{aligned}
\label{addition_formula}
f(x\ast y)=\sum_{i=1}^n f_i^{(1)}(x)f_i^{(2)}(y),\end{aligned}$$ where $\ast$ is a certain (associative) operation on the defining set of $f$ and $(f_i^{(1)},f_i^{(2)})_{i=1}^n$ be two (finite) families of functions on the same set (see exercises and on representative functions).\
The first examples are taken in the function space[^1] $\R^\R$ (with $*=+$, the ordinary addition of real numbers). The following functions admit “addition formulas” which can be expressed diagramatically as follows.
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Diagram Addition formula
--------------------------------------------------------------------------------------------------------- ---------------------------------------------------
$\xymatrix{ $\cos(x+y)=\cos(x)\cos(y)-\sin(x)\sin(y)$
\ar[r]^{\ \ +}\mathbb{R}\times\mathbb{R}\ar[dr]_{\cos_1\cos_2-\sin_1\sin_2\ }&\ar[d]^{\cos}\mathbb{R}\\
&\mathbb{R}\\
}$
$\xymatrix{ $\sin(x+y)=\sin(x)\cos(y)+\sin(y)\cos(x)$
\ar[r]^{\ \ +}\mathbb{R}\times\mathbb{R}\ar[dr]_{\sin_1\cos_2+\sin_2\cos_1\ }&\ar[d]^{\sin}\mathbb{R}\\
&\mathbb{R}\\
}$
$\xymatrix{ $\exp(x+y)=\exp(y)\exp(x)$
\ar[r]^{\ \ +}\mathbb{R}\times\mathbb{R}\ar[dr]_{\exp_1\exp_2 }&\ar[d]^{\exp}\mathbb{R}\\
&\mathbb{R}\\
}$
$\xymatrix{ $(x+y)^n=\sum_{j=0}^n {n \choose j} x^j\ y^{n-j}$
\ar[r]^{\ \ +}\mathbb{R}\times\mathbb{R}
\ar[dr]_{\sum_{j=0}^n {n \choose j} pr_1^{(j)}\ pr_2^{(n-j)}\ \ } & \ar[d]^{(\ )^n}\mathbb{R}\\
&\mathbb{R}\\
}$
-------------------------------------------------------------------------------------------------------------------------------------------------------------
Another example can be given where the domain set (source) is $\mathbb{C}^{n\times n}$, the algebra of square $n\times n$ matrices with complex coefficients. Let $a_{ij}:\mathbb{C}^{n\times n}\longrightarrow\mathbb{C}$ be the linear form which “takes” the coefficient of address $(i,j)$ (row $i$ and column $j$), that is to say $a_{ij}(M):=M[i,j]$. Then, the law of multiplication of matrices says that $MN[i,j]=\sum_{k=1}^n M[i,k]N[k,j]$, which can be represented in the style of Eq. by
$$\begin{aligned}
\label{matrixcoprod}
a_{ij}(MN)=\sum_{k=1}^n a_{ik}(M)a_{kj}(N).\end{aligned}$$
$$\begin{aligned}
\nonumber
\begin{array}{c}\xymatrix{
\ar[rr]^{\ \ \ \ product}\mathbb{C}^{n\times n}\times\mathbb{C}^{n\times n}\ar[drr]_{\sum_{k=1}^n (a_{ik})_1(a_{kj})_2\ \ \ \ \ \ }&&\ar[d]^{a_{ij}}\mathbb{C}^{n\times n}\\
&&\mathbb{C}\\
}
\end{array}\end{aligned}$$
Note that formula holds when the definition set (source) is a (multiplicative) semigroup of matrices (for example, the semigroup of unipotent positive matrices).
We now proceed to linear mappings that admit such “addition” or, rather, “multiplication” formulas.
***Derivations***: Let $\mathcal{A}$ be an arbitrary algebra with law of multiplication: $$\begin{aligned}
\mathcal{A}\otimes\mathcal{A}\stackrel{\mu}{\longrightarrow}\mathcal{A}.\end{aligned}$$ A derivation of $\mathcal{A}$ is an operator $D:\mathcal{A}\longrightarrow\mathcal{A}$ which follows the Leibniz rule, that is for all $x,y\in\mathcal{A}$, one has $$\begin{aligned}
D(xy)=D(x)y+xD(y)\ \ \ \ \text{(Leibniz rule)}.\end{aligned}$$ In the spirit of what has been represented above one has $$\begin{aligned}
\nonumber
\xymatrix{
\ar[r]^{\ \ \mu}\mathcal{A}\otimes\mathcal{A}\ar[dr]_{D_1Id_2+Id_1D_2}&\ar[d]^{D}\mathcal{A}\\
&\mathcal{A}
}\end{aligned}$$ which (as we have linear spaces and mappings) can be better represented by $$\begin{aligned}
\nonumber
\xymatrix{
\ar[r]^{\ \ \mu}\mathcal{A}\otimes\mathcal{A}\ar[d]_{D\otimes Id+Id\otimes D}&\ar[d]^{D}\mathcal{A}\\
\mathcal{A}\otimes\mathcal{A}\ar[r]^{\ \ \mu}&\mathcal{A}
}\end{aligned}$$
***Automorphisms***: An automorphism of $\mathcal{A}$ is an invertible linear mapping $g:\mathcal{A}\longrightarrow\mathcal{A}$ such that for all $x,y\in\mathcal{A}$, one has $$\begin{aligned}
g(xy)=g(x)g(y),\end{aligned}$$ which, in the spirit of what precedes can be represented by $$\begin{aligned}
\nonumber
\xymatrix{
\ar[r]^{\ \ \mu}\mathcal{A}\otimes\mathcal{A}\ar[d]_{g\otimes g}&\ar[d]^{g}\mathcal{A}\\
\mathcal{A}\otimes\mathcal{A}\ar[r]^{\ \ \mu}&\mathcal{A}
}\end{aligned}$$ Now remark that, classically, group representations act as automorphisms and representations of Lie algebras act as derivations. This immediately provides a scheme for constructing tensor products of two representations.
***Tensor product of two representations (groups and Lie algebras)***: First, take two representations of a group $G$, $\rho_i:G\longrightarrow End(V_i)$, $i=1,2$. The action of $g\in G$ on the tensor space $V_1\otimes V_2$ is given by $$\begin{aligned}
g(v_1\otimes v_2)=g(v_1)\otimes g(v_2).\end{aligned}$$ This means that the “tensor product” of the two (group) representations $\rho_i,\ i=1,2$ is given by the following data
- Space : $V_1\otimes V_2$
- Action : $\rho_1\sqtimes \rho_2 : g\ra \rho_1(g)\otimes \rho_2(g)$
Likewise, if we have two representations $\rho_i:\mathfrak{G}\longrightarrow End(V_i)$, $i=1,2$ of the Lie algebra $\mathfrak{G}$ the action of $g\in \mathfrak{G}$ on a tensor product $V_1\otimes V_2$ is given by $$\begin{aligned}
g(v_1\otimes v_2)=g(v_1)\otimes v_2+v_1\otimes g(v_2).\end{aligned}$$ Again, the “tensor product” of the two (Lie algebra) representations $\rho_i,\ i=1,2$ is given by the following data
- Space : $V_1\otimes V_2$
- Action : $\rho_1\sqtimes \rho_2 : g\ra \rho_1(g)\otimes Id_{V_2}(g)+Id_{V_1}(g)\otimes \rho_2(g)$
Roughly speaking, in the first case $g$ acts by $g\otimes g$ and in the second one by $g\otimes 1+1\otimes g$.
In view of two above cases it is convenient to construct linear mappings: $$\begin{aligned}
\mathcal{A}\stackrel{\Delta}{\longrightarrow}\mathcal{A}\otimes\mathcal{A}\ ,\end{aligned}$$ such that, in each case, $\rho_1\sqtimes\rho_2=(\rho_1\otimes\rho_2)\circ\Delta$.
In the first case ($\mathcal{A}=\mathbb{C}[G]$) one gets $$\begin{aligned}
\Delta\left(\sum_{g\in G}\alpha_g g\right)=\sum_{g\in G}\alpha_g g\otimes g.\end{aligned}$$ In the second case, one has first to construct the comultiplication on the monomials\
$g_1...g_n; g_i\in \mathfrak{G})$ as they span ($\mathcal{A}=\mathcal{U}_k(\mathfrak{G})$). Then, using the rule $\Delta(g)=g\otimes 1+1\otimes g$ (for $g\in \mathfrak{G}$) and the fact that $\Delta$ is supposed to be a morphism for the multiplication (the justification of this rests on the fact that the constructed action must be a representation see below around formula and exercise ), one has $$\begin{aligned}
\Delta(g_1...g_n)&=&(g_1\otimes 1+1\otimes g_1)(g_2\otimes 1+1\otimes g_2)...(g_n\otimes 1+1\otimes g_n)\cr
&=&\sum_{I+J=[1...n]}g[I]\otimes g[J].\end{aligned}$$ Where, for $I=\{i_1,i_2,\cdots ,i_k\}$ ($1\leq i_1<i_2<\cdots <i_k\leq n$), $g[I]$ stands for $g_{i_1}g_{i_2}\cdots g_{i_k}$ In each case (group algebra and envelopping algebra) one again gets a mapping $\Delta : \A\mapsto \A\otimes \A$ which will be expressed by $$\begin{aligned}
\Delta(a)=\sum_{i=1}^n a_i^{(1)}\otimes a_i^{(2)}\end{aligned}$$ which is rephrased compactly by $$\begin{aligned}
\label{sweedlers_notation}
\Delta(a)=\sum_{(1)(2)} a_{(1)}\otimes a_{(2)}.\end{aligned}$$
The action of $a\in \A$ on a tensor $v_1\otimes v_2$ is then, in both cases, given by $$\begin{aligned}
\label{tensor_act}
a.(v_1\otimes v_2)=\sum_{i=1}^n a_i^{(1)}. v_1\otimes a_i^{(2)}. v_2=\sum_{(1)(2)} a_{(1)}. v_1\otimes a_{(2)}. v_2.\end{aligned}$$
One can easily check, in these two cases, that $$\label{tensor_act1}
a.b.(v_1\otimes v_2)=(ab).(v_1\otimes v_2),$$ but in general does not guarantee ; this point will be discussed below in section .
Expression is very convenient for proofs and computations and known as Sweedler’s notation.
i\) In every case, we have extracted the “scheme of action” for building the tensor product of two representations. This scheme (a linear mapping $\Delta : \A\mapsto \A\otimes \A$) is independent of the considered representations and, in each case, $$\label{tensor1}
\rho_1\sqtimes \rho_2=(\rho_1\otimes \rho_2)\circ \Delta$$ ii) Sweedler’s notation becomes transparent when one speaks the language of “structure constants”. Let $\Delta: \c\mapsto \c\otimes \c$ be a comultiplication and $(b_i)_{i\in I}$ a (linear) basis of $\c$. One has $$\label{struct_const}
\Delta(b_i)=\sum_{j,k\in I}\lambda_i^{j,k}\ b_j\otimes b_k\ .$$ the family $(\lambda_i^{j,k})_{i,j,k\in I}$ is called the “structure constants” of the comultiplication $\Delta$. Note the duality with the notion the “structure constants” of a multiplication\
$\mu : \A\otimes \A \mapsto \A$ : if $(b_i)_{i\in I}$ is a (linear) basis of $\A$, one has $$\label{struct_const}
\mu(b_i\otimes b_j)=\sum_{k\in I}\lambda_{i,j}^k\ b_k\ .$$ For necessary and sufficient conditions for a family to be structure constants (see exercise ).
Then, the general construction for tensor products goes as follows.
: Let $\mathcal{A}$ be a vector space, a comultiplication $\Delta$ on $\mathcal{A}$ is a linear mapping $$\mathcal{A}\stackrel{\Delta}{\longrightarrow}\mathcal{A}\otimes\mathcal{A}.$$
Such a pair (vector space, comultiplication) without any prescription about the linear mapping “comultiplication” is called a coalgebra.
Now, imitating , if $\mathcal{A}$ is an algebra and $\rho_1$, $\rho_2$ are representations of $\mathcal{A}$ in $V_1$, $V_2$, for each $a\in \A$, we can construct an action of $a$ on $V_1\otimes V_2$ by $$\begin{aligned}
V_1\otimes V_2\stackrel{(\rho_1\otimes \rho_2)\circ \Delta(a)}{\longrightarrow}V_1\otimes V_2.\end{aligned}$$ This means that if $\Delta(a)=\sum_{(1)(2)} a_{(1)}\otimes a_{(2)}$, then, ($a$) acts on the tensor product by $$\begin{aligned}
\label{tensor_action}
\hspace{-1cm} a.(v_1\otimes v_2)=\sum_{(1)(2)} a_{(1)}. v_1\otimes a_{(2)}. v_2=\sum_{(1)(2)} \rho_1(a_{(1)})[v_1]\otimes \rho_2(a_{(2)})[v_2].\end{aligned}$$
But, at this stage, it is just an action and not (necessarily) a representation of $\mathcal{A}$. We shall later give the requirements on $\Delta$ for the construction of the tensor product to be reasonable (i. e. compatible with the usual tensor properties).
For the moment let us pause and consider some well known examples of comultiplication.
Combinatorics of some comultiplications
---------------------------------------
The first type of comultiplication is given by duality. This means by a formula of type $$\begin{aligned}
\label{dual_law}
\scal{\Delta(x)}{y\otimes z}^{\otimes2}=\scal{x}{y\ast z}\end{aligned}$$ for a certain law of algebra $V\otimes V\stackrel{\ast}{\mapsto} V$, where $\scal{\ }{\ }$ is a non degenerate scalar product in $V$ and $\scal{\ }{\ }^{\otimes2}$ stands for its extension to $V\otimes V$. In the case of words $\ast$ is the concatenation and $\scal{\ }{\ }$ is given by $\scal{u}{v}=\delta_{u,v}$. The comultiplication $\Delta_{Cauchy}$, dual to the concatenation, is given on a word $w$ by $$\begin{aligned}
\Delta(w)=\sum_{uv=w}u\otimes v.\end{aligned}$$ In the same spirit, one can define a comultiplication on the algebra of a finite group by
$$\Delta(g)=\sum_{g_1g_2=g}g_1\otimes g_2.$$
The second example is given by the multiplication law of elementary comultiplications, that is, if for each letter $x$ one has $\Delta(x)=x\otimes 1+1\otimes x$, then $$\begin{aligned}
\Delta(w)&=&\Delta(a_1...a_n)=\Delta(a_1)\Delta(a_2)...\Delta(a_n)\cr
&=&\sum_{I+J=[1...n]}w[I]\otimes w[J]\end{aligned}$$ where $w[\{i_1,i_2,\cdots i_k\}]=a_{i_1}a_{i_2}\cdots a_{i_k}$ (for $1\leq i_1<i_2<\cdots <i_k\leq n$). This comultiplication is dual to below for $q=0$ (shuffle product).\
Another example is a deformation (perturbation for small $q$) of the preceding. With $\Delta(a)=a\otimes 1+1\otimes a+ q a\otimes a$, one has $$\begin{aligned}
\Delta(w)&=&\Delta(a_1...a_n)=\Delta(a_1)\Delta(a_2)...\Delta(a_n)\cr
&=&
\sum_{I\cup J=[1...n]} q^{|I\cap J|} w[I]\otimes w[J].\end{aligned}$$
Note that this comultiplication is dual (in the sense of ) to the $q$-infiltration product given by the recursive formula (for general $q$ and with $1_{A^*}$ as the empty word)
$$\begin{aligned}
\label{infiltr}
\hspace{-15mm} w\ua 1_{A^*}=1_{A^*}\ua w=w\cr
\hspace{-15mm} au\ua bv=a(u\ua bv)+ b(au\ua v)+ q\delta_{a,b} (u\ua v).\end{aligned}$$
This product is an interpolation between the shuffle ($q=0$) and the (classical) infiltration ($q=1$) [@DFLL].
Requirements for a reasonable construction of tensor products {#reasonable_tensor}
=============================================================
We have so far constructed an action of $\A$ on tensors, but nothing indicates that this is a representation (see exercise ). So, the following question is natural.
If $\A$ is an algebra and $\Delta :\A\mapsto \A\otimes\A$, what do we require on $\Delta$ if we want the construction above to be a representation of $\A$ on tensor products ?
For $a,b\in \A$, $\rho_i$ representations of $\mathcal{A}$ in $V_i$, and $v_i\in V_i$ for $i=1,2$, we must have the following identity: $$\begin{aligned}
\label{action}
\!\!\!\!\!\!\!\!\!\! a.(b.v_1\otimes v_2)=(ab).v_1\otimes v_2\ \Longleftrightarrow\ \Delta(ab).v_1\otimes v_2=\Delta(a).(\Delta(b).v_1\otimes v_2).\end{aligned}$$ One can prove that, if this is true identically for all $a,b\in \mathcal{A}$ and all pairs of representations (see exercise ), one has $$\begin{aligned}
\label{Delta_morph}
\Delta(ab)=\Delta(a)\Delta(b).\end{aligned}$$ and, of course, if the latter holds, is true.\
This can be rephrased by saying that $\Delta$ is a morphism $\A\mapsto \A\otimes \A$.
ß Now, one would like to keep compatibility with the associativity of tensor products. This means that if we want to tensor $u\otimes v$ with $w$ it must give the same action as tensoring $u$ with $v\otimes w$. This means that we have to address the following question.\
If $\A$ is an algebra and $\Delta :\A\mapsto \A\otimes\A$ is a morphism of algebras, what do we require on $\Delta$ if we want the construction above to be associative ?
More precisely, for three representations $\rho_i,\ i=1,2,3$ of $\A$, we want $$\label{tensor_rep_assoc}
\rho_1\sqtimes(\rho_2\sqtimes\rho_3)=(\rho_1\sqtimes\rho_2)\sqtimes\rho_3$$ up to the identifications $(u\otimes v)\otimes w=u\otimes (v\otimes w)=u\otimes v\otimes w$ (if one is not satisfied with this identification, see exercise ).
Let us compute (up to the identification above) $$\begin{aligned}
\hspace{-1cm} a.[(u\otimes v)\otimes w]=\Delta(a).(u\otimes v)\otimes w=\Big((\Delta\otimes Id)\circ \Delta(a)\Big)\Big(u\otimes v\otimes w\Big)\end{aligned}$$ on the other hand $$\begin{aligned}
\hspace{-1cm} a.[u\otimes (v\otimes w)]=\Delta(a).u\otimes (v\otimes w)=\Big((Id\otimes \Delta)\circ\Delta(a)\Big)\Big(u\otimes v\otimes w\Big).\end{aligned}$$ Again, one can prove (see exercise ) that this holds identically (i. e. for every $a\in \A$ and triple of representations) iff $(Id\otimes \Delta)\circ\Delta=(\Delta\otimes Id)\circ\Delta$, *i.e.* $$\begin{aligned}
\label{co-associativity}
\xymatrix{
\ar[rr]^{\Delta}\mathcal{A}\ar[d]_{\Delta}&&\ar[d]^{Id\otimes \Delta}\mathcal{A}\otimes\mathcal{A}\\
\mathcal{A}\otimes\mathcal{A}\ar[rr]^{\Delta\otimes Id}&&\mathcal{A}\otimes\mathcal{A}\otimes\mathcal{A}
}\end{aligned}$$
The property is called co-associativity since if one reverses the arrows and replaces $\Delta$ by $\mu$, the multiplication in an algebra, the diagram expresses associativity (see also exercise on duals of co-algebras).
$$\begin{aligned}
\nonumber
\xymatrix{
\ar[rr]^{\mu\otimes Id}\mathcal{A}\otimes\mathcal{A}\otimes\mathcal{A}\ar[d]_{Id\otimes\mu}&&\ar[d]^{\mu}\mathcal{A}\otimes\mathcal{A}\\
\mathcal{A}\otimes\mathcal{A}\ar[rr]^{\mu}&&\mathcal{A}
}\end{aligned}$$
But the tensor product is not only associative, it has a “neutral” map which is “tensoring by the field of scalars”. This derives from the fact that the canonical mappings $$\begin{aligned}
\label{can_rl}
V\otimes_k k\stackrel{can_r}{\longrightarrow}V\stackrel{can_l}{\longleftarrow}k\otimes_k V.\end{aligned}$$ This can be summarized by the following question.
If $\A$ is an algebra and $\Delta :\A\mapsto \A\otimes\A$ a co-associative morphism of algebras, what do we require on $\Delta$ if we want the construction above to admit “tensoring by the field of scalars” as neutral ?
More precisely, we must have a representation of $\mathcal{A}$ in $\mathbb{C}$ (which means a morphism of algebras $\mathcal{A}\stackrel{\epsilon}{\longrightarrow}\mathbb{C}$) such that for a representation $\rho$ of $\A$, we want $$\rho\sqtimes \ep=\ep\sqtimes \rho=\rho$$ up to the identification $u\otimes 1=1\otimes u=u$ through the isomorphisms (if one is not satisfied with this identification, see exercise ).\
Hence, for all $a\in\mathcal{A}$ and $\rho$ representation on $V$, we should have: $$\begin{aligned}
can_r( a.(v\otimes 1))=a.v,\end{aligned}$$ and $$\begin{aligned}
a.(v\otimes 1)&=&\Big(\sum_{(1)(2)}\rho(a_{(1)})\otimes\epsilon(a_{(2)})\Big)[v\otimes 1]\nonumber\\
&=&(\rho\otimes Id_\C)\circ\Big(\sum_{(1)(2)}a_{(1)}\otimes\epsilon(a_{(2)})\Big)[v\otimes 1]\\\nonumber
&=&(\rho\otimes Id_\C)\circ(Id\otimes\epsilon)\circ\Delta(a)[v\otimes 1],\\\nonumber&&\\
\ \ \ \ a.v&=&\rho(a)[v]=can_r(\rho(a)[v]\otimes1)=can_r(\rho\otimes Id(a)[v\otimes1]).\end{aligned}$$ Similar computations could be made on the left, we leave them to the reader as an exercise.
ßThis means that one should require that $$\begin{aligned}
\nonumber
\xymatrix{
\ar[rr]^{\Delta}\mathcal{A}\ar[d]_{Id}&&\ar[d]^{Id\otimes\epsilon}\mathcal{A}\otimes\mathcal{A}\\
\mathcal{A}&&\ar[ll]_{can_r}\mathcal{A}\otimes\mathbb{C}
}
&\ \ \ \ \ \ \ \ \ \ \ &\xymatrix{
\ar[rr]^{\Delta}\mathcal{A}\ar[d]_{Id}&&\ar[d]^{\epsilon\otimes Id}\mathcal{A}\otimes\mathcal{A}\\
\mathcal{A}&&\ar[ll]_{can_l}\mathbb{C}\otimes\mathcal{A}
}\end{aligned}$$ Such a mapping $\epsilon:\mathcal{A}\longrightarrow\mathbb{C}$ is called a [*co-unit*]{}.
Again, one can prove (see excercice for details) that $$\begin{aligned}
\epsilon\text{ is a counit for }(\A,\Delta)\ \ \Longleftrightarrow\ \ \epsilon\text{ is a counit for } (\A^*,*_\Delta).\end{aligned}$$
Bialgebras
==========
Motivated by the preceding discussion, we define a [*bialgebra*]{} an algebra (associative with unit) endowed with a comultiplication (co-associative with counit) which allows for the two tensor properties of associativity and unit (see discussion above). More precisely
: $(\mathcal{A},\cdot,1_\mathcal{A},\Delta,\epsilon)$ is said to be a bialgebra iff
1. [$(\mathcal{A},\cdot,1_\mathcal{A})$ is an AAU,]{}
2. [$(\mathcal{A},\Delta,\epsilon)$ is a coalgebra coassociative with counit,]{}
3. [$\Delta$ is a morphism of AAU and $\epsilon$ is a morphism of AAU.]{}
The name bialgebra comes from the fact that the space $\A$ is endowed with two structures (one of AAU and one of co-AAU) with a certain compatibility between the two.
Examples of bialgebras
----------------------
Free algebra (word version : noncommutative polynomials) {#free-algebra-word-version-noncommutative-polynomials .unnumbered}
--------------------------------------------------------
Let $A$ be an alphabet (a set of variables) and $A^*$ be the free monoid constructed on $A$ (see Basic Structures ). For any field of scalars $k$ (one may think of $k=\R$ or $\C$), we call the algebra of noncommutative polynomials $\ncp{k}{A}$ (or free algebra), the algebra $k[A^*]$ of the free monoid $A^*$ constructed on $A$. This is the set of functions $f : A^*\mapsto k$ with finite support endowed with the convolution product $$f\ast g(w)=\sum_{uv=w}f(u)g(v)$$
Each word $w\in A^*$ is identified with its characteristic function (i.e. the Dirac function with value $1$ at $w$ and $0$ elsewhere). These functions form a basis of $\ncp{k}{A}$ and then, every $f\in \ncp{k}{A}$ can be written uniquely as a finite sum $f=\sum f(w)w$.\
The inclusion mapping $A\hookrightarrow \ncp{k}{A}$ will be denoted here by $can_A$.
**Comultiplications** The free algebra $\ncp{k}{A}$ admits many comultiplications (even with the two requirements of being a morphism and coassociative). As $A^*$ is a basis of $\ncp{k}{A}$, it is sufficient to define it on the words (if we require $\Delta$ to be a morphism it is enough to define it on letters).\
*Example 1*The first example is the dual of the Cauchy (or convolution) product $$\Delta(w)=\sum_{uv=w} u\otimes v$$ is not a morphism as $$\Delta(ab)=ab\otimes 1+a\otimes b+ 1\otimes ab$$ and $$\Delta(a)\Delta(b)=ab\otimes 1+a\otimes b+ b\otimes a+1\otimes ab$$ but it can be checked that it is coassociative (see also exercice for a quick proof of this fact).
ß*Example 2*A second example is given, on the alphabet $A=\{a,b\}$ by $$\Delta(a)=a\otimes b;\ \Delta(b)=b\otimes a$$ then $\Delta(w)=w\otimes \bar{w}$ where $\bar{w}$ stands for the word $w$ with $a$ (resp. $b$) changed in $b$ (resp. $a$). This comultiplication is a morphism but not coassociative as $$(I\otimes \Delta)\circ\Delta(a)=a\otimes b\otimes a\ ;\ (\Delta\otimes I)\circ\Delta(a)=a\otimes a\otimes b$$
*Example 3*The third example is given on the letters by $$\Delta(a)=a\otimes 1+1\otimes a+q a\otimes a$$ where $q\in k$. One can prove that $$\hspace{-1cm} \Delta(w)=\Delta(a_1...a_n)=\Delta(a_1)\Delta(a_2)...\Delta(a_n)=
\sum_{I\cup J=[1..|w|]} q^{|I\cap J|}w[I]\otimes w[J]$$ this comultiplication is coassociative.
For $q=0$, one gets a comultiplication given on the letters by $\Delta_s(a)=a\otimes 1+1\otimes a$. For every polynomial $P\in \ncp{k}{A}$, set $\ep(P)=P(1_{A^*})$ (the constant term). Then $(\ncp{k}{A},\ast,\Delta_s,\ep)$ is a bialgebra.
One has also another bialgebra structure with, for all $a\in A$ $$\label{letter_grouplike}
\Delta_h(a)=a\otimes a\ ;\ \ep_{aug}(a)=1$$ this bialgebra $(\ncp{k}{A},\ast,\Delta_h,\ep_{aug})$ is a substructure of the bialgebra of the free group.
Algebra of polynomials (commutative polynomials) {#algebra-of-polynomials-commutative-polynomials .unnumbered}
------------------------------------------------
We continue with the same alphabet $A$, but this time, we take as algebra $k[A]$. The construction is similar but the monomials, instead of words, are all the commutative products of letters i.e. $a_1^{\al_1}a_2^{\al_2}\cdots a_n^{\al_n}$ with $n$ arbitrary and $\al_i\in \N$. Denoting $\MON(A)$ the monoid of these monomials (comprising, as neutral, the empty one) and with $\Delta_s(a)=a\otimes 1+1\otimes a$, $\ep(P)=P(1_{\MON(A)})$, one can again check that $(k[A],\ast,\Delta_s,\ep)$ is a bialgebra.
Algebra of partially commutative polynomials {#algebra-of-partially-commutative-polynomials .unnumbered}
--------------------------------------------
For the detailed construction of a partially commutative monoid, the reader is referred to [@CF; @DK]. These monoids generalize both the free and free commutative monoids. To a given graph (non-oriented and without loops) $\vartheta\subset A\times A$, one can asssociate the monoid presented by generators and relations (see basic structures and diagram ) $$M(A,\vartheta)=\langle A;(xy=yx)_{(x,y)\in \vartheta}\rangle_{\bf Mon}\ .$$ This is exactly the monoid obtained as a quotient structure of the free monoid ($A^*$) by the smallest equivalence compatible with products (a congruence[^2]) which contains the pairs $(xy,yx)_{(x,y)\in \vartheta}$. A geometric model of this monoid using [*pieces*]{} was developped by X. Viennot [@Vi] where pieces are located on “positions” (drawn on a plane) two pieces “commute” iff they are on positions which do not intersect (see fig 1 below).
The partially commutative algebra $\ncp{k}{A,\vartheta}$ is the algebra $k[M(A,\vartheta)]$ [@DK]. Again, one can check that $(\ncp{k}{A,\vartheta},\ast,\Delta_s,\ep)$ (constructed as above) is a bialgebra.
Algebra of a group {#algebra-of-a-group .unnumbered}
------------------
Let $G$ be a group. The algebra under consideration is $k[G]$. We define, for $g\in G$, $\Delta(g)=g\otimes g$ and $\ep(g)=1$, then, one can check that $(k[G],.,\Delta,\ep)$ is a bialgebra.
Antipode and the problem of duals
=================================
Each vector space $V$ comes with its dual[^3] $$\begin{aligned}
V^*=Hom_\mathbb{C}(V,\mathbb{C}).\end{aligned}$$ The spaces $V^*$ and $V$ are in duality by $$\begin{aligned}
\langle p,\psi\rangle=p(\psi)\end{aligned}$$
Now, if one has a representation (on the left) of $\mathcal{A}$ on $V$, one gets a representation on the right on $V^*$ by $$\begin{aligned}
\langle p.a,\psi\rangle=\langle p,a.\psi\rangle\end{aligned}$$ If we want to have the action of $\A$ on the left again, one should use an anti-morphism $\alpha:\mathcal{A}\longrightarrow\mathcal{A}$ that is $\al\in End_k(\A)$ such that, for all $x,y\in \A$ $$\begin{aligned}
\alpha(xy)=\alpha(y)\alpha(x).\end{aligned}$$ In the case of groups, $g\longrightarrow g^{-1}$ does the job; in the case of Lie algebras $g\longrightarrow-g$ (extended by reverse products to the enveloping algebra) works.\
On the other hand, in the classical textbooks, the discussion of “complete reductibility” goes with the existence of an “invariant” scalar product $\phi : V\times V\mapsto k$ on a space $V$.\
For a group $G$, this reads $$\label{group_inv}
(\forall g\in G)(\forall x,y\in V)\Big(\phi(g.x,g.y)=\phi(x,y)\Big)$$ (think of unitary representations).\
For a Lie algebra $\mathfrak{G}$, this reads $$\label{Lie_inv}
(\forall g\in \mathfrak{G})(\forall x,y\in V)\Big(\phi(g.x,y)+\phi(x,g.y)=0\Big)$$ (think of the Killing form).\
Now, linearizing the situation by $\Phi(x\otimes y)=\phi(x,y)$ and remembering our unit representation, one can rephrase and in $$\label{phi_inv0}
(\forall a\in \A)(\forall x,y\in V)\Big(\Phi(\sum_{(1)(2)}a_{(1)}.x\otimes a_{(2)}.y)=\ep(a)\Phi(x\otimes y)\Big).$$ Likewise, we will say that a bilinear form $\Phi : U\otimes V\mapsto k$ is invariant is it satisfies $$\label{phi_inv}
(\forall a\in \A)(\forall x\in U,y\in V)\Big(\Phi(\sum_{(1)(2)}a_{(1)}.x\otimes a_{(2)}.y)=\ep(a)\Phi(x\otimes y)\Big).$$ which means that $\Phi$ is $\A$ linear for the structures being given respectively by $\rho_1\sqtimes \rho_2$ on $U\otimes_k V$ and $\ep$ on $k$.
Now, suppose that we have constructed a representation of a certain algebra $\A$ on $U^*$ by means of an antimorphism $\al: \A\mapsto \A$. To require that the natural contraction $c : U^*\otimes U\mapsto k$ be “invariant” means that
$$\begin{aligned}
\label{antipode0}
\hspace{-1.8cm}(\forall a\in \A)(\forall f\in U^*,y\in U)\Big(c(\sum_{(1)(2)}a_{(1)}.f\otimes a_{(2)}.y)=\ep(a)c(f\otimes y)=
\ep(a)f(y)\Big);\cr
(\forall a\in \A)(\forall f\in U^*,y\in U)\Big(\sum_{(1)(2)}f(\al(a_{(1)})a_{(2)}.y)=\ep(a)f(y)\Big).\end{aligned}$$
It is easy to check (taking a basis $(e_i)_{i\in I}$ and its dual family $(e_i^*)_{i\in I}$ for example) that is equivalent to $$\label{antipode1}
(\forall a\in \A)\Big(\rho_U(\sum_{(1)(2)} \al(a_{(1)})a_{(2)})=\ep(a)Id_U\Big).$$ Likewise, taking the natural contraction $c : U\otimes U^*\mapsto k$, one gets $$\label{antipode2}
(\forall a\in \A)\Big(\rho_U(\sum_{(1)(2)} a_{(1)}\al(a_{(2)}))=\ep(a)Id_U\Big).$$ Taking $U=\A$ and for $\rho_U$ the left regular representation, one gets $$\label{antipode3}
(\forall a\in \A)\Big(\sum_{(1)(2)} a_{(1)}\al(a_{(2)})=\sum_{(1)(2)} \al(a_{(1)})a_{(2)})=\ep(a)\Big).$$
Motivated by the preceding discussion, one can make the following definition.
Let $(\mathcal{A},\cdot,1_\mathcal{A},\Delta,\epsilon)$ be a bialgebra. A linear mapping $\al : \A\mapsto \A$ is called an antipode for $\A$, if for all $a\in\mathcal{A}$ $$\begin{aligned}
\label{antipode}
\sum_{(1)(2)}\alpha(a_{(1)})a_{(2)}=\sum_{(1)(2)}a_{(1)}\alpha(a_{(2)})=1_\mathcal{A}\epsilon(a).\end{aligned}$$
One can prove (see exercise ), that this means that $\al$ is the inverse of $Id_\A$ for a certain product of an algebra (AAU) on $End_k(\A)$ and this implies (see exercise ) that
1. [If $\alpha$ exists (as a solution of Eq.), it is unique.]{}
2. [If $\alpha$ exists, it is an antimorphism.]{}
: (Hopf Algebra) $(\mathcal{A},\cdot,1_\mathcal{A},\Delta,\epsilon ,S)$ is said to be a Hopf algebra iff
1. [$\mathcal{B}=(\mathcal{A},\cdot ,1_\mathcal{A},\Delta,\epsilon )$ is a bialgebra,]{}
2. [$S$ is an antipode (then unique) for $\mathcal{B}$]{}
In many combinatorial cases (see exercise on local finiteness ), one can compute the antipode by $$\begin{aligned}
\label{antipode_by_series}
\alpha(d)=\sum_{k=0}^\infty(-1)^{k+1} (I^+)^{(*k)} (d)).\end{aligned}$$ where $I^+$ is the projection on $\B^+=ker(\ep)$ such that $I^+(1_\B)=0$.
Hopf algebras and partition functions
=====================================
Partition Function Integrand
----------------------------
Consider the Partition Function $Z$ of a Quantum Statistical Mechanical System
$$Z = Tr\ exp(-\beta H)$$
whose hamiltonian is $H$ ($\beta\equiv 1/kT$, $k= \textrm{Boltzmann's constant}\ T=\textrm{absolute temperature}$). We may evaluate the trace over any complete set of states; we choose the (over-)complete set of coherent states $$|z\rangle = e^{-|z|^2/2}\sum_{n} (z^n / \sqrt{n!})a^{+n}|0\rangle$$ where $a^+$ is the boson creation operator satisfying $[a,a^+] = 1$ and for which the completeness or resolution of unity[^4] property is $$\label{res_unit}
\frac{1}{\pi} \int dz |z\rangle\langle z| = I \equiv \int d�(z) |z\rangle\langle z|.$$
The simplest, and generic, example is the free single boson hamiltonian $H =\ep a^+ a$ for which the appropriate trace calculation is $$\begin{aligned}
Z = \frac{1}{\pi} \int dz \langle z| exp(-\beta a^+ a) |z\rangle =\cr
= \frac{1}{\pi}\int dz \langle z| : exp(a^+ a(e^{-\beta\ep} - 1) : |z\rangle\ .\end{aligned}$$ Here we have used the following well known relation for the forgetful normal ordering operator $: f(a, a^+ ) :$ which means “normally order the creation and annihilation operators in $f$ forgetting the commutation relation $[a, a^+] = 1$”[^5]. We may write the Partition Function in general as $$\label{free_boson}
Z(x) = \int F (x, z) d�(z)$$ thereby defining the Partition Function Integrand (PFI) $F(x, z)$. We have explicitly written the dependence on $x= -\beta$, the inverse temperature, and $\ep$, the energy scale in the hamiltonian.
Combinatorial aspects: Bell numbers
-----------------------------------
The generic free boson example Eq. above may be rewritten to show the connection with certain well known combinatorial numbers. Writing $y = |z|^2$ and $x = -\beta\ep$, Eq. becomes $$Z = \int_0^\infty dy\ exp ( y(e^x - 1) )\ .$$ This is an integral over the classical exponential generating function for the Bell polynomials
$$exp(y(e^x - 1)) = \sum_{n=0}^\infty B_n(y) \frac{x^n}{n!}$$
where the Bell number is $B_n(1) = B(n)$, the number of ways of putting $n$ different objects into $n$ identical containers (some may be left empty). Related to the Bell numbers are the Stirling numbers of the second kind $S(n, k)$, which are defined as the number of ways of putting $n$ different objects into $k$ identical containers, leaving none empty. From the definition we have $B(n) = \sum^n_{k=1} S(n, k)$ (for $n\not=0$). The foregoing gives a combinatorial interpretation of the partition function integrand $F(x, y)$ as the exponential generating function of the Bell polynomials.
Graphs
------
We now give a graphical representation of the Bell numbers. Consider labelled lines which emanate from a white dot, the origin, and finish on a black dot, the vertex. We shall allow only one line from each white dot but impose no limit on the number of lines ending on a black dot. Clearly this simulates the definition of $S(n, k)$ and $B(n)$, with the white dots playing the role of the distinguishable objects, whence the lines are labelled, and the black dots that of the indistinguishable containers. The identification of the graphs for 1,2 and 3 lines is given in Figure 2. We have concentrated on the Bell number sequence and its associated graphs since, as we shall show, there is a sense in which this sequence of graphs is generic. That is, we can represent any combinatorial sequence by the same sequence of graphs as in the Figure 2
with suitable vertex multipliers (denoted by the $V$ terms in the same figure). Consider a general partition function $$\label{part_func8}
Z = Tr\ exp(-\beta H)$$
where the Hamiltonian is given by $H = \ep w(a, a^+)$, with $w$ a string ($=$ sum of products of positive powers) of boson creation and annihilation operators. The partition function integrand $F$ for which we seek to give a graphical expansion, is $$Z(x) = \int F (x, z) d�(z)$$ where $$\begin{aligned}
F (x, z) &=& \langle z| exp(xw)|z\rangle =\hspace{1cm} (x = -\beta\ep)\cr
&=& \sum_{n=0}^\infty \langle z|w^n|z\rangle \frac{x^n}{n!}\cr
&=& \sum_{n=0}^\infty W_n (z) \frac{x^n}{n!} \cr
&=& exp\Big(\sum_{n=1}^\infty V_n (z) \frac{x^n}{n!} \Big)\end{aligned}$$
with obvious definitions of $W_n$ and $V_n$. The sequences $\{W_n\}$ and $\{V_n\}$ may each be recursively obtained from the other. This relates the sequence of multipliers $\{V_n\}$ of Figure 2 to the Hamiltonian of Eq. . The lower limit $1$ in the $V_n$ summation is a consequence of the normalization of the coherent state $|z\rangle$.
The Hopf Algebra $\mathcal{L}_{\rm Bell}$
-----------------------------------------
We briefly describe the Hopf algebra $\LB$ which the diagrams of Figure 2 define.\
1. Each distinct diagram is an individual basis element of $\LB$ ; thus the dimension is infinite. (Visualise each diagram in a “box”.) The sum of two diagrams is simply the two boxes containing the diagrams. Scalar multiples are formal; for example, they may be provided by the V coefficients. Precisely, as a vector space, $\LB$ is the space freely generated by the diagrams of Figure 2 (see APPENDIX : Function Spaces).\
2. The identity element e is the empty diagram (an empty box).\
3. Multiplication is the juxtaposition of two diagrams within the same “box”. $\LB$ is generated by the connected diagrams; this is a consequence of the Connected Graph Theorem [@FU]. Since we have not here specified an order for the juxtaposition, multipli� cation is commutative.\
4. The comultiplication $\Delta : \LB\mapsto \LB\otimes \LB$ is defined by\
$$\begin{aligned}
\Delta(e) = e \otimes e\textrm{ (unit $e$, the empty box)}\cr
\Delta(x) = x \otimes e + e \otimes x \textrm{ (generator $x$)}\cr
\Delta(AB) = \Delta(A)\Delta(B) \textrm{ otherwise}\cr
\textrm{so that }\Delta \textrm{ is an algebra homomorphism.} \end{aligned}$$
The case of two modes
=====================
Let us consider an hamiltonian on two modes $H(a,a^+,b,b^+)$, with $$\begin{aligned}
[a,a^+]=1\cr
[b,b^+]=1\cr
[a^{\ep_1},b^{\ep_2}]=0,\ \ep_i \textrm{ being $+$ or empty (4 relations)}\end{aligned}$$
Suppose that one can express $H$ as $$H(a,a^+,b,b^+)=H_1(a,a^+)+H_2(b,b^+).$$
and that, $exp(\lambda H_1)$ and $exp(\lambda H_2)$ ($\lambda=-\beta$) are solved i.e. that we have expressions $$\hspace{-2cm}
F(\lambda)=exp(\lambda H_1)=\sum_{n=0}^\infty \frac{\lambda^n}{n!} H_1^{(n)}(a,a^+)\ ;\
G(\lambda)=exp(\lambda H_2)=\sum_{n=0}^\infty \frac{\lambda^n}{n!} H_2^{(n)}(b,b^+)$$
It is not difficult to check that $$exp(\lambda H)=exp\Big(\lambda (H_1+H_2)\Big)=F(\lambda \frac{d}{dx})G(x)\Big|_{x=0}$$
This leads us to define, in general, the “Hadamard exponential product”. Let
$$\begin{aligned}
F(z)=\sum_{n\geq 0} a_n\frac{z^n}{n!},\ G(z)=\sum_{n\geq 0}
b_n\frac{z^n}{n!} \end{aligned}$$
and define their product (the “Hadamard exponential product”) by
$$\begin{aligned}
\mathcal{H}(F,G):=\sum_{n\geq 0}
a_nb_n\frac{z^n}{n!}=\mathcal{H}(F,G)=\left.F\left(z\frac{d}{dx}\right)G(x)\right|_{x=0}.\end{aligned}$$
When $F(0)$ and $G(0)$ are not zero one can normalize the functions in this bilinear product so that $F(0)=G(0)=1$. We would like to obtain compact and generic formulas. If we write the functions as
$$\begin{aligned}
F(z)=\exp\left(\sum_{n=1}^\infty L_n\frac{z^n}{n!}\right) \textrm{ and } G(z)=\exp\left(\sum_{n=1}^\infty V_n\frac{z^n}{n!}\right).\end{aligned}$$
that is, as free exponentials, then by using Bell polynomials in the sets of variables $\L, \V$ (see [@GOF4; @OPG] for details), we obtain
$$\begin{aligned}
\mathcal{H}(F,G)=\sum_{n\geq 0} \frac{z^n}{n!} \sum_{P_1,P_2\in
UP_n} \L^{Type(P_1)}\V^{Type(P_2)}\end{aligned}$$
where $UP_n$ is the set of unordered partitions of $[1\cdots n]$. An unordered partition $P$ of a set $X$ is a collection of (nonempty) subsets of $X$, mutually disjoint and covering $X$ (i.e. the union of all the subsets is $X$, see [@GOF12] for details).
The type of $P\in UP_n$ (denoted above by $Type(P)$) is the multi-index $(\al_i)_{i\in \N^+}$ such that $\al_k$ is the number of $k$-blocks, that is the number of members of $P$ with cardinality $k$.
ßAt this point the formula entangles and the diagrams of the theory arise.\
Note particularly that
- the monomial $\L^{Type(P_1)}\V^{Type(P_2)}$ needs much less information than that which is contained in the individual partitions $P_1,\ P_2$ (for example, one can relabel the elements without changing the monomial),
- two partitions have an incidence matrix [*from which it is still possible to recover the types of the partitions.*]{}
ßThe construction now proceeds as follows.
1. Take two unordered partitions of $[1\cdots n]$, say $P_1,P_2$.
2. Write down their incidence matrix $\left(\card(Y\cap Z)\right)_{(Y,Z)\in P_1\times P_2}$.
3. Construct the diagram representing the multiplicities of the incidence matrix : for each block of $P_1$ draw a black spot (resp. for each block of $P_2$ draw a white spot).
4. Draw lines between the black spot $Y\in P_1$ and the white spot $Z\in P_2$; there are $\card(Y\cap Z)$ such.
5. Remove the information of the blocks $Y,Z,\cdots$.
In so doing, one obtains a bipartite graph with $p$ ($=\card(P_1)$) black spots, $q$ ($=\card(P_2)$) white spots, no isolated vertex and integer multiplicities. We denote the set of such diagrams by $\diag$.
The product formula now reads
$$\begin{aligned}
\mathcal{H}(F,G)=\sum_{n\geq 0} \frac{z^n}{n!} \sum_{d\in diag\atop
|d|=n} mult(d)\L^{\al(d)}\V^{\be(d)}\end{aligned}$$
where $\al(d)$ (resp. $\be(d)$) is the “white spot type” (resp. the “black spot type”) i.e. the multi-index $(\al_i)_{i\in \N^+}$ (resp. $(\be_i)_{i\in \N^+}$) such that $\al_i$ (resp. $\be_i$) is the number of white spots (resp. black spots) of degree $i$ ($i$ lines connected to the spot) and $mult(d)$ is the number of pairs of unordered partitions of $[1\cdots |d|]$ (here $|d|=|\al(d)|=|\be(d)|$ is the number of lines of $d$) with associated diagram $d$.
The diagrams as well as the product formula were introduced in [@BBM].
Diagrams
--------
One can design a (graphically) natural multiplicative structure on $\diag$ such that the arrow $$d \mapsto \L^{\al(d)}\V^{\be(d)}.$$ be a morphism. This is provided by the concatenation of the diagrams (the result, i.e. the diagram obtained in placing $d_2$ at the right of $d_1$, will be denoted by $[d_1|d_2]_D$). One must check that this product is compatible with the equivalence of the permutation of white and black spots among themselves, which is rather straightforward (see [@GOF4; @GOF12]). We have
[@GOF12] Let $\diag$ be the set of diagrams (including the empty one).\
i) The law $(d_1,d_2)\mapsto [d_1|d_2]_D$ endows $\diag$ with the structure of a commutative monoid with the empty diagram as neutral element(this diagram will, therefore, be denoted by $1_{\diag}$).\
ii) The arrow $d \mapsto \L^{\al(d)}\V^{\be(d)}$ is a morphism of monoids, the codomain of this arrow being the monoid of (commutative) monomials in the alphabet $\L\cup \V$ i.e. $$\hspace{-2.4cm}\mathfrak{MON}(\L\cup\V)=\{\L^\al \V^\be\}_{\al,\be\in (\N^+)^{(\N)}}=
\bigcup_{n,m\geq 1}\big\{L_1^{\al_1}L_2^{\al_2}\cdots
L_n^{\al_n}V_1^{\be_1}V_2^{\be_2}\cdots
V_m^{\be_m}\big\}_{\al_i,\be_j\in \N}.$$ iii) The monoid $(\diag,[-|-]_D,1_{\diag})$ is a free commutative monoid. The set on which it is built is the set of the connected (non-empty) diagrams.
The reader who is not familiar with the algebraic structure of $\MON(X)$ can find rigorous definitions in paragraph .
We denote $\phi_{mon,diag}$ the arrow $\diag\mapsto \mathfrak{MON}(\L\cup\V)$.
Labelled diagrams {#labelled_diagrams}
-----------------
We have seen the diagrams (of $\diag$) are in one-to-one correspondence with classes of matrices as in Fig. 3. In order to fix one representative of this class, we have to number the black (resp. white) spots from 1 to, say $p$ (resp. $q$). Doing so, one obtains a [*packed matrix*]{} [@DHT] that is, a matrix of integers with no row nor column consisting entirely of zeroes. In this way, we define the *labelled diagrams*.
A labelled diagram of size $p\times q$ is a bi-coloured (vertices are $p$ black and $q$ white spots) graph
- with no isolated vertex
- every black spot is joined to a white spot by an arbitrary quantity (but a positive integer) of lines
- the black (resp. white) spots are numbered from 1 to $p$ (resp. from 1 to $q$).
As in paragraph , one can concatenate the labelled diagrams, the result, i.e. the diagram obtained in placing $D_2$ at the right of $D_1$, will be denoted by $[D_1|D_2]_{L}$. This time we need not check compatibility with classes. We have a structure of free monoid (but not commutative this time)
[@GOF12] Let $\ldiag$ be the set of labeled diagrams (including the empty one).\
i) The law $(d_1,d_2)\mapsto [d_1|d_2]_L$ endows $\ldiag$ with the structure of a noncommutative monoid with the empty diagram ($p=q=0$) as neutral element (which will, therefore, be denoted by $1_{\ldiag}$).\
ii) The arrow from $\ldiag$ to $\diag$, which implies “forgetting the labels of the vertices” is a morphism of monoids.\
iii) The monoid $(\ldiag,[-|-]_L,1_{\ldiag})$ is a free (noncommutative) monoid which is constructed on the set of irreducible diagrams which are diagrams $d\not= 1_{\ldiag}$ which cannot be written $d=[d_1|d_2]_L$ with $d_i\not=1_{\ldiag}$.
i\) In a general monoid $(M,\star,1_M)$, the irreducible elements are the elements $x\neq 1_M$ such that $x=y\star z\Longrightarrow 1_M\in \{y,z\}$.\
ii) It can happen that an irreducible of $\ldiag$ has an image in $\diag$ which splits (i. e. is reducible), as shown by the simple example of the [*cross*]{} defined by the incidence matrix ${\pmatrix{0 & 1\cr 1 & 0}}$.
Hopf algebras $\DIAG$ and $\LDIAG$
----------------------------------
Let us first construct the Hopf algebra on the labelled diagrams (details can be found in [@GOF12]). In order to define the comultiplication, we need the notion of “restriction of a labelled diagram”. Consider $d\in \ldiag$ of size $p\times q$. For any subset $I\subset [1..p]$, we define a labelled diagram $d[I]$ (of size $k\times l$, $k=\card(I)$) by taking the $k=|I|$ black spots numbered in $I$ and the edges (resp. white spots) that are connected to them. We take this subgraph and relabel the black (resp. white) spots in increasing order.\
The construction of the Hopf algebra $\LDIAG$ goes as follows :\
1. the algebra structure is that of algebra of the monoid $\ldiag$ so that the elements of $\LDIAG$ are $$\label{gen_elem}
x=\sum_{d\in \ldiag} \al_d d$$ (the sum is finitely supported)
2. the comultiplication is given, on a labelled diagram $d\in \ldiag$ of size $p\times q$, by $$\Delta_L(d)=\sum_{I+J=[1..p]} d[I]\times d[J]$$
3. the counit is “taking the coefficient of the void diagram”,\
that is, for $x$ as in Eq. , $$\ep_L(x)= \al_{1_\ldiag}.$$
One can check that $(\LDIAG,[-|-]_L,1_{\ldiag},\Delta_L,\ep_L)$ is a bialgebra (for proofs see [@GOF12]). Now one can check that we satisfy the conditions of exercise question 3 and the antipode $S_L$ can be computed by formula of the same exercise.
ßWe have so far constructed the Hopf algebra $(\LDIAG,[-|-]_L,1_{\ldiag},\Delta,\ep,S_L)$.
ßThe constructions above are compatible with the arrow $$\phi_{\DIAG,\LDIAG}\ :\ \LDIAG\mapsto \DIAG$$ deduced from the class-map $\phi_{\diag,\ldiag}\ :\ \ldiag\mapsto \diag$ (a diagram is a class of labelled diagrams under permutations of black and white spots among themselves). So that, one can deduce “by taking quotients” a structure of Hopf algebra on the algebra of $\diag$. Denoting this algebra by $\DIAG$, one has a natural Hopf algebra structure $(\DIAG,[-|-]_D,1_{\diag},\Delta_D,\ep_D,S_D)$ and one can prove that this is the unique Hopf algebra structure such that $\phi_{\DIAG,\LDIAG}$ is a morphism for the algebra and coalgebra structures.
Link between $\LDIAG$ and other Hopf algebras
=============================================
The deformed case
-----------------
One can construct a three-parameter Hopf algebra deformation of $\LDIAG$ (see [@GOF12]), denoted $\LDIAG(q_c,q_s,q_t)$ such that $\LDIAG(0,0,0)=\LDIAG$ and\
$\LDIAG(1,1,1)=\MQS$ the algebra of Matrix Quasi Symmetric Functions [@DHT]). On the other hand, it was proved by L. Foissy [@F2; @F3] that one of the planar decorated trees Hopf algebra is isomorphic $\MQS$ and even to $\LDIAG(1,q_s,t)$ for every $q_s$ and $t\in\{0,1\}$. The complete picture is given below.
Duals of Hopf algebras
======================
The question of dualizing a Hopf algebra (i.e. endowing the dual - or a subspace of it - with a structure of Hopf algebra) is solved, in complete generality, by the machinery of Sweedler’s duals. The procedure consists in taking the “representative” linear forms (instead of all the linear forms) and dualize w.r.t. the following table
------------------ ------- ----------------------------
comultiplication $\ra$ multiplication
counit $\ra$ unit
multiplication $\ra$ comultiplication
unit $\ra$ counit
antipode $\ra$ trsnspose of the antipode.
------------------ ------- ----------------------------
In the case when the Hopf algebra is free as an algebra (which is often the case with noncommutative Hopf algebras of combinatorial physics), one can use rational expressions of Automata Theory to get a genuine calculus within this dual (see [@DT]).\
EXERCISES
=========
\[complex\_rep\_func\] **Representative functions on $\mathbb{R}$ (Complex valued)**. A function $\mathbb{R}\stackrel{f}{\longrightarrow}\mathbb{C}$ is said to be representative if there exist $(f_i^{(1)})_{i=1}^n$ and $(f_i^{(2)})_{i=1}^n$ such that, for all $x,y\in\mathbb{R}$ one has $$\begin{aligned}
\label{Repr}
f(x+y)=\sum_{i=1}^nf_i^{(1)}(x)f_i^{(2)}(y).\end{aligned}$$ 1) Show that $\cos$, $\cos^2$, $\sin$, $\exp$ and $a\ra a^n$ are representative. Provide minimal sums of type Eq.(\[Repr\]).\
2) Show that the following are equivalent\
i) $f$ is representative.\
ii) There exists a group representation $(\mathbb{R},+)\stackrel{\rho}{\longrightarrow}\C^{n\times n}$, a row vector $\lambda\in\C^{1\times n}$, a column vector $\gamma\in\C^{n\times 1}$ such that $f(x)=\lambda\rho(x)\gamma$.\
iii) $(f_t)_{t\in\mathbb{R}}$ is of finite rank in $\C^\mathbb{R}$ (here $f_t$, the shift of $f$ by $t$, is the function $x\longrightarrow f(x+t)$).
ß3) Show that the minimal $n$ such that formula Eq.(\[Repr\]) holds is also the rank (over $\C$) of $(f_t)_{t\in\mathbb{R}}$.
4\) a) If $f$ is continuous then $\rho$ can be chosen so and $\rho(x)=e^{xT}$ for a certain matrix $T\in\C^{n\times n}$.
b\) In this case show that representative functions are linear combinations of products of polynomials and exponentials.
5\) $f\in\mathbb{C}^\mathbb{R}$ is representative iff $\mathcal{R}e(f)=(f+\bar{f})/2$ and $\mathcal{I}m(f)=(f-\bar{f})/2i$ are representative in $\mathbb{R}^\mathbb{R}$.
6\) Show that the set of representative functions of $\C^\R$ is a $\C$-vector space. This space will be denoted $Rep_\C(\mathbb{R})$.
7\) Show that the functions $\varphi_{n,\lambda}=x^ne^{\lambda x}$ are a basis of $Rep_\mathbb{C}(\mathbb{R})\cap\mathcal{C}_o(\mathbb{R};\mathbb{C})$ (i.e. continuous complex valued representative functions on $\R$).
8\) Deduce from (7) that the following statement is false:\
“If a entire function $f:\R\mapsto \C$ is such that $(f_t)_{t\in \Z}$ is of finite rank, then it is representative”. [ : Consider $exp(exp(2i\pi x))$.]{}
\[rep\_func\] **Representative functions in general** (see also [@Ab; @CP]) Let $M$ be a monoid (semigroup with unit) and $k$ a field (one can first think of $k=\mathbb{R},\mathbb{C}$). For a function $M\stackrel{f}{\longrightarrow} k$ one defines the shifts: $$\begin{aligned}
f_z:x\longrightarrow f(zx),\cr
_yf:x\longrightarrow f(xy),\cr
_yf_z:x\longrightarrow f(zxy).\end{aligned}$$
1)a) Check the following formulas $$\begin{aligned}
\label{f1}
(f_{y_1})_{y_2}=f_{y_1y_2},\cr
_{y_2}(_{y_1}f_)=_{y_2y_1}f\cr
(_xf)_y=_x(f_y)=_xf_y.\end{aligned}$$
As for groups, if $M$ is a monoid, a $M$-module structure on a vector space $V$ is defined by a morphism or an anti-morphism $M\mapsto End_k(V)$.\
b) From Eqs.(\[f1\]) define two canonical $M$-module structures of $k^M$.\
2)a) Show that the following are equivalent
i\) $(f_z)_{z\in M}$ is of finite rank in $k^M$.
ii\) $(_yf)_{y\in M}$ is of finite rank in $k^M$.
iii\) $(_yf_z)_{y,z\in M}$ is of finite rank in $k^M$.
iv\) There exist two families $(f_i^{(1)})_{i=1}^n$ and $(f_i^{(2)})_{i=1}^n$ such that $$\begin{aligned}
\label{add_formula}
f(xy)=\sum_{i=1}^nf_i^{(1)}(x)f_i^{(2)}(y).\end{aligned}$$
v\) There exists a representation of $M$ $\rho:M\longrightarrow k^{n\times n}$, a row vector $\lambda\in k^{1\times n}$, a column vector $\gamma\in k^{n\times 1}$ such that $f(x)=\lambda\rho(x)\gamma$ for all $x\in M$.
b\) Using (v) above, show that the (pointwise) product of two representative functions is representative.
ßOne denotes $Rep_k(M)$ the space of ($k$-valued) representative functions on $M$.\
3) a) Recall briefly why the mapping $$k^{M}\otimes k^{M}\mapsto k^{M\times M}$$
defined by $$\sum_{i=1}^nf_i^{(1)}\otimes f_i^{(2)}\ra \Big((x,y)\ra \sum_{i=1}^nf_i^{(1)}(x) f_i^{(2)}(y)\Big)$$ is injective.\
b) Show that, if $n$ is minimal in , the families of functions $f_i^{(1)}$ and $f_i^{(2)}$ are representative and that the mapping $Rep_k(M)\mapsto Rep_k(M)\otimes Rep_k(M)$ defined by $$f\ra \sum_{i=1}^nf_i^{(1)}\otimes f_i^{(2)}$$ defines a structure of coassociative coalgebra on $Rep_k(M)$ Has it a co-unit ?\
We denote by $\Delta$ the compultiplication contructed above.\
4) Show that $(Rep_k(M),.,\mathbf{1},\Delta,\ep_{1_M})$ (with $\mathbf{1}$ the constant-valued function equal to $1$ on $M$ and $\ep$ the Dirac measure $f\ra f(1_M)$) is a bialgebra.
\[dissipation\] (Example of a monoid coming from dissipation theory [@S1]).\
Let $(\M_n,\times)$ be the monoid of $n\times n$ complex square matrices endowed with the usual product $$(\M_n,.)=(\C^{n\times n},\times)$$ We define a [state]{} (Von Neumann) as a positive semi-definite hermitian matrix of trace one ($=1$) i.e. a matrix $\rho$ such that
1. $\rho=\rho^*$
2. $(\forall x\in \C^{n\times 1})(x^*\rho x\geq 0)$
3. $Tr(\rho)=1$.
The set of such states will be denoted by $\St_n$.\
1) (Structure) a) Show that $\St_n$ is convex.\
b) Show that $\St_n$ is compact.\
[ : Consider the set of possible spectra i.e. the simplex $$S=\{(\lambda_1,\lambda_2,\cdots ,\lambda_n)\in (\R^+)^n | \sum_{i=1}^n \lambda_i=1\}$$ and show that $\St_n$ is the range of the continuous mapping $\phi : U(n)\times S\mapsto \M_n$ given by the formula $$\begin{aligned}
\phi(u,s)= u d u^* \ ;\ \textrm{with }
d=
\left(\begin{array}{cccc}
\lambda_1 & 0 & \cdots & 0\\
0 & \lambda_2 & \cdots & 0\\
& & \ddots & \\
0 & \cdots & 0 &\lambda_n
\end{array}\right)\end{aligned}$$ and where $s=(\lambda_1,\lambda_2,\cdots ,\lambda_n)$.\
]{} c) Show that the extremal elements [@B_TVS] of the compact $\St_n$ is the set of orthogonal projections of rank one and that this set is connected by arcs.\
2) (KS condition) We say that a finite family $(k_i)_{i\in I}$ ($I$ is finite) of elements in $\M_n$ fulfils the KS condition iff $\sum_{i\in I} k_i^*k_i=I_n$ ($I_n$ is the unity matrix, the unit of the monoid $\M_n$).\
On the other hand, given two finite families $A=(k_i)_{i\in I}$ and $B=(l_j)_{j\in J}$, we define $A\ast B$ as the family $$A\ast B=(k_il_j)_{(i,j)\in I\times J}$$ a) Show that, if $A$ and $B$ fulfil the KS condition then so too does $A\ast B$.
ßTo every (finite) family $A=(k_i)_{i\in I}$ which fulfils the KS condition, we attach a transformation $\phi_A : \M_n\mapsto \M_n$, given by the formula $$\phi_A(M)=\sum_{i\in I} k_i M k_i^*$$ b) Show that $\phi_A$ is linear and preserves $\St_n$ (i.e. $\phi_A(\St_n)\subset \St_n$).\
c) Show that if $A$ and $B$ fulfil the KS condition, one has $$\label{composition_phi}
\phi_A\circ\phi_B=\phi_{A\ast B}$$ Conclude that the $\phi_A$ form a semigroup of transformations (with unit).\
d) (Example) Let $(E_{i,j})_{i,j\in\{1,2\}}$ be the set of the four matrix units in $\M_2$. Show that the following families fulfil KS condition $$\begin{aligned}
A&=&(E_{11},\frac{1}{\sqrt(2)}E_{12},\frac{1}{\sqrt(2)}E_{22})\cr
B&=&(\frac{1}{\sqrt(2)}E_{11},\frac{1}{\sqrt(2)}E_{21},E_{22}) \end{aligned}$$ Compute $A*B$.\
3) (Description of the semigroup at the level of multisets). In order to pull-back the formula at the level of multisets, we remark that the order or the labelling of the elements $k_i$ is irrelevant; all that counts is their multiplicities.\
a) (Example showing that the “set” structure is too weak). Let $(E_{i,j})_{i,j\in\{1,2\}}$ be the set of the four matrix units in $\M_2$ and set $$\begin{aligned}
A&=&(\frac{1}{\sqrt(2)}E_{11},\frac{1}{\sqrt(2)}E_{12},\frac{1}{\sqrt(2)}E_{21},\frac{1}{\sqrt(2)}E_{22})\end{aligned}$$ compute $A\ast A$ and check that it has (non-zero) repeated elements and thus corresponds to a multiset (see appendix).\
b) Show that the multisets of $\M_n$ are exactly the elements $\sum_{M\in\M_n}\lambda_M [M]$ (here we note $[M]$ the image of $M\in \M_n$ in the algebra $\R[\M_n]$ in order to forbid matrix addition) of the algebra $\R[\M_n]$ such that $$\label{multisets_in_algebra}
(\forall M\in\M_n)(\lambda_M\in\N)$$ the set of elements fulfilling will be denoted $\N[\M_n]$.\
b) To every finite family of matrices (in $\M_n$) $A=(M_i)_{i\in I}$ one may associate its sum (in $\R[\M_n]$) $S(A)=\sum_{i\in I}M_i$, check that it is an element of $\N[\M_n]$ and that every element of $\N[\M_n]$ is obtained so.\
c) Show that $S(A\ast B)=S(A).S(B)\in\N[\M_n]\subset \R[\M_n]$ and deduce that $(\N[\M_n],.)$ is a monoid.\
d) To every multiset of matrices (in $\M_n$) $A=\sum_{M\in\M_n}\al(M) [M]$, one associates $$T(A)=\sum_{M\in\M_n}\al(M) [M^*M]$$ show that, if $A=(M_i)_{i\in I}$ is a finite set of matrices, one has $$\sum_{i\in I} M_i^*M_i=T(S(A))$$ e) Show that, if $T(A)=I_n$, $T(A.B)=T(B)$.\
We denote $\N[\M_n]^{KS}$ the set of $A\in \N[\M_n]$ such that $T(A)=I_n$.\
f) Check that the mapping $\phi_A$ defined previously depends only on $S(A)$ and denote, for $A\in \N[\M_n]^{KS}$ the mapping $\Phi_A$ deduced from this property.\
g) Prove that the mapping $(\N[\M_n]^{KS}, .)\mapsto (End_{\C}(\C^{n\times n}),\circ)$ is a morphism of semigroups (preserving the units).
ß 4) (Invertible elements) To every $A=\sum_{M\in\M_n}\al(M) [M]\in \N[\M_n]^{KS}$ one associates $\ep(A)=\sum_{M\in\M_n}\al(M)\in \N$.\
a) Prove that $$\ep(A.B)=\ep(A)\ep(B)$$ b) Prove that $$\ep(A)=1\Longrightarrow |supp(\al)|=1$$ and, thus, in this condition, $A=[M]$ (a single matrix).\
c) Deduce from (a) and (b) that the set of invertible elements of $\N[\M_n]^{KS}$ is exactly the unitary group $U(n)$.
\[quotient\_monoid\] Let $(M,\ast,1_M)$ be a monoid, $\equiv$ an equivalence relation on $M$ and $Cl : M\mapsto M/\equiv$ the class function (which, to every element of $M$ associates its class $Cl(x)$).\
a) Suppose that there is an (internal) law $\perp$ on $M/\equiv$ such that $Cl$ is a morphism i.e. one has $$\label{quotient_monoid1}
(\forall x,y\in M) (Cl(x*y)=Cl(x)\perp Cl(y))$$ then prove that $\equiv$ is compatible with the right and left “translations” of the monoid this means that $$\label{congruence}
(\forall\ x,y,t,s\in S)(x\equiv y\Longrightarrow [s*x*t\equiv s*y*t])$$ b) Conversely, we suppose that $\equiv$ is an equivalence on $M$ satisfying condition , show that the result $Cl(x'*y')$ does not depend on the choice of $x'\in Cl(x);\ y'\in Cl(x)$ and therefore construct a law $\perp$ on $M/\equiv$ such that the class function is a morphism (i.e. ).\
c) Show moreover that, in the preceding conditions, $\Big(M/\equiv,\perp,Cl(1_M)\Big)$ is a monoid.
\[tensor\_rep\] Let $\A$ be an AAU and $\Delta : \A\mapsto \A\otimes \A$ a comultiplication. We build (tensor) products of two representations by the formula , more precisely by
$$\label{tensor2}
can\circ (\rho_1\otimes \rho_2)\circ\Delta$$
where $can : End(V_1)\otimes End(V_2)\mapsto End(V_1\otimes V_2)$ is the canonical mapping.\
a) Prove that, if $\Delta$ is a morphism of algebras, then the linear mapping $$\rho_1\sqtimes \rho_2 : \A\mapsto End(V_1\otimes V_2)$$ defined by the composition is a morphism of AAU (and hence a representation).\
b) Prove that, if $\rho_1\sqtimes \rho_2$ is a representation for any pair $\rho_1,\rho_2$ of representations of $\A$, then $\Delta$ is a morphism of AAU (use $\rho_1=\rho_2$, the - regular - representation of $\A$ on itself by multiplications on the left).
\[tensor\_assoc\] We consider the canonical isomorphisms $$\begin{aligned}
\label{trilinear_ident}
can_{1|23} : V_1\otimes (V_2\otimes V_3)\mapsto V_1\otimes V_2\otimes V_3\cr
can_{12|3} : (V_1\otimes V_2)\otimes V_3\mapsto V_1\otimes V_2\otimes V_3\end{aligned}$$
Show that, in order to have for every triple $\rho_i,\ i=1,2,3$ of representations it is necessary and sufficient that $$can_{12|3}\circ(\Delta\otimes Id_\A)\circ \Delta=can_{1|23}\circ(Id_\A\otimes \Delta)\circ \Delta$$ (for the necessary condition, consider again the left regular representations).
\[coalgebra\_dual\] Let $(\A,\Delta)$ be a coalgebra ($\Delta$ is an arbitrary - but fixed - linear mapping) and $(\A^*,\ast_\Delta)$ be its dual algebra. Explicitely, for $f,g\in \A^*$ and $x\in \A$ (for convenience, the law is written in infix denotation) $$\scal{f\ast_\Delta g}{x}=\scal{f\otimes g}{\Delta(x)}$$ Prove the following equivalences $$\begin{aligned}
\Delta \text{ is co-associative }\Longleftrightarrow \ast_\Delta \text{ is associative }\\
\hspace{-2cm} (\forall \ep\in \A^*)\Big(\ep \text{ is a unity for }(\A^*,\ast_\Delta)\Longleftrightarrow
\ep \text{ is a co-unity for }(\A,\Delta)\Big)\end{aligned}$$
\[tensor\_unit\] The mappings $can_l,can_r$ are as in . Prove that, in order that for any representation $\rho$ of $\A$, one has $$can_l\circ(\ep\otimes \rho)\circ\Delta=can_r\circ(\rho\otimes \ep)\circ\Delta$$ it is necessary and sufficient that $\ep$ be a counit.
\[convolution\] Let $(\c,\Delta)$ be a coalgebra and $(\A,\mu)$ ba an algebra on the same (commutative) field of scalars $k$. We define a multiplication (called convolution) on $Hom_k(\c,\A)$ by $$f*g=\mu\circ (f\otimes g)\circ \Delta$$ so that, if $\Delta(x)=\sum_{(1)(2)}x_{(1)}x_{(2)}$, $$f*g(x)=\sum_{(1)(2)}f(x_{(1)})g(x_{(2)}).$$ 1) If $\A$ is associative and $\c$ coassociative, show that the algebra $(Hom_k(\c,\A),*)$ is associative.\
2) We suppose moreover that $\c$ admits a counit $\ep : \c\mapsto k$ and $\A$ a unit $1_\A$ (identified with the linear mapping $k\mapsto \A$ given by $\lambda\ra \lambda 1_\A$).\
Show that $1_\A\circ \ep$ (traditionnally denoted $1_\A\ep$) is the unit of the algebra $(Hom_k(\c,\A),*)$.\
3) Let $(\B,\ast,1,\Delta,\ep)$ be a bialgebra. The convolution under consideration will be that constructed between the coalgebra $(\B,\Delta,\ep)$ and the algebra $(\B,\ast,1)$.\
a) Let $S\in End(\B)$. Show that the following are equivalent\
i) $S$ is an antipode for $\B$\
ii) $S$ is the inverse of $Id_\B$ in $(End_k(\B),*)$.\
b) Deduce from (b) that the antipode, if it exists, is unique.\
c) Prove that the bialgebra $(\ncp{k}{A},\ast,\Delta_h,\ep_{aug})$ defined around equation admits no antipode (if the alphabet $A$ is not empty).\
4) Let $(\c,\Delta,\ep)$ be a coalgebra coassociative with counit. We define $\Delta_2$ by $T_{2,3}\circ \Delta\otimes \Delta$ where $T_{2,3} : \c^{\otimes 4}\mapsto \c^{\otimes 4}$ is the flip between the 2nd and the 3rd component $$T_{2,3}(x_1\otimes x_2\otimes x_3\otimes x_4)=x_1\otimes x_3\otimes x_2\otimes x_4$$ a) Show that $(\c\otimes\c,\Delta_2,\ep\otimes\ep)$ (with $\ep\otimes\ep(x\otimes y)=\ep(x)\ep(y)$) is coassociative coalgebra with counit.\
Let $(\H,\mu,1,\Delta,\ep,S)$ be a Hopf algebra. The convolution $*$ here will be that constructed between the coalgebra $(\H\otimes\H,\Delta_2,\ep\otimes\ep)$ and the algebra $(\H,\mu,1)$. We consider the two elements $\nu_i\in (Hom_k(\H\otimes\H,\H),*)$ defined by $\nu_1=S\circ\mu$ and $\nu_2(x\otimes y)=S(y)S(x)$. b) Show that the elements $\nu_i$ are the convolutional inverses of $\mu$. Deduce from this that $S : \H\mapsto \H$ is an antimorphism of algebras.
\[locally\_finite\] 1) Let $(\B,\ast,1,\Delta,\ep)$ be a bialgebra, we denote by $\B^+$ the kernel of $\ep$.\
a) Prove that $\B=\B^+\oplus k.1_\B$.\
We denote $I^+$ the projection $\B\mapsto\B^+$ with respect to the preceding decomposition.\
b) Prove that, for every $x\in\B^+$, one can write $$\Delta(x)=x\otimes 1+ 1\otimes x+ \sum_{(1)(2)}x_{(1)}\otimes x_{(2)} \text{ with } x_{(i)}\in \B^+$$ 2) Define for $x\in \B^+$, $$\Delta^+(x)=\Delta(x)-(x\otimes 1+ 1\otimes x)=\sum_{(1)(2)}x_{(1)}\otimes x_{(2)}$$ a) Check that $(\B^+,\Delta^+)$ is a coassociative coalgebra.\
Define $$(\B^*)^+=\{f\in \B^*|f(1)=0\}$$ b) Prove that $(\B^*)^+$ is a subalgebra of $(\B,\ast_\Delta)$ and that its law is dual of $\Delta^+$.\
c) Prove that the algebra $(\B^*,\ast_\Delta)$ is obtained from $\Big((\B^*)^+,\ast_{\Delta^+}\Big)$ by adjunction of the unity $\ep$.\
3) The bialgebra is called locally finite if $$(\forall x\in \B)(\exists k\in \N^*)(\Delta^{+(k)}(x)=0).$$ The projection $I^+$ being as above, show that, in case $\B$ is locally finite, $$(\forall x\in \B)(\exists N\in \N^*)(\forall k\geq N)((I^+)^k(x)=0)$$ and that $$\label{antipode_by_series1}
S=\sum_{n\in \N}(-I^+)^n$$ is an antipode for $\B$.
1\) Let $G$ be a group and $\H=(\C[G],.,1_G,\Delta,\ep,S)$ be the Hopf algebra of $G$.\
a) Show that $\{(g-1)\}_{g\in G-\{1\}}$ is a basis of $\H^+$ (defined as above) and that $\Delta^+(g-1)=(g-1)\otimes (g-1)$.\
b) Show that, if $G\not=\{1\}$, $\H^+$ is not locally finite, but $\H$ admits an antipode.\
2) Prove that, if the coproduct of $H$ is graded (i.e. there exists a decomposition $H=\oplus_{n\in \N}H_n$ with $\Delta(H_n)\subset \sum_{a+b=n} H_a\otimes H_b$) and $H_0=k.1_H$, then the comultiplication is locally finite.\
3) Define the degree of a labelled diagram as its number of edges and $\LDIAG_n$ as the vector space generated by the diagrams of degree $n$ and check that we satisfy the conditions of exercise question 5.
\[struct\_const\_ex\] 1) Show that, in order that a family $(\lambda_{i,j}^k)_{i,j,k\in I}$ be the family of structure constants of some algebra it is necessary and sufficient that $$\label{struct_const_alg}
(\forall (i,j)\in I^2)\Big((\lambda_{i,j}^k)_{k\in I}\textrm { is finitely supported}\Big)$$ 2) Similarly show that in order that a family $(\lambda_{i}^{j,k})_{i,j,k\in I}$ be the family of structure constants of some coalgebra it is necessary and sufficient that $$\label{struct_const_coalg}
(\forall i\in I)\Big((\lambda_{i}^{j,k})_{(j,k)\in I^2}\textrm { is finitely supported}\Big)$$ 3) Give examples of mappings $\lambda : I^3\mapsto k$ such that the corresponding families satisfy
- and
- and not
- and not
- none of and
4\) Give further examples such as those in 3) i-iii but now defining associative (resp. coassociative) multiplications (resp. comultiplications).
APPENDIX
========
Function spaces {#function_spaces}
---------------
Throughout the text, we use the basic constructions of set theory and algebra (see [@B_Set; @B_Alg_III]).\
The set of mappings between two sets $X$ and $Y$ is denoted by $Y^X$. Thus if $k$ is a field $$k^X=\{f:X\longrightarrow k\}$$ the vector space of all functions defined on $X$ with values in $k$. For each function $f\in k^X$, we call the “support of $f$” the set of points $x\in X$ such that $f(x)$ is not zero[^6]. $$\label{support}
supp(f)=\{x\in X:f(x)\neq 0\}$$ the set of functions with finite support is a vector subspace of $k^X$ which is denoted by $k^{(X)}$.
ßAn interesting extension of this notion to other sets of coefficients is the combinatorial notion of (finite) multisets.\
Recall that a multiset is a [(set with repetitions)]{} [@K2]. For example, the first multisets with elements from $\{a,b\}$ are $$\begin{aligned}
\label{first_multisets}
\{\},\{a\},\{b\},\{a,a\},\{a,b\},\{b,b\},\{a,a,a\},\{a,a,b\},\{a,b,b\},\{b,b,b\},\cr
\{a,a,a,a\},\{a,a,a,b\},\{a,a,b,b\},\{a,b,b,b\},\{b,b,b,b\},\cdots .\end{aligned}$$ A multiset with elements in $X$ is then described equivalently by a multiplicity function $\al : X\mapsto \N$ with finite support. The support of such a mapping is defined as in and the set of finite multiplicity functions will be denoted by $\N^{(X)}$ (see below the free commutative monoid ). For example, the multiplicity functions $\{a,b\}\mapsto \N$ corresponding to the multisets given in are, in the same order (we characterize $\al$ by the pair $(\al(a),\al(b))$) $$\begin{aligned}
\label{first_multisets}
(0,0),(1,0),(0,1),(2,0),(1,1),(0,2),(3,0),(2,1),(1,2),(0,3),\cr
(4,0),(3,1),(2,2),(1,3),(0,4),\cdots \end{aligned}$$
Basic structures {#basic_structures}
----------------
(Semigroup) A semigroup $(S,*)$ is a set $S$ endowed with a closed binary operation $*$ satisfying an associative law, this means that, for all $x,\ y,\ z\in S$ one has $x*(y*z)=(x*y)*z$.
[http://en.wikipedia.org/wiki/Semigroup]{}
(Monoid) A monoid $(M,\ast)$ is a semigroup which possesses a neutral element, i.e. an element $e\in M$ such that, for all $x\in M$: $$\begin{aligned}
e\ast x=x\ast e=x.\end{aligned}$$ Such an element, if it exists is unique. The neutral element is often denoted $1_M$.
[http://en.wikipedia.org/wiki/Monoid]{}
\[f\_monoid\] (Free Monoid) The free monoid of alphabet $X$ is the set of [*strings*]{} $x_1x_2\cdots x_n$ with letters $x_i\in X$ (comprising the empty string). This set is denoted $X^*$, its law is the concatenation and its neutral element is the empty string.
It is easily seen that this monoid is free in the following sense. For any “set-theoretical” mapping $\phi : X\mapsto M$, where $(M,\ast)$ is a monoid, $\phi$ can be extended to strings so that $$\label{f_monoid_universal}
\begin{array}{c}\xymatrix{
\ar[r]^{\ \phi}X\ar[dr]_{can}& M\\
&\ar[u]^{\bar\phi}X^*\\
}
\end{array}$$
\[fc\_monoid\] (Free Commutative Monoid) The free commutative monoid of the alphabet $X$ is the set of [*monomials*]{} $X^\al$ ($\al\in \N^{(X)}$). This set is denoted by $\MON(X)$ and its law is the multiplication of monomials $$X^\al X^\be=X^{\al+\be}$$
It is easily seen that this monoid is free in the following sense. For any “set-theoretical” mapping $\phi : X\mapsto M$, where $(M,\ast)$ is a commutative monoid, $\phi$ can be extended to monomials so that $$\begin{array}{c}\xymatrix{
\ar[r]^{\ \phi}X\ar[dr]_{can}& M\\
&\ar[u]^{\bar\phi} \MON(X)\\
}
\end{array}$$
An interesting application of the free monoid is the explicit construction of a monoid defined “by generators and relations”.
Let $X$ be a set (of generators) and $R=(u_i,v_i)_{i\in I}$ a family of pairs of words, then one can construct explicitely the smallest (i.e. the intersection of) congruence $\equiv_R$ for which $$(\forall i\in I) (u_i\equiv v_i).$$ Let say that two words $U,V\in X^*$ are “related” by $\equiv_R$ if there is a chain of replacements of the type $pu_is\ra pv_is$ or $pv_is\ra pu_is$ $p,s\in X^*$ leading from $U$ to $V$. Formally, there exists a chain $$U=U_0,U_1,\cdots U_n=V$$ such that for each $j<n$ $U_j=p_jAq_j\ ;\ U_{j+1}=p_jBq_j$ with $(A,B)=(u_i,v_i)$ or $(B,A)=(u_i,v_i)$ for some $i$ (depending on $j$. One can show that the constructed $\equiv_R$ is a congruence (see exercise ) and we define $$\langle X;R\rangle_{\bf Mon}$$ as the quotient $X^*/\equiv_R$. This monoid has the following property : if $\phi X^* \mapsto M$ is a morphism (of monoids) such that, for all $i\in I$ one has $\phi(u_i)=\phi(v_i)$, then $\phi$ factorises uniquely through $\langle X;R\rangle_{\bf Mon}$
$$\label{gen_rel_universal}
\begin{array}{c}\xymatrix{
\ar[r]^{\ \phi}X\ar[dr]_{can}& M\\
&\ar[u]^{\bar\phi} \langle X;R\rangle_{\bf Mon}\\
}
\end{array}$$
(Group) A group $(G,\ast)$ is a monoid such that for each $x\in G$ there exists $y$ such that $$\begin{aligned}
x\ast y=y\ast x=e.\end{aligned}$$ For fixed $x$ such an element is unique and is usually denoted by $x^{-1}$ and called the inverse of $x$.
(Algebra of a monoid) Let $k$ be a field (scalars, for example $k=\mathbb{R}$ or $\mathbb{C}$). The algebra $k[M]$ of a monoid $M$ (with coefficients in $k$) is the set of mappings $k^{(M)}$ endowed with the convolution product $$f\ast g(w)=\sum_{uv=w}f(u)g(v).$$ The algebra $(k[M],\ast)$ is an AAU.
Each $m\in M$ may be identified with its characteristic function (i.e. the Dirac function $\delta_m$ with value $1$ at $m$ and $0$ elsewhere). These functions form a basis of $k[M]$ and then, every $f\in k[M]$ can be written as a finite sum $f=\sum_w f(w)w$. Through this identification the unity of $M$ and $k[M]$ coincide.
The algebra of a monoid solves the following univeral problem.\
Let $M$ be a monoid and $j:M\mapsto k[M]$ the embedding described above. For any AAU $\A$ and any morphism of monoids $\phi: M\mapsto \A$ ($(\A,.)$ is a monoid), one has an unique factorization
$$\begin{aligned}
\nonumber
\ \ \ \ \ \ \xymatrix{
M\ar[r]^{\phi}\ar[d]^{j}& \A\\
k[M]\ar[ur]_{\bar{\phi}} &\\
}\end{aligned}$$
where $\bar{\phi}$ is a morphism of AAU.
Likewise, the enveloping algebra $\mathcal{U}_k(\mathfrak{G})$ of a Lie algebra $\mathfrak{G}$ with the canonical mapping $can: \mathcal{U}_k(\mathfrak{G})\mapsto \mathfrak{G}$ is the solution of a universal problem. The specifications are the following
1. $\mathcal{U}_k(\mathfrak{G})$ is an AAU
2. $can$ is a morphism of Lie algebras (for this, $\mathcal{U}_k(\mathfrak{G})$ is endowed of the structure of Lie algebra given by the bracket $[X,Y]=XY-YX$).
For any morphism of Lie algebras $\phi : \mathfrak{G}\mapsto \A$ (where $\A$ is endowed with the structure of a Lie algebra given by the bracket $[X,Y]=XY-YX$), one has a unique factorization
$$\begin{aligned}
\xymatrix{
\mathfrak{G}\ar[r]^{\phi}\ar[d]^{can}& \A\\
\mathcal{U}_k(\mathfrak{G})\ar[ur]_{\bar{\phi}} &\\
}\end{aligned}$$
where $\bar{\phi}$ is a morphism of AAU.
References {#references .unnumbered}
==========
[ABC]{} , [*Hopf Algebras*]{}, Cambridge University Press (2004). , [*Deformation and Quantization*]{}, Ann. of Phys. [**111**]{}, (1978), pp. 61-151. , [ *Quantum field theory of partitions*]{}, J. Math. Phys. [**40**]{}, 3239 (1999) , [*Rational series and their languages EATCS Monographs on Theoretical Computer Science*]{}, Springer (1988). , [*Boson normal ordering via substitutions and Sheffer-Type Polynomials*]{}, Phys. Lett. A [**338**]{} (2005) 108 , [*Theory of sets*]{}, Springer (2004) , [*Algebra, chapter 1-III*]{}, Springer (2006) , [*Algebra, chapter VI*]{}, Springer (2006) , [*Topological Vector Spaces*]{}, Springer (2006) , [*Integration I*]{}, Springer (2006) , [*Towards cohomology of renormalization: bigrading the combinatorial Hopf algebra of rooted trees*]{}, Comm. Math. Phys. [**215**]{} (2000), no. 1, 217�236, hep-th/00 01202. , [*Renormalization of QED with planar binary trees*]{}, [arXiv:hep-th/0003202v1]{}. , [*Quantum field theory and Hopf algebra cohomology*]{}, [arXiv:hep-th/0311253v2]{}. , [*Trees, renormalization and differential equations*]{}, BIT 44 (2004), no. 3, 425�438. <span style="font-variant:small-caps;">Cartier P.</span>, [*A primer of Hopf algebras*]{}, Septembre (2006), IHES preprint IHES/M/06/40. , [*Hopf algebras, Renormalization and Noncommutative geometry*]{}, Comm. Math. Phys 199 (1998), [**1**]{}, hep-th/98 08042. , [*Hopf algebras, cyclic cohomology and the transverse index theorem*]{}, Comm. Math. Phys. 198 (1998), [**1**]{}, math.DG/98 06109. <span style="font-variant:small-caps;">Cartier P., Foata D.</span>, [*Problèmes combinatoires de commutation et r�arrangements*]{}, 1969, Springer-Verlag\
Free electronic version available at\
http://www-irma.u-strasbg.fr/ foata/paper/pub11.html , [*An operational calculus for the Mould operad*]{}, [arXiv:0710.0349v3]{} , A guide to quantum groups. Cambridge (1994). , [*Feynman graphs and related Hopf algebras*]{}, Journal of Physics: Conference Series, SSPCM’05, Myczkowce, Poland. arXiv : cs.SC/0510041 , [*Direct and dual laws for automata with multiplicities*]{}, T.C.S. [**267**]{}, 105-120 (2001).\
[arXiv]{} : math.CO0607412 , [*Free partially commutative structures*]{}, Journal of Algebra [**156**]{}, (1993) 318-361 , [*Congruences Compatible with the Shuffle Product*]{}, Proceedings, (FPSAC 2000), D. Krob, A.A. Mikhalev Eds., Springer.\
[arXiv]{} : math.CO0607419 , [*Non commutative functions VI: Free quasi-symmetric functions and related algebras*]{}, International Journal of Algebra and Computation [**12**]{}, No 5 (2002). , [*One-parameter groups and combinatorial physics*]{}, Proceedings of the Symposium Third International Workshop on Contemporary Problems in Mathematical Physics (COPROMAPH3) (Porto-Novo, Benin, Nov. 2003), J. Govaerts, M. N. Hounkonnou and A. Z. Msezane (eds.), p. 436 (World Scientific Publishing 2004)\
arXiv: [quant-ph/04011262]{} , [*A multipurpose Hopf deformation of the algebra of Feynman-like diagrams*]{}, Proceedings of the 26th International Colloquium on Group Theoretical Methods in Physics, New York 2006, Editor: S. Catto (2007). , [*A Three Parameter Hopf Deformation of the Algebra of Feynman-like Diagrams*]{}, [arXiv:0704.2522]{}, Subject: Mathematical Physics (math-ph). <span style="font-variant:small-caps;">Duchamp G. H. E., Tollu C.</span>, *Sweedler’s duals and Schützenberger’s calculus* `arXiv:0712.0125` (to be published). , [*Combinatorics of rooted trees and Hopf algebras*]{}, Trans. Amer. Math. Soc. 355 (2003), no. 9, 3795�3811, math.CO/02 01253. , [*Finite-dimensional comodules over the Hopf algebra of rooted trees*]{}, J. Algebra 255 (2002), no. 1, 85�120, math.QA/01 05210. , [*Les algèbres de Hopf des arbres enracinés I*]{}, Bull. Sci. Math. 126 (2002), 193�239. , [*Les algèbres de Hopf des arbres enracinś II*]{}, Bull. Sci. Math. 126 (2002), 249�288. , Proc. Nat. Acad. [**42**]{}, 122,1956. , [*Hopf algebras of endomorphisms of Hopf algebras*]{}, (Oct 2004) [ArXiv : math.QA/0410364]{} , [*Lectures on Quantum Theory: Mathematical and Structural Foundations*]{}, Imperial College Press, London, 1995. , [*Quasi-shuffle products*]{}, J. Algebraic Combin. (2000), 49-68. , [*The art of computer programming*]{} Volume II. Addison-Wesley (1981). , [*Knots and Feynman Diagrams*]{}, Cambridge Lecture Notes in Physics (2000). , [*On the Hopf algebra structure of pertubative quantum field theories*]{}, Adv. Theor. Math. Phys. [**2**]{} (1998), [**2**]{}, [q-alg/97 07029]{}. , [Duality between quasi-symmetric functions and the Solomon descent algebra]{}, J. Algebra [**177**]{} (1995), [**3**]{}. , [*Binomialkoeffitzenten und Shuffle-Zahlen*]{}, Technischer Bericht, Fachbereich Informatik, T. H. Darmstadt, 1981. , [Symmetry-Adapted Polynomial Basis for Global Potential Energy Surfaces - Applications to XY4 Molecules]{}. J. Math. Chem. To appear. , [*Quantum Theory: Concepts and Methods*]{}, Kluwer Academic Publishers, New York, 2002. , [*Free Lie algebras*]{}, Oxford University Press (1993). , [*Hopf algebra structure of a model quantum field theory*]{}, Proceedings of the 26th International Colloquium on Group Theoretical Methods in Physics, New York 2006, Editor: S. Catto (2007). , [*Thermalization of squeezed states*]{}, J. Opt. B: Quantum Semiclass. Opt. 7 (2005) doi:10.1088/1464-4266/7/12/015. , [*Heaps of pieces, I: Basic definitions and combinatorial lemmas*]{}. In Labelle and Leroux, editors, Combinatoire Énumérative, [**1234**]{} in Lect. Notes in Math., 321-350. Springer, 1986.
[^1]: As usual, $Y^X$ is the set of all mappings from $X$ to $Y$, see appendix.
[^2]: See exercise and the presentation of monoids “by generators and relations in paragraph ”.
[^3]: In general, for two $k$-vector spaces $V$ and $W$, $Hom_k(V,W)$ is the set of all linear mappings $V\mapsto W$.
[^4]: Sometimes physicists write $d^2z$ to emphasize that the integral is two dimensional (over $\R$) but here, the l. h. s. of is the integration of the operator valued function $z\ra |z\rangle\langle z|$ - see [@B_Int1] Chap. III Paragraph 3 - w.r.t. the Haar mesure of $\C$ which is $dz$.
[^5]: Of course, this procedure may alter the value of the operator to which it is applied.
[^6]: In integration theory, the support of a function is the closure of what we define as the (algebraic) support.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $X$ be a max-stable random vector with positive continuous density. It is proved that the conditional independence of any collection of disjoint subvectors of $X$ given the remaining components implies their joint independence. We conclude that a broad class of tractable max-stable models cannot exhibit an interesting Markov structure.'
author:
- |
Ioannis Papastathopoulos[^1],\
Kirstin Strokorb[^2]
title: 'Conditional independence among max-stable laws'
---
[*Keywords*: [Conditional independence, exponent measure, Markov structure, max-stable random vector, M[ö]{}bius inversion]{}\
*2010 MSC*: [Primary 60G70;]{}\
]{}
Introduction {#sect:intro}
============
As pointed out by [@daw79] *independence* and *conditional independence* are key concepts in the theory of probability and statistical inference. A collection of (not necessarily real-valued) random variables $Y_{1},\dots,Y_{k}$ on some probability space $(\Omega,\mathcal{A},{\ensuremath{\mathbb{P}}})$ are called [conditionally independent]{} given the random variable $Z$ (on the same probability space) if $$\begin{aligned}
{\ensuremath{\mathbb{P}}}(Y_1\in A_1,\dots,Y_k \in A_k \mid Z) = \prod_{i=1}^k
{\ensuremath{\mathbb{P}}}(Y_i \in A_i \mid Z) \qquad {\ensuremath{\mathbb{P}}}\text{-a.s.},\end{aligned}$$ for any measurable sets $A_1,\dots,A_k$ from the respective state spaces. The conditioning is meant with respect to the $\sigma$-algebra generated by $Z$. A particularly important example for the conditional independence to be an omnipresent attribute are the *Gaussian Markov random fields* that have evolved as a useful tool in spatial statistics [@ruehel05; @lau96]. Here, the zeroes of the *precision matrix* (the inverse of the covariance matrix) of a Gaussian random vector represent precisely the conditional independence of the respective components conditioned on the remaing components of the random vector. Hence, sparse precision matrices are desirable for statistical inference.
In the analysis of the extreme values of a distribution (rather than fluctuations around mean values) *max-stable* models have been frequently considered. We refer to [@engetal14; @navetal09; @bladav11; @buisetal08] for some spatial applications among many others. Their popularity originates from the fact that max-stable distributions arise precisely as possible limits of location-scale normalizations of i.i.d. random elements. A random vector $X$ is called [max-stable]{} if it satisfies the distributional equality $a_n X + b_n \stackrel{{\ensuremath{\mathcal{D}}}}{=}
\max(X^{(1)},\dots,X^{(n)})$ for independent copies $X^{(1)},\dots,X^{(n)}$ of $X$ for some appropriate normalizing sequences $a_n>0$ and $b_n \in {\ensuremath{\mathbb{R}}}$, where all operations are meant componentwise. If the components $X_i$ of $X$ are *standard Fr[é]{}chet* distributed, i.e. ${\ensuremath{\mathbb{P}}}(X_i \leq
x)=\exp(-1/x)$ for $x \in (0,\infty)$, we have $a_n=n$ and $b_n=0$ and the random vector $X$ will be called *simple max-stable*.
Let $I$ be a non-empty finite set. It is well-known (cf. e.g. [@res08]) that the distribution functions $G$ of simple max-stable random vectors $X=(X_i)_{i \in I}$ are in a one-to-one correspondence with Radon measures $H$ on some reference sphere $S_+=\{ \omega \in [0,\infty)^I : \lVert \omega \rVert = 1\}$ that satisfy the moment conditions $\int \omega_i H({\ensuremath{\text{\normalfont d}}}\omega)=1$, $i
\in I$. The correspondence between $G$ and $H$ is given by the relation $$\begin{aligned}
G(x)={\ensuremath{\mathbb{P}}}(X_i \leq x_i, \, i \in I) = \exp\left(- \int_{S_+}
\max_{i \in I} \frac{\omega_i}{x_i}~H({\ensuremath{\text{\normalfont d}}}\omega)\right), \qquad
x \in (0,\infty)^I.\end{aligned}$$ Here, $\lVert \cdot \rVert$ can be any norm on ${\ensuremath{\mathbb{R}}}^I$ and $H$ is often called *angular* or *spectral measure*.
In general, neither does independence imply conditional independence nor does conditional independence imply independence of the subvectors of a random vector. Consider the following two simple examples which illustrate this fact in the case of Gaussian random vectors (Example \[ex:gauss\]) and max-stable random vectors (Example \[ex:max\]). For notational convenience, we write if $X$ and $Y$ are independent and if $X$ and $Y$ are conditionally independent given $Z$ and likewise use the instructive notation and if more than two random elements are involved.
\[ex:gauss\] Let $X_1,X_2,X_3$ be three independent standard normal random variables and, moreover, $X_4=X_1+X_2$ and $X_5=X_1+X_2+X_3$. Then all subvectors of $(X_i)_{i=1}^5$ are Gaussian and $$\begin{aligned}
\label{eq:indep:example} &X_1 \ci X_2, && \hspace{-2cm} \text{but not} \qquad X_1 \ci X_2 \mid X_5,\\
\label{eq:cindep:example} \text{whereas} \qquad &X_1 \ci X_5 \mid
X_4, && \hspace{-2cm} \text{but not} \qquad X_1 \ci X_5.
\end{aligned}$$
\[ex:max\] Let $X_1,X_2,X_3$ be three independent standard Fr[é]{}chet random variables and, moreover, $X_4= \max(X_1, X_2)$ and $X_5= \max(X_1 ,X_2,X_3)$. Then all subvectors of $(X_i)_{i=1,\dots,5}$ are max-stable and both relations and hold true also in this setting.
However, if the distribution of a max-stable random vector has a [positive continuous density]{}, then conditional independence of any two subvectors conditioned on the remaining components implies already their independence. To be precise, when we say that a random vector has a *positive continuous density*, we mean that the joint distribution of its components has a positive continuous density. The following theorem is the main result of the present article. If $X=(X_i)_{i \in I}$ is a random vector, we write $X_A$ for the subvector $(X_i)_{i \in A}$ if $A \subset I$. The same convention applies to non-random vectors $x=(x_i)_{i \in I}$.
\[thm:CImaxstabledensity\] Let $X=(X_i)_{i \in I}$ be a simple max-stable random vector with positive continuous density. Then, for any disjoint non-empty subsets $A$ and $B$ of $I$, the conditional independence implies the independence .
A proof of this theorem is given in Section \[sec:proofs\]. Beforehand, some comments are in order.
\(a) First, the requirement of a positive continuous density for $X$ is much less restrictive than requiring the spectral measure $H$ of $X$ to admit such a density, [cf. ]{}[@beirletal04] pp.262-264 and references therein. For instance, fully independent variables $X=(X_i)_{i \in I}$ have a discrete spectral measure, while their density exists and is positive and continuous. A more subtle example is, for instance, the asymmetric logistic model [@tawn90], which admits a continuous positive density and whose spectral measure carries mass on all faces of $S_+$, [cf. ]{}also Example \[ex:examples\].
\(b) Secondly, both random vectors $(X_i)_{i=1,2,5}$ and $(X_i)_{i=1,4,5}$ that were considered in the Gaussian case in Example \[ex:gauss\] have a positive continuous density on ${\ensuremath{\mathbb{R}}}^d$. Hence, there exists no version for Theorem \[thm:CImaxstabledensity\] for the Gaussian case.
\(c) Note that the implication of Theorem \[thm:CImaxstabledensity\] is the independence of $X_A$ and $X_B$, not the independence of all three subvectors $X_A,X_B,X_{I \setminus (A \cup B)}$.
\(d) By means of the same argument that shows that pairwise independence of the components of a max-stable random vector implies already their joint independence, we may deduce a version of Theorem \[thm:CImaxstabledensity\], in which more than two subvectors are considered.
\[cor:CImaxstabledensity\] Let $X=(X_i)_{i \in I}$ be a simple max-stable random vector with positive continuous density. Then, for any disjoint non-empty subsets $A_1,\dots,A_k$ of $I$, the conditional independence implies the independence .
\(e) The non-degenerate univariate max-stable laws are classified up to location and scale by the one parameter family of extreme value distributions indexed by $\gamma \in {\ensuremath{\mathbb{R}}}$ $$\begin{aligned}
F_\gamma(x)=\exp(-(1+\gamma x)^{-1/\gamma}), \qquad x \in \left\{
\begin{array}{ll}
(-1/\gamma, \infty) & \gamma > 0,\\
{\ensuremath{\mathbb{R}}}& \gamma = 0,\\
(-\infty, -1/\gamma) & \gamma <0.
\end{array} \right.\end{aligned}$$ Any other (not necessarily simple) max-stable random vector is obtained through a transformation of the marginals that is differentiable and strictly monotone on the respective sub-domain on ${\ensuremath{\mathbb{R}}}^d$ (cf. e.g. [@res08] Prop. 5.10). Hence, the above results remain valid for the general class of max-stable random vectors. (f) [@domeyi14] show that, up to time reversal, only max-autoregressive processes of order one can appear as discrete time stationary max-stable processes that satisfy the first order Markov property. This result indicates already that the conditional independence assumption is to some extent unnatural in presence of the max-stability property.\
\[ex:examples\] Various classes of tractable max-stable distributions admit a positive continuous density, such that Theorem \[thm:CImaxstabledensity\] and Corollary \[cor:CImaxstabledensity\] apply. Popular models that are commonly used for statistical inference include the asymmetric logistic model [@tawn90], the asymmetric Dirichlet model [@ct91], the pairwise beta model [@cooletal10] and its generalizations involving continuous spectral densities [@ballschl11] in the multivariate case. Moreover, most marginal distributions of spatial models such as the Gaussian max-stable model [@smit90b; @gentetal11] or the Brown-Resnick model [@kabletal09; @hr89] possess a positive continuous density if the parameters are non-degenerate. Hence, if any of the components of the previously mentioned extreme value models exhibit conditional independence given any of the remaining components, they must be independent.
In the remaining article we subsume auxiliary arguments in Section \[sec:exponentmeasures\] and give all proofs in Section \[sec:proofs\].
Preparatory results on max-stable random vectors {#sec:exponentmeasures}
================================================
Throughout this section let $G$ be the distribution function of a simple max-stable random vector $X=(X_i)_{i \in I}$ that has a positive continuous density. We denote its exponent function by $$\begin{aligned}
V(x):= -\log G(x)= \int_{S_+} \max_{i \in I}
\left(\frac{\omega_i}{x_i}\right)\, H({\ensuremath{\text{\normalfont d}}}\omega), \qquad x \in
(0,\infty)^I.\end{aligned}$$ Lower order marginals $G^A$ that refer to a subset $A$ of $I$ are obtained as $\min_{i \in A^c}(x_i) \to \infty$, where $A^c=I\setminus A$. We write $x_{A^c} \to \infty$ for $\min_{i \in A^c}(x_i) \to \infty$, and with this notation $$\begin{aligned}
G^A(x_A):= \lim_{x_{A^c} \to \infty} G(x)
\quad \text{and} \quad
V^A(x_A):= -\log G^A(x_A).\end{aligned}$$ Since $G$ is absolutely continuous, the partial derivatives $$\begin{aligned}
G^A_B(x_A):= \frac{\partial^{|B|}}{\partial x_{B}} G^A(x_A)
\quad \text{and} \quad
V^A_B(x_A):= \frac{\partial^{|B|}}{\partial x_{B}} V^A(x_A)\end{aligned}$$ exist and are continuous for $B \subset A$, and the latter $V^A_B$ are homogeneous of order $-(|B|+1)$ [@ct91]. An elementary computation shows that $$\begin{aligned}
G^{A}_B(x_{A}) &= {W^{A}_B(x_{A})} \exp\left(-V^{A}(x_{A})\right), \end{aligned}$$ where $$\begin{aligned}
W^N_M(x_M)= \sum_{\pi \in \Pi(M)} (-1)^{|\pi|} \prod_{J \in \pi}
V^N_J(x_N), \end{aligned}$$ and $\Pi(M)$ stands for the set of partitions of $M$ for $M \subset N \subset I$.
Let us further denote the set of non-empty subsets of $I$ by ${\ensuremath{\mathcal{C}}}(I)$. The collection of exponent functions $(V^A)_{A \in
{\ensuremath{\mathcal{C}}}(I)}$ is in a one-to-one correspondence with its Möbius inversion $(d_A)_{A \in {\ensuremath{\mathcal{C}}}(I)}$, i.e., if we set $$\begin{aligned}
d_A(x):= \sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B} (-1)^{|B \cap
A|+1}V^B(x_B),
$$ it follows that $V^A$ can be recovered via $$\begin{aligned}
\label{eqn:moeb}
V^A(x_A)=\sum_{B \in {\ensuremath{\mathcal{C}}}(I) : B \cap A \neq \emptyset} d_B(x)\end{aligned}$$ (cf. [@paptawn14], Theorem 2 and [@schlathertawn_02], Theorem 4 or, more generally, [@berge71] Chapter 3, Section 2 for the Möbius inversion). Finally, we define $$\begin{aligned}
\chi_A(x_A)& := \lim_{x_{A^c} \to \infty}
d_A(x) =\sum_{B \in {\ensuremath{\mathcal{C}}}(I) : B \subset A} (-1)^{|B|+1}V^B(x_B) =\sum_{B
\in {\ensuremath{\mathcal{C}}}(I) : A \subset B} d_B(x).\end{aligned}$$ Then the collection of functions $(\chi_A)_{A \in {\ensuremath{\mathcal{C}}}(I)}$ is also in a one-to-one correspondence with $(V^A)_{A \in {\ensuremath{\mathcal{C}}}(I)}$ as well as $(d_A)_{A \in
{\ensuremath{\mathcal{C}}}(I)}$ and the inversions are given by $$\begin{aligned}
d_A(x)&= \sum_{B \in {\ensuremath{\mathcal{C}}}(I): A \subset B} (-1)^{|B \setminus A|} \chi_{B}(x_B),\\
V^A(x_A)&= \sum_{B \in {\ensuremath{\mathcal{C}}}(I) : B \subset A} (-1)^{|B|+1}
\chi_B(x_B).\end{aligned}$$ Further expressions for $V^A$, $d_A$ and $\chi_A$ are collected in Lemma \[lemma:d\_spectralmeasure\]. Note that $\chi_A (x_A) \geq d_A(x)$ and thus, $$\begin{aligned}
\label{eqn:zeroequivalence}
d_A = 0 \quad \Leftrightarrow \quad \chi_A = 0.\end{aligned}$$
\[lemma:d\_spectralmeasure\] The functions $V^A$ and $d_A$ and $\chi_A$ (with $A \in {\ensuremath{\mathcal{C}}}(I)$) can be expressed in terms of the spectral measure $H$ as follows: $$\begin{aligned}
V^A(x_A)& =\int_{S_+} \max_{i \in A} \left(\frac{\omega_i}{x_i}\right)\, H({\ensuremath{\text{\normalfont d}}}\omega),\\
d_A(x) &= \int_{S^+} \left[ \min_{i \in A}
\left(\frac{\omega_i}{x_i}\right) - \max_{j \in A^c}
\left(\frac{\omega_j}{x_j}\right) \right]_+
\, H({\ensuremath{\text{\normalfont d}}}\omega),\\
\chi_A(x_A) &= \int_{S^+} \min_{i \in A}
\left(\frac{\omega_i}{x_i}\right) \, H({\ensuremath{\text{\normalfont d}}}\omega).
\end{aligned}$$ Here $z_+=\max(0,z)$ and $\max(\emptyset)=0$.
It turns out that the following two quantities are closely linked to conditional independence and independence of subvectors of $X$, respectively. For non-empty disjoint subsets $A,B$ of $I$ and $C=I\setminus(A \cup B)$, we set for $x \in (0,\infty)^I$ $$\begin{aligned}
d_{A,B}(x) &:= V^{A \cup C}(x_{A \cup C}) + V^{B \cup C}(x_{B \cup C}) - V(x)-V^C(x_C)\\
&\,= \sum_{L \in {\ensuremath{\mathcal{C}}}(I): L \cap A \neq \emptyset, L \cap B \neq \emptyset, L \cap C = \emptyset} d_L(x),\\
\chi_{A,B}(x_{A \cup B}) &:= \lim_{x_C \to \infty} d_{A,B}(x) = V^{A}(x_{A}) + V^{B}(x_{B}) - V^{A\cup B}(x_{A \cup B}) \\
&\,= \sum_{L \in {\ensuremath{\mathcal{C}}}(I): L \cap A \neq \emptyset, L \cap B \neq \emptyset} d_L(x),\end{aligned}$$ where the last equalities follow from . The following lemma is an analogue of Lemma \[lemma:d\_spectralmeasure\].
\[lemma:d\_extended\_spectralmeasure\] The functions $d_{A,B}$ and $\chi_{A,B}$ can be expressed in terms of the spectral measure $H$ as follows $$\begin{aligned}
d_{A,B}(x) &= \int_{S^+} \left[ \min \left(\max_{i \in A}
\left(\frac{\omega_i}{x_i}\right), \max_{i \in B}
\left(\frac{\omega_i}{x_i}\right) \right) - \max_{j \in C}
\left(\frac{\omega_j}{x_j}\right) \right]_+
\, H({\ensuremath{\text{\normalfont d}}}\omega),\\
\chi_{A,B}(x_{A \cup B}) &= \int_{S^+} \min \left( \max_{i \in A}
\left(\frac{\omega_i}{x_i}\right), \max_{i \in B}
\left(\frac{\omega_i}{x_i}\right) \right) \, H({\ensuremath{\text{\normalfont d}}}\omega).
\end{aligned}$$
Note that $\chi_{A,B}(x_{A \cup B}) \geq d_{A,B}(x)$ implies similarly to (\[eqn:zeroequivalence\]) that $$\begin{aligned}
\label{eqn:extzeroequivalence}
d_{A,B} = 0 \quad \Leftrightarrow \quad \chi_{A,B} = 0.\end{aligned}$$
[General expressions for the regular conditional distributions for the distribution of a max-stable process conditioned on a finite number of sites that are based on hitting scenarios of Poisson point process representations have been computed in [@domeyi13; @oestschl14; @oest15] under mild regularity assumptions or in [@wansto11] for spectrally discrete max-stable random vectors.]{}
Let again $A$ and $B$ be non-empty disjoint subsets of $I$. Since we assumed a positive continuous density for $G$ (and hence also for its marginals), the numerators and denominators in $$\begin{aligned}
G(x_A|x_B)
:= \frac{G^{A \cup B}_B(x_{A \cup B})}{G^B_B(x_B)} = \exp\left(-\left[V^{A \cup B}(x_{A \cup B})-V^B(x_B) \right]
\right) \frac{W^{A \cup B}_B(x_{A \cup B})}{W^B_B(x_B)}\end{aligned}$$ are non-zero and continuous for $x \in (0,\infty)^I$ and the expression $G(x_A | x_B)$ constitutes a regular version of the conditional probability ${\ensuremath{\mathbb{P}}}(X_A \leq x_A | X_B = x_B)$.
\[prop:independence\] The functions $\chi_{A,B}$ and $d_{A,B}$ are connected with the independence and conditional independence of the respective subvectors of $X$ as follows.
a) $X_A \ci X_B \mid X_{I\setminus (A \cup B)}\quad \Rightarrow \quad d_{A,B}=0$.
b) $X_A \ci X_B \quad \Leftrightarrow \quad \chi_{A,B}=0$.
The assumption that $G$ admits a positive continuous density on $(0,\infty)^I$ is crucial for part a) to hold true. It fails in Example \[ex:max\].
Moreover, it is a simple consequence of [@ber61] and [@dehaan78] that the pairwise independence of any disjoint subvectors of the simple max-stable random vector $X$ implies already their joint independence.
\[lemma:pwindep\] If $X_{A_1},\dots,X_{A_k}$ are pairwise independent subvectors of a simple max-stable random vector $X$ (for necessarily disjoint $A_i
\subset I$), then they are jointly independent.
Proofs {#sec:proofs}
======
The first equation is clear from the definition of $V^A$. The relation for $d_A$ can be obtained as follows. $$\begin{aligned}
d_A(x)
&= \sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B} (-1)^{|B \cap A|+1}V^B(x_B)\\
&= \int_{S_+} \sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B} (-1)^{\lvert B \cap A\rvert +1} \max_{i \in B} \left(\frac{\omega_i}{x_i}\right)\, H({\ensuremath{\text{\normalfont d}}}\omega)\\
&= \int_{S^+}
\left[
\min_{i \in A} \left(\frac{\omega_i}{x_i}\right)
- \max_{j \in A^c} \left(\frac{\omega_j}{x_j}\right)
\right]_+
\, H({\ensuremath{\text{\normalfont d}}}\omega).\end{aligned}$$ In order to obtain the last equality, we denote $a_i=\omega_i/x_i$ and distinguish two cases:\
1st case: $A=I$. Then $$\begin{aligned}
\sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B} (-1)^{\lvert B \cap A\rvert +1} \max_{i \in B} \left(a_i\right)
= \sum_{B \in {\ensuremath{\mathcal{C}}}(I)} (-1)^{\lvert B\rvert +1} \max_{i \in B} \left(a_i\right) = \min_{i \in I} \left(a_i\right).\end{aligned}$$ 2nd case: $A \neq I$. Then set $b:=\max_{i \in A^c} a_i$ and $c_i:=\max(a_i,b)$, such that $$\begin{aligned}
&\sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B} (-1)^{\lvert B \cap A\rvert +1} \max_{i \in B} \left(a_i\right)
= \sum_{B \in {\ensuremath{\mathcal{C}}}(I) : A^c \subset B, B \neq A^c} (-1)^{\lvert B \cap A\rvert +1} \max_{i \in B \cap A} \left(c_i\right) - b\\
&= \sum_{U \subset A : U\neq \emptyset} (-1)^{\lvert U \rvert +1} \max_{i \in U} \left(c_i\right) - b
= \min_{i \in A} \left(c_i\right) - b\\
&= \min_{i \in A} \left(\max(a_i,b)\right) - b
= \max\left(\min_{i \in A}(a_i),b\right) - b
= \left(\min_{i \in A}(a_i) - b\right)_+ .\end{aligned}$$ The expression for $\chi_A$ follows immediately.
Similar to the proof of Lemma \[lemma:d\_spectralmeasure\], the relation for $d_{A,B}$ follows from $$\begin{aligned}
d_{A,B}(x) &= V^{A \cup C}(x_{A \cup C}) + V^{B \cup C}(x_{B \cup C}) - V(x)-V^C(x_C)\\
&= \int_{S_+}
\max_{i \in A\cup C} \left(\frac{\omega_i}{x_i}\right)
+ \max_{i \in B\cup C} \left(\frac{\omega_i}{x_i}\right)
- \max_{i \in I} \left(\frac{\omega_i}{x_i}\right)
- \max_{i \in C} \left(\frac{\omega_i}{x_i}\right)
\, H({\ensuremath{\text{\normalfont d}}}\omega)\\
&= \int_{S^+} \left[ \min \left(\max_{i \in A}
\left(\frac{\omega_i}{x_i}\right), \max_{i \in B}
\left(\frac{\omega_i}{x_i}\right) \right) - \max_{j \in C}
\left(\frac{\omega_j}{x_j}\right) \right]_+
\, H({\ensuremath{\text{\normalfont d}}}\omega),\end{aligned}$$ where the last equality is obtained from $$\begin{aligned}
&\max_{i \in A\cup C} \left(a_i\right)
+ \max_{i \in B\cup C} \left(a_i\right)
- \max_{i \in I} \left(a_i\right)
- \max_{i \in C} \left(a_i\right) \\
&=
\min\left(\max_{i \in A\cup C} \left(a_i\right),
\max_{i \in B\cup C} \left(a_i\right)\right)
- \max_{i \in C} \left(a_i\right) \\
&= \max\left [\min\left(\max_{i \in A} \left(a_i\right),
\max_{i \in B} \left(a_i\right)\right) , \max_{i \in C}\left(a_i\right)\right]
- \max_{i \in C} \left(a_i\right) \\
&= \left [\min\left(\max_{i \in A} \left(a_i\right),
\max_{i \in B} \left(a_i\right)\right) - \max_{i \in C}\left(a_i\right)\right]_+\end{aligned}$$ if we denote $a_i=\omega_i/x_i$. The expression for $\chi_{A,B}$ follows immediately.
\[Proof of Proposition \[prop:independence\]\]
a) As before, let $C=I\setminus (A \cup B)$. Since $G(x)=\exp(-V(x))$ has a positive continuous density, we have that the conditional independence for $C=I \setminus (A \cup B)$ implies that for all $x \in (0,\infty)^I$ $$\begin{aligned}
G(x_A | X_C) \, G(x_B | X_C) = G(x_{A \cup B} | X_C) \qquad {\ensuremath{\mathbb{P}}}\text{-a.s.} \,.\end{aligned}$$ Since $X_C$ has a positive continuous density with respect to the Lebesgue-measure on $(0,\infty)^C$, it follows that $$\begin{aligned}
G(x_A | x_C) \, G(x_B | x_C) = G(x_{A \cup B} | x_C) \qquad \text{for all } x \in Q,\end{aligned}$$ where $Q$ is a dense subset of $(0,\infty)^I$. By the continuity of these expressions in $x \in (0,\infty)^I$, the equality holds for all $x \in (0,\infty)^I$ and is equivalent to $$\begin{aligned}
\label{eqn:proofCImaxstabledensity}
\exp\left(d_{A,B}(x) \right)
= \frac{W^{A \cup C}_C(x_{A \cup C}) W^{B \cup C}_C(x_{B \cup C})}{W^{A \cup B \cup C}_C(x_{A \cup B \cup C}) W^C_C(x_C)}, \qquad x \in (0,\infty)^I.\end{aligned}$$ Here, $d_{A,B} \geq 0$ and $d_{A,B}$ is homogeneous of order $-1$, while the components $V^N_J$ that build the terms $W^N_M$ are homogeneous of order $-(|J|+1)$. Now, replacing $x$ by $t^{-1} x$ for $t > 0$ in (\[eqn:proofCImaxstabledensity\]), we see that the left-hand side grows exponentially in the variable $t$ as $t$ tends to $\infty$ if $d_{A,B}(x)>0$, while the right-hand side exhibits at most polynomial growth. Therefore, $d_{A,B}(x) = 0$ for $x \in (0,\infty)^I$.
b) Both sides are equivalent to $G^A(x_A) G^B(x_B) = G^{A \cup B}(x_{A \cup B})$ for all $x \in (0,\infty)^I$.
The hypothesis follows from Proposition \[prop:independence\] and .
It suffices to show that for $x_{A_i} \in (0,\infty)^{A_i}$, $i=1,\dots,k$ and $r \in (0,\infty)$ $$\begin{aligned}
{\ensuremath{\mathbb{P}}}\left(X_{A_1} \leq x_{A_1},\dots,X_{A_k}\leq x_{A_k}\right)
= \prod_{i=1}^k{\ensuremath{\mathbb{P}}}\left(X_{A_i} \leq x_{A_i}\right).\end{aligned}$$ Using the notation $r_i=\sum_{j_i \in A_i} x^{-1}_{j_i}$, $u_{j_i}=(r_ix_{j_i})^{-1}$ for $j_i \in A_i$ and $Y_{i}=\max_{j_i \in A_i} u_{j_i} X_{j_i}$, $i=1,\dots,k$, we can rewrite this equality in the form $$\begin{aligned}
{\ensuremath{\mathbb{P}}}\left(Y_1 \leq r_1^{-1},\dots, Y_k \leq r_k^{-1} \right)
= \prod_{i=1}^k{\ensuremath{\mathbb{P}}}\left(Y_i \leq r_i^{-1}\right),\end{aligned}$$ where the random vector $(Y_1,\dots,Y_k)$ is simple max-stable [@dehaan78] and has pairwise independent components due to our assumptions. Hence, by [@ber61] Theorem 2, the $Y_i$ are jointly independent, which entails the relation above.
implies for $i_1 \neq i_2$ and hence by Theorem \[thm:CImaxstabledensity\]. The hypothesis follows if we apply Lemma \[lemma:pwindep\] to the $X_{A_i}$, $i=1,\dots,k$.
#### Acknowledgments
The authors would like to thank an anonymous referee for helpful comments improving the clarity of the paper. IP acknowledges funding from the SuSTaIn program - Engineering and Physical Sciences Research Council grant EP/D063485/1 - at the School of Mathematics of the University of Bristol. Part of this paper was written during a visit of KS at the University of Bristol funded by the SuSTaIn program and during a visit of IP at the University of Mannheim funded by the German Research Foundation DFG through the RTG 1953. IP and KS thank them for their generous hospitality.
[xx]{}
Ballani, F. Schlather, M. 2011 , ‘A construction principle for multivariate extreme value distributions’, [*Biometrika*]{} [**98**]{}(3), 633–645.
Beirlant, J., Goegebeur, Y., Teugels, J. Segers, J. 2004, [*Statistics of [E]{}xtremes*]{}, Wiley Series in Probability and Statistics, John Wiley & Sons, Ltd., Chichester.
Berge, C. 1971, [*Principles of combinatorics*]{}, Translated from the French. Mathematics in Science and Engineering, Vol. 72, Academic Press, New York-London.
Berman, S. M. 1961/1962, ‘Convergence to bivariate limiting extreme value distributions’, [*Ann. Inst. Statist. Math.*]{} [**13**]{}, 217–223.
Blanchet, J. Davison, A. C. 2011, ‘Spatial modeling of extreme snow depth’, [*Ann. Appl. Stat.*]{} [**5**]{}(3), 1699–1725.
Buishand, T. A., de Haan, L. Zhou, C. 2008, ‘On spatial extremes: with application to a rainfall problem’, [*Ann. Appl. Stat.*]{} [**2**]{}(2), 624–642.
Coles, S. G. Tawn, J. A. 1991, ‘Modelling extreme multivariate events’, [*J. Roy. Statist. Soc. Ser. B*]{} [**53**]{}(2), 377–392.
Cooley, D., Davis, R. A. Naveau, P. 2010, ‘The pairwise beta distribution: a flexible parametric multivariate model for extremes’, [*J. Multivariate Anal.*]{} [ **101**]{}(9), 2103–2117.
Dawid, A. P. 1979, ‘Conditional independence in statistical theory’, [*J. Roy. Statist. Soc. Ser. B*]{} [ **41**]{}(1), 1–31.
de Haan, L. 1978, ‘A characterization of multidimensional extreme-value distributions’, [*Sankhyā Ser. A*]{} [ **40**]{}(1), 85–88.
Dombry, C. Eyi-Minko, F. 2013, ‘Regular conditional distributions of continuous max-infinitely divisible random fields’, [*Electron. J. Probab.*]{} [**18**]{}, no. 7, 21.
Dombry, C. Eyi-Minko, F. 2014, ‘Stationary max-stable processes with the [M]{}arkov property’, [*Stochastic Process. Appl.*]{} [**124**]{}(6), 2266–2279.
Engelke, S., Malinowski, A., Oesting, M. Schlather, M. 2014, ‘Statistical inference for max-stable processes by conditioning on extreme events’, [*Adv. in Appl. Probab.*]{} [**46**]{}(2), 478–495.
Genton, M. G., Ma, Y. Sang, H. 2011, ‘On the likelihood function of [G]{}aussian max-stable processes’, [*Biometrika*]{} [**98**]{}(2), 481–488.
H[ü]{}sler, J. Reiss, R.-D. 1989, ‘Maxima of normal random vectors: between independence and complete dependence’, [*Statist. Probab. Lett.*]{} [ **7**]{}(4), 283–286.
Kabluchko, Z., Schlather, M. de Haan, L. 2009, ‘Stationary max-stable fields associated to negative definite functions’, [*Ann. Probab.*]{} [**37**]{}(5), 2042–2065.
Lauritzen, S. L. 1996, [*Graphical models*]{}, Vol. 17 of [*Oxford Statistical Science Series*]{}, The Clarendon Press, Oxford University Press, New York. Oxford Science Publications.
Naveau, P., Guillou, A., Cooley, D. Diebolt, J. 2009, ‘Modelling pairwise dependence of maxima in space’, [*Biometrika*]{} [**96**]{}(1), 1–17.
Oesting, M. 2015, ‘On the distribution of a max-stable process conditional on max-linear functionals’, [*Statist. Probab. Lett.*]{} [**100**]{}, 158–163.
Oesting, M. Schlather, M. 2014 , ‘Conditional sampling for max-stable processes with a mixed moving maxima representation’, [*Extremes*]{} [**17**]{}(1), 157–192.
Papastathopoulos, I. Tawn, J. A. 2014, ‘Dependence properties of multivariate max-stable distributions’, [*J. Multivariate Anal.*]{} [**130**]{}, 134–140.
Resnick, S. I. 2008, [*Extreme values, regular variation and point processes*]{}, Springer Series in Operations Research and Financial Engineering, Springer, New York. Reprint of the 1987 original.
Rue, H. Held, L. 2005, [ *Gaussian [M]{}arkov [R]{}andom [F]{}ields*]{}, Vol. 104 of [*Monographs on Statistics and Applied Probability*]{}, Chapman & Hall/CRC, Boca Raton, FL. Theory and applications.
Schlather, M. Tawn, J. 2002, ‘Inequalities for the extremal coefficients of multivariate extreme value distributions’, [*Extremes*]{} [**5**]{}(1), 87–102.
Smith, R. L. 1990, Max-stable processes and spatial extremes, Technical report, University of North Carolina.
Tawn, J. A. 1990, ‘Modelling multivariate extreme value distributions’, [*Biometrika*]{} [**77**]{}, 245–53.
Wang, Y. Stoev, S. 2011, ‘Conditional sampling for spectrally discrete max-stable random fields’, [ *Adv. in Appl. Probab.*]{} [**43**]{}(2), 461–483.
[^1]: School of Mathematics and Maxwell Institute, University of Edinburgh, Edinburgh, EH9 3FD Email: [email protected]
[^2]: Institute of Mathematics, University of Mannheim, D-68131 Mannheim, Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that every real analytic action of a connected supersoluble Lie group on a compact surface with nonzero Euler characteristic has a fixed point. This implies that E. Lima’s fixed point free $C^{\infty}$ action on $S^2$ of the affine group of the line cannot be approximated by analytic actions. An example is given of an analytic, fixed point free action on $S^2$ of a solvable group that is not supersoluble.'
author:
- |
[**Morris W. Hirsch[^1] & Alan Weinstein[^2]** ]{}\
Department of Mathematics\
University of California at Berkeley
title: '**Fixed points of analytic actions of supersoluble Lie groups on compact surfaces**'
---
Introduction {#introduction .unnumbered}
============
Let $M$ denote a compact connected surface, with possibly empty boundary $\partial M$, endowed with a (real) analytic structure. $T_pM$ is the tangent space to $M$ at $p\in M$. The Euler characteristic of $M$ is denoted by $\chi (M)$.
Let $G$ be a Lie group with Lie algebra ${\ensuremath{{\cal L}}} (G)={\ensuremath{{\cal G}}}$; all groups are assumed connected unless the contrary is indicated. An [*action*]{} of $G$ on $M$ is a homomorphism $\alpha$ from $G$ to the group $\mathsf{H} (M)$ of homeomorphisms of $M$ such that the evaluation map $$\mathsf{ev}^\alpha=\mathsf{ev}{\colon\thinspace}G\times M\to M,\;(g,x)\to\alpha (g)(x)$$ is continuous. We usually suppress notation for $\alpha$, denoting $\alpha (g)(x)$ by $g(x)$. The action is called $C^r,\:r\in
\{1,2\dots;\omega\}$ if $\mathsf{ev}$ is a $C^r$ map, where $C^\omega$ means analytic.
The set ${\ensuremath{{\cal A}}} (G, M)$ of actions of $G$ on $M$ is embedded in the space of continuous maps $G\times M\to M$ by the correspondence $\alpha\mapsto\mathsf{ev}^\alpha$. We endow ${\ensuremath{{\cal A}}} (G, M)$ with the topology of uniform convergence on compact sets.
A point $p\in M$ is a [*fixed point*]{} for an action $\alpha$ of $G$ if $\alpha(g)(p)=p$ for all $g\in G$. The set of fixed points is denoted by ${\ensuremath{\mathsf {Fix}\,(G)}}$ or ${\ensuremath{\mathsf {Fix}\,(\alpha(G))}}$.
In this paper we consider the problem of finding conditions on solvable group actions that guarantee existence of a fixed point.
When $\chi(M)\ne 0$, every flow (action of the real line ${\ensuremath{{\bf R}}}$) on $M$ has a fixed point; this was known to Poincaré for flows generated by vector fields, and for continuous actions it is a well known consequence of Lefschetz’s fixed point theorem. E. Lima [@Li64] showed that every abelian group action on $M$ has a fixed point, and J. Plante [@Pl86] extended this to nilpotent groups.
These results do not extend to solvable groups: Lima [@Li64] constructed a fixed point free action on the 2-sphere of the solvable group $A$ of homeomorphisms of ${\ensuremath{{\bf R}}}$ having the form $x\mapsto ax+b,
\;a>0, b\in{\ensuremath{{\bf R}}}$; and Plante [@Pl86] constructs fixed point free action of $A$ on all compact surfaces. These actions are not known to be analytic; but Example \[th:ex1\] below describes a fixed point free, analytic action of a 3-dimensional solvable group on $S^2$.
Recall that $G$ is [*supersoluble*]{} if every element of ${\ensuremath{{\cal G}}}$ belongs to a codimension one subalgebra (see Barnes [@Ba67]). Our main result is the following theorem:
Let $G$ be a connected supersoluble Lie group and $M$ a compact surface $M$ such that $\chi (M)\ne 0$. Then every analytic action of $G$ on $M$ has a fixed point.
Since the group $A$ described above is supersoluble, Lima’s $C^\infty$ action cannot be improved to a fixed point free analytic action. The following result shows it cannot be approximated by analytic actions:
Let $G$ and $M$ be as in Theorem \[th:main\]. If $\alpha \in {\ensuremath{{\cal A}}}
(G, M)$ has no fixed point, then $\alpha$ has a neighborhood in ${\ensuremath{{\cal A}}}
(G, M)$ containing no analytic action.
By Theorem \[th:main\] and compactness of $M$, it suffices to prove the following: For all convergent sequences $\beta_n\to \beta$ in ${\ensuremath{{\cal A}}}
(G,M)$ and $p_n \to p$ in $M$, with $p_n\in {\ensuremath{\mathsf {Fix}\,(\beta_n(G))}}$, we have $p\in{\ensuremath{\mathsf {Fix}\,(\beta (G))}}$. Being a connected locally compact group, $G$ is generated by a compact neighborbood $K$ of the identity. Then $\beta_n (g)\to \beta (g)$ uniformly for $g\in K$, so $\beta (g)
(p)=p$ for all $g\in K$. Since $K$ generates $G$, this implies that $p\in{\ensuremath{\mathsf {Fix}\,(\beta (G))}}$.
In Theorem \[th:main\], the hypothesis that $G$ is connected is essential: the abelian group of rotations of $S^2$ generated by reflections in the three coordinate axes is a well known counterexample. And every Lie group with a nontrivial homomorphism to the group of integers acts analytically without fixed point on every compact surface admitting a fixed point free homeomorphism, thus on every surface except the disk and the projective plane.
The following example shows that supersolubility is essential:
Let $Q$ be the 3-dimensional Lie group obtained as the semidirect product of the real numbers ${\ensuremath{{\bf R}}}$ acting on the complex numbers ${\ensuremath{{\bf C}}}$ by $t\cdot z=e^{it}z$; this group is solvable but not supersoluble. Identify $Q$ with the space ${\ensuremath{{\bf R}}}\times {\ensuremath{{\bf C}}}\approx{\ensuremath{{\bf R}^{3}}}$ and note that left multiplication defines a linear action of $Q$ on ${\ensuremath{{\bf R}^{3}}}$. The induced action on the 2-sphere $S$ of oriented lines in ${\ensuremath{{\bf R}^{3}}}$ through the origin has no fixed point, and $\chi
(S)=2$. Geometrically, one can see this as the universal cover of the proper euclidean motions of the plane, acting on two copies of the plane joined along a circle at infinity.
We thank F.-J. Turiel for pointing out a small error in an earlier version of our manuscript. He has also obtained some interesting results complementary to ours in [@tu-analytic].
Proof of Theorem \[th:main\] {#proof-of-theorem-thmain .unnumbered}
============================
We assume given an action $\alpha{\colon\thinspace}G\to \mathsf{H} (M)$. The [*orbit*]{} of $p\in M$ is $G(p)= \{ g(x) {\colon\thinspace}g \in G \}$. The [*isotropy group*]{} of $p\in M$ is the closed subgroup $I_p=\{g\in G{\colon\thinspace}\alpha(g)(p)=p\}$. The [*evaluation map*]{} $\mathsf{ev}_p{\colon\thinspace}G\to M$ at $p\in M$ is defined by $g\mapsto g(p)$.
Suppose that the action is $C^r, \:r\ge 1$. Then $\mathsf{ev}_p$ induces a bijective $C^r$ immersion $i_p{\colon\thinspace}G/I(p) \to G(p)$. The tangent space $E(p)\subset T_p M$ to this immersed manifold at $p$ is the image of $T_e G$ under the differential of $\mathsf{ev}_p$ at the identity $e\in G$.
For $j=0,1,2$, let $V_j=V_j(G)\subset M$ denote the union of the $j$-dimensional orbits. Then $M=V_2\cup V_1\cup V_0$. Each $V_j$ is invariant, $V_2$ is open, $V_1\cup V_0$ is compact, and $V_0={\ensuremath{\mathsf {Fix}\,(G)}}$.
Assume that $G$ is solvable and that $G(p)$ is a compact 1-dimensional orbit. Then there is a closed normal subgroup $H\subset G$ of codimension 1 such that every point of $G(p)$ has isotropy group $H$.
Choose a homeomorphism $f{\colon\thinspace}G(p)\approx S^1$ (the circle). Let $\beta{\colon\thinspace}G\to \mathsf{H}(S^1)$ be the action defined by $\beta(g) =
f\circ \alpha (g)\circ f^{-1}$. Because $G$ is solvable, by a result of Plante ([@Pl86], Theorem 1.2) there exists a homeomorphism $h$ of $S^1$ conjugating $\beta(G)$ to the rotation group $\mathrm{SO}(2)$. Since $\beta(G)$ is abelian and acts transitively on $S^1$, all points of $S^1$ have the same isotropy group for $\beta$; this isotropy group is the required $H$.
Analyticity is used to establish the following useful property:
Assume that $G$ acts analytically and that ${\ensuremath{\mathsf {Fix}\,(G)}}=\varnothing$. Then either $V_1=M$ and $\chi (M)=0$, or else $V_1$ is the (possibly empty) union of a finite family of orbits, each of which is a smooth Jordan curve contained in $\partial M$ or in $M\setminus \partial M$.
Since there are no orbits of dimension $0$, $V_1$ is a compact set comprising the points $p$ such that $\dim E_p \le 1$. It is easy to see that $V_1$ is a local analytic variety.
If $V_1=M$ then the map $p\mapsto E_p$ is a continuous field of tangent lines to $M$, tangent to $\partial M$ at boundary points. The existence of such a field implies that $\chi (M)= 0$.
Assume that $V_1\ne M$. Note that $\dim_p V_1\ge 1$ at each $p\in V_1$. Since $M$ is connected and $V_1$ is a variety, $V_1$ must have dimension $1$ at each point. The set of points where $V_1$ is not smooth is a compact, invariant 0-dimensional subvariety, i.e., a finite set of fixed points, hence empty. Since $V_1$ consists of 1-dimensional orbits, $V_1$ must be a compact, smooth invariant 1-manifold without boundary, i.e. each component of $V_1$ is a Jordan curve. Since $\partial M$ is the union of invariant Jordan curves, any component of $V_1$ that meets $\partial M$ is a component of $\partial M$.
In view of Lemma \[th:lem2\], it suffices to prove the following more general result:
Let $G$ be a connected supersoluble Lie group acting continuously on the compact connected surface $M$. Assume that
(a)
: there are no fixed points
(b)
: for each closed subgroup $H,\,$ $V_1(H)$ is the union (perhaps empty) of finitely many disjoint Jordan curves.
Then $\chi
(M)=0$.
By passing to a universal covering group we assume that $G$ is simply connected. This implies that every closed subgroup is simply connected (see Hochschild [@Ho65], Theorem XII.2.2.)
We proceed by induction on $\dim G$, the case $G={\ensuremath{{\bf R}}}$ having been covered in the introduction. Henceforth assume inductively that $\dim
G = n\ge 2$ and that the proposition holds for all supersoluble groups of lower dimension. With this hypothesis in force, we first rule out the case that $M$ is a disk:
If $M$ is as in Proposition \[th:main2\], then $\chi (M)\ne 1$
Suppose not; then $M$ is a closed 2-cell. Since there are no fixed points, $\partial M$ is an orbit, hence a component of $V_1$. Every component of $V_1$ bounds a unique 2-cell in $M$, and there are only finitely many such 2-cells. Let $D$ be one that contains no other. Then $D$ is invariant under $G$, and the action of $G$ on $D$ is fixed point free. Therefore we may assume that $M=D$, so that $V_1=\partial M$.
By Lemma \[th:plante\] there exists a closed normal subgroup $H$ of codimension one with $\partial M\subset {\ensuremath{\mathsf {Fix}\,(H)}}$. Let $R\subset G$ be a 1-parameter subgroup transverse to $H$ at the identity; then $RH=G$.
Because $G$ is supersoluble, there is a codimension one subalgebra ${\ensuremath{{\cal K}}}\subset {\ensuremath{{\cal G}}}$ containing the Lie algebra ${\ensuremath{{\cal R}}}$ of $R$. Because $G$ is simply connected and solvable ${\ensuremath{{\cal K}}}$ is the Lie algebra of a closed subgroup $K\subset G$ of dimension $n-1$, and $KH=G$. By the induction hypothesis there exists $p\in{\ensuremath{\mathsf {Fix}\,(K)}}$. Then $\dim G(p)
\le \dim G -\dim K =1$. Therefore $p\in V_1=\partial D$. We now have $p\in{\ensuremath{\mathsf {Fix}\,(K)}}\cap{\ensuremath{\mathsf {Fix}\,(H)}}={\ensuremath{\mathsf {Fix}\,(G)}}$, a contradiction.
We return now to the case of general $M$.
Denote the connected components of $M\setminus V_1$ by $U_i,\dots,U_r,\; r\ge 1$. Each $U_i$ is an open orbit, whose set theoretic boundary $\mathsf {bd}\,U_i$ is a (possibly empty) union of components of $V_1$. The closure ${\ensuremath{\overline{U}}}_i$ is a compact surface invariant under $G$, whose boundary as a surface is $\partial
U_i=\mathsf {bd}\,U_i$.
We show that $U_i$ is an open annulus. Let $H\subset G$ be the isotropy subgroup of $p\in U_i$. Evaluation at $p$ is a surjective fibre bundle projection $G\to U_i$ with standard fibre $H$. Therefore there is an exact sequence of homotopy groups $$\cdots\to\pi_j (G)\to\pi_j (U_i)\to \pi_{j-1}(H)\to \pi_{j-1} (G)\to\dots\to
\pi_0 (G)=\{0\}$$ ending with the trivial group $\pi_0 (G)$ of components of $G$. The component group $\pi_0(H)$ is solvable (see Raghunathan [@Ra72], Proposition III.3.10), so taking $j=1$ shows that $\pi_1 (U_i)$ is solvable. Therefore $U_i$ is a sphere, torus, open 2-cell, or open annulus. If $U_i$ is a torus then $U_i=M$, contradicting $\chi (M)\ne 0$. The sphere is ruled out by the exact sequence $\pi_2(G)\to\pi_2(U_i)\to\pi_1(H)$, because $\pi_2
(G)={0}$ for every Lie group and $\pi_1(H)={0}$. Proposition \[th:propD\] rules out the 2-cell.
It follows that ${\ensuremath{\overline{U}}}_i$ is a closed annulus, so $\chi ({\ensuremath{\overline{U}}}_i)=0$. By the additivity property $\chi(A\cup B) = \chi(A) + \chi(B) -
\chi(A\cap B)$ of the Euler characteristic, any space $M$ built by gluing annuli along their boundary circles must have $\chi(M)=0$.
[9999]{} D.W. Barnes, [*On the cohomology of soluble Lie algebras*]{}, Math. Z. [**101**]{} (1967), 343–349.
G. Hochschild, [*The structure of Lie groups*]{}, San Francisco, Holden-Day, 1965.
E. Lima, [*Commuting vector fields on $2$-manifolds*]{}, Bull. Amer. Math. Soc. [**69**]{} (1963), 366–368
E. Lima, [*Common singularities of commuting vector fields on $2$-manifolds*]{}, Comment. Math. Helv. [**39** ]{} (1964), 97–110
E. Lima, [*Commuting vector fields on $S^2$*]{}, Proc. Amer. Math. Soc. [**15**]{} (1964) 138–141
J. Plante, [*Fixed points of Lie group actions on surfaces*]{}, Ergod. Th. Dynam. Sys. [**6**]{} (1986), 149–161
M. Raghunathan, [*Discrete subgroups of Lie groups*]{}, Berlin, New York, Springer-Verlag, 1972.
F.-J. Turiel, [*Analytic actions on compact surfaces and fixed points*]{}, preprint, 2000.
[^1]: Partially supported by NSF grant DMS-9802182
[^2]: Partially supported by NSF Grant DMS-9971505
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The Wiener polarity index of a graph is defined as the number of unordered pairs of vertices at distance three. In recent years, this topological index was extensively studied since it has many known applications in chemistry and also in network theory. In this paper, we generalize the result of Behmaram, Yousefi-Azari, and Ashrafi proved in 2012 for calculating the Wiener polarity index of a graph. An important advantage of our generalization is that it can be used for graphs that contain $4$-cycles and also for graphs whose different cycles have more than one common edge. In addition, using the main result a closed formula for the Wiener polarity index is derived for phenylenes and recalculated for catacondensed benzenoid graphs. The catacondensed benzenoid graphs and phenylenes attaining the extremal values with respect to the Wiener polarity index are also characterized.'
author:
- Niko Tratnik
date: 'Received: / Accepted: date'
title: Formula for calculating the Wiener polarity index with applications to benzenoid graphs and phenylenes
---
[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 and preliminaries
==============================
Throughout the paper, all the considered graphs are simple, finite, and connected. For a graph $G$, $V(G)$ and $E(G)$ denote the sets of all the vertices and edges, respectively. The [*distance*]{} $d_G(x,y)$ between vertices $x$ and $y$ of a graph $G$ is the length of a shortest path between vertices $x$ and $y$ in $G$. We sometimes write $d(x,y)$ instead of $d_G(x,y)$. For a vertex $x$, we will denote by $N(x)$ the set of all the vertices in $G$ that are adjacent to $x$. The *degree* of $x$, denoted as $\deg_G(x)$ or shortly $\deg(x)$, is the cardinality of the set $N(x)$. For any $k \geq 3$, a cycle on $k$ vertices will be called a *$k$-cycle*. Moreover, the set of all $k$-cycles in $G$ is denoted as $C_k(G)$. A cycle or a path in graph $G$ will be often denoted as the sequence of its vertices. In addition, a *star graph* $S_k$, $k \geq 1$, is a complete bipartite graph $K_{1,k}$.
The Wiener polarity index is a molecular descriptor defined as the number of unordered pairs of vertices at distance three. More precisely, for a graph $G$, the *Wiener polarity index* of $G$, denoted by $W_p(G)$, is defined as $$W_p(G) = | \lbrace \lbrace u,v \rbrace \subseteq V(G) \, | \, d(u,v)=3 \rbrace |.$$ The Wiener polarity index was introduced in [@Wiener], where H.Wiener used this index and the classical Wiener index to predict boiling points of paraffins. Later, Lukovits and Linert [@luko] demonstrated quantitative structure-property relationships for the Wiener polarity index in a series of acyclic and cycle-containing hydrocarbons. Furthermore, Hosoya and Gao analysed very good correlation of this index with liquid density [@hosoya]. In the same work, remarkably good correlations with refractive index and molar volume are also mentioned. In recent years, the Wiener polarity index was mostly studied on trees [@ali; @ashrafi; @deng; @deng1; @lei; @liu2] and also on unicyclic graphs [@hou; @liu]. Moreover, the Nordhaus-Gaddum-type results for this index were considered [@hua; @zhang] and the index was used to study properties of networks [@lei1]. Also, in [@tratnik] a cut method for the Wiener polarity index was developed for nanotubes and benzenoid graphs that can be embedded into the regular hexagonal lattice. Some other recent investigations on the Wiener polarity index can be found in [@hua1; @ilic; @il-il].
The Zagreb indices have been introduced in 1972 by Gutman and Trinajsti' c [@gut_tri]. For a graph $G$ with at least one edge, the *first Zagreb index*, $M_1(G)$, and the *second Zagreb index*, $M_2(G)$, are defined as $$\begin{aligned}
M_1(G) & = & \sum_{v \in V(G)} \deg(v)^2, \\
M_2(G) & = & \sum_{e=uv \in E(G)} \deg(u) \cdot \deg(v).\end{aligned}$$ If $G$ has no edges, we set $M_1(G)=M_2(G)=0$. Since the introduction, many mathematical and chemical properties were found for these topological indices. For example, it is easy to check that $$\label{zag1}
M_1(G) = \sum_{e=uv \in E(G)} (\deg(u) + \deg(v)).$$
In [@beh] the Wiener polarity index was expressed by using the Zagreb indices and the following theorem was proved.
[@beh] \[behas\] Suppose $G$ is a connected triangle- and quadrangle-free graph such that its different cycles have at most one common edge. If $N_p = N_p(G)$ and $N_h = N_h(G)$ denote the number of pentagons and hexagons of $G$, then $W_p(G) = M_2(G) - M_1(G) -
5N_p - 3N_h + |E(G)|$.
However, many chemical graphs contain $4$-cycles and consequently, Theorem \[behas\] can not be used in such cases. Therefore, in Theorem \[glavni\] we generalize this result such that it can be used also for graphs that contain $4$-cycles. Moreover, in our generalization the number of common edges of different cycles is limited only for small cycles.
To show the application of the main result, in the final section we derive a closed formula for the Wiener polarity index of phenylenes, which are important and extensively studied molecular graphs that contain $4$-cycles. For catacondensed benzenoid graphs similar formula was obtained in [@beh], but it does not coincide with our computations. Therefore, we compute the Wiener polarity index also for these graphs. Moreover, the graphs attaining the extremal values with respect to the Wiener polarity index are characterized.
Formula for calculating the Wiener polarity index
=================================================
In this section we prove the main result of the paper and it enables us to easily compute the Wiener polarity index for a large class of graphs. We start with two definitions.
Let $G$ be a graph and $C$ a $4$-cycle in $G$. An edge $e$ **exits** $C$ if $e$ and $C$ have exactly one common vertex. See Figure \[exit\].
![\[exit\] A cycle $C$ and an edge $e$ such that $e$ exits $C$.](exit.eps)
Obviously, if $G$ does not contain $3$-cycles, then for a given $4$-cycle $C=u_1,u_2,u_3,u_4,u_1$ the number of edges exiting $C$ is exactly $\deg(u_1) + \deg(u_2)+ \deg(u_3)+ \deg(u_4) - 8$.
Let $G$ be a graph without 3-cycles. If $G$ has at least one $4$-cycle, we define
$$f(G) = \sum_{C=u_1,u_2,u_3,u_4,u_1 \in C_4(G)} (\deg(u_1) + \deg(u_2)+ \deg(u_3)+ \deg(u_4) - 8).$$ Otherwise, $f(G)=0.$
It is easy to notice that $f(G)$ is exactly the number of ordered pairs $(C,e)$, where $C$ is a cycle of $G$ and $e$ exits $C$. Now we are able to state the main result of the paper. In the proof, vertices $u,v$ of a 6-cycle $C$ are called *diametrically opposite* if $d_C(u,v)=3$.
\[glavni\] Let $G$ be a connected graph without $3$-cycles. Moreover, suppose that any two distinct cycles $C_1,C_2 \in C_4(G) \cup C_5(G) \cup C_6(G)$ have at most two common edges and any two cycles $C_1,C_2 \in C_4(G)$ have at most one common edge. Then it holds $$W_p(G) = M_2(G) - M_1(G) - f(G) - 4|C_4(G)| - 5|C_5(G)| - 3|C_6(G)| + |E(G)|.$$
If $G$ is a star graph or a graph with only one vertex, it can be easily checked that the theorem holds. Therefore, in the rest of the proof we assume that $G$ has more than one vertex and that $G$ is not a star graph. Let $e=xy \in E(G)$ be an edge with $\deg(x) > 1$ and $\deg(y)>1$. Moreover, let $u,v \in V(G) \setminus \lbrace x,y \rbrace$ be two distinct vertices such that $ux \in E(G)$ and $vy \in E(G)$. Obviously, $P=u,x,y,v$ is a path of length three in $G$, see Figure \[potP\].
![\[potP\] Path $P$.](potP.eps)
We define $$p_3(G) = \sum_{e=xy \in E(G)}(\deg(x)-1)(\deg(y)-1).$$
One can notice that $p_3(G)$ is exactly the number of all the paths of length three in $G$. By a simple calculation we obtain
$$\begin{aligned}
p_3(G) & = & \sum_{e=xy \in E(G)} \Big[ \deg(x)\deg(y) - (\deg(x) + \deg(y))+ 1) \Big] \\
& = & \sum_{e=xy \in E(G)}\deg(x)\deg(y) - \sum_{e=xy \in E(G)}(\deg(x) + \deg(y)) + \sum_{e=xy \in E(G)} 1 \\
& = & M_2(G) - M_1(G) + |E(G)|,\end{aligned}$$
where the last equality follows by Equation .
However, for two vertices $u,v \in V(G)$ there may exists more than one path of length three between them. Suppose that there is some other path (distinct from $P$) $P'=u,a,b,v$ of length three between $u$ and $v$. Consider the following options:
- $x \notin V(P')$ and $y \notin V(P')$:\
In this case, $u$ and $v$ lie on a $6$-cycle $x,u,a,b,v,y,x$, see Figure \[sita\]. Moreover, this is the only $6$-cycle in $G$ containing $u$ and $v$ as diametrically opposite vertices (otherwise we have two distinct $6$-cycles with three edges in common).
![\[sita\] Paths $P$ and $P'$ is case (a).](sita.eps)
- $x=a$ and $y \notin V(P')$:\
In this case, we obtain a $4$-cycle $C=x,b,v,y,x$ such that the edge $xu$ exits $C$, see Figure \[sitb\]. Moreover, $P'$ is the only such path (otherwise we obtain two distinct $4$-cycles with two edges in common).
![\[sitb\] Paths $P$ and $P'$ in case (b).](sitb.eps)
- $y=b$ and $x \notin V(P')$:\
In this case, we obtain a $4$-cycle $C=u,x,y,a,u$ such that the edge $yv$ exits $C$. Moreover, $P'$ is the only such path (otherwise we obtain two distinct $4$-cycles with two edges in common).
Next we show that cases (a) and (b) or (a) and (c) can not happen simultaneously. Suppose that there is a path $P'=u,x,b,v$ of length $3$ between $u$ and $v$ such that $y\neq b$ (case (b)). Moreover, let $P''=u,a',b',v$ be another path between $u$ and $v$ such that $x \notin V(P'')$ and $y \notin V(P'')$ (case (a)).
- If $b'=b$, then we obtain a $4$-cycle $x,y,v,b,x$ and a $6$-cycle $u,x,y,v,b,a',u$ with three common edges - a contradiction.
- If $b'\neq b$, then we obtain a $6$-cycle $u,a',b',v,y,x,u$ and a $6$-cycle $u,a',b',v,b,x',u$ with three common edges - a contradiction.
Since we get a contradiction in every case, (a) and (b) can not both happen. In a similar way we show that (a) and (c) can not both happen. However, cases (b) and (c) can both happen.
On the other hand, any $6$-cycle contains exactly three pairs of diametrically opposite vertices and every $4$-cycle $u_1,u_2,u_3,u_4,u_1$ has exactly $\deg(u_1) + \deg(u_2)+\deg(u_3)+\deg(u_4) - 8$ exiting edges. Therefore, $$p_3(G) - f(G) - 3|C_6(G)|$$ represents the number of unordered pairs of distinct vertices $u,v$ for which there exists a path of length three between them.
However, it can happen that $d(u,v) < 3$. In such a case we consider two possibilities:
- $d(u,v)=1$:\
In this case, $u$ and $v$ lie on a $4$-cycle $C=x,y,v,u,x$, see Figure \[siti\]. Obviously, it holds $d_C(u,v)=1$ and $C$ is the only $4$-cycle with such property (otherwise we obtain $4$-cycles with two common edges or a 4-cycle and a $6$-cycle with three common edges).
![\[siti\] $4$-cycle in case $(i)$.](siti.eps)
- $d(u,v)=2$:\
In this case, $u$ and $v$ lie on a $5$-cycle $C=x,y,v,a,u,x$ such that $a \in V(G) \setminus \lbrace x,y,u,v \rbrace$, see Figure \[sitii\]. Obviously, it holds $d_C(u,v)=2$ and $C$ it the only $5$-cycle with such property (otherwise we obtain 5-cycles with three common edges or a 5-cycle and a 6-cycle with three common edges).
![\[sitii\] $5$-cycle in case $(ii)$.](sitii.eps)
On the other hand, for every $4$-cycle there are exactly four pairs of vertices at distance one and for every $5$-cycle there are exactly five pairs of vertices at distance two. Therefore, $$p_3(G) - f(G) - 3|C_6(G)| - 4|C_4(G)| - 5|C_5(G)|$$ represents the number of unordered pairs of vertices at distance three, which is exactly the Wiener polarity index of $G$. This completes the proof.
Catacondensed benzenoid graphs and phenylenes
=============================================
In this section we apply the main result of the paper to find explicit formulas for the Wiener polarity index of catacondensed benzenoid graphs and phenylenes. First, we need to introduce some additional definitions and notation.
In the existing (both mathematical and chemical) literature, there is inconsistency in the terminology pertaining to (what we call here) “benzenoid graph". In this paper, a *benzenoid graph* is a 2-connected plane graph in which the boundary of any inner face is a 6-cycle (and any 6-cycle is the boundary of some inner face), such that two 6-cycles are either disjoint or have exactly one common edge, and no three 6-cycles share a common edge.
However, in some literature it is assumed that a benzenoid graph can be embedded into the regular hexagonal lattice [@gucy-89]. Let $G$ be a benzenoid graph. Any inner face of $G$ is called a *hexagon* of $G$ and the number of all the hexagons of $G$ is denoted by $h(G)$. A vertex shared by three hexagons of $G$ is called an *internal* vertex of $G$. The number of all internal vertices of $G$ will be denoted by $n_i(G)$. A benzenoid graph is said to be *catacondensed* if it does not possess internal vertices. Otherwise it is called *pericondensed*.
We say that two faces of a plane graph are *adjacent* if they have at least one edge in common. The *dualist graph* of a given benzenoid graph $G$ consists of vertices corresponding to hexagons of $G$; two vertices are adjacent if and only if the corresponding hexagons are adjacent. Obviously, the dualist graph of $G$ is a tree if and only if $G$ is catacondensed. For a catacondensed benzenoid graph $G$, its dualistic tree has $h(G)$ vertices and none of its vertices have degree greater than three, see Figure \[benzenoid\].
![\[benzenoid\] A catacondensed benzenoid graph $G$ with its dualistic tree.](benzenoid.eps)
Let $G$ be a catacondensed benzenoid graph with at least two hexagons. If we add *quadrilaterals* (faces whose boundary is a $4$-cycle) between all pairs of adjacent hexagons of $G$, the obtained graph $G'$ is called a *phenylene*, see Figure \[phenylene\]. We then say that $G$ is the *hexagonal squeeze* of $G'$. Obviously, graph $G$ in Figure \[benzenoid\] is the hexagonal squeeze of phenylene $G'$ from Figure \[phenylene\]. For a phenylene $G'$ the *dualist tree* of $G'$ is defined as the dualist tree of its hexagonal squeeze. Moreover, a *hexagon* of $G'$ is any inner face of $G'$ whose boundary is a $6$-cycle.
![\[phenylene\] A phenylene $G'$.](phenylene.eps)
Let $G$ be a catacondensed benzenoid graph or a phenylene and let $h_0$ be a hexagon of $G$. Hexagon $h_0$ is called *terminal* if it is adjacent to exactly one other inner face of $G$ and it is called *branched* if it is adjacent to exactly three other inner faces of $G$. If $h_0$ is adjacent to exactly two other inner faces, it has two vertices of degree two. Then $h_0$ is called *angular* if these two vertices are adjacent and *linear* it they are not adjacent. The number of terminal, branched, angular, and linear hexagons of $G$ will be denoted by $t(G)$, $b(G)$, $a(G)$, and $l(G)$, respectively. Moreover, a catacondensed benzenoid graph or a phenylene is called a *linear chain* if it does not contain angular or branched hexagons. A *segment* in a catacondensed benzenoid graph or a phenylene $G$ is any maximal linear chain which is a subgraph of $G$. Finally, the number of all the segments of $G$ will be denoted by $s(G)$. Obviously, for the graph from Figure \[benzenoid\] it holds $t(G)=3$, $b(G)=a(G)=l(G)=1$, $s(G)= 4$ and the same is true for the graph from Figure \[phenylene\].
For any $k \geq 0$, the number of vertices of a graph $G$ which have degree $k$ will be denoted by $n_k(G)$. We can now state the following proposition.
[@gut] \[gut-ben\] If $G$ is a benzenoid graph, then $|V(G)| = 4h(G) + 2 - n_i(G)$, $ |E(G)| = 5h(G) + 1 - n_i(G)$, $n_3(G) = 2h(G) - 2$, and $n_2(G) = 2h(G) + 4 - n_i(G)$.
In [@gut] Proposition \[gut-ben\] was proved just for benzenoid graphs that can be embedded into the regular hexagonal lattice. However, the same proof works for any benzenoid graph.
Using Proposition \[gut-ben\] it is possible to derive similar formulas also for phenylenes.
\[phen\] If $G$ is a phenylene, then $|V(G)| = 6h(G)$, $|E(G)| = 8h(G) - 2$, $n_3(G) = 4h(G) - 4$, and $n_2(G) = 2h(G) + 4$.
Let $G'$ be the hexagonal squeeze of $G$. Since $h(G)=h(G')$ and $G$ has exactly $h(G)-1$ quadrilaterals, $|V(G)| = |V(G')| + 2(h(G)-1)$ and $|E(G)|= |E(G')| + 3(h(G)-1)$. Moreover, $G$ and $G'$ have the same number of vertices of degree two. Therefore, the result follows by Proposition \[gut-ben\].
In the next lemma, we derive relations between the numbers of different types of hexagons and the number of segments.
\[pomoc\] If $G$ is a catacondensed benzenoid graph or a phenylene, then it holds $2b(G) + a(G) = s(G) -1$ and $b(G) + 2 = t(G)$.
Let $T$ be a dualistic tree of $G$. For any $v \in V(T)$ with $\deg(v)=2$ and $N(v)=\lbrace a, b \rbrace$, let $T_v$ be a tree obtained from $T$ by removing $v$, $va$, $vb$ and adding an edge $ab$. In such a case, we say that $T_v$ is obtained from $T$ by *skipping* vertex $v$. Moreover, let $T'$ be a tree obtained from $T$ by successively skipping all the vertices corresponding to the linear hexagons of $G$. Obviously, $|V(T')| = b(G) + a(G) + t(G)$ and $|E(T')| = s(G)$. Since $T'$ is a tree, we obtain $$\label{en1}
b(G) + a(G) + t(G) = s(G) +1.$$ Moreover, since the sum of all the degrees of vertices in $T'$ is equal to $2|E(T')|$ (by the handshaking lemma) one can conclude $$\label{en2}
3b(G) + 2a(G) + t(G) = 2s(G).$$
Finally, the results follow from and .
In the next lemma, formulas for the first Zagreb index are given.
\[firzag1\] For a catacondensed benzenoid graph $G_1$ and a phenylene $G_2$ it holds $M_1(G_1)= 26h(G_1)-2$ and $M_1(G_2)=44h(G_2)-20 $.
The proof for catacondensed benzenoid graph $G_1$ can be found in [@beh; @doslic]. By Proposition \[phen\], $G_2$ has $4h(G_2)-4$ vertices of degree three and $2h(G_2)+4$ vertices of degree two. Therefore, $$M_1(G_2) = 9(4h(G_2)-4) + 4(2h(G_2)+4) = 44h(G_2) - 20$$ and the proof is complete.
In the following proposition we obtain formulas for the second Zagreb index with respect to the number of hexagons, the number of segments, and the number of branched hexagons. Note that for catacondensed benzenoid graphs a similar formula was obtained in [@doslic]. In addition, for benzenoid graphs this index can be expressed also by using inlets (for the details see Proposition 5.5 in [@gut2]).
\[second\_zagreb\] If $G_1$ is a catacondensed benzenoid graph and $G_2$ a phenylene, then it holds $$\begin{aligned}
M_2(G_1) & = & 33h(G_1) + s(G_1) + b(G_1) - 10, \\
M_2(G_2) & = & 60h(G_2) + s(G_2) + b(G_2) - 37.\end{aligned}$$
In this proof, an edge $e=xy$ will be called an edge of *type 1* if $x$ and $y$ both have degree two. Moreover, if one of $x$ and $y$ has degree two and the other one has degree three, we say that $e$ is of *type 2*. Finally, $e$ is an edge of *type 3* if $x$ and $y$ both have degree three.
We notice that a terminal hexagon of $G_1$ or $G_2$ has three edges of type 1, two edges of type 2, and one edge of type 3. A linear hexagon has four edges of type 2 and two edges of type 3. An angular hexagon has one edge of type 1, two edges of type 2, and three edges of type 3. Finally, a branched hexagon has six edges of type 3.
If $G_1$ has only one hexagon, the result obviously holds. Otherwise, in a catacondensed benzenoid graph $G_1$ there are $h(G_1)-1$ edges of type 3 that appear on two different hexagons. Therefore, by counting contributions of hexagons and contributions of edges that belong to two hexagons, we get $$\begin{aligned}
M_2(G_1) & = & 24t(G_1) + 24l(G_1) + 25a(G_1) + 27b(G_1) + 9(h(G_1)-1) \\
& = & 24h(G_1) + a(G_1) + 3b(G_1) + 9h(G_1) - 9 \\
& = & 33h(G_1) + s(G_1) + b(G_1) - 10,\end{aligned}$$ where the last equality follows from Lemma \[pomoc\].
In a phenylene $G_2$ there are $4(h(G_1)-1)$ edges of type 3 that appear on a quadrilateral. Therefore, by counting contributions of hexagons and quadrilaterals we get $$\begin{aligned}
M_2(G_2) & = & 24t(G_2) + 24l(G_2) + 25a(G_2) + 27b(G_2) + 36(h(G_2)-1)
\\ & = & 60h(G_2) + s(G_2) + b(G_2) - 37,\end{aligned}$$ where the computation is done in a similar way as before. This completes the proof.
The following theorem is the main result of this section.
\[Wie\_pol\] If $G_1$ is a catacondensed benzenoid graph and $G_2$ a phenylene, then it holds $$\begin{aligned}
W_p(G_1) & = & 9h(G_1) + s(G_1) + b(G_1) - 7, \\
W_p(G_2) & = & 13h(G_2) + s(G_2) + b(G_2) - 11.\end{aligned}$$
Since $G_1$ does not contain $4$-cycles and $5$-cycles, we have $f(G_1)=|C_4(G_1)|=|C_5(G_1)|=0$ and $|C_6(G_1)| = h(G_1)$. On the other hand, for $G_2$ we have $|C_4(G_2)|=h(G_2)-1$, $|C_5(G_2)|=0$, $f(G_2) = 4h(G_2)-4$, and $|C_6(G_2)|=h(G_2)$. Also, it is easy to check that benzenoid graphs and phenylenes suffice the conditions of Theorem \[glavni\]. Therefore, the desired results follow by Theorem \[glavni\], Proposition \[gut-ben\], Proposition \[phen\], Lemma \[firzag1\], and Proposition \[second\_zagreb\].
To show one example, let $G$ be a benzenoid graph from Figure \[benzenoid\] and $G'$ a phenylene from Figure \[phenylene\]. As already noticed, $h(G)=h(G')=6$, $b(G)=b(G')=1$, and $s(G)=s(G')= 4$. Therefore, by Theorem \[Wie\_pol\] we compute $W_p(G) = 52$ and $W_p(G') = 72$.
In the rest of the section we will characterize the catacondensed benzenoid graphs and phenylenes with exactly $h$ hexagons for which the Wiener polarity index attains the minimum and the maximum value. Note that the catacondensed benzenoid graphs attaining the lower bound coincide with the graphs obtained in [@beh]. However, the result for the maximum value requires some additional insights.
If $G$ is a catacondensed benzenoid graph or a phenylene with exactly $h$ hexagons, we denote by $L_h$ the linear benzenoid chain or the linear phenylene chain, respectively, that also contains $h$ hexagons. Then $W_p(G) \geq W_p(L_h)$ and the equality holds if and only if $G \cong L_h$.
By Theorem \[Wie\_pol\], the Wiener polarity index attains its minimum value if and only if $s(G)=1$ and $b(G)=0$, which is true if and only if $G$ is a linear chain.
To find the graphs that attain the maximum value, we introduce special families of catacondensed benzenoid graphs and phenylenes. These families will be defined by the following extensions.
- **Extension 1:** *Let $G_0$ be a catacondensed benzenoid graph (or a phenylene) and let $h_0$ be an arbitrary terminal hexagon of $G_0$. We denote by $G$ a graph obtained from $G_0$ by attaching exactly two new hexagons to $h_0$ such that $h_0$ becomes a branched hexagon (and in the case of phenylenes, we also add two quadrilaterals between $h_0$ and the two new hexagons). See Figure \[dodas\_dva\].*
![\[dodas\_dva\] A catacondensed benzenoid graph (or a phenylene) $G$ obtained from $G_0$ by Extension 1.](dodas_dva.eps)
- **Extension 2:** *Let $G_0$ be a catacondensed benzenoid graph (or a phenylene) and let $h_0$ be an arbitrary terminal hexagon of $G_0$. We denote by $G$ a graph obtained from $G_0$ by attaching exactly one new hexagon to $h_0$ such that $h_0$ becomes an angular hexagon (and in the case of phenylenes, we also add a quadrilateral between $h_0$ and the new hexagon). See Figure \[dodas\_enega\].*
![\[dodas\_enega\] A catacondensed benzenoid graph (or a phenylene) $G$ obtained from $G_0$ by Extension 2.](dodas_enega.eps)
For any $h \geq 2$, we define families ${\mathcal{B}}_h$ and ${\mathcal{P}}_h$ as follows.
- If $h$ is an even number, then ${\mathcal{B}}_h$ (or ${\mathcal{P}}_h$) is the set of all the graphs that can be obtained from the catacondensed benzenoid graph with exactly two hexagons (or from the phenylene with exactly two hexagons) by performing Extension 1 exactly $\left(\frac{h}{2} - 1\right)$ times.
- If $h$ is an odd number, then ${\mathcal{B}}_h$ (or ${\mathcal{P}}_h$) is the set of all the graphs that can be obtained from the catacondensed benzenoid graph with exactly two hexagons (or from the phenylene with exactly two hexagons) by performing Extension 1 exactly $\left(\frac{h-1}{2}-1\right)$ times and performing Extension 2 exactly once (in an arbitrary order).
For an example, Figure \[graf\_iz\_druzine\] shows a catacondensed benzenoid graph with nine hexagons that belongs to ${\mathcal{B}}_9$. Obviously, all the graphs in ${\mathcal{B}}_h \cup {\mathcal{P}}_h$, $h\geq 2$, have exactly $h$ hexagons.
![\[graf\_iz\_druzine\] A graph from ${\mathcal{B}}_9$.](graf_iz_druzine.eps)
The next lemma gives an upper bound on the number of segments.
\[st\_segmentov\] For a catacondensed benzenoid graph or a phenylene $G$ with $h(G) \geq 2$ it holds $s(G) \leq h(G)-1$. Moreover, for all the graphs in ${\mathcal{B}}_h \cup{\mathcal{P}}_h$, $h \geq 2$, the equality holds.
If $G$ has exactly two hexagons, then $s(G)=1$ and the statement holds. Whenever we add exactly one hexagon to a graph, the number of segments stays the same or increases by one. Therefore, by induction we obtain $s(G) \leq h(G)-1$ for any $G$.
From the definition of the families ${\mathcal{B}}_h$ and ${\mathcal{P}}_h$ it follows that whenever we add two hexagons to a smaller graph, the number of segments increases by two, and whenever we add one hexagon, the number of segments increases by one. Therefore, the equality holds for these graphs.
Finally, everything is prepared for the following theorem.
If $G$ is a catacondensed benzenoid graph (or a phenylene) with $h$ hexagons, $h \geq 2$, and if $G'$ is an arbitrary graph from ${\mathcal{B}}_h$ (or ${\mathcal{P}}_h$), then $W_p(G) \leq W_p(G')$ and the equality holds if and only if $G \in {\mathcal{B}}_h$ (or $ G \in {\mathcal{P}}_h$).
By Theorem \[Wie\_pol\], the maximum value of the Wiener polarity index is attained by a graph $G$ which attains also the maximum value with respect to the number $s(G) + b(G)$.
By the construction, the graphs in ${\mathcal{B}}_h \cup{\mathcal{P}}_h$ obviously have the maximum number of branched hexagons among all the catacondensed benzenoid graphs or phenylenes with $h$ hexagons. If $h$ is even, these graphs are the only graphs attaining the upper bound with respect to the number of branched hexagons. If $h$ is odd, the upper bound with respect to the number of branched hexagons is attained also by some other graphs. To describe them, we first introduce Extension 3.
**Extension 3:** *Let $G_0$ be a catacondensed benzenoid graph (or a phenylene) and let $h_0$ be an arbitrary terminal hexagon of $G_0$. We denote by $G$ a graph obtained from $G_0$ by attaching exactly one new hexagon to $h_0$ such that $h_0$ becomes a linear hexagon (and in the case of phenylenes, we also add a quadrilateral between $h_0$ and the new hexagon). See Figure \[dodas\_ravno\].*
![\[dodas\_ravno\] A catacondensed benzenoid graph (or a phenylene) $G$ obtained from $G_0$ by Extension 3.](dodas_ravno.eps)
If $h$ is an odd number, then ${\mathcal{B}}_h'$ (or ${\mathcal{P}}_h'$) is the set of all the graphs that can be obtained from the catacondensed benzenoid graph with exactly two hexagons (or from the phenylene with exactly two hexagons) by performing Extension 1 exactly $\left(\frac{h-1}{2}-1\right)$ times and performing Extension 3 exactly once (in an arbitrary order). It is clear that for any odd $h$ the upper bound with respect to the number of branched hexagons is attained exactly by the graphs in ${\mathcal{B}}_h \cup{\mathcal{B}}_h'$ (or ${\mathcal{P}}_h \cup{\mathcal{P}}_h'$).
Moreover, by Lemma \[st\_segmentov\] the graphs in ${\mathcal{B}}_h \cup{\mathcal{P}}_h$ also have the maximum possible number of segments. On the other hand, by applying the same reasoning as in the proof of Lemma \[st\_segmentov\], we deduce that the graphs in ${\mathcal{B}}_{h}'$ (or ${\mathcal{P}}_{h}'$) do not attain the upper bound with respect to the number of segments since the number of segments does not increase when Extension 3 is performed. Therefore, it is clear that $G$ attains the maximum value with respect to the Wiener polarity index if and only if $G \in {\mathcal{B}}_h$ (or $ G \in {\mathcal{P}}_h$).
Concluding remarks
==================
In the paper we have generalized a formula for calculating the Wiener polarity index of a graph. The obtained result can be used for many (molecular) graphs since the conditions in the main theorem do not prohibit $4$-cycles in a considered graph. As an example of the main result, the closed formulas for the Wiener polarity index are calculated for phenylenes and recalculated for catacondensed benzenoid graphs. Moreover, the closed formulas are used to characterize the graphs with the minimum and the maximum Wiener polarity index. Regarding the future work, it would be interesting to generalize the main result such that it could be used even for graphs sufficing weaker conditions. However, in such a case the formula would probably become much more complicated.
Acknowledgment {#acknowledgment .unnumbered}
==============
The author was financially supported by the Slovenian Research Agency (research core funding No. P1-0297 and J1-9109).
[99]{} A. Ali, Z. Du, M. Ali, A note on chemical trees with minimum Wiener polarity index. Appl. Math. Comput. 335 (2018) 231–236.
A. R. Ashrafi, A. Ghalavand, Ordering chemical trees by Wiener polarity index, Appl. Math. Comput. 313 (2017) 301–312.
A. Behmaram, H. Yousefi-Azari, A. R. Ashrafi, Wiener polarity index of fullerenes and hexagonal systems, Appl. Math. Lett. 25 (2012) 1510–1513.
H. Deng, H. Xiao, The maximum Wiener polarity index of trees with $k$ pendants, Appl. Math. Lett. 23 (2010) 710–715.
H. Deng, On the extremal Wiener polarity index of chemical trees, MATCH Commun. Math. Comput. Chem. 66 (2011) 305–314.
T. Došli' c, On discriminativity of Zagreb indices, Iranian J. Math. Chem. 3 (2012) 25–34.
I. Gutman, Hexagonal systems. A chemistry-motivated excursion to combinatorial chemistry, Teach. Math. 10 (2007) 1–10.
I. Gutman, Selected Theorems in Chemical Graph Theory, University of Kragujevac and Faculty of Science Kragujevac, Kragujevac, 2017.
I. Gutman, S. J. Cyvin, Introduction to the Theory of Benzenoid Hydrocarbons, Springer-Verlag, Berlin, 1989.
I. Gutman, N. Trinajsti' c, Graph theory and molecular orbitals. Total $\pi$-electron energy of alternant hydrocarbons, Chem. Phys. Lett. 17 (1972) 535–538.
H. Hosoya, Y. Gao, Mathematical and chemical analysis of Wiener’s polarity number, in: D. H. Rouvray, R. B. King (Eds.), Topology in Chemistry – Discrete Mathematics of Molecules, Horwood, Chichester, 2002, pp. 38–57.
H. Hua, K. C. Das, On the Wiener polarity index of graphs, Appl. Math. Comput. 280 (2016) 162–167.
H. Hua, M. Faghani, A. R. Ashrafi, The Wiener and Wiener polarity indices of a class of fullerenes with exactly 12n carbon atoms, MATCH Commun. Math. Comput. Chem. 71 (2014) 361–372.
H. Hou, B. Liu, Y. Huang, The maximum Wiener polarity index of unicyclic graphs, Appl. Math. Comput. 218 (2012) 10149–10157.
A. Ili' c, M. Ili' c, Generalizations of Wiener polarity index and terminal Wiener index, Graphs Combin. 29 (2013) 1403–1416.
A. Ili' c, M. Ili' c, On some algorithms for computing topological indices of chemical graphs, MATCH Commun. Math. Comput. Chem. 78 (2017) 665–674.
H. Lei, T. Li, Y. Ma, H. Wang, Analyzing lattice networks through substructures, Appl. Math. Comput. 329 (2018) 297–314.
H. Lei, T. Li, Y. Shi, H. Wang, Wiener polarity index and its generalization in trees, MATCH Commun. Math. Comput. Chem. 78 (2017) 199–212.
M. Liu, B. Liu, On the Wiener polarity index, MATCH Commun. Math. Comput. Chem. 66 (2011) 293–304.
G. Liu, G. Liu, Wiener polarity index of dendrimers, Appl. Math. Comput. 322 (2018) 151–153.
I. Lukovits, W. Linert, Polarity-numbers of cycle-containing structures, J. Chem. Inf. Comput. Sci. 38 (1998) 715–719.
N. Tratnik, The Wiener polarity index of benzenoid systems and nanotubes, Croat. Chem. Acta 91(3) (2018).
H. Wiener, Structural determination of paraffin boiling points, J. Amer. Chem. Soc. 69 (1947) 17–20.
Y. Zhang, Y. Hu, The Nordhaus-Gaddum-type inequality for the Wiener polarity index, Appl. Math. Comput. 273 (2016) 880–884.
| {
"pile_set_name": "ArXiv"
} |
**[Study of the Nuclear Activity of the Seyfert Galaxy ]{}**
**[NGC 7469 over the Period of Observations 2008–2014]{}**
L. Ugol’kova$^{1*}$, B. Artamonov$^1$, E. Shimanovskaya$^1$,
V. Bruevich$^1$, O. Burhonov$^2$, Sh. Egamberdiev$^2$, N. Metlova$^1$
[*$^1$ M. V. Lomonosov Moscow State University, Sternberg Astronomical Institute, Moscow*]{}
[*$^2$ Ulugh Beg Astronomical Institute of the Uzbek Academy of Sciences, Tashkent*]{}
[**Abstract.**]{} We present results of multicolor UBVRI observations of the type 1 Seyfert galaxy (SyG 1) NGC 7469 carried out at the 1.5-meter telescope of the Maidanak Observatory (Uzbekistan) in 2008–2014. Analysis of the light curves indicates the presence of another slow flare of a long-term variability in 2009-2014 with a maximum in 2011-2012. We investigate properties of the long-term variability in 2009-2014, present (U-B)–(B-V) color diagrams for maxima and minima of NGC 7469 nuclear variability using various apertures and compare them with the black-body gas radiation which models the accretion disk radiation. Color-index measurements shows that the color becomes bluer at maximum brightness, indicating a higher temperature of the accretion disk. We have analysed the relation of X-ray and optical variability of NGC 7469 in 2008 and 2009 in comparison with the activity minimum in 2003. In 2008 the correlation coefficient between the X-ray and optical radiation is close to 0.5. Such poor correlation can be explained by the influence of an SN 1a explosion in close proximity to the nucleus of NGC 7469. The SN manifests itself in the optical band but does not affect the X-ray variability pattern. Comparison of the variability data in 2009 reveals a good correlation between the optical (U band) and the X-ray (7-10 keV) variability with the correlation coefficient of about 0.93. The correlation coefficient and the lag depend on the wavelength in the optical and X-ray bands. The lag between the X-ray and optical fluxes in 2009 is 2–4 days. In 2003 the lag is almost zero.
[**Keywords:**]{} active galaxy nuclei, Seyfert galaxies, NGC 7469, photometry
Introduction {#introduction .unnumbered}
============
Extensive observations of active galactic nuclei showed that AGNs vary in all wavelength bands on different time scales. Study of the variability can provide information about structure and energy-generation mechanisms in AGNs. Long-term observations of the Seyfert galaxy NGC 7469 (Arp 298 = MCG 1-58-25) revealed the presence of variability with time scales from minutes to decades.
NGC 7469 is a spiral galaxy of the SBa type, slightly inclined to the line of sight. The object coordinates are RA 23h 03m 15.75s, DEC $+08œ$ $52'$ $25''.9$. The distance to the galaxy is $D=68$ Mpc for $H_0=75$ km/s/Mpc, $z = 0.01639$. A physical companion, the irregular galaxy IC 5283, is situated at the distance of $80''$. In IR and optical bands, a star-burst region is observed around the nucleus in the form of a ring with the diameter of $1.5-2.5''$.
The central part of the galaxy is variable in the X-ray, UV, optical and IR ranges. In the radio-frequency range, NGC 7469 is a weak source. The variability is also observed in spectral lines. In the optical range, the photometry of NGC 7469 was performed by Doroshenko, Lyuty and Rahimov (1989) in 1967–1987. The variability is confirmed in observations by Merkulova (2000) in 1990–1998 and Sergeev, Doroshenko et al. (2005, 2010) in 2002–2009 at the Crimean Astrophysical Observatory. Results of Maidanak observations of NGC 7469 in 1990–2007 are presented by Artamonov et al. (2010).
As for many other AGNs, the variability of the NGC 7469 nucleus can be represented as a superposition of a slow component with a duration of several years and a flare, or fast, one with a duration from several days to tens of days. We will refer them as S and F components, respectively. That model was suggested by Lyuty and Pronik (1975, 2006). The maximum of the S component is usually attributed to an active state of a Seyfert galaxy nucleus. The F component is probably connected to an accretion rate.
The lightcurve of NGC 7469 over observational period 1990–2014 has two stages of the long-term variability: in 1994–2001 with a maximum radiation in 1997-1998, and in 2003–2007 with a maximum radiation in 2005. Combined lightcurves of NGC 7469 based on various observations in 1990–2007 were published in the proceedings of the Odessa Gamov Conference in 2011 (Ugol’kova, Artamonov (2011)). Chesnok et al. (2009) investigated the optical-X ray cross-correlation in the activity minimum in 2003 and discovered almost zero time delay between X-ray emission with respect to optical variability in the B band. Doroshenko et al. (2010) presented lightcurves of NGC 7469 based on Crimean observations and calculated time delays between optical and X-ray emission in different activity periods, they found the correlation with optical variability in B band in 2003–2007.
Observations {#observations .unnumbered}
============
In 2008–2014, CCD observations of NGC 7469 were carried out at 1.5-m telescope AZT–22 of the Maidanak Observatory (Ulugh Beg Astronomical Institute of the Uzbek Academy of Sciences) during 1.5-3 months every year. The SNUCAM 4096È4096 CCD array and UBVRI filters of the Cousin’s system were used. The frames were processed with the reduction software developed based on ESO-MIDAS (European Southern Observatory Munich Image Data Analysis System) package. The main image reduction stages were correction for bias and flat field, removal of cosmic-ray traces, determining the sky background, then subtracting it from each image frame. Photometry measurements were taken with apertures $10''$, $15''$, $20''$ and $30''$. As far as the intra-night variability goes beyond the scope of this work, the data were averaged for every night. Reference stars are the same as in the article by Doroshenko et al. (2005), they are on the same frame with the object, so the atmospheric extinction and variable air mass effects were not taken into account. Typical uncertainties of the CCD observations are in the range from 0.001 to 0.01 mag. The lightcurves of NGC 7469 nucleus variability in 2008–2014 are presented in Fig. 1,2. We did not subtract the constant galaxy background because it is model dependent.
Analysis of the long-term variability of NGC 7469 in 1990–2007 (Artamonov et al. (2010)) and in 2008–2014 revealed the presence of one more cycle of variability of the S component with duration of 6 years (2009–2014) and fast flux changes of the F component with durations from several days to 2 months. An amplitude of the fast variability (F component) is always less than that of the S component. The smaller a flare duration is, the smaller a flare amplitude is. The same conclusion is true for the S-component too. The brightness of NGC 7469 in 2009–2014 activity cycle is a bit weaker, than in 1994–2001, but is higher than in the 2004–2007 activity cycle. A maximum of that S component cycle is between 2010 and 2012.
[\[UBVRI20082014\]]{}
The UBVRI photometry (the aperture $d=10''$) of NGC 7469 is presented in Fig. 1 for the monitoring period 2008–2014. The duration of the rise part is shorter than that of the descending part of the light curves in the U and B bands. The lightcurve gradient rise from the I band to the U band. The variability amplitude also increases from the I band to the U band, that is suggestive of indicative emission burst in the blue part of the spectrum in active galactic nuclei. This is most clearly seen in Fig. \[BIlc\], where B and I light curves in relative flux units in 2008–2014 are presented for apertures with different diameters (10, 15, 20 É 30 arcsec). For apertures with $d>15''$, the brightness maximum occurs in 2011 in all bands, ascending and descending arms are symmetrical in all bands (see Fig.\[BIlc\]). An area, where brightness variations of a central part of the galaxy occur, is much less than the smallest aperture diameter ($10''$) that we use for obtaining photometry. We did not investigate variability in apertures with $d<10''$, as the PSF analysis based on stars in the field of the galaxy suggested that a size of PSF wings is about $5-7''$. The contribution of wings is negligible, but it can still lead to photometry errors when measuring the star-like nucleus of the galaxy.
For more detailed analysis of NGC 7469 variability, we present the galaxy UBV brightness in apparent magnitudes for the aperture $d=10''$ in time windows corresponding to actual observation periods in Fig. 3. In that figure a flux change in different bands in the minimum and maximum of the S variability is clearly pronounced. The lightcurve in the U pass band has more steep gradient in comparison to other bands on exit from the minimum of the cycle. On the down grade of the lightcurve in 2013, a new flare (of the F-component) is observed, and it also has more steep gradient in the U pass band.
\
\
\
\
\
\
\
\
\
\
\
\
\
Quantitative estimates of the U-B and B-V color indexes behaviour are presented on the two-color diagrams (Fig. 4a,b,c for three apertures $d=10''$, $20''$, $30''$. On the basis of the common assumption that the continuum spectrum of active galactic nuclei is comparable to the black body radiation, we present a color index of the black body radiation vs the black body temperature. The curve is from Straizhis book (1992). The color indexes are for minimum and maximum, ascending and descending branches of the lightcurve, and marked with different symbols (see details in figure legends). Main features of the color index distribution are as follows:
1\. The color becomes bluer along the lightcurves from minimum to maximum of radiation (Fig. 4a-d). Color indexes are situated on the line that is parallel to the black-body radiation curve.
2\. The smaller the aperture is, the bigger the offset of the color indexes to the blue part on the two-color diagram. The same phenomenon for Seyfert galaxies in contrast to normal galaxies was noted e. g. in works by Zasov (1989), Artamonov et al. (2010).
3\. With increase of the aperture, threshold color indexes U-B and B-V change towards lower black-body temperature, and the color becomes redder for the constant part of the galaxy radiation.
The galaxy color change trajectory is shifted with respect to the black-body curve towards the blue part of the diagram. To perform more correct comparison, it is necessary to take into account the galaxy contribution to the radiation of the central part of the galaxy. Furthermore, the radiation can be not quite the black body one. In the framework of the black-body radiation model, the accretion disk temperature in the maximum of the S-component approaches almost 8000œK.
\
\
\
\
Optical–X-ray correlation {#opticalx-ray-correlation .unnumbered}
=========================
To compare optical and X-ray variability of NGC 7469, we used data from Rossi X-Ray Timing Explorer (RXTE) in the 2-10 keV range (http://cass.ucsd.edu/ rxteagn/NGC7469/NGC7469.html). The RXTE data were obtained for NGC 7469 in 1996 and 2003-2009 (Rivers et al. 2013).
Optical and X-ray lightcurves NGC 7469 in 2009 are presented in Fig. 5, where their similarity can be clearly observed: both have a minimum, an ascending branch and a plateau. The shape of the X-ray lightcurve more closely coincides with that in U and slightly more poorly with those in B and V, so the X-ray variability better correlates with the optical variability in the U band. In 2009 the optical–X-ray correlation coefficient increases from 0.7 in the I band to 0.9 in the U band. Cross-correlation functions (CCF) were calculated for UBVRI and X-ray data. In this work we calculate the CCFs using the technique which is a modification of the Gaskell-Spark method (1986) and is described in articles by Oknyanskiy (1993), Koptelova et al. (2010), Shimanovskaya et al. (2015). In Fig. 6a, the óóFs for the X-ray (7–10 keV) and optical variability in 2003 based on RXTE data and data from Artamonov et al. (2010) and Doroshenko et al. (2009) are presented. Before CCF calculation we averaged data for every night. In Fig. 6b, CCFs between X-ray data (7-10 keV) and UBVRI data in 2009 are presented.
\
\
Comparison of CCFs for 2003 and 2009 reveals that in the NGC 7469 activity minimum (2003) the time delay between X and U is almost zero. The exact value depends on the method that is chosen to calculate it – centroid or maximum of CCF. But number of correlated pairs is insufficient for even reliable sign determination. In 2009, a noticeable delay of the X-ray variability with respect to the optical one is observed. The delay value increases from the U band to the I band (Fig. 6b), from 2 to 4 days according to preliminary estimates. It is noteworthy that the number of correlated pairs is not sufficient for reliable statistics.
In 2008, a minimum of the NGC 7469 activity is observed. The correlation of the optical radiation (B band) with the X-ray variability is weak ($k\sim 0.6$) for 7-10 keV and even weaker ($k\sim0 .5$) for 2-4 keV range. In 2008, a long-term flare on the lightcurve of NGC 7469 (about two months) is observed, the maximum of that flare occurs in several dozens days after a SN explosion in the galaxy at $~16''$ distance from its nucleus.
Similar studies were also performed for other active galactic nuclei: Maoz et al.(2000, 2002) investigated connection between optical and X-ray data for NGC 3516. They smoothed lightcurves with a 30 day boxcar running mean and found that the X-ray emission follows the optical one but the deep minimum in the optical variability corresponds to the minimum in the X band with almost zero lag. Observations of NGC 3516 during 5 years (1997–2002) reveals no correlation between optical and X-ray bands except in 1997-1998 when 100 day lag of X-ray behind optical variations was detected. Thus, the correlation is observed when NGC 3516 is in a high state, and there is no correlation when it is in a low state. Doroshenko et al. (2009) compared variability of the 3ó120 in optical bands (UBVRI) and in the X-ray band (RXTE) for 1996–2008 and found that sometimes the X-ray variability is followed by the optical one, sometimes a zero lag is observed, sometimes the optical variability is leading.
In different activity periods, the optical-X ray correlation is different because of complex and diverse physical nature of processes near the NGC 7469 nucleus. As an addition to well-known reprocessing mechanism, Gaskell (2006, 2007) suggested the anisotropic high-energy emission model for explanation of the optical-X-ray variability and impact of IR dust emission around an accretion disk to a visible band when estimating time delays between optical passbands from B to I. Detailed discussion of processes near the NGC 7469 nucleus can be found in Doroshenko et al. (2010), Chesnok et al. (2009). They compare X-ray luminosity and optical luminosity and conclude that the direct Compton re-processing is not likely to dominate in NGC 7469, it can generate only part of apparent optical luminosity. Other processes, such as the thermal radiation from the accretion disk, the star formation processes, the inverse Compton scattering by hot coronal electrons etc. may also contribute to the optical emission.
Conclusion {#conclusion .unnumbered}
==========
At present, the universally accepted view on the nature of variability in the nuclei of Seyfert galaxies is an accretion of matter onto a supermassive compact object, and all manifestations of the variability are associated with the accretion disk. The increase in the duration of the slow component of variability, associated with the accretion disk, indicates possible change in accretion conditions.
In this paper we present results of our UBVRI observations of the Seyfert galaxy NGC 7469 at the Maidanak Observatory in 2008–2014. We obtained light curves for all passbands with various apertures: $d = 10''$, $15''$, $20''$, and $30''$. Analysis of the light curves reveals the presence of a fast variability component with a duration from several days to several tens of days, and a slow component with the duration of about 6 years. This is the third activity cycle of the NGC 7469 nucleus over the period of its observations at the Maidanak Observatory from 1990 to 2014.
Our observations showed the following:
1\. In 2008–2014, the characteristic behaviour of the light curves in all passbands are similar; the differences refer mainly to the variability amplitude.
2\. The relative variability amplitude decreases with increasing the wavelength (from U to I) for both components (S and F). In the observation period, the amplitude of U brightness variations reached its maximum value of 0.9 mag.
3\. The brightening time (ascending branch) is shorter than the fading time (descending branch) in all passbands. The brightness gradient increases from I to U. The variability amplitude increases from I to U, indicating a characteristic burst of light in the blue part of the spectrum in active galactic nuclei.
4\. We analysed the color variations of NGC 7469 in different activity periods. On the (U-B)–(B-V ) diagram we compared the color characteristics of the slow component with the blackbody radiation of a gas modeling the accretion disk. The color becomes bluer along the light curves from minimum to maximum light, and the color indices lie on a line that is parallel to the blackbody radiation. The color of the galaxy NGC 7469 becomes bluer as the nucleus is approached, in contrast to normal galaxies.
5\. We analysed the X-ray variability of NGC 7469 in 2008 and 2009 in comparison with the 2003 activity minimum. In 2008, there is a weak correlation between optical and X-ray variability (the correlation coefficient is 0.5–0.6), which can be explained by an SN Ia explosion that manifests itself in the optical band but does not change the pattern of X-ray variability. In 2009, a U-X correlation with a correlation coefficient that is close to 0.9 is observed. The correlation coefficient and the lag depend on the wavelength in the optical and X-ray bands. The results of these studies and the results of such an analysis in other papers show that the correlation between the optical and X-ray variability is different in different activity periods due to the complex and diverse physics of the processes near the nucleus of a Seyfert galaxy.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank B. Khafizov for the technical support of the Seyfert galaxies monitoring program at the Maydanak observatory. The work is supported by Russian Foundation for Basic Research, the project ID is 14-02-01274. This work has made use of light curves provided by the University of California, San Diego Center for Astrophysics and Space Sciences, X-ray Group (R.E. Rothschild, A.G. Markowitz, E.S. Rivers, and B.A. McKim), obtained at http://cass.ucsd.edu/$\sim$rxteagn/.
[99]{}
Artamonov, B.P. et al., ARep., [**54**]{}, 767A (2010).
Chesnok, N.G., Sergeev, S.G. et al., Kinematics and Physics of Celestial Bodies, **25**, 2, pp. 107?113 (2009)
Doroshenko, V.T., Lyuty, V.M. and Rahimov, V. Yu., PAZh (Astronomy Letters), **15**, 6, 48 (1989)
Doroshenko, V.T. et al., Astrophysics, **48**, No. 3 (2005)
Doroshenko, V.T. et al., PAZh (Astronomy Letters), **35**, 403 (2009)
Doroshenko, V.T., Sergeev, S.G. et al., PAZh (Astronomy Letters), **36**, 9, 643 (2010)
Gaskell, C.M. & Spark, L.S., ApJ, **305**, 175 (1986)
Gaskell, C.M., AGN Variability from X-rays to Radio Waves,ASP Conference Series, **360**, 111 (2006)
Gaskell, C.M., The Central Engine of Active Galactic Nuclei, ASP Conference Series, **373**, p.596 (2007)
Hughes, P.A., Aller, H. D. and Aller, V. F., Astrophys. J., **396**, 469 (1992).
Koptelova E. et al., MNRAS, **401**, 2805 (2010)
Lyuty V.M., Pronik V.I., IAUS, **67**, 591 (1975)
Lyuty V.M., ASPC, **360**, 3 (2006)
Maoz et al., AJ **119**, 119-125 (2000)
Maoz et al., AJ **124** (2000)
Merkulova N.I., AJ, **119**, 631-643 (2000)
Nandra K., Le T. et al., ApJ, **544**, 734-746 (2000)
Nandra K. and Papadakis I. E., ApJ, **554**, 710 (2001)
Oknyanskij V.L., PAZh (Astronomy Letters), **19**, 1021 (1993)
Perez-Torres M. A. et al. MNRAS, **399**, 1641 (2009)
Papadakis, I.E., Nandra, K., Kazanas, ApJ, **554L**, 133 (2001)
Petrucci et al., AA, **413**, 477 (2004)
Rivers, Markowitz, & Rothschild, ApJ, 772, 114 (2013)
Sergeev S.G., Doroshenko V.T. et al., Aph.J, **622**,129 (2005)
Simmonetti, J.H., Cordes, J.M., and Heeschen, D.S., Astrophys. J. **296**, 46 (1985).
Straizhis V., Multicolor stellar photometry, Tucson : Pachart Pub. House (1992)
Shimanovskaya E.V. et al., ARep, **59**, 12 (2015)
Ugol’kova L.S., Artamonov B.P., Odessa Astronomical Publications, **24**, 78 (2011)
Ulrich, M-H., Maraschi, L., Urry, C. M., Annual Review of Astronomy and Astrophysics, **35**, 1997, pp. 445-502 (1997)
Wanders, I. et al, ApJ. Suppl. Series, [**113**]{}, 69 (1997).
Zasov, A. V.; Neizvestnyj, S. I., Soviet Astronomy Letters, **15**, NO. 6/NOV, P. 419 (1989)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present the results of detailed N-body simulations of clusters moving in a realistic Milky Way (MW) potential. The strong interaction with the bulge and the disk of the Galaxy leads to the formation of tidal tails, emanating from opposite sides of the cluster. Some characteristic features in the morphology and orientation of these streams are recognized and intepreted. The tails have a complex morphology, in particular when the cluster approaches its apogalacticon, showing multiple “arms” in remarkable similarity to the structures observed around NGC 288 and Willman 1.\
Actually, the tails are generally good tracers of the cluster path quite far from the cluster center ($>7$–$8$ tidal radii), while on the smaller scale they are mainly pointing in the direction of the Galaxy center. In particular, the orientation of the inner part of the tails is highly correlated to the cluster orbital phase and to the local orbital angular acceleration. This implies that, in general, the orbital path cannot be estimated directly from the orientation of the tails, unless a sufficient large field around the cluster is available.
author:
- 'M. Montuori'
- 'R. Capuzzo-Dolcetta'
- 'P. Di Matteo'
- 'A. Lepinette'
- 'P. Miocchi'
title: |
Tidal tails around globular clusters.\
Are they a good tracer of cluster orbits?
---
Introduction
============
It is commonly accepted that the present globular cluster (GC) population in the Milky Way represents the survivor of an initially more numerous one, depauperated by many disruptive processes [@mw97a; @mw97b; @fz01].\
The observational results obtained in the last decade have clearly confirmed the role played by the Milky Way environment on the GC evolution. Up to now, about $30$ galactic GCs show evidences of star depletion due to the tides caused by the field of the Galaxy.\
The first evidence of the existence of tails surrounding GCs were achieved by @grill95. Using colour-magnitude selected star counts in a dozen of galactic GCs, these authors showed that in the outer parts the stellar surface density profiles exceeded significantly the prediction of King models, extending also outside the tidal radius. They identified this surrounding material as made up of stars in the act of being tidally stripped from the cluster field and pointed out the importance of defining the large-scale distribution of extra-tidal stars on the sky to obtain constraints and traces for the GC orbits. More recently, other works confirmed and enlarged Grillmair et al.’s findings [@ls97; @testa00; @lmc00; @sieg00; @lee03] giving evidence of the existence of many GCs surrounded by haloes or tails. For M92, @lee03 found, also, that the extratidal material is not randomly distributed, being the density profile in the tails shallower for bright stars than for fainter ones.\
This was the state of the art until the spectacular findings of two tidal tails emanating from the outer part of the Palomar 5 globular cluster and covering an arc of 10 degrees on the sky, corresponding to a (projected) length of 4 kpc at the distance of the cluster [@oden01; @oden02; @oden03], obtained in the framework of the Sloan Digital Sky Survey (see also *http://www.sdss.org*). The stellar mass in the tails of this sparse, low-mass, halo cluster (with an estimated concentration parameter $c=0.7$) adds up to 1.2 times the mass of stars in the cluster, estimated in the range between $4.5\times 10^3 M_{\odot}$ and $6\times 10^3 M_{\odot}$. More recently, @grill06, still using SDSS data, have detected a continuation of Pal 5’s trailing tidal stream out to almost 19 degrees from the cluster. Combining this with the already known southern tail of Pal 5 yields a stream some $9 $ kpc long on the sky.\
Substantial tidal streams have recently been found associated with another low-mass and low-concentration GC in the SDSS area: NGC 5466. For this cluster @belok06 reported giant tails extended for about 1 kpc in length and @grillj06, still using SDSS data, suggested that a $\approx 13$ kpc tidal stream of stars, extending from Bootes to Ursa Major, could also be associated with this system.\
Together with the so-called Sagittarius stream [@mateo98; @yanny00; @mart01; @maj03; @mart04], which emerges from a dwarf galaxy that is currently being accreted by the Milky Way, Palomar 5 and NGC 5466 represent outstanding examples of ongoing tidal erosion of stellar systems in the Milky Way, being, up to now, also the only two globulars known for which such extended stream-like structures have been detected in the Galactic halo.\
One of the first numerical investigations of the role played by a galactic tidal field on spherical stellar systems was that of @keenan75, who studied the effect of realistic, time-varying tidal fields on the stellar orbits in a star cluster. They numerically integrated the equations of motion of three-bodies in models of spherically symmetric clusters which, in turn, move in eccentric orbits in the field of a model galaxy. One of the main conclusions of this work, which extended previous investigations made by @king62, was that star clusters rotating in a retrograde sense are more stable in a tidal field than clusters with either direct rotation or no rotation due to the contribution of the Coriolis acceleration, acting in the same direction as the gravitational attraction for retrograde motion.\
More recent works on weak tidal encounters based on a Fokker-Planck approach [@ohlin92; @leegood95] and self-consistent N-body techniques [@grill98] confirmed that the interaction with external tidal field, combined with two-body relaxation in the core of the cluster and following replenishment of stars near the tidal radius, causes a flow of stars away from the cluster. The stripped stars remain in the vicinity of the cluster for several orbital periods, migrating either ahead of the cluster or falling behind, giving rise to a slow growth of the tidal tails.\
A semi-analytic study of the development of tidal streams in galactic satellites was done by @john98, who gave estimates for the rate of growth of tidal tails. The effects of a realistic galactic tidal field (including both bulge, halo and disk) on GCs were investigated few years later by @clm99. The main findings of the work were the following: stars escaped from the system go to populate two giant tidal tails along the cluster orbit; these tails present substructures, or clumps, attributed to strong shocks suffered by the cluster, and are preferentially formed by low mass stars.\
@yimlee02 performed N-body simulations of GCs orbiting in a two-component galaxy model (with bulge and halo and no disk), using the direct-summation NBODY6 code [@aarseth99] and focusing their attention, in particular, to the correlation between tidal tail elongation (described by mean of a ’position angle’, defined as the angle between the direction of the tail and the galactic center direction) and the cluster orbit. They found that, on circular orbits, tidal tails mantain an almost constant position angle ($\sim 60^{\circ}$), while GCs on non circular orbits show a variation of the position angle, according to orbital path and phase. The position angle increases when the cluster heads for perigalacticon. On the other hand, it tends to decreases when the cluster heads for apogalacticon.
Finally, some authors investigated also the dynamical evolution of some globular clusters in the tidal field of the Galaxy. In this context @dehnen04 modelled the disruption of the globular cluster Pal 5 by galactic tides. Pal 5 is remarkable not only for its extended and massive tidal tails, but also for its very low mass and velocity dispersion. In order to understand these extreme properties, they performed many simulations aiming at reproducing the Pal 5 evolution along its orbit across the Milky Way. They explained the very large size of Pal 5 as the result of an expansion following the heating induced by the last strong disk shock about 150 Myr ago. The clumpy substructures detected in the tidal tails of Pal 5 are not reproduced in their simulations, so that they argued that these overdensities have been probably caused by interaction with Galactic substructures, such as giant molecular clouds, spiral arms, and dark matter clumps, which were not condsidered in their modeling. These simulations also predict the destruction of Pal 5 at its next disk crossing in about 110 Myr, suggesting that many more similar systems have once populated the inner parts of the Milky Way but have been transformed into debris streams by the Galactic tidal field.\
In this context, it may be interesting mentioning the recent numerical work devoted to the study of smaller size systems (open clusters) in the MW tidal field [@chum06], which confirms already known results in the case of the external part of GC tidal tails) about the alignement of stars of the tidal stream around a common orbit in the external field [@grill98; @clm99; @cdm05].
In the above sketched theoretical and observational background, this work –which is inserted in a wider study on the dynamics of globular clusters in external tidal fields [@cdm05; @dm04; @mioc; @mioc2]– is devoted to clarifying the connection among tidal tails and cluster orbit. We will describe and discuss the mechanisms that determines the tails morphology and how they depend on cluster trajectory and orbital phase.\
For this purpose, we performed detailed $N-$body simulations (with $N=10^5$) of GCs moving in a realistic three-components (bulge, disk and halo) Milky Way potential.\
The paper is organized as follows: in Sect.\[Cluster\] the Galaxy (Sect.\[Galaxy\]) and cluster (Sect.\[Cluster1\]) models adopted are presented, so as the numerical approach used; in Sect.\[results\] we deal with the main results of our work, showing the formation and development of tidal tails around the cluster (Sect. \[tails\]), giving a qualitative approach for describing the tail morphology (Sect.\[qualit\]), presenting the numerical procedure adopted to fit tails direction (Sect.\[fit\]) and, finally, discussing the tail-orbit alignement and its dependence on the orbital phase. In Sect.\[concl\], all the results are summarized and discussed.
Models and methods {#Cluster}
==================
Galaxy model {#Galaxy}
------------
The model adopted for the Galactic mass distribution is that of @allsant91. It consists of a three component system: a spherical central bulge and a flattened disk, both of @Miyamoto-Nagai form, plus a massive spherical halo. The gravitational potential is time indipendent, axysimmetric and given in an analytical form which is continuous together with its spatial derivatives.\
Choosing a reference frame where the ([*x,y*]{}) plane coincides with the MW equatorial plane, the three components of the potential have, in cylindrical coordinates, the form: $$\begin{aligned}
\Phi_B(\omega,z)&=&-\frac{GM_B}{\sqrt{R^2+z^2+{b_B}^2}}\\
\Phi_D(\omega,z)&=&-\frac{GM_D}{\sqrt{R^2+\left(a_D+\sqrt{z^2+{b_B}^2}
\right)^2}}\\
\Phi_H(r)&=&-\frac{GM(r)}{r}-
\frac{GM_H}{1.02a_H}\left[-\frac{1.02}{1+(\frac{r}{a_H})^{1.02}}+ln
\left(1+(\frac{r}{a_H})^{1.02}\right)\right]^{100}_r,\\\end{aligned}$$ where square brackets indicate the difference of the function evaluated at the 100 kpc and the generic $r=\sqrt{R^2+z^2}$ extremes. The parameters in the formulas above are listed in Table \[tbl1\].
Cluster model and numerical method {#Cluster1}
----------------------------------
The total initial mass of the cluster was chosen $M_{GC}=4.7 \times 10^5$ M$_{\odot}$, i.e. a value compatible with masses of galactic globular clusters lying inside 4.5 kpc from the MW center (see @har96 and @pry93). The stellar mass spectrum of our cluster initial model was chosen as a Kroupa IMF sampled in the range $0.1M_{\odot}-20M_{\odot}$ and “evolved” (in the sense of accounting for mass loss on the base of stellar evolution, according to @scl97 [@dom99]) up to $3$ Gyr (which corresponds to our assumed cluster age). In this interval of time all masses greater than $1.2$ M$_\odot$ go into the $0.5M_{\odot}-1.2M_{\odot}$ range. As in @cdm05, we sampled this mass function into 12 mass classes, equally spaced in a linear scale, whose space and velocity distribution were evaluated by the adoption of a multimass King distribution [@king66; @costa76].\
Obviously, the choice of $N=10^5$ as total number of cluster stars, together with the given value of $M_{GC}$ and of the above described mass function implied a rescaling of the star masses. Finally, to investigate the role played by the degree of cluster concentration, we considered clusters with two different values for the King concentration parameter $c$ (listed in Table \[tbl2\]).\
The clusters move on the $y-z$ coordinate plane (the ($x,y$) plane corresponds to the Galactic disk), along orbits of different eccentricity, defined as $$e=\frac{r_a-r_p}{r_a+r_p}, \label{ecc}$$ being $r_p$ and $r_a$, respectively, the GC pericentric and apocentric orbital distances. See Table \[tbl3\] for GC orbital parameters and Fig.\[orbits\] for a plot of the different simulated orbits.
All the simulations were performed by means of the TreeATD code, developed by two of us [@bib2] which, as the original code by [@bh], relies upon a tree-algorithm for the gravitational force evaluation on the large scale and on a direct summation on the small scale. The code was parallelized to run on high performance computers (via MPI routines) employing an original parallelization approach (see again [@bib2 see]). The time-integration of the ‘particles’ trajectories is performed by a leap-frog algorithm that uses individual and variable time-steps according to the block-time scheme [@aars; @bib5]. See @mioc for further details. The only constant of motion that can be significantly used to check the quality of the orbital time–integration is the total energy of the cluster, i.e. including the contribution of the external field to the GC potential energy. We saw that the upper bound of the relative error in the energy conservation, $\Delta E/E$, is $10^{-4}$ over the whole simulation time. This is no more than one order of magnitude worse than the error we got in a comparison simulation for the same GC in abscence of external field.
Results
=======
Formation and evolution of tidal tails {#tails}
--------------------------------------
In Figs. \[orbitaI\],\[orbitaII\] and \[orbitaIII\] the formation and subsequent evolution of tidal tails around the GC is shown, for all the simulations performed in the case of the least concentrated of the two GC models (the one with $c=0.81$ in Table \[tbl2\]). In the case of the more concentrated initial model cluster, the development of the tidal tails shows the same time dependence being just slightly less populated.\
After about $5\times10^7$ yr, tidal tails have clearly formed. They continuously accrete by the stars leaving the cluster, so that after $1.6\times 10^8$ yr, in the case of the quasi-circular orbit (orbit III), they are elongated for more than 1.5 kpc each. In general, the extention of the two cluster tails depends on the velocity of the cluster and its variation along the orbit. For the quasi-circular orbit (orbit III) this velocity is almost constant and the two tails have nearly the same length. For more eccentric orbits (orbit I and II), the tail extending between the GC and the pericenter is more elongated than the opposite tail, because the stars in the latter have smaller velocities than those belonging to the former one. In any case, the tail that precedes the cluster always extends slightly below the orbit, whereas the trailing one lies slightly above the orbit in agreement with what was observed for Pal 5. This feature can be explained considering the role played by the Coriolis acceleration, as we will see later.
As regards the tails orientation, these figures (Figs \[orbitaI\],\[orbitaII\],\[orbitaIII\]) clearly show the alignment of the outer part of the tails (at distances from the GC center $>7$–$8$ tidal radii) with the orbit. On the contrary, the alignment of the inner part is strongly correlated to both the orbit eccentricity and the GC location. Notice, indeed, that the inner tails are roughly aligned with the orbital path only when the cluster is near the perigalacticon of the more eccentric orbits (I and II, respectively shown in Fig.\[orbitaI\] and Fig.\[orbitaII\]). This confirms previous results in @cdm05, where analogous features were found for GCs movings in a triaxial potential.
It is also worth noting the peculiar morphology in the streams when the cluster is approaching the apocenter in the eccentric orbits (orbits I and II): tidal tails divide into “multiple arms”, two nearly elongated along the GC path, while the other two point to the galactic center (and anticenter) direction. At this regard, the tails observed around NGC288 [@lmc00] have a similar complex morphology, showing three different “arms”. Interestingly, this cluster has been estimated to be near its apogalacticon [@dinescu] in agreement with our results. Also the galactic satellite Willman 1 shows multi-directional stellar tails [@wil], which according to our results, could represent evidence that this system is approaching the apocenter on an eccentric orbit.
Tidal tails elongation: a qualitative description {#qualit}
-------------------------------------------------
As anticipated in @dm04, the simplified scheme of the formation and shape of tidal tails around GCs can be understood by the motion of a star escaping from a cluster moving in a spherical potential. Let us consider a rotating reference frame with the origin in the cluster center-of-mass, the $(x',y')$ plane coinciding with the orbital plane and $x'$ pointing to the galactic center. The equation of motion of the $i_{th}$ star belonging to the cluster that moves around the galaxy center with variable angular velocity $\mbox{\boldmath{$\omega$}}$ is (see also Appendix \[app\] and Fig.\[coriolis\]): $$\ddot{\bf r}'_i=\ddot{\bf r}_i-\ddot{\bf r}_{GC}-2 \mbox{\boldmath{$\omega$}}
\times \dot{\bf r}'_i-\mbox{\boldmath{$\omega$}} \times
\left(\mbox{\boldmath{$\omega$}} \times \bf
{r}'_i\right)-\dot{\mbox{\boldmath{$\omega$}}}\times \bf {r}'_i,
\label{equat}$$ where ${\bf r}'_i$ is the position vector in the rotating frame, while ${\bf r}_i$ and the position vector of the GC center-of-mass$,
{\bf r}_{GC}$, refer to the inertial frame. For a star escaping through one of the unstable Lagrangian points along the $x'$ axis, the first, second and fourth terms on the right-hand side of Eq. \[equat\] are directed along the $x'$ direction, while the third term (the Coriolis acceleration) and the fifth term are along $y'$. The latter is parallel to the Coriolis term when $|\mbox{\boldmath{$\omega$}}|$ increases (i.e. moving from the apogalacticon to perigalacticon) and antiparallel when $|\mbox{\boldmath{$\omega$}}|$ decreases (moving from the perigalacticon to the apogalacticon). Consequently, these latter two terms ere responsible of the initial deviation of the tails from the radial direction and of the formation of the S-shape profile [see @cdm05]).
The tail direction along the orbit {#fit}
----------------------------------
As a reliable quantitative definition of the tail orientation, we used the direction of the eigenvector corresponding to the greatest eigenvalue[^1] of the inertia tensor of a suitable portion of the cluster. More precisely, we considered the stars contained in an annulus centered at the cluster center-of-density (CD), as defined by @ch85, and extending from $\approx 30$ pc up to a distance of $7$–$8$ tidal radii ($\sim 170$ pc). This region includes, indeed, the tail portion we call “inner” part of the tails; this part is that usually detected in the observations. Note that in the study of @yimlee02 the direction of the tails was determined by eye, without any mathematically defined procedure.\
For each configuration of the system:
- we evaluated the GC center-of-density;
- we selected all the stars distributed between $30$ pc and $170$ pc from the cluster CD (excluding stars lying inside $r<30$ pc, because this region corresponds to the spherical part of the system);
- we evaluated the inertia tensor and the corresponding eigenvalues and eigenvectors of the selected stars;
- we determined the eigenvector ${\bf u}$ corresponding to the maximum eigenvalue.
This eigenvector is parallel to the direction of the two tails having opposite orientation. In the following, we always refer to the tail internal to the cluster orbit.
The evolution of the tidal tail orientation is shown in Figs. \[tail-radI\], \[tail-radII\], \[tail-radIII\].\
These figures show the GC orbit and some kinematic quantities plotted as functions of time. From top to bottom we plot: the GC orbit (first panel), the distance $r$ of the cluster from the Galaxy center (second panel), the magnitude of the angular velocity vector $\mbox{\boldmath{$\omega$}}$ (third panel), the time derivative of $|\mbox{\boldmath{$\omega$}}|$ (fourth panel). Finally, the fifth panel shows the evolution with time of two quantities: the angle
$$\alpha\equiv
\cos^{-1}\left({\frac{ {\bf u} \cdot {\bf r_{GC}}}{ |{\bf u}||{\bf r_{GC}}| }}\right)$$
between the tail direction and that of the galactic center and the angle $$\beta \equiv \cos^{-1}\left({\frac { {\bf u}\cdot \dot{\bf r}_{GC}} {|{\bf u}|
|\dot{\bf r}_{GC}|}}\right)$$ between the tail direction and the GC velocity vector (the angles assume values in the $[0^\circ,180^\circ[$ interval).\
### Tail-Galaxy center alignement {#tailgal}
In Figure \[tail-radI\], which refers to the orbit I, $\alpha$ is represented by two, barely distinguishable, solid curves, corresponding respectively to the GC with $c=1.02$ and $c=0.8$. The first result is that the time evolution of $\alpha$ is independent of the GC concentration (in the interval of $c$ studied); this means that the morphology and orientation of the tails depend on GC orbit, rather than on GC internal parameters.\
Moreover, the tail is elongated toward the galactic center for most of the time: more in particular, if in Fig.\[tail-radI\] we exclude the regions around the pericenter, the average value of $\alpha$ is $\sim 10^\circ$, i.e. the inner tail has only a slight deviation respect to the galactic center.
When approaching the pericenter, instead, the Coriolis and the $\dot{\mbox{\boldmath{$\omega$}}}$ terms (in Eq.\[equat\]) are parallel and grow up. This induces a rapid alignement of the tails with the GC orbit. Beyond the pericenter, the Coriolis acceleration decreases and $\dot{\mbox{\boldmath{$\omega$}}}$ becomes antiparallel to the former. The net effect yields a realignment of the tails with the direction of the galactic center. Figure \[tail-radII\], which refers to orbit II, confirms the results found for the previous orbit.
Finally, Figure \[tail-radIII\] shows the case of the quasi-circular orbit (orbit III). Along this orbit, both the distance of the GC from the galactic center and the orbital angular velocity are nearly constant. This leads to a roughly constant $\alpha\sim
40 ^\circ$. This suggests that the change in time of the orbital angular velocity is what determine the tails orientation.
### Tail-orbit alignement {#tailvel}
The dashed curve shown in the bottom panel of Figures \[tail-radI\],\[tail-radII\] and \[tail-radIII\] represents the angle $\beta$, previously defined. In the case of eccentric orbits, the first striking feature is that when approaching the pericenter the angle $\beta$ decreases, reaches a minimum at the pericenter and then increases moving away from the pericenter. The maximum value of $\beta$ is reached between the pericenter and the apocenter, depending on the shape of the GC orbit. However, in correspondence of the apocenter, $\beta$ is $\approx 90^\circ$ in agreement with the value $\alpha \approx 0^\circ$. At the apocenter the tail is aligned along the radial direction and roughly perpendicular to the GC velocity.
It is worth noting that, in all the cases considered here, the tidal tails deviate considerably by the GC velocity direction; only for very eccentric orbits and close to the pericenter, the angle $\beta$ reaches a minimum value of $\approx 17^\circ$. This indicates that the extrapolation of the cluster orbit from the elongation of this part of the tidal streams leads, in general, to predict GC paths with large errors.
Conclusions {#concl}
===========
This work is devoted to the study of the morphology and orientation of tidal tails surrounding globular clusters, in order to understand to what extent they trace the GC orbit and if some correlations exist between their orientation and the GC orbital phase.\
The main findings of our work can be resumed as follows:
1. Tidal tails are good tracers of GC orbit only on the large scale, being the [*inner part of the streams never aligned with the GC path*]{}, unless the system moves on very eccentric orbit and close to the orbital pericenter.
2. A [*strong correlation*]{} exists between the orientation of the inner tails and GC orbital position: they tend to be more elongated along the GC orbit when the cluster approaches the pericenter, while, in turn, they tend to align toward the galactic center when the GC approaches the apocenter.
3. We have shown that a key role in determining the tails morphology is played by the orbital angular velocity of the system and its variation with time. In the case of the least eccentric orbit, a nearly constant angular velocity determines an almost constant orientation of tails with respect to the cluster orbit and the galactic center direction. The amplitude of $\dot{\mbox{\boldmath{$\omega$}}}$ contributes, too, to the alignement of tails with the GC path, in particular just beyond the pericenter.
4. The existing correlation between tails elongation and GC position along its orbit can be easily understood when referring to a non-inertial frame centered on the GC center, due to the role played by non inertial accelerations, as the Coriolis acceleration and that related to $\dot{\mbox{\boldmath{$\omega$}}}\times \bf {r}'_i$.
We want to outline that our findings are in good agreement with some observational results. In particular we refer to the galactic globular clusters sample studied by @lmc00, considering the 7 clusters for which the reliability level of the observed tidal streams is highest and orbital parameters are available [@dinescu]. Six clusters of this subsample [see @lmc00 Table 5] have tidal tails clearly elongated along the galactic center direction, and eccentricities greater than 0.62. This in agreement with our results that GCs, on eccentric orbits, have tidal tails elongated towards the galactic center for most of their path. The only exception is represented by NGC5139 ($\omega$ Cen), which has tidal tails deviating from the galactocentric direction and extending towards the galactic plane; this situation does not contradict our results considering that NGC5139 is presently only 1.3 kpc above the galactic disk and is most probably undergoing a strong disk-shocking.\
The simulations also reproduce the formation of multiple tails around globular clusters, as observed for NGC288 [@lmc00] and for the galactic satellite Willman 1 [@wil]. Our simulations show that these features are expected for GCs on eccentric orbits near their apogalacticon, while they are absent in the case of GCs moving on less eccentric orbits. Actually, their formation can be explained by the behaviour of the Coriolis acceleration along the orbit: when its value is large, stars escaping from the cluster are pushed toward the orbital path. This is particularly evident around the pericenter; at the apocenter, where the Coriolis acceleration is weaker, stars tend to escape from the cluster more radially. At every pericenter passage, this effect produce a new tidal tail roughly aligned to the cluster path.
Acknowledgements
================
The authors acknowledge the Centro de Astrobiologiá (Madrid, Spain) for the use of its computational facilities. We also thank prof. D. Heggie for useful comments and for pointing us an error in one of the figure catpions.
Equations of motion in the non-inertial reference frame {#app}
=======================================================
Using a no-inertial reference frame $(x^{\prime},y^{\prime},z^{\prime})$ with the origin in the cluster center-of-mass, with the $(x^{\prime},y^{\prime})$ plane coinciding with the cluster orbital plane and $x^{\prime}$ always pointing to the galactic center (so that this reference frame rotates with the GC angular velocity $\mbox{\boldmath{$\omega$}}$ respect to the inertial reference system $(x,y,z)$), the position of the $i-th$ star in the galactocentric $(x,y,z)$ reference frame can be expressed simply as: $${\bf r}_i={\bf r}_{GC}+{\bf r'}_i$$ being ${\bf r}_{GC}$ the position of the cluster center-of-mass in the inertial system. Rewriting the previous equation in terms of the $(x^{\prime},y^{\prime})$ components (we omit the $z'$ one for simplicity) we have: $$\label{pos}
x_i \mbox{\boldmath{$\hat{x}$}}+y_i \mbox{\boldmath{$\hat{y}$}}=x_{GC}\mbox{\boldmath{$\hat{x}$}}+y_{GC}\mbox{\boldmath{$\hat{y}$}}+x'_i \mbox{\boldmath{$\hat{x'}$}}+y'_i \mbox{\boldmath{$\hat{y'}$}}$$ being $\mbox{\boldmath{$\hat{x}$}}$,$\mbox{\boldmath{$\hat{y}$}}$ and $\mbox{\boldmath{$\hat{x'}$}}$,$\mbox{\boldmath{$\hat{y'}$}}$ the unity vectors in the two frames, respectively. Derivating Eq.\[pos\] with respect to time $t$, one obtains: $$\begin{aligned}
\label{vel}
\dot{x}_i \mbox{\boldmath{$\hat{x}$}}+\dot{y}_i \mbox{\boldmath{$\hat{y}$}}&=&\dot{x}_{GC}\mbox{\boldmath{$\hat{x}$}}+\dot{y}_{GC}\mbox{\boldmath{$\hat{y}$}}+\dot{x'}_i\mbox{\boldmath{$\hat{x'}$}}+\dot{y'}_i\mbox{\boldmath{$\hat{y'}$}}+x'_i\mbox{\boldmath{$\dot{\hat{x}'}$}}+y'_i\mbox{\boldmath{$\dot{\hat{y}'}$}}=\\
& &\dot{x}_{GC}\mbox{\boldmath{$\hat{x}$}}+\dot{y}_{GC}\mbox{\boldmath{$\hat{y}$}}+\dot{x'}_i\mbox{\boldmath{$\hat{x'}$}}+\dot{y'}_i\mbox{\boldmath{$\hat{y'}$}}+\mbox{\boldmath{$\omega$}}\times(x'_i \mbox{\boldmath{$\hat{x'}$}}+y'_i \mbox{\boldmath{$\hat{y'}$}})\end{aligned}$$ being $$\begin{aligned}
\mbox{\boldmath{$\dot{\hat{x}'}$}}&=&\mbox{\boldmath{$\omega$}}\times\mbox{\boldmath{$\hat{x'}$}}\label{xdot}\\
\mbox{\boldmath{$\dot{\hat{y}'}$}}&=&\mbox{\boldmath{$\omega$}}\times\mbox{\boldmath{$\hat{y'}$}}\label{ydot}\end{aligned}$$ Derivating Eq.\[vel\] with respect to time $t$ and using Eqs. \[xdot\]-\[ydot\], one obtains: $$\begin{aligned}
\ddot{x}_i \mbox{\boldmath{$\hat{x}$}}+\ddot{y}_i \mbox{\boldmath{$\hat{y}$}}&=&\ddot{x}_{GC}\mbox{\boldmath{$\hat{x}$}}+\ddot{y}_{GC}\mbox{\boldmath{$\hat{y}$}}+\ddot{x'}_i \mbox{\boldmath{$\hat{x'}$}}+\ddot{y'}_i \mbox{\boldmath{$\hat{y'}$}}+2\mbox{\boldmath{$\omega$}}\times(\dot{x'}_i\mbox{\boldmath{$\hat{x'}$}}+\dot{y'}_i\mbox{\boldmath{$\hat{y'}$}})+\\
& &\mbox{\boldmath{$\omega$}}\times\left(\mbox{\boldmath{$\omega$}}\times(x'_i \mbox{\boldmath{$\hat{x'}$}}+y'_i \mbox{\boldmath{$\hat{y'}$}})\right)+\dot{\mbox{\boldmath{$\omega$}}}\times(x'_i \mbox{\boldmath{$\hat{x'}$}}+y'_i \mbox{\boldmath{$\hat{y'}$}})\end{aligned}$$ and so, finally, Eq.\[equat\].
Allen, C., Santillan, A. 1991, RevMexAA, 22, 255 Aarseth, S.J. 1985, in ‘Multiple time scales’, Acad. Press, 378 Aarseth, S. J. 1999, PASP, 111, 1333 Barnes, J. & Hut, P. 1986, , 324, 446 Baumgardt, H., Makino, J. 2003, , 340, 227 Belokurov, V., Evans, N. W., Irwin, M. J., Hewett, P. C., Wilkinson, M. I. 2006, ApJL,637,29 Capuzzo Dolcetta, R., Di Matteo, P. & Miocchi, P. 2005, AJ, 129, 1906 Casertano, S., Hut, P. 1985, , 298, 80 Chumak, Ya. O., Rastorguev, A. S. 2006, Astronomy Letters, 32, 446 Combes, F., Leon, S, Meylan, G. 1999, , 352, 149 Da Costa, G. S., Freeman, K. C. 1976, , 206, 128 Dehnen, W., Odenkirchen, M., Grebel, E. K., Rix, H. W. 2004, , 127, 2753 Di Matteo, P., Capuzzo Dolcetta, R., Miocchi, P., 2005, Celest. Mech., 91, 59 Dinescu, D.I., Girard, T.M. & van Altena, W.F. 1999, , 117, 1792 Dominguez, I., Chieffi, A., Limongi, M. & Straniero, O. 1997, , 524-1, 226 Fall, S. M., Zhang, Q. 2001, , 561, 751 Grillmair, C. J., Freeman, K. C., Irwin, M., Quinn, P. J. 1995, , 109, 2553 Grillmair, C. J. 1998, in D. Zaritsky ed., ASP Conf. Ser. vol.136, Galactic Halos: A UC Santa Cruz Workshop, Astron. Soc. Pac., San Francisco, p.45 Grillmair, C. J., Dionatos, O. 2006, , 641, L37 Grillmair, C. J., Johnson, R. 2006, , 639, L17 Harris, W.E. 1996, AJ, 112, 1487 Hernquist, L. & Katz, N. 1989, , 70, 419 Johnston, K. V. 1998, , 495, 297 Keenan, D. W., Innanen, K. A. 1975, , 80, 290 King, I. R. 1962, AJ, 67, 471 King, I. R. 1966, , 71, 276 Kroupa, P., 2001, MNRAS, 322, 231 Lee, H. M., Goodman, J. 1995, , 443, 109 Lehmann, I., Scholz, R.D. 1997, , 320, 776 Lee, K. H., Lee, H. M., Fahlman, G. G., Lee, M. G. 2003, , 126, 815 Leon, S., Meylan, G., Combes, F. 2000, , 359, 907 Majewski, S. R., Skrutskie, M. F., Weinberg, M. D., Ostheimer, J. C. 2003, ApJ, 599, 1082 Martínez-Delgado, D., Aparicio, A., Gómez-Flechoso, M. A., Carrera, R. 2001, ApJ, 549, L199 Martínez-Delgado, D., Gómez-Flechoso, M. A., Aparicio, A., Carrera, R. 2004, ApJ, 601, 242 Mateo, M., Olszewski, E. W., Morrison, H. L. 1998, ApJ, 508, L55 Miocchi, P. & Capuzzo-Dolcetta, R. 2002, , 382, 758 Miocchi, P., Capuzzo-Dolcetta, R., Di Matteo, P., & Vicari, A. 2006, ApJ, 644, 940 Miocchi, P., Capuzzo-Dolcetta, R., Di Matteo, P., contribution to “Globular Clusters: Guides to Galaxies”, March 6th-10th, 2006, astro-ph/0605008 Murali, C., Weinberg, M. D. 1997a, , 291, 717 Murali, C., Weinberg, M. D. 1997b, , 288, 749 Miyamoto, M. & Nagai, R. 1975, Pub.Astr.Soc.Japan, 27, 533 Odenkirchen, M., Grebel, E. K., Rockosi, C. M., Dehnen, W., Ibata, R., Rix, H. W., Stolte, A., Wolf, C., Anderson, J. E. Jr., Bahcall, N. A., Brinkmann, J., Csabai, I., Hennessy, G., Hindsley, R. B., Ivezic, Z., Lupton, R. H., Munn, J. A., Pier, J. R., Stoughton, C., York, D. G. 2001, , 548, L165 Odenkirchen, M., Grebel, E. K., Dehnen, W., Rix, H. W., Cudworth, K. M. 2002, , 124, 1497 Odenkirchen, M., Grebel, E, K., Dehnen, W, Rix, H. W., Yanny, B., Newberg, H. J., Rockosi, C. M., Martínez-Delgado, D., Brinkmann, J., Pier, J. R. 2003, , 126, 2385 Oh, K. S., Lin, D. N. C. 1992, ApJ, 386, 51 Pryor, T., Meylan, G. 1993, in Structure and Dynamics of Globular Clusters, ed. S. G. Djorgovski, & G. Meylan (San Fransisco), ASP Conf. Ser., 50, 357 Siegel, M. H., Majewski, S. R., Cudworth, K. M., Takamiya, M. 2001, , 121, 935 Straniero, O., Chieffi, A. & Limongi, M. 1997, , 490, 425. Testa, V., Zaggia, S. R., Andreon, S., Longo, G., Scaramella, R., Djorgovski, S. G., de Carvalho, R. 2000, , 356, 127 Yanny, B., Newberg, H.J., Kent, S., Laurent-Muehleisen, S. A., Pier, J. R., Richards, G. T., Stoughton, C., Anderson, J. E., Jr., Annis, J., Brinkmann, J., Chen, B., Csabai, I., Doi, M., Fukugita, M., Hennessy, G. S.; Ivezić, $\breve{Z}$., Knapp, G. R., Lupton, R., Munn, J. A., Nash, T., Rockosi, C. M., Schneider, D. P., Smith, J. A., York, D. G.2000, ApJ, 540, 825 Yim, K., Lee, H. M. 2002, JKAS, 35, 75 Willman, B., Masjedi, M., Hogg, D. W., Dalcanton, J. J., Martinez-Delgado, D., Blanton, M., West, A. A., Dotter, A., Chaboyer, B. 2006, astro-ph/0603486
![ Orbits of a point mass with the initial position and velocity given in Table \[tbl3\] in the MW potential (see sect. \[Galaxy\]). Top panel: orbit I; central panel: orbit II; bottom panel: orbit III. \[orbits\]](f1.eps)
![ Some snapshots of the GC moving along the orbit I (see Table \[tbl3\] for the orbital parameters). One time unit corresponds to about $1$ Myr. \[orbitaI\]](f2.eps)
![Some snapshots of the GC moving along the orbit II (see Table \[tbl3\] for the orbital parameters). One time unit corresponds to about $1$ Myr. \[orbitaII\]](f3.eps)
![Some snapshots of the GC moving along the orbit III (see Table \[tbl3\] for the orbital parameters). One time unit corresponds to about $1$ Myr. \[orbitaIII\]](f4.eps)
![Interpretation of the S-shape of the inner tidal tail around a globular cluster. The different terms in the rhs of the displayed equation (Eq. 7) are represented in the plot as arrows of different line styles. Note that the last term in the equation is here plotted antiparallel to the Coriolis term as it occurs when the GC moves from pericenter to apocenter (see text). The Galactic potential (included in the first and second term) is assumed, for simplicity, spherical. The cross in the lower part of the figure represents the Galaxy center. \[coriolis\]](f5.eps)
![ Tails direction for the GC moving on orbit I. From top to the bottom: Plot of the GC orbit (some points along the orbit are marked with different symbols) ; the cross indicates the Galaxy center. 2: distance in kpc of the GC from the Galaxy center, as a function of time. The different symbols correspond to that in the previous panel. 3: GC orbital angular velocity in [*rad/Myr*]{}, as a function of time. 4: angular acceleration in [*rad/$Myr^2$*]{}.5: solid curve: angle $\alpha$ formed by the inner part of the tails and the galactic center direction, vs time; dashed curve: angle $\beta$ formed by the inner part of the tails and the cluster velocity, vs time. Both angles are in degrees. \[tail-radI\]](f6.eps)
![Same as Fig.\[tail-radI\], but for the GC moving on orbit II. \[tail-radII\]](f7.eps)
![Same as Fig.\[tail-radI\], but for the GC moving on orbit III. \[tail-radIII\]](f8.eps)
[rcc]{} Bulge & $M_B$ & $1.41 \times 10^{10} M_{\odot}$\
& $b_B$ & 387.3 pc\
Disk & $M_D$& $8.56\times 10^{10} M_{\odot}$\
& $a_D$ & 5317.8 pc\
& $b_D$ & 250.00 pc\
Halo & $M_H$ &$10.7\times10^{10} M_{\odot}$\
& $a_H$& 12000 pc
[cccccc]{} $M_{GC}[M_{\odot}]$ & $r_c$ \[pc\] & $r_t$ \[pc\] & $c=log(r_t/r_c)$ & $\sigma$\[km/s\]\
& & & & &\
$4.7 \times 10^5$ &2.2 & 23.5 & 1.03 & 10.5\
$4.7 \times 10^5$ &3.5 & 22.1 & 0.81 & 10.5
[cccccccc]{} Orbit ID & $x$\[pc\] & $y$\[pc\] & $z$\[pc\] & $v_x$\[km/s\] & $v_y$\[km/s\] & $v_z$\[km/s\] & $e$\
& & & & & & &\
I &0. & 0. & 4000. & 0. & 53.3 & 0. & 0.8\
II & 0. & 0. & 2500. & 0. & 116.7 & 0. & 0.7\
III & 0. & 0. & 1500. & 0. & 200.0 & 0. & 0.1
[^1]: One of the three eigenvectors of the inertia tensor is parallel to the $z$-axis, while the other two lye in the ($x,y$) plane. Among these latter two, the eigenvector associated to the greatest eigenvalue gives the direction of maximum elongation of the stellar stream.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We give a uniform construction of free pseudospaces of dimension $n$ extending work in [@BP]. This yields examples of $\omega$-stable theories which are $n$-ample, but not $n+1$-ample. The prime models of these theories are buildings associated to certain right-angled Coxeter groups.'
author:
- Katrin Tent
title: 'The free pseudospace is $n$-ample, but not $(n+1)$-ample '
---
Introduction
============
In the investigation of geometries on strongly minimal sets the notion of ampleness plays an important role. Algebraically closed fields are $n$-ample for all $n$ and it is not known whether there are strongly minimal sets which are $n$-ample for all $n$ and do not interpret an infinite field. Obviously, one way of proving that no infinite field is interpretable in a theory is by showing that the theory is *not* $n$-ample for some $n$.
In [@BP], Baudisch and Pillay constructed a free pseudospace of dimension $2$. Its theory is ${\omega}$-stable (of infinite rank) and $2$-ample. F. Wagner posed the question whether this example was $3$-ample or not.
In Section \[sec:construction\] we give a uniform construction of a free pseudospace of dimension $n$ and show that it is $n$-ample, but not $n+1$-ample. It turns out that the theory of the free pseudospace of dimension $n$ is the first order theory of a Tits-building associated to a certain Coxeter diagram and we will investigate this connection in Section \[sec:building\].
In the final section we show that there are exactly two orthogonality classes of regular types.
Construction and results {#sec:construction}
========================
Fix a natural number $n\geq 1$. Let $L_n$ be the language for $n+1$-coloured graphs containing predicates $V_i,i=0,\ldots n$ and an edge relation $E$.
By an $L_n$-graph we mean an $n+1$-coloured graph with vertices of types $V_i, i=0,\ldots n$ and an edge relation $E\subseteq \bigcup_{i=1,\ldots n}V_{i-1}\times V_i$. We say that a path in this graph is of type $E_i$ if all its vertices are in $V_{i-1}\cup V_i$ and of type $E_i\cup \ldots \cup E_{i+j}$ if all its vertices are in $V_{i-1}\cup\ldots\cup V_{i+j}$
The free pseudospaces will be modeled along the lines of a projective space, i.e. we will think of vertices of type $V_i$ as $i$-dimensional spaces in a . Therefore we extend the notion of incidence as follows:
\[d:incidence\]
1. \[d:incidence1\] We say that a vertex $x_i$ of type $V_i$ is *incident* to a vertex $x_j$ of type $V_j$ if there are vertice $x_l$ of type $V_l, l=i+1\ldots j$ such that $E(x_{l-1},x_l)$ holds. In this case the sequence $(x_i,\ldots x_j)$ is called a *dense flag*. A *flag* is a sequence of vertices $(x_1,\ldots x_k)$ in which any two vertices are incident.
2. The *residue* $R(x)$ of a vertex $x$ is the set of vertices incident with $x$.
3. We say that two vertices $x$ and $y$ *intersect in the vertex* $z$ and write $z=x\wedge y$ if the set of vertices of type $V_0$ incident with $x$ and $y$ is exactly the set of vertices of type $V_0$ incident with $z$. If there is no vertex of type $V_0$ incident to $x$ and $y$, we say that $x$ and $y$ intersect in the empty set.
4. We say that two vertices $x$ and $y$ *generate the vertex* $z$ and write $z=x\vee y$, if the set of vertices of type $V_n$ incident with $x$ and $y$ is exactly the set of vertices of type $V_n$ incident with $z$. If there is no vertex of type $V_n$ incident to $x$ and $y$, we say that $x$ and $y$ generate the empty set.
5. A *simple cycle* is a cycle without repetitions.
We now give an inductive definition of a of dimension $n$:
A of dimension $1$ is a free pseudoplane, i.e. an $L_1$-graph which does not contain any cycles and such that any vertex has infinitely many neighbours.
Assume that a free pseudospace of dimension $n-1$ has been defined. Then a of dimension $n$ is an $L_n$-graph such that the following holds:
1. 1. The set of vertices of type $V_0\cup \ldots \cup V_{n-1}$ is a .
2. The set of vertices of type $V_1\cup \ldots \cup V_n$ is a free pseudospace of dimension $(n-1)$.
2. 1. For any vertex $x$ of type $V_0$, $R(x)$ is a free pseudospace of dimension $(n-1)$.
2. For any vertex $x$ of type $V_n$, $R(x)$ is a free pseudospace of dimension $(n-1)$.
3. 1. Any two vertices $x$ and $y$ intersect in some vertex $z$ or the emptyset.
2. Any two vertices $x$ and $y$ generate some vertex $z$ or the emptyset.
4. 1. If $a$ is a vertex of type $V_n$ and $\gamma=(a,b,\ldots , b',a)$ is a simple cycle of length $k$, then there is an $E_1\cup\ldots \cup E_{n-1}$-path from $b$ to $b'$ of length at most $k-1$ in $R(a)$ all of whose $V_0$-vertices appear in $\gamma$.
2. If $a$ is a vertex of type $V_0$ and $\gamma=(a,b,\ldots , b',a)$ is a simple cycle of length $k$, then there is an $E_2\cup\ldots \cup E_n$-path from $b$ to $b'$ of length at most $k-1$ in $R(a)$ all of whose $V_n$-vertices appear in $\gamma$.
Let $T_n$ denote the $L_n$-theory expressing these axioms.
Note that the inductive nature of the definition immediately has the following consequences:
1. The induced subgraph on $V_j\cup\ldots\cup V_{j+m}$ is a free pseudospace of dimension $m$.
2. If $a$ is a vertex of type $V_{j+m}$ and $\gamma=(a,b,\ldots , b',a)$ is a simple cycle of length $k$ contained in $V_j\cup\ldots\cup V_{j+m}$, then there is an $E_{j+1}\cup\ldots \cup E_{j+m-2}$-path from $b$ to $b'$ of length at most $k-1$ in $R(a)$ all of whose $V_j$-vertices appear in $\gamma$.
3. The notion of a is *self-dual*: if we put $W_i=V_{n-i}, i=0,\ldots n$, then $W_0,\ldots W_n$ with the same set of edges is again a .
Our first goal is to show that $T_n$ is consistent and complete.
Let $A$ be a finite $L_n$-graph. The following extensions are called elementary strong extensions of $A$:
1. add a vertex of any type to $A$ which is connected to at most one vertex of $A$ of an appropriate type.
2. If $(x,y,z)$ is a dense flag in $A$, add a vertex of the same type as $y$ to $A$ which is connected to both $x$ and $z$.
We write $A\leq B$ if $B$ arises from $A$ by finitely many elementary strong extensions.
Let $\K_n$ be the class of finite $L_n$-graphs $A$ such that the following holds
1. $A$ does not contain any $E_i$-cycles for $i=1,\ldots n$.
2. If $a\neq a'$ are in $A$, they intersect in a vertex of $A$ or the emptyset.
3. If $a\neq a'$ are in $A$, they generate a vertex of $A$ or the emptyset.
4. If $(b,a,b')$ is a path with $a\in V_i,b,b'\in V_{i-1}$, and $\gamma=(a,b,\ldots , b',a)$ is an $E_i\cup E_{i-j}$ path of length $k$, then there is some $E_{i-1}\cup \ldots E_{i-j}$-path from $b$ to $b'$ of length at most $k-1$ in $R(a)$ with all $V_{i-j}$-vertices occurring in $\gamma$.
5. If $(b,a,b')$ is a path with $a\in V_i,b,b'\in V_{i+1}$, and $\gamma=(a,b,\ldots , b',a)$ is an $E_i\cup\ldots\cup E_{i+j}$ path of length $k$, then there is some $E_{i+1}\cup\ldots \cup E_{i+j}$-path from $b$ to $b'$ of length at most $k-1$ in $R(a)$ with all $V_{i+j}$-vertices occurring in $\gamma$.
Note that $\K_n$ is closed under finite substructures. We next show that $(\K,\leq)$ has the amalgamation property for strong extensions.
For any finite $L_n$-graphs $A\subseteq B,C$ we denote by $B\otimes_AC$ the *free amalgam* of $B$ and $C$ over $A$, i.e. the graph on $B\cup C$ containing no edges between elements of $B\setminus A$ and $C\setminus A$.
If $A\leq B,C$ are in $\K_n$, then $D:=B\otimes_AC\in\K$ and $B,C\leq D$.
Clearly, $B,C\leq D$. To see that $D\in\K_n$, note that if $B\in \K_n$ and $B'$ is an elementary strong extension of $B$, then also $B'\in\K_n$. This is clear for strong extensions of type 1. For strong extensions of type 2. suppose that $(b,a,b')$ is a path with $a\in V_i,b,b'\in V_{i-1}$, and $\gamma=(a,b,\ldots , b',a)\subset B'$ is an $E_i\cup\ldots\cup E_{i-j}$-path of length $k$ containing the new vertex $y$. Since the new vertex has exactly two neighbours $y_1,y_2$, this implies that the vertex is of type $V_m$ for some $i-j\leq m\leq i$ and $(y_1,y,y_2)$ is contained in $\gamma$. By construction of strong extensions, there is some $z\in B$ such that $(y_1,z,y_2)$ is a path. Hence we may replace all occurrences of $y$ in $\gamma$ by $z$. Then $\gamma$ is contained in $B$ and we find the required path in $R(a)$ with all $V_{i-j}$-vertices occurring in $\gamma$.
This shows that the class $(\K_n,\leq)$ has a Hrushovski limit $M_n$, i.e. a countable $L_n$-structure $M_n$ whose strong subsets are exactly the $L_n$-graphs in $\K_n$ and which is homogeneous for strong subsets: if $A,B\leq M_n$ then any isomorphism from $A$ to $B$ extends to an automorphism of $M_n$. Here we say as usual that a subset $A$ of $M_n$ is strong in $M_n$ if $A\cap B\leq B$ for any finite set $B\subset M_n$.
\[p:pseudospace\] The Hrushovski limit $M_n$ is a model of $T_n$.
By construction, $V_i\cup\ldots\cup V_{i+j}$ satisfies $(\Sigma 3)_j$ and $(\Sigma 4)_j$ for any $i,j$. In particular, $M_n$ satisfies $(\Sigma 3)_n$ and $(\Sigma 4)_n$.
[**$(\Sigma 1)_n$:**]{} In order to show that $M$ satisfies $(\Sigma 1)_n$, we first note that $V_i\cup V_{i+1}$ is a free pseudoplane for all $i=0,\ldots n-1$. Assume inductively that $V_j\cup\ldots \cup V_{j+i}$ is a free pseudospace of dimension $i$. To see that $V_j\cup\ldots \cup V_{j+i+1}$ is a free pseudospace of dimension $i+1$, we need only verify $(\Sigma 2)_{i+1}$. Hence we have to show that for $a\in V_j$ the residue $R(a)\cap (V_j\cup\ldots \cup V_{j+i+1})$ is a free pseudospace of dimension $i$. We know by induction that $R(a)\cap (V_j\cup\ldots \cup V_{j+i})$ is a free pseudospace.
Clearly, $$R(a)\cap (V_{j+1}\cup\ldots \cup V_{j+i+1})=\bigcup \{R(b)\cap (V_{j+1}\cup\ldots \cup V_{j+i+1})\colon b\in V_{j+1}, E(a,b) \}.$$ For each neighbour $b\in V_{j+1}$ of $a$, the set $R(b)\cap (V_{j+1}\cup\ldots \cup V_{j+i+1})$ is a free pseudospaces of dimension $i-1$ by induction. Since $(V_{j+1}\cup\ldots \cup V_{j+i+1})$ is a free pseudospace of dimension $i$, $(\Sigma 2)_{i+1}$ follows from the induction hypothesis. Hence $V_0\cup\ldots \cup V_{n-1}$ and $V_1\cup\ldots \cup V_n$ are free pseudospaces of dimension $n-1$.
[**$(\Sigma 2)_n$:**]{} The proof of $(\Sigma 2)_n$ is similar.
We say that a model $M$ of $T_n$ is $\K_n$-saturated if for all finite $A\leq M$ and strong extensions $C$ of $A$ with $C\in\K_n$ there is a strong embedding of $C$ into $M$ fixing $A$ elementwise. Clearly, by construction, $M_n$ is $\K_n$-saturated.
\[l:K-sat\] A model $M$ of $T_n$ is $\omega$-saturated if and only if $M$ is $\K_n$-saturated.
Let $M$ be an $\omega$-saturated model of $T_n$. To show that $M$ is $\K_n$-saturated, let $A\leq M$ and $A\leq B\in\K_n$. By induction we may assume that $F$ is an elementary strong extension of $A$ and it is easy to see that $F$ can be imbedded over $A$ into $M$.
The theory $T_n$ is complete.
Let $M$ be a model of $ T_n$. In order to show that $M$ is elementarily equivalent to $M_n$ choose an $\omega$-saturated $M'\equiv M$. By Lemma \[l:K-sat\], $M'$ is $\K_n$-saturated. Now $M'$ and $M_n$ are partially isomorphic and therefore elementarily equivalent.
We will see in Section \[sec:building\] that $T_n$ is the theory of the building of type $A_{\infty, n+1}$ with infinite valencies.
Following *[@BP]* we call a subset $A$ of a model $M$ of $T_n$ *nice* if
1. any $E_i$-path between elements of $A$ lies entirely in $A$ and
2. if $a,b\in A$ are connected by a path in $M$ there is a path from $a$ to $b$ inside $A$.
\[l:strong=nice\]\[r:qftp\] Note that a subset $A$ of $M_n$ is strong in $M_n$ if and only if it is nice. (This follows immediately from the definition of strong extension.)
Since $M_n$ is homogeneous for strong subsets, for any nice subset of $M_n$ the quantifier-free type determines the type, i.e. if $\bar a,\bar b$ are nice in $M_n$ and such that $\operatorname{qftp}(\bar a)=\operatorname{qftp}(\bar b)$, then $\operatorname{tp}(\bar a)=\operatorname{tp}(\bar b)$.
We now work in a very saturated model $\overline{M}$ of $T_n$.
\[l:strongext\] If $A$ is a finite set, there is a nice finite set $B$ containing $A$.
Since single vertices are nice it suffices to prove the following
[**Claim:**]{} *If $A$ is nice and $a$ arbitrary, then there is a nice finite set $B$ containing $A\cup\{a\}$.*
*Proof of Claim:* Of course we may assume $a\notin A$. If there is no path from $a$ to $A$, clearly $A\cup\{a\}$ is nice. Hence we may also assume that there is some path $\gamma=(a=x_0,\ldots b)$ for some $b\in A$ and $\gamma\cap A=\{b\}$. It therefore suffices to prove the claim for the case where $a$ has a neighbour in $A$. If $a$ has two neighbours $x,y\in A$ then $(x,a,y)$ is a dense flag and $A\cup\{a\}$ is nice.
Now assume that $a\in V_i$ has a unique neighbour of type $V_{i+1}$ in $A$. (The other case then follows by self-duality.) If the $E_i$-connected component of $a$ does not intersect $A$, then again $A\cup\{a\}$ is nice. Otherwise there is some $E_i$-path $\gamma=(x_0=a,\ldots x_m=b)$ in $M_n$ with $\gamma\cap A=\{b\}$. If for some $V_{i-1}$-vertex $x_k$ of $\gamma$ there is an $E_{i-1}$-path to some $c\in A$, then the $E_i$-path from $c$ to $b$ extends $(x_k,\ldots x_m=b)$ and is entirely contained in $A$ since $A$ is nice. Since $\gamma\cap A=\{b\}$, no such $x_k$ exists implying that $A\cup\gamma$ is nice.
Let us say that $\gamma$ *changes direction in $x_i$* if $x_i\in V_j$ and either $x_{i-1},x_{i+1}\in V_{j-1}$ or $x_{i-1},x_{i+1}\in V_{j-1}$ for some $j$. Clearly a path which doesn’t change direction is a dense flag.
We will use the following easy consequence of $(\Sigma 4)_{j\leq n}$.
\[l:cycle\] If $\gamma=(a,b,\ldots, b',a)$ is a simple cycle, then any two vertices $x,y\in \gamma\cap (V_{j-1}\cup V_j)$ are joined by an $E_j$-path.
Suppose that $j,j+m$ are minimal and maximal, respectively, with $\gamma\cap V_j, V_{j+m}\neq\emptyset$. Apply $(\Sigma 4)_m$ to all vertices of $\gamma\cap V_{j+m}$ and replace any path of the form $(x,y,z)$ with $y\in V_{j+m}$ by a path from $x$ to $z$ contained in $V_j\cup\ldots \cup V_{j+m-1}$. This yields a new path contained in $V_j\cup\ldots \cup V_{j+m-1}$. Next apply $(\Sigma 4)_{m-1}$ to each vertex of type $V_j$ to obtain a path in $V_{j+1}\cup\ldots \cup V_{j+m-1}$. Alternating between replacing the top and bottom extremal peaks by a path using fewer levels, we end up with an $E_k$-path for some $j\leq k\leq j+m-1$ between vertices of $\gamma$. Since an $E_k$-path changes direction at every vertex, we can apply the same procedure in order to replace this path by an $E_{k'}$-path for any $j\leq k'\leq j+m-1$.
\[def:reduced\] We call a path $\gamma=(x_0,\ldots x_k)\subseteq V_j\cup\ldots \cup V_{j+m}$ *reduced* if $\gamma$ is a flag or if the following holds:
1. if $m=1$ the path $\gamma$ is reduced if it does not contain any repetition.
2. for any simple cycle $\gamma'$ containing $(x_i,\ldots, x_{i+q})$ replacing all $(y,x,z)$ with $x\in (x_{i+1},\ldots, x_{i+q-1})\cap V_{j+m}$ by an $E_{j+m-1}$-path in $R(x)$ yields a reduced path $(x_i,\ldots, y,\ldots z,\ldots, x_{i+q})\subseteq V_j\cup\ldots \cup V_{j+m-1}$
Note that the definition and Lemma \[l:cycle\] immediately imply the following:
\[r:reducedpath\] Suppose that every reduced path from $a$ to $b$ contains $x$ and let $\gamma_1,\gamma_2$ be paths from $a$ to $x$ and from $x$ to $b$ respectively. Then the path $\gamma_1\gamma_2$ is reduced if and only if $\gamma_1$ and $\gamma_2$ are.
If $\gamma=(a,\ldots, b)$ is a reduced path and changes direction in $x\in V_k$, then either every reduced path from $a$ to $b$ contains $x$ or the path $(y,x,z)\subseteq \gamma$ can be replaced by an $E_{k-1}$-path (or by an $E_k$-path) $(y,\ldots z)$ such that $(a,\ldots, y,\ldots, z,\ldots b)$ is still reduced.
\[l:nicereplacement\] Suppose $\gamma=(a=x_0,\ldots, x_s=b)$ is a reduced path from $a$ to $b$ which can be completed into a simple cycle $(a,x_1,\ldots, b,\ldots,a)$. If $\gamma$ changes direction in $x\in V_k$, then there is a reduced path $\gamma'=(a=y_0,\ldots, y_t=b)$ not containing $x$ such that $x$ is on the $E_{k}$-(or $E_{k+1}$-) path between appropriate elements of $\gamma'$.
To fix notation let us assume that the neighbours of $x$ in $\gamma$ are $y,z\in V_{k-1}$. By Lemma \[l:cycle\] we may replace the path $(y,x,z)$ by an $E_k$-path. This path is obviously still reduced and $x$ is on the $E_{k+1}$-path from $y$ to $z$.
\[l:nicepath\] Suppose $D$ is a nice set containing an $E_j$-path from $a$ to $b$. Suppose furthermore that $\gamma=(a=y_0,\ldots, y_t=b)$ is a reduced path such that the composition with the $E_j$-path is a simple cycle. If $\gamma$ changes direction in $c$, then $c\in D$.
Using the fact that $M_n$ is $\omega$-saturated and homogeneous for strong substructures we can now describe the algebraic closure:
\[l:acl2\] The algebraic closure $\operatorname{acl}(a,b)$ is the intersection of all nice sets containing $a,b$.
Since there are finite nice sets containing $a,b$, the intersection of all nice sets containing $a,b$ is certainly finite and invariant over $a,b$, hence contained in $\operatorname{acl}(a,b)$.
Conversely let $D$ be a nice set containing $\{a,b\}$ and $c\notin D$. If there is no reduced path between $a$ and $b$ changing direction in $c$, then $c$ has infinitely many conjugates over $ab$, so $c\notin\operatorname{acl}(ab)$.
So suppose there is a reduced path $\gamma'$ from $a$ to $b$ changing direction in $c$. Let $\gamma\subset D$ be a path from $a$ to $b$. Composing these paths we obtain a simple cycle $\gamma"$ through $c$ consisting of subpaths $\gamma_1'\subseteq \gamma_1$ of $\gamma'\subseteq\gamma$, respectively, from some $a'$ to $b'$. In particular $\gamma_1$ is not a flag. We may assume inductively that $\gamma_1'\cap D=\{a',b'\}$. If $a'\in V_j$, then possibly after exchanging the roles of $a',b'$, there is some $x\in\gamma\cap (V_{j-1}\cup V_j)$ closest to $b'$. Then $a',x$ are connected by an $E_j$-path entirely contained in $D$ and $(x,\ldots b')$ is a flag. Since the $E_j$-path changes direction in every vertex, we also find an $E_{j-1}$-path inside $D$ to the next vertex of the flag $(x,\ldots b')$. Since $\gamma_1'$ doesn’t meet $D$ except in $a',b'$, we see that ${\gamma'_1}^\smallfrown (b',\ldots, x)$ is a reduced path. Now Lemma \[l:nicepath\] implies $c\in D$, a contradiction. Hence $c\notin\operatorname{acl}(ab)$.
Note that the proof shows in fact the following:
\[r:flagcl\]A vertex $x\neq a,b$ is in $\operatorname{acl}(ab)$ if and only if there is a reduced path from $a$ to $b$ that changes direction in $x$. Hence $\operatorname{acl}(ab)=\{a,b\}$ if and only if $a,b$ is a flag or $a$ and $b$ are not connected. In fact, we have $\operatorname{dcl}(ab)=\operatorname{acl}(ab)$.
\[l:treeacl\] If $g\in\operatorname{acl}(A)$, there exist $a,b\in A$ with $x\in\operatorname{acl}(ab)$.
We may assume that $A$ is finite. By induction it suffices to prove that if $d\in\operatorname{acl}(bc),g\in\operatorname{acl}(ad)$, then $g\in\operatorname{acl}(ab)\cup\operatorname{acl}(bc)\cup\operatorname{acl}(ac)$.
By Corollary \[r:flagcl\] there is a reduced path $\gamma_1=(b,\ldots,d,\ldots, c)$ changing direction in $d$ and a reduced path $\gamma_2=(d,\ldots ,g,\ldots, a)$ changing direction in $g$. If $\gamma_1\cup\gamma_2\in V_i\cup V_{i+1}$ for some $i$, then clearly either $(a,\ldots,g\ldots,d,\ldots,b)$ or $(a,\ldots,g\ldots,d,\ldots,c)$ is reduced.
Now assume that $\gamma_1\cup\gamma_2\in V_i\cup\ldots\cup V_{i+k}$. Clearly we may assume that $d$ is not contained in every reduced path from $a$ to $b$ or in every reduced path from $a$ to $c$. Furthermore, we may reduce to the case where no vertex of $\gamma_1$ is contained in every reduced path from $b$ to $c$ and no vertex of $\gamma_2$ is contained in every reduced path from $a$ to $d$.
By Remark \[r:reducedpath\] we may therefore replace any element $x\in V_{i+k}$ in the interior of $\gamma_1$ and $\gamma_2$ by an $E_{i+k-1}$ path between its neighbours. If $d\in V_{i+k}$, we may replace $d$ as follows: the neighbour $w$ of $d$ in $\gamma_2$ either appears in the $E_{i+k-1}$ path between the neighbours of $d$ in $\gamma_1$ or that path extends to $w$ on one of its ends. We may thus replace the paths $(b,\ldots,d,\ldots, c)$ and $(a,\ldots g,\ldots, d)$ by reduced paths $(b,b_1,\ldots,w, \ldots,c_1,c)$ and $(a,a_1,\ldots w)$ contained in $V_i\cup\ldots\cup V_{i+k-1}$, except possibly for the endpoints. We may assume that at most $a\in V_{i+k}$ since we may exchange the roles of $V_i$ and $V_{i+k}$ and replace the elements of $V_i$ instead.
By induction assumption, at least one of $(a_1,\ldots w,\ldots b_1)$ and $(a_1,\ldots w,\ldots c_1)$ is reduced, and replacing the new $E_{i+k-1}$-paths by the old $E_{i+k}$-paths, this path remains reduced and changes direction in $g$. If $a,b\in V_{i+k}$, then we may exchange the roles of $V_i$ and $V_{i+k}$ and replace the elements of $V_i$ instead. Thus we see that the path remains reduced when adding the endpoint.
\[r:acl\] For any set $A$, the algebraic closure $\operatorname{acl}(A)$ is the intersection of all nice sets containing it.
Clearly, we may restrict ourselves to finite sets $A$. As before we see that the intersection of all nice sets containing $A$ is contained in $\operatorname{acl}(A)$.
For the converse, assume that $c$ is not in the intersection of all nice sets containing $A$. If $c$ was in $\operatorname{acl}(A)$, then by Lemma \[l:treeacl\] there are $a,b\in A$ with $c\in\operatorname{acl}(ab)$. Then by Lemma \[l:acl2\] $c\in\bigcap_{a,b\in D}D\subset\bigcap_{A\subseteq D}D$, a contradiction.
For any vertex $a$ and set $A$, there is a flag $C\in\operatorname{acl}(A)$ such that for any $b\in\operatorname{acl}(A)$ there is a reduced path from $a$ to $b$ passing through one of the elements of $C$.
The flag $C$ is called the *projection* from $a$ to $A$ and we write $C=\operatorname{proj}(a/A)$. Note that $\operatorname{proj}(a/A)=\emptyset$ if and only if $a$ is not connected to any vertex of $\operatorname{acl}(A)$.
Let $b_1,b_2\in\operatorname{acl}(A)$ and let $\gamma_1=(x_0=a,\ldots ,x_m=b_1)$ and $\gamma_2=(y_0=a,\ldots ,y_l=b_2)$ be reduced paths such that $i,j$ are minimal possible with $x_i,y_j\in\operatorname{acl}(A)$. By composing the initial segments of $\gamma_1$ and $\gamma_2$ and reducing we obtain a reduced path from $x_i$ to $y_j$ intersecting $\operatorname{acl}(A)$ only in $x_i,y_j$ since $x_i, y_j$ were chosen at minimal distance from $a$. By Corollary \[r:flagcl\] $x_i,y_j$ is a flag. Thus the set of such vertices forms a flag $C$.
It is now easy to show the following:
\[T:stable\] The theory $T_n$ is $\omega$-stable.
Let $M$ be a countable model and let $\bar d$ be a tuple from $\overline{M}$. Let $C\in M$ be the finite set of projections from $\bar d$ to $M$. Then the type $\operatorname{tp}(\bar d/M)$ is determined by $\operatorname{tp}(\bar d/C)$. By Lemmas \[l:strong=nice\] and \[l:strongext\], $\bar d\cup C$ is contained in a finite strong subset of $M_n$ and for such subsets the quantifier-free type determines the type by Remark \[r:qftp\]. Hence there are only countably many types over a countable model.
In fact, it is easy to see directly without counting types that $T_n$ is superstable (see Remark \[r:superstable\]).
The has weak elimination of imaginaries.
Let $a$ be a vertex and $A$ any set. Then we can choose $\operatorname{Cb}(stp(a/A))$ as the projection of $a$ on $A$. This is a finite set.
The following immediate corollary will be very useful:
\[c:forking\] The vertex $a$ is independent from $A$ over $C$ if $\operatorname{proj}(a/AC)\subseteq\operatorname{acl}(C)$. In particular, $a$ is independent from $A$ over $\emptyset$ if and only if $a$ is not connected to any vertex of $\operatorname{acl}(A)$.
\[r:superstable\] As in [@TZU] we could have defined a notion of independence on models of $T_n$ by saying $${
A\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{C}B
}$$ if and only if $\operatorname{proj}(a/BC)\subseteq\operatorname{acl}(C)$ for all $a\in\operatorname{acl}(A)$. It is easy to see that this notion of independece satisfies the characterizing properties of forking in stable theories (see [@TZ] Ch. 8) and hence agrees with the usual one. Note that the existence of nonforking extensions follows from the construction of $M_n$ as a Hrushovski limit. Since we have just seen that for any type $\operatorname{tp}(a/A)$ there is a finite set $A_0$ such that ${
a\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{A_0}A
}$ this shows directly (without counting types) that $T_n$ is superstable.
Using this description of forking it is easy to give a list of regular types such that any nonalgebraic type is non-orthogonal to one of these. This is entirely similar to the list given in [@BP] and we omit the details but will return to this point in Section \[sec:types\]. It is also clear from this description of forking that the geometry on these types is trivial.
Ampleness
=========
We now recall the definition of a theory being $n$-ample:
A theory $T$ eliminating imaginaries is called $n$-ample if possibly after naming parameters there are tuples $a_0,\ldots a_n$ in $M$ such that the following holds:
1. for $i=0,\ldots n-1$ we have $$\operatorname{acl}(a_0,.\ldots a_{i-1},a_i)\cap \operatorname{acl}(a_0,.\ldots a_{i-1},a_{i+1})=\operatorname{acl}(a_0,.\ldots a_{i-1});$$
2. ${
a_n\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\mathchardef
\nn=12854\hss$\nn$\kern1.4\wd0\hss}\hbox to 0pt{\hss$\mid$\hss}\lower.9\ht0
\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{}a_0
},$ and
3. ${
a_n\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{a_i}a_0\ldots a_i
}$ for $i=0,\ldots n-1$.
\[T:n-ample\] The theory $T_n$ is $n$-ample and any maximal flag $(x_0,\ldots x_n)$ in $M_n$ is a witness for this.
This follows immediately from the description of $\operatorname{acl}$ in Lemma \[r:acl\] and of forking in Corollary \[c:forking\].
\[T:notample\] The is not $n+1$-ample.
Suppose towards a contradiction that $a_0,\ldots a_{n+1}$ are witnesses for $T_n$ being $n+1$-ample over some set of parameters $A$. We have $${
a_{n+1}\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\mathchardef
\nn=12854\hss$\nn$\kern1.4\wd0\hss}\hbox to 0pt{\hss$\mid$\hss}\lower.9\ht0
\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{A}a_0
},$$ $${
a_{n+1}\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{Aa_i}a_0\ldots a_i
},i=0,\ldots n.$$
By the first condition there are vertices in $\operatorname{acl}(a_0)$ and in $ \operatorname{acl}(a_{n+1})$ which are in the same connected component. Put $f_0=\operatorname{proj}(a_{n+1}/a_0A)\in\operatorname{acl}(a_0A)$ and $f_{n+1}=\operatorname{proj}(f_0/a_{n+1}A)\in\operatorname{acl}(a_{n+1}A)$.
Since $${
a_{n+1}\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{Aa_i}a_0\ldots a_i
}, i=1,\ldots n$$ using Corollary \[c:forking\] we inductively find flags $$f_i=\operatorname{proj}(f_{n+1}/f_0f_1\ldots f_{i-1}a_iA)=\operatorname{proj}(f_{n+1}/a_iA)\in\operatorname{acl}(a_iA),i=1,\ldots n$$ such that $${
f_{n+1}\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{f_i}f_0f_1\ldots f_i
}.$$
For $i=1,\ldots n$ we clearly have $$\operatorname{acl}(f_0,f_1.\ldots f_{i-1},f_i)\cap \operatorname{acl}(f_0,.\ldots f_{i-1},f_{i+1})\subseteq \operatorname{acl}(a_0,.\ldots a_{i-1}A).$$
By construction, there is a reduced path $\gamma=(f_0,x_1,\ldots x_k=f_{n+1})$ containing a vertex of each of the $f_i$ in ascending order. Since we cannot have a flag containing more than $n$ elements, there must be some vertex $x$ in $\gamma$ where $\gamma$ changes direction. For some $i$ we then have $x\in f_{i+1}$ or $x$ occurs in $\gamma$ between an element of $f_i$ and an element of $f_{i+1}$. By Corollary \[r:flagcl\] we have $x\in\operatorname{acl}(f_if_{i+1})\cap\operatorname{acl}(f_if_{i+2})$. Then $${
x\mathop{\mathpalette{
\setbox0=\hbox{$x$}\kern\wd0\hbox to 0pt{\hss$
\mid$\hss}\lower.9\ht0\hbox to 0pt{\hss$\smile$\hss}\kern\wd0
}{}}_{f_i}a_0a_1,\ldots a_iA
},$$ so $x\notin\operatorname{acl}(a_0a_1,\ldots a_iA)$, a contradiction.
The proof shows that in fact the following stronger ampleness result holds:
If $a_0,\ldots a_n$ are witnesses for $T_n$ being $n$-ample, then there are vertices $b_i\in\operatorname{acl}(a_i)$ such that $(b_0,\ldots b_n)$ is a flag.
Buildings and the prime model of $T_n$ {#sec:building}
======================================
We now turn towards constructing the prime model $M_n^0$ of $T_n$ as a Hrushovski-limit. We will show that $M_n^0$ is the building associated to a right-angled Coxeter group.
For this purpose we now consider an expansion $L_n'$ of the language $L_n$ by binary function symbols $f^i_k$. For an $L_n$-graph $A$ we put $f^i_k(x,y)=z$ if $z$ is the $k^{th}$ element on a unique shortest $E_i$-path of length at least $k$ from $x$ to $y$ and $z=x$ otherwise.
We say that an $L_n$-graph $A$ is $E_i$-connected if the set $V_{i-1}(A)\cup V_i(A)$ is connected.
Let $\K'$ be the class of finite $L_n'$-graphs $A\in\K$ which are $E_i$-connected for $i=1,\ldots n$ and additionally satisfy the following condition:
- If $a\in A$ is of type $V_j$, then the residue $R(a)$ is $E_i$-connected for $i=1,\ldots n$.
Note that $\K'$ is closed under finitely generated substructures by the choice of language.
Let $A$ be a finite $L_n'$-graph which is $E_i$-connected for $i=1,\ldots n$. The following extensions are called elementary strong extensions of $A$:
1. add a vertex of type $P$ or $E$ to $A$ which is connected to at most one vertex of $A$ and such that the extension is still $E_i$-connected for all $i=1,\ldots n$.
2. If $(x,y,z)$ is a dense flag in $A$, add a vertex $y'$ of the same type as $z$ to $A$ such that $(x,y',z)$ is a flag.
3. if $|A|\leq 1$ contains no line, add a vertex of appropriate type which is connected to the vertex of $A$ if $A\neq\emptyset$.
Again we write $A\leq B$ if $B$ arises from $A$ by finitely many elementary strong extensions.
We next show that $(\K_n',\leq)$ has the amalgamation property for $\leq$-extensions.
If $A$ contains a flag of type $(V_1,\ldots V_{n-1})$ and $A\leq B,C$ are in $\K_n'$, then $D:=B\otimes_AC\in\K_n'$ and $B,C\leq D$.
Clearly, $B,C\leq D$ and $D$ is $E_i$-connected for all $i=1,\ldots n$ since $A$ contains a flag. To see that $D\in\K$, note that if $B\in \K$ and $B'$ is an elementary strong extension of $B$, then also $B'\in\K$.
This shows that the class $(\K,\leq)$ has a Hrushovski limit $M^0_n$. Clearly, $M^0_n$ is $E_i$-connected for $i=1,\ldots n$ and since any two vertices of $M_n^0$ lie in a maximal flag, it follows that $M_n^0$ is in fact $n$-connected. Note that an $L_n'$-substructure of $M_n^0$ is automatically nice, see Remark \[r:qftp\].
The same proof as in the case of $M_n$ shows the first part of the following proposition:
\[p:primmodell\] The Hrushovski limit $M^0_n$ is a model of $T_n$ and $M^0_n$ is the unique countable model of $T_n$ which is $E_i$-connected for $i=1,\ldots n$ and such that every vertex is contained in a maximal flag.
(Note that in [@BP] the corresponding Remark 3.6 of uses Lemma 3.2, which is not correct as phrased there: $M_n^0$ and $M_n^0\cup\{a\}$ with $a$ an isolated point are not isomorphic, but satisfy the assumptions of Remark 3.6.)
The uniqueness part of Proposition \[p:primmodell\] follows directly from the following theorem and Proposition 5.1 of [@HP] which states that this type of building is uniquely determined by its associated Coxeter group and the cardinality of the residues.
\[t:building\] $M^0_n$ is a building of type $A_{\infty,n+1}$ all of whose residues have cardinality $\aleph_0$.
Recall the following definitions (see e.g. [@G]).
Let $W$ be the Coxeter group
$$W=\langle t_0,\ldots t_n\colon t_i^2=(t_it_k)^2=1, i,k=0\ldots n, |k-i|\geq 2 \rangle,$$ whose associated diagram we call $A_{\infty,n+1}$.
\[d:building\] A building of type $A_{\infty,n+1}$ is a set $\Delta$ with a *Weyl distance function* $\delta: \Delta^2\to W$ such that the following two axioms hold:
1. For each $s\in S:=\{t_i,i=0,\ldots n\}$, the relation $x\sim_s y$ defined by $\delta(x,y)\in\{1,s\}$ is an equivalence relation on $\Delta$ and each equivalence class of $\sim_s$ has at least $2$ elements.
2. Let $w=r_1r_2\ldots r_k$ be a shortest representation of $w\in W$ with $r_i\in S$ and let $x,y\in\Delta$. Then $\delta(x,y)=w$ if and only if there exists a sequence of elements $x,x_0,x_1,\ldots, x_k=y$ in $\Delta$ with $x_{i-1}\neq x_i$ and $\delta(x_{i-1},x_i)=r_i$ for $i=1,\ldots, k$.
A sequence as in 2. is called a *gallery of type* $(r_1,r_2,\ldots, r_k)$. The gallery is called reduced if the word $w=r_1r_2,\ldots, r_k$ is reduced, i.e. a shortest representation of $w$.
We now show how to consider $M_n^0$ as a building of type $A_{\infty,n+1}$.
(of Theorem \[t:building\]) We extend the set of edges of the $n+1$-coloured graph $M_n^0$ by putting edges between any two vertices that are incident in the sense of Definition \[d:incidence\]\[d:incidence1\]. In this way, flags of $M_n^0$ correspond to a complete subgraph of this extended graph, which thus forms a simplicial complex. A maximal simplex consists of $n+1$ vertices each of a different type $V_i$. (Such a simplex is called a *chamber*.) Let $\Delta$ be the set of maximal simplices in this graph. Define $\delta:\Delta^2\to W$ as follows:
Put $\delta(x,y)=t_i$ if and only if the flags $x$ and $y$ differ exactly in the vertex of type $V_i$. Extend this by putting $\delta(x,y)=w$ for a reduced word $w=r_1r_2\ldots r_k$ if and only if there exists a sequence of elements $x=x_0,x_1,\ldots, x_k=y$ in $\Delta$ with $x_{i-1}\neq x_i$ and $\delta(x_{i-1},x_i)=r_i$ for $i=1,\ldots, k$.
Clearly, with this definition of $\delta$, the set $\Delta$ satisfies the first condition of Definition \[d:building\]. In fact, for all $s\in S$ every equivalence class $\sim_s$ has cardinality $\aleph_0$.
We still need to show that $\delta$ is well-defined, i.e. if we have to show the following for any $x,y\in\Delta$: if there are reduced galleries $x_0=x,x_1,\ldots, x_k=y$ and $y_0=x,y_1,\ldots, y_m=y$ of type $(r_1,r_2,\ldots,r_k)$ and $(s_1,\ldots s_m)$, respectively, then in $W$ we have $r_1r_2\ldots r_k=s_1\ldots s_m$. Equivalentely, we will show the following, which completes the proof of Theorem \[t:building\]:
[**Claim:**]{} There is no reduced gallery $a_0,a_1,\ldots, a_k=a_0$ for $k>0$ in $M_n^0$.
*Proof of Claim.* Suppose otherwise. Let $a_0,a_1,\ldots, a_k=a_0$ be a reduced gallery of type $(r_1,\ldots r_k)$ for some $k>0$. Note that the flags $a_i$ and $a_{i+1}$ contain the same vertex of type $V_j$ as long as $r_i\neq t_j$.
Now consider the sequence of vertices of type $V_n$ and $V_{n-1}$ occurring in this gallery. Since $V_n\cup V_{n-1}$ contains no cycles, the sequence of vertices of type $V_n$ and $V_{n-1}$ occurring in this gallery will be of the form $$\label{eq:gallery}
(x_1,y_1,x_2,y_2,\ldots ,x_i,y_i,x_i,y_{i-1},\ldots y_1,x_1)$$ with $x_i\in V_n,y_i\in V_{n-1}$ and $x_i$ a neighbour of $y_i$ and $y_{i-1}$ in the original graph. This implies that at some place in the gallery type there are two occurrences of $t_n$ which are not separated by an occurrence of $t_{n-1}$ (or conversely). Since $t_n$ commutes with all $t_i$ for $i\neq n-1$ and the word $r_1\ldots r_k$ is reduced, there are two occurrences of $t_{n-1}$ which are not separated by an occurrence of $t_n$, say $r_j,r_{j+m}=t_{n-1}$ with $r_{j+1},\ldots, r_{j+m-1}\neq t_n$.
We now consider the gallery $a_j,\ldots a_{j+m}$ of type $(r_j=t_{n-1},r_{j+1},\ldots, r_{j+m}=t_{n-1})$. Notice that by (\[eq:gallery\]), the flags $a_j$ and $a_{j+m}$ have the same $V_n$ and the same $V_{n-1}$ vertex. Since $M_n^0$ does not contain any $E_{n-1}$-cycles, the sequence of $V_{n-1}$- and $V_{n-2}$-vertices appearing in this sequence must again be of the same form as in (\[eq:gallery\]). Exactly as before we find two occurrences[^1] of $t_{n-2}$ in the gallery type of $a_j,\ldots a_{j+m}$ which are not separated by an occurrence of $t_{n-1}$. Continuiung in this way, we eventually find two occurrences of $t_1$ which are not separated by any $t_i$. Since $t_1^2=1$ this contradicts the assumption that the gallery be reduced.
The proof shows in fact the following:
A model of $T_n$ is a building of type $A_{\infty,n+1}$ if and only if it is $E_i$-connected for all $i$ and every vertex is contained in a maximal flag.
The building $M_n^0$ is the prime model of $T_n$
To see that $M_n^0$ is the prime model of $T_n$ note that for any flags $C_1,C_2\in M_n^0$ and gallery $C_1,=x_0,\ldots ,x_k=C_2$ the set of vertices occuring in this gallery is $E_i$-connected for all $i$. Hence by Remark \[r:qftp\] its type is determined by the quantifier-free type.
Thus, given a maximal flag $M$ in any model of $T_n$ and a maximal flag $c_0$ of $M_n^0$ we can embed $M_n^0$ into $M$ by moving along the galleries of $M_n^0$.
Ranks and types {#sec:types}
===============
Recall that for vertices $x,y\in M_n^0$ with $x\in V_i,y\in V_j$ the *Weyl-distance* $\delta(x,y)$ equals $w\in W$ if there are flags $C_1,C_2$ containing $x,y$, respectively, with $\delta(C_1,C_2)=w'$ and such that $w$ is the shortest representative of the double coset $\langle t_k\colon k\neq i\rangle^W w'\langle t_k\colon k\neq j\rangle^W$ (where as usual $\langle X\rangle^W$ denotes the subgroup of $W$ generated by $X$).
The following is clear:
\[p:qe\] The theory $T_n$ has quantifier elimination in a language containing predicates $\delta_w^{i,j}$ for Weyl distances between vertices of type $V_i$ and of type $V_j$.
For any small set $A$ in a large saturated model we have the following kinds of regular types:
1. $\operatorname{tp}(a/A)$ where $a\in V_i$ is not connected to any element in $\operatorname{acl}(A)$
2. $\operatorname{tp}(a/A)$ where $a\in V_i$ is incident with some $b\in\operatorname{acl}(A)\cap V_j$ but not connected in $R(b)$ to any vertex in $\operatorname{acl}(A)\cap R(b)$.
3. $\operatorname{tp}(a/A)$ where $a\in V_i$ is incident with some $x,y\in \operatorname{acl}(A)$ such that $(x,a,y)$ is a flag with $x\in V_k,y\in V_j$; and as a special case of this we have
4. $\operatorname{tp}(a/A)$ where $a\in V_i$ has neighbours $x,y\in \operatorname{acl}(A)$ such that $(x,a,y)$ is a (necessarily dense) flag.
By quantifier elimination any of these descriptions determines a complete type. Using the description of forking in Corollary \[c:forking\] one sees easily that each of these types is regular and trivial.
Clearly, any type in (IV) has $U$-rank $1$ and in fact Morley rank $1$ by quantifier elimination. It also follows easily that $\operatorname{MR}(a/A)=\omega^n$ if $\operatorname{tp}(a/A)$ is as in (I). In case (II) we find that $\operatorname{MR}(a/A)=\omega^{n-j-1}$ or $\operatorname{MR}(a/A)=\omega^{j-1}$ depending on whether or not $i<j$. In case (II) we have $\operatorname{MR}(a/A)=\omega^{|k-j|-2}$.
Just as in [@BP] we obtain:
Any regular type in $T_n$ is non-orthogonal to a type as in (I) or as in (IV).
Let $p=\operatorname{tp}(b/\operatorname{acl}(B))$. If $b$ is not connected to $\operatorname{acl}(B)$, then $p$ is as in (I), so we may assume that $\operatorname{proj}(b/B)=C\neq\emptyset$. Let $a$ be a vertex on a short path from $b$ to $C$ incident with an element of $C$. Then by Corollary \[c:forking\] we see that $p$ is non-orthogonal to $\operatorname{tp}(a/C)$.
[**Acknowledgement:**]{} I would like to thank Linus Kramer, in particular for providing reference [@HP].
[WWW]{} A. Baudisch, A. Pillay, A free pseudospace. J. Symbolic Logic 65 (2000), no. 1, 443 – 460.
Th. Grundhöfer, Basics on buildings. Tits buildings and the model theory of groups (Würzburg, 2000), 1 – 21, London Math. Soc. Lecture Note Ser., 291, Cambridge Univ. Press, Cambridge, 2002.
F. Haglund, F. Paulin, Constructions arborescentes d’immeubles. Math. Ann. 325 (2003), no. 1, 137 – 164.
K. Tent, M. Ziegler, A Course in Model theory, to appear in ASL Lecture Notes in Logic, Cambridge University Press, 2012.
K. Tent, M. Ziegler, On the isometry group of the Urysohn space, submitted.
[^1]: If $t_{n-2}$ does not occur in the type of the gallery, this would contradict the assumption that the type is reduced since $t_{n-1}$ commutes with all $t_i$ for $i\neq n,n-2$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Remo Garattini
title: 'Self Sustained Traversable Phantom Wormholes and Gravity’s Rainbow'
---
Introduction {#sec:1}
============
In recent years, several attempts to modify gravity at Planckian or Transplanckian energy have been considered. The hope is to include quantum gravitational effects in the description of physical phenomena keeping under control the usual Ultraviolet (UV) divergences, avoiding the usual regularization/renormalization scheme. To this purpose, Noncommutative geometry, Gravity’s Rainbow and Generalized Uncertainty Principle (GUP) represent some possibilities to cure the divergences that appear in general relativity[@RemoGRw; @RGPN]. In particular, when Gravity’s Rainbow is considered, spacetime is endowed with two arbitrary functions $g_{1}\left(
E/E_{P}\right) $ and $g_{2}\left( E/E_{P}\right) $ having the following properties$$\lim_{E/E_{P}\rightarrow 0}g_{1}\left( E/E_{P}\right) =1\qquad \text{and}%
\qquad \lim_{E/E_{P}\rightarrow 0}g_{2}\left( E/E_{P}\right) =1.
\label{g1g2}$$$g_{1}\left( E/E_{P}\right) $ and $g_{2}\left( E/E_{P}\right) $ appear into the solutions of the modified Einstein’s Field Equations[@MagSmo]$$G_{\mu \nu }\left( E/E_{P}\right) =8\pi G\left( E/E_{P}\right) T_{\mu \nu
}\left( E/E_{P}\right) +g_{\mu \nu }\Lambda \left( E/E_{P}\right) ,$$where $G\left( E/E_{P}\right) $ is an energy dependent Newton’s constant, defined so that $G\left( 0\right) $ is the low-energy Newton’s constant and $%
\Lambda \left( E/E_{P}\right) $ is an energy dependent cosmological constant. Usually $E$ is the energy associated to the particle deforming the spacetime geometry. Since the scale of deformation involved is the Planck scale, it is likely that spacetime itself begins to fluctuate in such a way to produce a Zero Point Energy (ZPE). In absence of matter fields, the only particle compatible with the deformed Einstein’s gravity is the graviton. What makes a ZPE calculation interesting is that it is strictly related to the Casimir effect. Casimir effect has many applications and it can be considered under different points of view, but it can also be used as a tool to probe another appealing production of the gravitational field theory: a wormhole. A wormhole is often termed Einstein-Rosen bridge because a *bridge* connecting two sheets was the result obtained by A. Einstein and N. Rosen in attempting to build a geometrical model of a physical elementary “particle” that was everywhere finite and singularity free[@ER]. It was J.A. Wheeler who introduced the term wormhole[@JAW], although his wormholes were at the quantum scale. We have to wait for M. S. Morris and K. S. Thorne[@MT] to see the subject of wormholes seriously considered by the scientific community. To exist, traversable wormholes must violate the null energy conditions, which means that the matter threading the wormhole’s throat has to be *exotic*. Usually classical matter satisfies the usual energy conditions, while Casimir energy on a fixed background has the correct properties to substitute the exotic matter: indeed, it is known that, for different physical systems, Casimir energy is negative. Usually one considers some matter or gauge fields which contribute to the Casimir energy necessary to the traversability of the wormholes, nevertheless nothing forbids to use the Casimir energy of the graviton on a background of a traversable wormhole. In this way, one can think that the quantum fluctuations of the gravitational field of a traversable wormhole are the same ones which are responsible to sustain traversability. Note that in Ref.[@DBGL], the ZPE was used as an indicator for a topology change without a Gravity’s Rainbow scheme, while in Ref.[@RGFL], it has been shown that a topology change is a ZPE consequence induced by Gravity’s Rainbow. In this contribution we would like to probe ZPE with the help of Gravity’s Rainbow and an equation of state.
Gravity’s Rainbow and the Equation of State {#p1}
===========================================
In Schwarzschild coordinates, the traversable wormhole metric can be cast into the form$$ds^{2}=-\exp \left( -2\phi \left( r\right) \right) dt^{2}+\frac{dr^{2}}{1-%
\frac{b\left( r\right) }{r}}+r^{2}d\Omega ^{2}. \label{metric}$$where $\phi \left( r\right) $ is called the redshift function, while $%
b\left( r\right) $ is called the shape function and where $d\Omega
^{2}=d\theta ^{2}+\sin ^{2}\theta d\phi ^{2}$ is the line element of the unit sphere. Using the Einstein field equation$$G_{\mu \nu }=8\pi GT_{\mu \nu },$$in an orthonormal reference frame, we obtain the following set of equations$$\rho \left( r\right) =\frac{1}{8\pi G}\frac{b^{\prime }}{r^{2}},
\label{rhob}$$$$p_{r}\left( r\right) =\frac{1}{8\pi G}\left[ \frac{2}{r}\left( 1-\frac{%
b\left( r\right) }{r}\right) \phi ^{\prime }-\frac{b}{r^{3}}\right] ,
\label{pr}$$$$p_{t}\left( r\right) =\frac{1}{8\pi G}\left( 1-\frac{b\left( r\right) }{r}%
\right) \left[ \phi ^{\prime \prime }+\phi ^{\prime }\left( \phi ^{\prime }+%
\frac{1}{r}\right) \right] -\frac{b^{\prime }r-b}{2r^{2}}\left( \phi
^{\prime }+\frac{1}{r}\right) , \label{pt}$$in which $\rho \left( r\right) $ is the energy density, $p_{r}\left(
r\right) $ is the radial pressure, and $p_{t}\left( r\right) $ is the lateral pressure. Using the conservation of the stress-energy tensor, in the same orthonormal reference frame, we get$$p_{r}^{\prime }=\frac{2}{r}\left( p_{t}-p_{r}\right) -\left( \rho
+p_{r}\right) \phi ^{\prime }.$$When Gravity’s Rainbow comes into play, the line element $\left( \ref{metric}%
\right) $ becomes[@MagSmo] $$ds^{2}=-\exp \left( -2\phi \left( r\right) \right) \frac{dt^{2}}{%
g_{1}^{2}\left( E/E_{P}\right) }+\frac{dr^{2}}{\left( 1-\frac{b\left(
r\right) }{r}\right) g_{2}^{2}\left( E/E_{P}\right) }+\frac{r^{2}}{%
g_{2}^{2}\left( E/E_{P}\right) }d\Omega ^{2}\, \label{dS}$$and the Einstein’s Field Equations $\left( \ref{rhob}\right) $, $\left( \ref{pr}%
\right) $ and $\left( \ref{pt}\right) $ can be rearranged to give$$b^{\prime }=\frac{8\pi G\rho \left( r\right) r^{2}}{g_{2}^{2}\left(
E/E_{P}\right) }, \label{b'g2}$$$$\phi ^{\prime }=\frac{b+8\pi Gp_{r}r^{3}/g_{2}^{2}\left( E/E_{P}\right) }{%
2r^{2}\left( 1-\frac{b\left( r\right) }{r}\right) }. \label{phig2}$$Now, we introduce the equation of state $p_{r}=\omega \rho $, and using Eq.$%
\left( \ref{b'g2}\right) $, Eq.$\left( \ref{phig2}\right) $ becomes$$\begin{aligned}
\phi ^{\prime } &=&\frac{b+8\pi G\left( \omega g_{2}^{2}\left(
E/E_{P}\right) b^{\prime }\left( r\right) /\left( 8\pi Gr^{2}\right) \right)
r^{3}/g_{2}^{2}\left( E/E_{P}\right) }{2r^{2}\left( 1-\frac{b\left( r\right)
}{r}\right) } \notag \\
&=&\frac{b+\omega b^{\prime }r}{2r^{2}\left( 1-\frac{b\left( r\right) }{r}%
\right) }.\end{aligned}$$It is immediate to see that the equation related the redshift is unchanged and can be set to a constant with respect to the radial distance if$$b+\omega b^{\prime }r=0. \label{bb'}$$The integration of this simple equation leads to$$b\left( r\right) =r_{0}\left( \frac{r_{0}}{r}\right) ^{\frac{1}{\omega }},
\label{shape}$$where we have used the condition $b\left( r_{t}\right) =r_{t}$. In this situation, the line element $\left( \ref{dS}\right) $ becomes$$ds^{2}=-\frac{A}{g_{1}^{2}\left( E/E_{P}\right) }dt^{2}+\frac{dr^{2}}{%
1-\left( \frac{r_{0}}{r}\right) ^{1+\frac{1}{\omega }}g_{2}^{2}\left(
E/E_{P}\right) }+\frac{r^{2}}{g_{2}^{2}\left( E/E_{P}\right) }d\Omega ^{2},
\label{line}$$where $A$ is a constant coming from $\phi ^{\prime }=0$ which can be set to one without loss of generality. The parameter $\omega $ is restricted by the following conditions$$b^{\prime }\left( r_{0}\right) <1;\qquad \frac{b\left( r\right) }{r}\underset%
{r\rightarrow +\infty }{\rightarrow 0}\qquad \Longrightarrow \qquad \left\{
\begin{array}{c}
\omega >0 \\
\omega <-1%
\end{array}%
\right. .$$
Self-sustained Traversable Wormholes, Gravity’s Rainbow and Phantom Energy {#p2}
==========================================================================
In this Section we shall consider the formalism outlined in detail in Refs. [@Remo; @Remo1], where the graviton one loop contribution to a classical energy in a wormhole background is used. A traversable wormhole is said to be *self sustained* if$$H_{\Sigma }^{(0)}=-E^{TT}, \label{SS}$$where $E^{TT}$ is the total regularized graviton one loop energy and $%
H_{\Sigma }^{(0)}$ is the classical term. When we deal with spherically symmetric line element, the classical Hamiltonian reduces to$$\begin{aligned}
H_{\Sigma }^{(0)}& =\int_{\Sigma }\,d^{3}x\left[ \,\left( 16\pi G\right)
G_{ijkl}\pi ^{ij}\pi ^{kl}-\frac{\sqrt{g}}{16\pi G}\!{}\!\,\ R\right] \notag
\\
& =-\frac{1}{16\pi G}\int_{\Sigma }\,d^{3}x\,\sqrt{g}\,R\,=-\frac{1}{2G}%
\int_{r_{0}}^{\infty }\,\frac{dr\,r^{2}}{\sqrt{1-b(r)/r}}\,\frac{b^{\prime
}(r)}{r^{2}g_{2}\left( E/E_{P}\right) }\,,\end{aligned}$$where we have used the explicit expression of the scalar curvature in three dimensions in terms of the shape function. $G_{ijkl}$ is the super-metric and $\pi ^{ij}$ is the super-momentum. Note that, in this context, the kinetic term disappears. Note also that boundary terms become important when one compares different configurations like Wormholes and Dark Stars[DBGL]{} or Wormholes and Gravastars[@JHEP]. With the help of Eq.$\left( %
\ref{bb'}\right) $, the classical energy becomes$$H_{\Sigma }^{(0)}=\frac{1}{2G}\int_{r_{0}}^{\infty }\,\frac{dr\,r^{2}}{\sqrt{%
1-b(r)/r}}\,\frac{b(r)}{r^{3}g_{2}\left( E/E_{P}\right) \omega }$$and following Ref.[@RGFSNL], the self-sustained equation $\left( \ref{SS}%
\right) $ becomes $$-\frac{b(r)}{2Gr^{3}g_{2}\left( E/E_{P}\right) \omega }=\frac{2}{3\pi ^{2}}%
\left( I_{1}+I_{2}\right) \,, \label{ETT}$$where the r.h.s. of Eq.$\left( \ref{ETT}\right) $ is represented by$$I_{1}=\int_{E^{\ast }}^{\infty }E\frac{g_{1}\left( E/E_{P}\right) }{%
g_{2}^{2}\left( E/E_{P}\right) }\frac{d}{dE}\left( \frac{E^{2}}{%
g_{2}^{2}\left( E/E_{P}\right) }-m_{1}^{2}\left( r\right) \right) ^{\frac{3}{%
2}}dE\, \label{I1}$$and$$I_{2}=\int_{E^{\ast }}^{\infty }E\frac{g_{1}\left( E/E_{P}\right) }{%
g_{2}^{2}\left( E/E_{P}\right) }\frac{d}{dE}\left( \frac{E^{2}}{%
g_{2}^{2}\left( E/E_{P}\right) }-m_{2}^{2}\left( r\right) \right) ^{\frac{3}{%
2}}dE\,, \label{I2}$$respectively. $E^{\ast }$ is the value which annihilates the argument of the root while $m_{1}^{2}\left( r\right) $ and $m_{2}^{2}\left( r\right) $ are two r-dependent effective masses. Of course, $I_{1}$ and $I_{2}$ are finite for appropriate choices of the Rainbow’s functions $g_{1}\left(
E/E_{P}\right) $ and $g_{2}\left( E/E_{P}\right) $. With the help of the EoS, one finds$$\left\{
\begin{array}{c}
m_{1}^{2}\left( r\right) =\frac{6}{r^{2}}\left( 1-\frac{b\left( r\right) }{r}%
\right) +\frac{3}{2r^{3}\omega }b\left( r\right) \left( \omega +1\right) \\
\\
m_{2}^{2}\left( r\right) =\frac{6}{r^{2}}\left( 1-\frac{b\left( r\right) }{r}%
\right) +\frac{3}{2r^{3}\omega }b\left( r\right) \left( \frac{1}{3}-\omega
\right)%
\end{array}%
\right.$$and on the throat, the effective masses reduce to$$\begin{tabular}{c}
$m_{1}^{2}\left( r_{0}\right) =\frac{3}{2r_{0}^{2}\omega }\left( \omega
+1\right) \qquad \left\{
\begin{array}{l}
>0\qquad \mathrm{when\qquad }\omega >0\qquad \mathrm{or\qquad }\omega <-1 \\
<0\qquad \mathrm{when\qquad }-1<\omega <0%
\end{array}%
\right. $ \\
\\
$m_{2}^{2}\left( r_{0}\right) =\frac{3}{2r_{0}^{2}\omega }\left( \frac{1}{3}%
-\omega \right) \qquad \left\{
\begin{array}{l}
>0\qquad \mathrm{when\qquad }1/3>\omega >0 \\
<0\qquad \mathrm{when\qquad }\omega >1/3\qquad \mathrm{or\qquad }\omega <0%
\end{array}%
\right. $%
\end{tabular}%
. \label{m1m2}$$However, to have values of $\omega $ compatible with the traversability condition, only the cases with $\omega >0$ and $\omega <-1$ are allowed. It is easy to see that if we assume$$g_{1}\left( E/E_{P}\right) =1\qquad g_{2}\left( E/E_{P}\right) =\left\{
\begin{array}{c}
1\qquad \mathrm{when}\qquad E<E_{P} \\
\\
E/E_{P}\qquad \mathrm{when}\qquad E>E_{P}\qquad%
\end{array}%
\right. \,, \label{rel}$$Eq.$\left( \ref{ETT}\right) $ becomes, close to the throat,$$-\frac{1}{2Gr_{0}^{2}\omega }=\frac{2}{\pi ^{2}}\left( \int_{\sqrt{%
m_{1}^{2}\left( r\right) }}^{E_{P}}E^{2}\sqrt{E^{2}-m_{1}^{2}\left(
r_{0}\right) }dE\,+\int_{\sqrt{m_{2}^{2}\left( r\right) }}^{E_{P}}E^{2}\sqrt{%
E^{2}-m_{2}^{2}\left( r_{0}\right) }dE\right) \,, \label{ETT1}$$where $m_{1}^{2}\left( r_{0}\right) $ and $m_{2}^{2}\left( r_{0}\right) $ have been defined in $\left( \ref{m1m2}\right) $. Since the r.h.s. is certainly positive, in order to have real solutions compatible with asymptotic flatness, we need to impose $\omega <-1$, that it means that we are in the Phantom regime. With this choice, the effective masses $\left( \ref{m1m2}\right) $ become, on the throat$$m_{1}^{2}\left( r_{0}\right) =\frac{3}{2r_{0}^{2}\omega }\left( \omega
+1\right) \qquad m_{2}^{2}\left( r_{0}\right) =-\frac{3}{2r_{0}^{2}\omega }%
\left( \frac{1}{3}-\omega \right)$$and Eq.$\left( \ref{ETT1}\right) $ simplifies into$$1=-\frac{4r_{0}^{2}\omega }{\pi ^{2}E_{P}^{2}}\left( \int_{\sqrt{%
m_{1}^{2}\left( r_{0}\right) }}^{E_{P}}E^{2}\sqrt{E^{2}-\frac{3}{%
2r_{0}^{2}\omega }\left( \omega +1\right) }dE+\int_{0}^{E_{P}}E^{2}\sqrt{%
E^{2}+\frac{3}{2r_{0}^{2}}\left\vert \frac{1}{3\omega }-1\right\vert }%
dE\right) \, \label{ETT2}$$The solution can be easily computed numerically and we find$$\begin{aligned}
-1& \geq \omega \geq -4.5 \notag \\
2.038& \geq x\geq 1.083.\end{aligned}$$ Therefore we can conclude that a wormhole which is traversable in principle, but not in practice, can be produced joining Gravity’s Rainbow and phantom energy. Of course, the result is strongly dependent on the rainbow’s functions which, nonetheless must be chosen in such a way to give finite results for the one loop integrals $\left( \ref{I1}\right) $ and $\left( \ref{I2}\right) $.
[99]{}
R. Garattini and G. Mandanici, *Phys. Rev.* **D 83** (2011) 084021; ArXiv:1102.3803 \[gr-qc\]. R. Garattini, *JCAP* ** 017** (2013) 1306; ArXiv:1210.7760 \[gr-qc\]. R. Garattini and B. Majumder, *Nucl. Phys.* **B 884** (2014) 125 ArXiv:1311.1747 \[gr-qc\]. R. Garattini, *Phys.Lett.* **B** 685 (2010) 329; ArXiv:0902.3927 \[gr-qc\]. R. Garattini and G. Mandanici, *Phys. Rev.* **D 85** (2012) 023507; ArXiv:1109.6563 \[gr-qc\]. R. Garattini and B. Majumder, *Nucl. Phys.* **B 883** (2014) 598; ArXiv:1305.3390 \[gr-qc\].
R. Garattini and P. Nicolini, *Phys. Rev.* **D 83**, 064021 (2011); arXiv:1006.5418 \[gr-qc\].
J. Magueijo and L. Smolin, Class. Quant. Grav. 21, 1725 (2004); arXiv:gr-qc/0305055.
A. Einstein and N. Rosen, *Phys. Rev.* **48**, 73, (1935).
J. A. Wheeler, *Phys. Rev.* **97**, 511-536 (1955).
M. S. Morris and K. S. Thorne, *Am. J. Phys.* **56**, 395 (1988).
R. Garattini, *Class. Quant. Grav.* **22**, 1105 (2005); arXiv:gr-qc/0501105.
R. Garattini, *Class. Quant. Grav.* **24**, 1189 (2007); arXiv: gr-qc/0701019.
A. DeBenedictis, R. Garattini and F. S. N. Lobo, *Phys. Rev.* **D 78**, 104003 (2008); arXiv:0808.0839 \[gr-qc\].
R. Garattini and F. S. N. Lobo, *Eur.Phys.J.* **C 74**, 2884 (2014); arXiv:1303.5566 \[gr-qc\].
R. Garattini, *JHEP* **052**, 1309 (2013); arXiv:1001.3831 \[gr-qc\].
Visser M 1995 *Lorentzian Wormholes: From Einstein to Hawking* (American Institute of Physics, New York).
R. Garattini and F. S. N. Lobo, *Class. Quant. Grav.* **24**, 2401 (2007); arXiv:gr-qc/0701020.
R. Garattini and F. S. N. Lobo, *Phys. Rev.* **D 85** (2012) 024043; ArXiv: 1111.5729 \[gr-qc\].
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Konstantinos Pelechrinis$^*$, Ioannis Broustis$^*$, Srikanth V. Krishnamurthy$^*$, Christos Gkantsidis$^{\dagger}$\
\
\
bibliography:
- 'main.bib'
title: ' A measurement driven, 802.11 anti-jamming system\'
---
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank Ralink Technologies for providing us the Linux source driver for the RT2860 AP and Intel Research, for providing us with the prototype firmware of [*ipw2200*]{} AP.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Improved general slow-roll formulae giving the primordial gravitational wave spectrum are derived in the present work. Also the first and second order general slow-roll inverse formulae giving the Hubble parameter $H$ in terms of the gravitational wave spectrum are derived. Moreover, the general slow-roll consistency condition relating the scalar and tensor spectra is obtained.'
---
\
0.5cm
1.2cm
Introduction
============
Inflation inevitably leads to scalar curvature perturbations and gravitational waves caused by the tensor perturbations to the spatial metric [@Mukha]. The relative contribution of scalar and tensor fluctuations to the Cosmic Microwave background (CMB) anisotropy depends upon the details of the inflationary potential. Also, the scalar and tensor fluctuations generate different patterns of polarisation and contribute independently to the observed value [@obs] of the angular power spectrum, $\mathcal{C}_l$. The exploration of gravitational wave might be possible from the detection of the B-mode polarisation in the CMB anisotropy [@B] and it is hoped to make more progress in that direction once we get the more precise data from $\textit{Planck}$ [@planck] and $\textit{Big Bang Observer}$ [@BBO].
The standard slow-roll approximations used for inflationary scenarios make some strong assumptions about the properties of inflation, which have not yet been fully confirmed observationally. Hence, a more general slow-roll approximation has been put forward [@gsr] which lifts the extra, unjustified, assumptions of the standard slow-roll approximation. The advantages of the general slow-roll approximation, compared with the standard approximation, are clearly discussed in Ref. [@ewanrecon].
The reconstruction of the inflationary potential from the observed scalar density perturbation spectrum [@recon1] and also from the gravitational wave spectrum have been discussed by many authors [@recon2; @recon3]. The nearly scale-invariant spectrum, $\mathcal{P_\psi}$ of gravitational wave, traces the evolution of the Hubble parameter during inflation and it has been shown that one can reconstruct the time dependence of the very early Hubble parameter and matter energy density from the relic gravitational wave spectrum [@GWrecon]. Recently, we proposed a general inverse formula [@inverse] for extracting inflationary parameters from the scalar power spectrum. There, we inverted the single field, general slow-roll formula for the curvature perturbation spectrum to obtain a formula for inflationary parameters. Here, our inverse formalism is applied to the gravitational wave spectrum so as to estimate the Hubble parameter $
H $.
First order general slow-roll formulae
======================================
Gravitational wave spectrum
---------------------------
The formalism of general slow-roll approximation and the power spectrum calculation are clearly described in [@gsr]. With this general formalism, the first order gravitational wave spectrum can be given by [@JGW] $$\label{FOGWJO}
\ln\mathcal{P_\psi}(\ln k) = \int_0^\infty \frac{d\xi}{\xi} \left[
- k\xi \,W'(k\xi) \right] \left[ \ln\frac{1}{p^2} + \frac{2}{3}
\frac{p'}{p} \right]$$ where $ p =2\pi a \xi $ and $\xi = - \int \frac{dt}{a} =
\frac{1}{aH} \left( 1 - \frac{\dot{H}}{H^2} + ... \right)$ is minus the conformal time. The window function $- x \, W'(x)$ is given by, $$\label{W}
W(x) = \frac{3\sin(2x)}{2x^3} - \frac{3\cos(2x)}{x^2} -
\frac{3\sin(2x)}{2x} - 1$$ It has the asymptotic behavior $$\lim_{x \rightarrow 0} W(x) = \frac{2}{5} x^2 + \mathcal{O}(x^4)$$ and the window property, $$\label{Wwin}
\int_0^\infty \frac{dx}{x} \left[ - x \, W'(x) \right] = 1$$ An alternative form for the gravitational wave spectrum with a particularly simple window function is $$\label{FOGW1}
\ln\mathcal{P_\psi}(\ln k) = \int_0^\infty \frac{d\xi}{\xi} \left[
- k\xi \, v'(k\xi) \right] \left[ \ln\frac{1}{p^2} - 2 \frac{p'}{p}
\right]$$ where $$\label{v}
v(x) = \frac{\sin(2x)}{2x} - 1$$ $v(x)$ has the asymptotic behavior $$\lim_{x \rightarrow 0} v(x) = -\frac{2}{3} x^2 + \mathcal{O}(x^4)$$ and the window property $$\label{vwin}
\int_0^\infty \frac{dx}{x} \left[ - x \, v'(x) \right] = 1 \,.$$ The following figures give the window functions as the function of $-\ln (k\xi)$. Large $k\xi$ corresponds to earlier times, when the mode of interest is within the horizon and oscillates rapidly. The window function starts to vanish for $k\xi$ $<1$ once the mode leaves the horizon because it freezes out.
![The window function $-k\xi\,v'(k\xi)$ as the function of $-\ln (k\xi)$[]{data-label="VP"}](GWVP.eps){width="10cm" height="6cm"}
![$v(k\xi)$ as the function of $-\ln (k\xi)$[]{data-label="VP"}](GWV.eps){width="10cm" height="6cm"}
Here, $p$ is taken as a function of $\ln\xi$ so that $$p' \equiv \frac{dp}{d\ln\xi} = \xi \frac{dp}{d\xi} =
-\frac{p}{2\pi}\,\frac{dp}{dt}$$ and since $ \frac{d\xi}{dt} = -\frac{1}{a} $ and $H =
\frac{\dot{a}}{a} $ we get, $$\label{pH}
\frac{p'}{p} = 1 - \left(\frac{H}{2\pi}\right) p$$ Therefore, $$\label{H}
\ln\left(\frac{H}{2\pi}\right)^2 = \ln\left(\frac{1}{p^2}\right) - 2
\left(\frac{p'}{p}\right) - \left(\frac{p'}{p}\right)^2 - \cdots$$ Substituting the above form in equation (\[FOGW1\]) we get the general slow-roll formula for gravitational wave spectrum in terms of the Hubble parameter $H$, up to the first order correction terms, as $$\label{FOGW}
\ln\mathcal{P_\psi}(\ln k) = \int_0^\infty \frac{d\xi}{\xi} \left[ -
k \, \xi \, v'(k\,\xi)\right] \ln\left(\frac{H}{2\pi}\right)^2$$ Also note that $\xi = \frac{1}{a H}$ upto this order.
Inverse
-------
Using the inverse identity, $$\label{id1}
\int_0^\infty \frac{dk}{k} \, m(k\zeta) \, v(k\xi) =
\frac{1}{2\zeta} \left[(\zeta-\xi)
\,\mathrm{sgn}(\zeta-\xi)+(\zeta+\xi)\, \mathrm{sgn}(\zeta+\xi)
\right] - \frac{1}{\zeta} \left[\zeta \,\mathrm{sgn}(\zeta)+ \xi
\,\mathrm{sgn}(\xi) \right]$$ and its derivative with respect to $\xi$, $$\label{id2}
\int_0^\infty \frac{dk}{k} \, m(k\zeta) \left[ - k\xi \, v'(k\xi)
\right] = \frac{\xi}{2\zeta} \left[ \mathrm{sgn}(\zeta-\xi) -
\mathrm{sgn}(\zeta+\xi) \right]+\frac{\xi}{\zeta}\mathrm{sgn}(\xi)$$ where $\mathrm{sgn}(x)=-1$ for $x<0$ and $\mathrm{sgn}(x)=1$ for $x>0$ and also where $$m(x) = \frac{2}{\pi} \left[ \frac{1}{x} - \frac{\cos(2x)}{x} -
\sin(2x) \right]$$ we get the first order inverse expression for the gravitational wave spectrum as $$\label{invH}
\ln\left(\frac{H}{2\pi}\right)^2 = \int_0^\infty \frac{dk}{k} \,
m(k\,\xi) \left[\ln\mathcal{P_\psi} -
\frac{\mathcal{P_\psi}'}{\mathcal{P_\psi}} \right] \,.$$ It will be interesting to note the asymptotic behaviour $$\lim_{x \rightarrow 0} m(x) = \frac{4}{3\pi} x^3 + \mathcal{O}(x^5)$$ and the window properties $$\label{mwin}
\int_0^\infty \frac{dx}{x} \, m(x) = 1 \, ,$$ $$\label{mx^2}
\int_0^\infty \frac{dx}{x} \, \frac{1}{x} \, m(x) = \frac{2}{\pi} \,
.$$ The equation (\[invH\]) gives an explicit formula for $ H $ in terms of the gravitational wave spectrum $ \mathcal{P_\psi} $.
Second order general slow-roll formulae
=======================================
Under the general slow-roll formalism [@gsr], the second order spectrum for the scalar [@JY] and tensor perturbations [@JGW] have been calculated. An improved second order general slow-roll formula for the gravitational wave spectrum in terms of the Hubble parameter $H$ is presented in this section.
Gravitational wave spectrum
---------------------------
The general slow-roll gravitational wave spectrum can be derived up to second order terms as [@JGW] $$\begin{aligned}
\label{SOGW1}
\ln\mathcal{P_\psi}(\ln k) & = & \int_0^\infty \frac{d\xi}{\xi}
\left[ - k\xi \, W'(k\xi) \right] \left[ \ln\frac{1}{p^2} +
\frac{2}{3} \frac{p'}{p} \right] + \frac{\pi^2}{2} \left[
\int_0^\infty \frac{d\xi}{\xi} m(k\xi) \frac{p'}{p} \right]^2
\nonumber \\ && \mbox{} - 2\pi \int_0^\infty \frac{d\xi}{\xi}
m(k\xi) \frac{p'}{p} \int_\xi^\infty \frac{d\zeta}{\zeta}
\frac{1}{k\zeta} \frac{p'}{p}\end{aligned}$$ Using equation (\[pH\]) and its derivative, we get $$\label{epsilon}
- \frac{\dot{H}}{H^2} = - \frac{\frac{p''}{p} + \frac{p'}{p} \left(
1 - 2 \frac{p'}{p} \right)}{\left( 1 - \frac{p'}{p} \right)^2}$$ Now, we can rewrite the equation (\[SOGW1\]) to get the second order general slow-roll gravitational wave spectrum in terms of $H$ as, $$\begin{aligned}
\label{SOGWH}
\ln\mathcal{P_\psi}(\ln k) & = &\int_0^\infty \frac{d\xi}{\xi}
\left[ - k\xi \, v'(k\xi) \right] \,
\ln\left(\frac{H}{2\pi}\right)^2 + \frac{\pi^2}{2} \left[
\int_0^\infty \frac{d\xi}{\xi} n(k\xi) \frac{\dot{H}}{H^2} \right]^2
\nonumber \\ && \mbox{} + 2\pi \int_0^\infty \frac{d\xi}{\xi}
n(k\xi) \frac{\dot{H}}{H^2} \left\{\frac{1}{k\xi}
\frac{\dot{H}}{H^2} - \int_\xi^\infty \frac{d\zeta}{\zeta}
\frac{1}{k\zeta} \frac{\dot{H}}{H^2} \right\} \nonumber \\\end{aligned}$$ where $$n(x) = \frac{1}{\pi} \left[ \frac{1}{x} - \frac{\cos(2x)}{x} \right]$$ $n(x)$ has the asymptotic behavior $$\lim_{x \rightarrow 0} n(x) = - \frac{2}{\pi} x + \mathcal{O}(x^3)
\, ,$$ window property $$\int_0^\infty \frac{dx}{x} \, n(x) = 1 \, ,$$ and is related to $m(x)$ by $$m(x) = n(x) - x \, n'(x) \, .$$
Inverse
-------
Substituting first order inverse expression, equation (\[invH\]), into equation (\[SOGWH\]) and following the same formalism of Ref. [@inverse] we get the second order inverse formula for gravitational wave spectrum as, $$\begin{aligned}
\label{SOGWinv}
\ln\left(\frac{H}{2\pi}\right)^2 &=& \int_0^\infty \frac{dk}{k} \,
m(k\xi) \left[\ln\mathcal{P_\psi}(\ln k) -
\frac{\mathcal{P_\psi}'(\ln k)}{\mathcal{P_\psi}(\ln k)} \right]
\nonumber \\ && \mbox{} - \frac{1}{2\pi^2} \int_0^\infty
\frac{dk}{k} \, m(k\xi) \int_0^\infty \frac{dl}{l}
\ln\left|\frac{k+l}{k-l}\right| \frac{\mathcal{P_\psi}' (\ln
l)}{\mathcal{P_\psi}(\ln l)} \int_0^\infty \frac{dq}{q}
\left[\ln\left|\frac{k+q}{k-q}\right| - \frac{2 q k}{q^2-k^2}
\right] \frac{\mathcal{P_\psi}' (\ln q)}{\mathcal{P_\psi}(\ln q)}
\nonumber
\\ && \mbox{} + \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \, N(l\xi,q\xi) \, \frac{\mathcal{P_\psi}'(\ln
l)}{\mathcal{P_\psi}(\ln l)} \frac{\mathcal{P_\psi}'(\ln
q)}{\mathcal{P_\psi}(\ln q)}\end{aligned}$$ where $ N(x,y) = M_1(x,y) - M_2(x,y) $ with $$\label{M1}
\int_0^\infty \frac{d\zeta}{\zeta} \, m(l\zeta) \, \int_0^\infty
\frac{dk}{k^2} \, m(k\xi) \, m(k\zeta) \, \frac{\sin^2(q\zeta)}
{q\zeta^2} = M_1(l\xi,q\xi)$$ $$\label{M2}
\int_0^\infty \frac{d\zeta}{\zeta} \, m(l\zeta) \, \int_0^\infty
\frac{dk}{k^2} \, [(-k\xi)m'(k\xi)] \, m(k\zeta) \,
\frac{\sin^2(q\zeta)} {q\zeta^2} = M_2(l\xi,q\xi)$$ Carrying out the integrations in the above expression, $$\begin{aligned}
\frac{M_j(x,y)}{j} & = & \frac{2}{\pi^2 x y} \left[g_j(x) + g_j(y) -
\frac{1}{2} \, g_j(x-y) - \frac{1}{2} \, g_j(x+y) \right]\end{aligned}$$ for which the index $ j $ takes values 1 and 2. $$g_j(x) = x \, \mathrm{Si}(2x) + \frac{j}{2}\,\left( \cos(2x) - 1
\right)$$ where we denote $$\mathrm{Si}(x) \equiv \int_0^x \frac{\sin t}{t} \, dt$$ $M_j(x,y)$ has the window property $$\label{Mwin}
\int_0^\infty \frac{dx}{x} \int_0^\infty \frac{dy}{y} \,
\frac{M_j(x,y)}{j} = 1$$ and the asymptotic behaviour $$\lim_{x,y \rightarrow 0} \frac{M_j(x,y)}{j} = \frac{4 x y}{3\pi^2}
\left[ 1 + \mathcal{O}\left(x^2+y^2 \right)\right]$$
General slow-roll consistency condition
=======================================
The standard slow-roll consistency condition relating the scalar and tensor spectra is detailed in [@recon2] for the single field case and [@MisaEwan] discusses the same for the multi-component scalar field inflation models. This section describes the general slow-roll generalisation of the constraint on the spectra.\
\
The inverse formula for the scalar power spectrum [@inverse], $\mathcal{P}_s$ is given by, $$\begin{aligned}
\label{finv}
\ln\frac{1}{f^2} & = & \int_0^\infty \frac{dk}{k} m(k\xi)
\ln\mathcal{P}_s (\ln k) \nonumber \\ && \mbox{} - \frac{1}{2\pi^2}
\int_0^\infty \frac{dk}{k} \, m(k\xi) \left[ \int_0^\infty
\frac{dl}{l} \ln\left|\frac{k+l}{k-l}\right|
\frac{\mathcal{P}_s' (\ln l)}{\mathcal{P}_s(\ln l)} \right]^2
\nonumber \\ && \mbox{} + \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \, M_1(l\xi,q\xi) \,
\frac{\mathcal{P}_s'(\ln l)}{\mathcal{P}_s(\ln l)} \frac{\mathcal{P}_s'(\ln q)}{\mathcal{P}_s(\ln q)}\end{aligned}$$ where $ f = \frac{2\pi a \xi \dot\phi}{H} $. Also, for the tensor spectrum we have, $$\begin{aligned}
\label{pinv}
\ln\frac{1}{p^2} & = & \int_0^\infty \frac{dk}{k} \, m(k\xi)
\ln\mathcal{P_\psi}(\ln k) \nonumber \\ && \mbox{} -
\frac{1}{2\pi^2} \int_0^\infty \frac{dk}{k} \, m(k\xi) \left[
\int_0^\infty \frac{dl}{l} \ln\left|\frac{k+l}{k-l}\right|
\frac{\mathcal{P_\psi}' (\ln l)}{\mathcal{P_\psi}(\ln l)} \right]^2
\nonumber \\ && \mbox{} + \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \, M_1(l\xi,q\xi) \, \frac{\mathcal{P_\psi}'(\ln
l)}{\mathcal{P_\psi}(\ln l)} \frac{\mathcal{P_\psi}'(\ln
q)}{\mathcal{P_\psi}(\ln q)}\end{aligned}$$ Combining the above two inverse formulae we can write, $$\begin{aligned}
\label{phiH}
\ln \left(\frac{\dot\phi}{H} \right)^2 & = & \int_0^\infty
\frac{dk}{k} \, m(k\xi) \ln
\left(\frac{\mathcal{P_\psi}}{\mathcal{P}_s}\right)\nonumber \\ &&
\mbox{}- \frac{1}{2\pi^2} \int_0^\infty \frac{dk}{k} m(k\xi)
\int_0^\infty \frac{dl}{l} \ln\left|\frac{k+l}{k-l}\right| \left(
\frac{\mathcal{P_\psi}' }{\mathcal{P_\psi}} + \frac{\mathcal{P}_s'
}{\mathcal{P}_s} \right) \int_0^\infty \frac{dq}{q}
\ln\left|\frac{k+q}{k-q}\right| \left( \frac{\mathcal{P_\psi}'
}{\mathcal{P_\psi}} + \frac{\mathcal{P}_s' }{\mathcal{P}_s} \right)
\nonumber
\\ && \mbox{} + \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \, M_1(l\xi,q\xi) \, \left\{
\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln l)}
\frac{\mathcal{P_\psi}'(\ln q) }{\mathcal{P_\psi}(\ln q)} -
\frac{\mathcal{P}_s'(\ln l) }{\mathcal{P}_s(\ln l)}
\frac{\mathcal{P}_s'(\ln q) }{\mathcal{P}_s(\ln q)} \right\}\end{aligned}$$ Defining, $$m_1(x)= \int\frac{dx}{x}\,m(x) =
\frac{2}{\pi}\left[\frac{\cos(2x)}{x} - \frac{1}{x}
+\mathrm{Si}(2x)\right]$$ we can rewrite, $$\begin{aligned}
\label{phiH}
\left(\frac{\dot\phi}{H} \right)^2 & \simeq & \left(
\frac{\mathcal{P_\psi}}{\mathcal{P}_s}\right)_\diamond \left[1 -
\int_0^\infty \frac{dk}{k} (m_1(k\xi) - \theta(k\xi - k_\diamond
\xi)) \left(\frac{\mathcal{P_\psi}' }{\mathcal{P_\psi}} -
\frac{\mathcal{P}_s' }{\mathcal{P}_s} \right) \right. \nonumber \\
&& \left. \mbox{} \hspace{1.5cm} + \frac{1}{2}\left(\int_0^\infty
\frac{dk}{k} (m_1(k\xi) - \theta(k\xi - k_\diamond \xi))
\left(\frac{\mathcal{P_\psi}' }{\mathcal{P_\psi}} -
\frac{\mathcal{P}_s' }{\mathcal{P}_s}\right)\right)^2 \right. \nonumber \\
&& \left. \mbox{}\hspace{1.5cm} - \frac{1}{2\pi^2} \int_0^\infty
\frac{dk}{k} m(k\xi) \int_0^\infty \frac{dl}{l}
\ln\left|\frac{k+l}{k-l}\right| \left(\frac{\mathcal{P_\psi}'(\ln l)
}{\mathcal{P_\psi}(\ln l)} +
\frac{\mathcal{P}_s'(\ln l) }{\mathcal{P}_s(\ln l)}\right) \right. \nonumber \\
&& \left. \mbox{} \hspace{2.5cm} \times \int_0^\infty \frac{dq}{q}
\ln\left|\frac{k+q}{k-q}\right| \left(\frac{\mathcal{P_\psi}'(\ln q)
}{\mathcal{P_\psi}(\ln q)} +
\frac{\mathcal{P}_s'(\ln q) }{\mathcal{P}_s(\ln q)}\right) \right. \nonumber \\
&& \left. \mbox{}\hspace{1.3cm} + \int_0^\infty \frac{dl}{l}
\int_0^\infty \frac{dq}{q} \, M_1(l\xi,q\xi) \, \left\{
\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln l)}
\frac{\mathcal{P_\psi}'(\ln q) }{\mathcal{P_\psi}(\ln q)} -
\frac{\mathcal{P}_s'(\ln l) }{\mathcal{P}_s(\ln l)}
\frac{\mathcal{P}_s'(\ln q) }{\mathcal{P}_s(\ln q)} \right\} \right]
\nonumber \\\end{aligned}$$ where $k_\diamond$ is some reference wavenumber. Now, taking the derivative of equation (\[SOGWinv\]) with respect to $\ln \xi $ we get, $$\begin{aligned}
\label{HdotH}
2\frac{\dot{H}}{H^2} & = & \int_0^\infty \frac{dk}{k} \,
m(k\xi)\left[\frac{\mathcal{P_\psi}'(\ln k) }{\mathcal{P_\psi}(\ln
k)} - \left(\frac{\mathcal{P_\psi}'(\ln k) }{\mathcal{P_\psi}(\ln
k)}\right)'\,\right] \nonumber \\ && \mbox{} - \frac{1}{2\pi^2}
\int_0^\infty \frac{dk}{k}\,(k\xi) m'(k\xi) \int_0^\infty
\frac{dl}{l} \ln\left|\frac{k+l}{k-l}\right|
\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln l)}
\int_0^\infty \frac{dq}{q} \left[\ln\left|\frac{k+q}{k-q}\right| +
\frac{2 q k}{k^2 - q^2} \right] \frac{\mathcal{P_\psi}'(\ln q)
}{\mathcal{P_\psi}(\ln q)} \nonumber
\\ && \mbox{}+ \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \,[ M'_1(l\xi,q\xi)- M'_2(l\xi,q\xi)] \,
\left(\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln
l)}\right)' \left(\frac{\mathcal{P_\psi}'(\ln q)
}{\mathcal{P_\psi}(\ln q)}\right)'\end{aligned}$$ Using the definition of the slow-roll parameter $\epsilon \equiv
\frac{1}{2}\left(\frac{\dot\phi}{H} \right)^2 = -
\frac{\dot{H}}{H^2}$, one can equate the equations (\[phiH\]) and (\[HdotH\]) to get the general slow roll constraint on the spectra as, $$\begin{aligned}
\label{cons}
\left(\frac{\mathcal{P_\psi}}{\mathcal{P}_s}\right)_\diamond &=&
-\int_0^\infty \frac{dk}{k} \,
m(k\xi)\left[\frac{\mathcal{P_\psi}'(\ln k) }{\mathcal{P_\psi}(\ln
k)}- \left(\frac{\mathcal{P_\psi}'(\ln k) }{\mathcal{P_\psi}(\ln
k)}\right)'\,\right] \nonumber \\ && \mbox{} \hspace{1cm}
\times\left\{1 + \int_0^\infty \frac{dl}{l} \, (m_1(k\xi) -
\theta(l\xi - l_\diamond \xi))
\left[\left(\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln
l)}\right) - \left(\frac{\mathcal{P}_s'(\ln l) }{\mathcal{P}_s(\ln
l)}\right)\right]\right\} \nonumber \\ && \mbox{}+ \frac{1}{2\pi^2}
\int_0^\infty \frac{dk}{k}\,(k\xi) m'(k\xi) \int_0^\infty
\frac{dl}{l} \ln\left|\frac{k+l}{k-l}\right|
\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln l)}\nonumber
\\ && \mbox{} \hspace{1.3cm} \times \int_0^\infty \frac{dq}{q}
\left[\ln\left|\frac{k+q}{k-q}\right| + \frac{2 q k}{k^2 - q^2}
\right] \frac{\mathcal{P_\psi}'(\ln q) }{\mathcal{P_\psi}(\ln q)}
\nonumber
\\ && \mbox{}- \int_0^\infty \frac{dl}{l} \int_0^\infty
\frac{dq}{q} \,[ M'_1(l\xi,q\xi)- M'_2(l\xi,q\xi)] \,
\left(\frac{\mathcal{P_\psi}'(\ln l) }{\mathcal{P_\psi}(\ln
l)}\right)' \left(\frac{\mathcal{P_\psi}'(\ln q)
}{\mathcal{P_\psi}(\ln q)}\right)'\end{aligned}$$
Single field approximations
---------------------------
In single field approximation[^1], $ - \frac{\dot{H}}{H^2} \simeq constant$ and thus we can simplify the equation (\[SOGWH\]) to get an easy forward formula as, $$\label{SOGWH1}
\ln\mathcal{P_\psi}(\ln k) = \int_0^\infty \frac{d\xi}{\xi} \left[
- k\xi \, v'(k\xi) \right] \, \ln\left(\frac{H}{2\pi}\right)^2 +
\frac{\pi^2}{2} \left( \frac{\dot{H}}{H^2} \right)^2$$ Now integrating the above equation by parts and then differentiating with respect to $\ln k$ we get, $$\label{nconst}
\frac{\mathcal{P'_\psi}}{\mathcal{P_\psi}} = \int_0^\infty
\frac{d\xi}{\xi} \left[(-k\xi)v'(k\xi)\right] \,
\left(2\frac{\dot{H}}{H^2}\right) - \frac{\dot{H}}{H^2}$$ Thus, we can find that the gravitational wave spectral index $n_\psi \equiv \frac{d \ln\mathcal{P_\psi}}{d \,ln k} =
\frac{\mathcal{P'_\psi}}{\mathcal{P_\psi}} \simeq constant $ for the single filed approximation.\
Also from equation (\[SOGWinv\]) we get the inverse for single field approximation as, $$\ln\left(\frac{H}{2\pi}\right)^2 = \int_0^\infty \frac{dk}{k} \,
m(k\xi) \left[\ln\mathcal{P_\psi} -
\frac{\mathcal{P_\psi}'}{\mathcal{P_\psi}}\, \right] -
\frac{\pi^2}{8} \left(\frac{\mathcal{P_\psi}'}{\mathcal{P_\psi}}
\right)^2$$ Again, since $n_\psi \simeq constant $ and neglecting the terms containing ${n'_\psi}^2$ also, we can simplify the consistency condition in equation (\[cons\]) to get the single field general slow roll constraint on the spectra, $$\begin{aligned}
\label{Consiml}
\left(\frac{\mathcal{P_\psi}}{\mathcal{P}_s}\right)_\diamond &=&
-\int_0^\infty \frac{dk}{k} m(k\xi)\,\left(n_\psi - {n_\psi}'\right)
+ \left[ \frac{\pi}{2} - \frac{5}{2}+ \ln(k_\diamond )\right]
n^2_\psi \nonumber
\\ && \mbox{} + n_\psi \int_0^\infty \frac{dk}{k}
(m_1(k\xi) - \theta(k\xi - k_\diamond \xi)) (n_s-1)\end{aligned}$$
Multi-field consistency condition {#multi}
---------------------------------
For the multi-component scalar field inflation models, since the scalar power spectrum has the contribution from both parallel and orthogonal components, $P_S \geq P_{single field}$. Thus the tensor to scalar power spectra ratio will be smaller than the single field case and so we can write the inequality upto first order corrections as, $$\label{Conmulti}
\frac{\mathcal{P_\psi}}{\mathcal{P}_S} < -\int_0^\infty \frac{dk}{k}
m(k\xi) \,\left(n_\psi - {n_\psi}'\right)$$ Note that in this case we can not assume $\epsilon \simeq constant $ as it is for the single field.
Standard slow-roll approximation {#secssr}
================================
In the context of the standard approximation for slow-roll, the gravitational wave spectrum has the form $$\label{gwssr}
\ln\mathcal{P_\psi} = \ln\mathcal{P_\psi}_\diamond +
{n_\psi}_\diamond \ln \left( \frac{k}{k_\diamond} \right) +
\frac{1}{2} {n_\psi}'_\diamond \ln^2
\left(\frac{k}{k_\diamond}\right) + \cdots$$ where $n_\psi \equiv \frac{d \ln\mathcal{P_\psi}}{d ln k}$, is the gravitational wave spectral index and $k_\diamond$ is some reference wavenumber. Applying our inverse formula given by equation (\[SOGWinv\]), using the window properties given by equations. (\[mwin\]) and (\[Mwin\]), and the following results $$\int_0^\infty \frac{dx}{x} \, m(x) \, \ln(x) = \alpha$$ $$\int_0^\infty \frac{dx}{x} \, m(x) \, \ln^2(x) = \alpha^2 +
\frac{\pi^2}{12}$$ $$\int_0^\infty \frac{dy}{y} \ln\left|\frac{x+y}{x-y}\right| =
\frac{\pi^2}{2}\, , \ \ \texttt{for}\ \ x > 0$$ where $\alpha = 2 - \ln 2 - \gamma \simeq 0.7296$, we get $$\label{ssrinv1}
\ln\left(\frac{H}{2\pi}\right)^2 = \ln\mathcal{P_\psi}_\diamond +
(\alpha_\diamond-1)\,{n_\psi}_\diamond + \frac{1}{2} \left(
\alpha_\diamond^2 -2\alpha_\diamond + \frac{\pi^2}{12} \right)
{n_\psi}'_\diamond + \left( 1 - \frac{\pi^2}{8} \right)
{n_\psi}_\diamond^2 + \cdots$$ where $\alpha_\diamond = \alpha - \ln(k_\diamond \xi)$. Equation (\[ssrinv1\]) reproduces the standard slow-roll inverse, which is trivially obtained from the standard slow-roll formula, $$\ln\mathcal{P_\psi} = \ln\frac{1}{p_\star^2} - 2 \alpha_\star
\frac{p_\star'}{p_\star} - \left( \alpha^2_\star - \frac{\pi^2}{12}
\right) \frac{p_\star''}{p_\star} + \left(\alpha^2_\star - 4 +
\frac{5\pi^2}{12} \right) \left(\frac{p_\star'}{p_\star}\right)^2 +
\cdots$$ where $\alpha_\star = \alpha - \ln(k \xi_\star)$ and $\xi_\star$ is an arbitrary evaluation point usually taken to be around horizon crossing.\
\
Now let us deduce the consistency conditions for the standard slow-roll approximation case. Substituting the standard slow roll gravitational wave spectrum given by equation (\[gwssr\]) in the general constraint on the spectra given by equation (\[cons\]), we get the constraint, with the standard approximation as, $$\begin{aligned}
\label{conssr}
\left(\frac{\mathcal{P_\psi}}{\mathcal{P}_s}\right) &=& -n_{\psi
*} - (\alpha_* - 1) n'_{\psi *} + (2+\ln k_*) (n_{s
*} -1) n_{\psi *} \nonumber \\ && \mbox{} +
\left(\frac{\pi}{2} - \frac{5}{2} + \ln k_* \right) n^2_{\psi *}\end{aligned}$$ Comparison of the above expression with the general slow roll condition shows that the second order correction terms are different. Also we can write the standard slow roll multi-field constraint as $$\label{conssr}
\left(\frac{\mathcal{P_\psi}}{\mathcal{P}_S}\right) < -n_{\psi*}$$ It is clear from the general multi-field constraint in equation (\[Conmulti\]) that the sign of second term on right hand side will depend on that of $ n'_\psi $ and thus the bound could be varied depending on the tilt of $ n_\psi $.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The author is indebted to Ewan D Stewart for the timely suggestions and valuable discussions on this work.
[99]{} V. F. Mukhanov and G. V. Chibisov, *JETP Lett.* **33**, 532 (1981) ; A. A. Starobinsky, *Phys. Lett.* **B 117**, 175 (1982) ; V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, *Phys. Rep.* **215**, 203 (1992)
M. Tegmark *et al*., *Phys. Rev. D* **69**, 103501 (2004) `astro-ph/0310723` ; D. N. Spergel *et al*., *Astrophys. J. Suppl.* **170**, 377 (2007) `astro-ph/0603449`
U. Seljak and M. Zaldarriaga, *Phys. Rev. Lett* **78**, 2054 (1997) ; M. Kamionkowski, A.Kosowsky and A. Stebbins, *Phys. Rev. Lett.* **78**, 2058 (1997) ; N. Barolo, E. W. Kolb and A. Riotto, *Mod. Phys. Letts.*, **A20**, 3077, 2005 `astro-ph/0507573`
http://astro.estec.esa.nl/Planck
http://universe.nasa.gov/program/bbo.html
S. Dodelson and E. D. Stewart, *Phys. Rev.* **D 65**, 101301 (2002) `astro-ph/0109354` ; E. D. Stewart, *Phys. Rev.* **D 65**, 103508 (2002) `astro-ph/0110322`
K. Kadota, S. Dodelson, W. Hu and E. D. Stewart, *Phys. Rev.* **D 72**, 023510 (2005) `astro-ph/0505158`
H. M. Hodges and G. R. Blumenthal, *Phys. Rev.* **D 42**, 3329 (1990)
E. J. Copeland et al., *Phys. Rev.* **D 48**, 2529 (1993) `hep-ph/9303288` ; E. J. Copeland et at., *Phys. Rev.* **D 49**, 1840 (1994) `astro-ph/9308044` ; J. E. Lidsey et al., *Rev. Mod. Phys.* **69**, 373 (1997) `astro-ph/9508078`
S. Habib, K. Heitmann and G. Jungman, *Phys. Rev. Lett.* **94**, 061303 (2005) `astro-ph/0409599`
L. P. Grishchuk and M. Solokhin, *Phys. Rev.* **D 43**, 2566 (1991)
M. Joy, E. D. Stewart, J. Gong and H. C. Lee, *JCAP* **04**, 012 (2005) `astro-ph/0501659` ; M. Joy and E. D. Stewart, *JCAP* **02**, 005 (2006) `astro-ph/0511476`
J.-O. Gong, *Classical and Quantum Garvity* **21**, 5555 (2004) `gr-qc/0408039`
J. Choe, J.-O. Gong and E. D. Stewart, *JCAP* **07**, 012 (2004) `hep-ph/0405155`
M. Sasaki, E. D. Stewart, *Prog. Theor. Phys* **95**, 71 (1996) `astro-ph/9507001`
H. C. Lee, M. Sasaki, E. D. Stewart, T. Tanaka and S. Yokoyama, *JCAP* **10**, 004 (2005) `astro-ph/0506262`
[^1]: The general form of Eq.(\[FOGWJO\]) will be identical if we consider the multi-scalar field inflation model also, but with a different interpretation of the quantity ’p’. Comparison of the single and multifield general slow-roll formulae in the case of scalar perturbations is discussed in [@HCL].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The spectroscopic parameters and decay channels of the doubly charged scalar, pseudoscalar and axial-vector charm-strange tetraquarks $Z_{%
\overline{c}s}=[sd][\overline u \overline c]$ are explored within framework of the QCD sum rule. The masses and current couplings of these diquark-antidiquark states are calculated by means of two-point correlation functions and taking into account the vacuum condensates up to eight dimensions. To compute the strong couplings of $Z_{\overline{c}s}$ states with $D,\ D_{s},\ D^{\ast},\ D_{s}^{\ast},\ D_{s1}(2460),\
D_{s0}^{\ast}(2317),\ \pi$ and $K$ mesons we use QCD light-cone sum rules and evaluate width of their $S$- and $P$-wave decays to a pair of negatively charged conventional mesons: For the scalar state $Z_{\overline{c}s}\to D_s
\pi,\ DK, \ D_{s1}(2460)\pi$, for the pseudoscalar state $Z_{\overline{c}s}
\to D_{s}^{\ast}\pi,\ D^{\ast}K, \ D_{s0}^{\ast}(2317)\pi,$ and for the axial-vector state $Z_{\overline{c}s} \to D_{s}^{\ast}\pi,\ D^{\ast}K,\
D_{s1}(2460)\pi$ decays are investigated. Obtained predictions for the spectroscopic parameters and decay widths of the $Z_{\overline{c}s}$ tetraquarks may be useful for experimental investigations of the doubly charged exotic hadrons.
author:
- 'S. S. Agaev'
- 'K. Azizi'
- 'H. Sundu'
title: 'Testing the doubly charged charm-strange tetraquarks'
---
Introduction
============
During last decade tetraquarks, i.e. bound states of four quarks are in the center of intensive experimental and theoretical investigations. Starting from discovery of the famous resonance $X(3872)$ in $B$ meson decay $%
B\rightarrow KX\rightarrow KJ/\psi \rho \rightarrow KJ/\psi \pi ^{+}\pi ^{-}$ by Belle [@Choi:2003ue], and after observation of the same state by other groups [@Abazov:2004kp; @Acosta:2003zx; @Aubert:2004ns] experimental collaborations collected valuable information on the spectroscopic parameters and decay channels of the exotic states. They were discovered in various inclusive and exclusive hadronic processes. In this connection it is worth to note $B$ meson decays, $e^{+}e^{-}$ and $\overline{p}p$ annihilations and $pp$ collisions. Theoretical studies of exotic hadrons, apart from tetraquark states, include pentaquarks and hybrid mesons and encompass variety of models and calculational methods claiming to explain the internal structure of these states and calculate their experimentally measured parameters. Comprehensive information on collected experimental data and detailed analysis of theoretical achievements and existing problems can be found in latest review works Refs. [Chen:2016qju,Chen:2016spr,Esposito:2016nozN,Esposito:2014rxa,Meyer:2015eta]{}.
The great success in physics of the exotic hadrons is connected with discovery of charged multiquark resonances. The first charged tetraquarks, namely $Z^{\pm}(4430)$ states were observed by the Belle Collaboration in $B$ meson decays $B \to K\psi^{\prime} \pi^{\pm}$ as resonances in the $%
\psi^{\prime}\pi^{\pm}$ invariant mass distributions [@Choi:2007wga]. The resonances $Z^{+}(4430)$ and $Z^{-}(4430)$ were detected and studied by Belle in the processes $B \to K\psi^{\prime} \pi^{+}$ [@Mizuk:2009da] and $B^{0} \to K^{+}\psi^{\prime} \pi^{-}$ [@Chilikin:2013tch], as well. These states constitute an important subclass of multiquark systems, because charged resonances can not be explained as excited charmonium or bottomonium states, and therefore, are real candidates to genuine tetraquarks.
Hadrons built of four quarks of different flavors form another intriguing class in the tetraquark family. Depending on a quark content these states may be neutral or charged particles. Among the observed tetraquarks the $%
X(5568)$ resonance remains a unique candidate to a hadron composed of four different quarks. At the same time it is a particle containing $b$-quark, i.e. is an open bottom tetraquark. The evidence for $X(5568)$ was first reported by the D0 Collaboration in Ref. [@D0:2016mwd]. Later it was observed again by D0 in the $B_{s}^{0}$ meson’s semileptonic decays [@D0]. But other experimental groups, namely the LHCb and CMS collaborations could not find this resonance from analysis of their experimental data [Aaij:2016iev,CMS:2016]{}, which make the experimental situation around $%
X(5568)$ unclear and controversial. Numerous theoretical works devoted to investigation of $X(5568)$ resonance’s structure and calculation of its parameters led also to contradictory conclusions. The results of these studies are in a reasonable agreement with measurements carried out by the D0 Collaboration, while in other works an existence of the $X(5568)$ state is an object of discussions [Agaev:2016mjb,Chen:2016mqt,Wang:2016mee,Wang:2016tsi,Zanetti:2016wjn,Agaev:2016ijz,Dias:2016dme,Wang:2016wkj,Xiao:2016mho,Agaev:2016urs,Burns:2016gvy,Guo:2016nhb,Lu:2016zhe,Albaladejo:2016eps,Lang:2016jpk,Esposito:2016itg,Kang:2016zmv]{}. The detailed analysis of problems related to the status of the $X(5568)$ resonance can be found in original papers (see for instance, Ref. [Chen:2016spr]{} and references therein).
The tetraquarks which might carry double electric charge constitute another interesting class of exotic hadrons [@Esposito:2013fma]. These hypothetical particles if observed can be interpreted as diquark-antidiquark states: Formation of molecular states from two mesons of same charge is almost impossible due to repulsive forces between them. The doubly charged particles may exist, for example, as double charmed tetraquarks $[cc][\bar d \bar s]$ or $[cc][\bar s \bar s]$. In other words, they may contain two or three quark flavors. The phenomenology of these states, their decay modes and production mechanisms were investigated in Ref. [@Esposito:2013fma]. In the context of the lattice QCD the mass spectra of these particles were evaluated in the paper [@Guerrieri:2014nxa]. As it was revealing recently the tetraquarks containing quarks of four different flavors may also carry double electric charge [@Chen:2017rhl]. In fact, it is not difficult to see that tetraquarks $Z_{\overline{c}s}=[sd][\overline u
\overline c]$ and $Z_{c \overline{s}}= [uc][\overline {s} \overline {d}]$ belong to this category of particles, and at the same time, are open charm states. Authors of Ref. [@Chen:2017rhl] wrote down also possible $S$- and $P$-wave decay channels of these states. Strictly speaking, the open charm tetraquarks were previously investigated in the literature (see, for example Refs. [@Galkin; @Agaev:2016lkl; @Liu:2016ogz]). The spectroscopic parameters and decay widths of the open charm tetraquark containing three different light quarks were calculated in Ref. [@Agaev:2016lkl]. In this study the open charm tetraquark was considered as a partner of the $%
X(5568)$ state. In other words, the quark content of $X_c=[su][\overline{c}%
\overline{d}]$ was obtained from $X_b=[su][\overline{b}\overline{d}]$ by $b
\to c$ replacement. Due to differences in the charges of $b$ and $c$ quarks the partner state $X_c$ does not bear the same charge as $X_b$. This conclusion is true in the case of $Z_{\overline{c}s}$, as well. If the state $Z_{\overline{c}s}$ bears the charge $-2|e|$, its $b$-partner $Z_{\overline{b%
}s}$ has $-|e|$. In general, there do not exist doubly charged tetraquarks composed of $b$ and three different light quarks. The genuine doubly charged tetraquarks with $b$ belong to a subclass of open charm-bottom particles and should contain also $c$-quark. For example, the state $Z_{b\overline{c}%
}=[bs][\overline{u}\overline{c}]$ has the charge $-2|e|$.
In the present work we are going to concentrate on features of doubly charged charm-strange tetraquarks $Z_{\overline{c}s}$ with spin-parity $%
J^{P}=0^{+},\ 0^{-}$ and $1^{+}$, and calculate their masses, current couplings and decay widths. To this end, we use QCD two-point sum rule approach by including into analysis quark, gluon and mixed vacuum condensates up to eight dimensions, and evaluate their spectroscopic parameters. Obtained results are employed to reveal kinematically allowed decay channels of the tetraquarks $Z_{\overline{c}s}$. They also enter as input parameters to expressions of the corresponding decay widths. We calculate the width of decay channels $Z_{\overline{c}s}\rightarrow
D_{s}\pi,\ DK$, and $D_{s1}(2460)\pi$ (for $J^{P}=0^{+}$), $Z_{\overline{c}%
s}\rightarrow D_{s}^{\ast }\pi ,\ D^{\ast }K$ and $D_{s0}^{\ast}(2317)\pi$ (for $J^{P}=0^{-}$), as well as $Z_{\overline{c}s}\rightarrow D_{s}^{\ast
}\pi ,\ D^{\ast }K$ and $D_{s1}(2460)\pi $ (in the case of $J^{P}=1^{+}$). For these purposes, we analyze vertices of the tetraquarks $Z_{\overline{c}%
s} $ with the conventional mesons, and evaluate the corresponding strong couplings using QCD sum rules on the light-cone. The QCD light-cone sum rule method is one of the powerful nonperturbative tools to explore parameters of the conventional hadrons [@Balitsky:1989ry]. In the case of vertices built of a tetraquark and two conventional mesons the standard methods of the light-cone sum rules should be supplemented by a technique of an approach known as the “soft-meson” approximation [Belyaev:1994zk,Ioffe:1983ju]{}. For investigation of the exotic states the light-cone sum rules method was adapted in Ref. [@Agaev:2016dev], and successfully applied for analysis of various tetraquarks’ decays [Agaev:2016dsg,Agaev:2017uky,Agaev:2017foq,Agaev:2017tzv]{}.
This article is organized in the following manner. In Sec. \[sec:Masses\] we calculate the masses and current couplings of the doubly charged scalar, pseudoscalar and axial-vector charm-strange tetraquarks $Z_{\overline{c}%
s}=[sd][\overline{u}\overline{c}]$ by treating them as diquark-antidiquark systems. In Sec. \[sec:ScalDec\] we consider the decays of the doubly charged scalar tetraquark to $D_{s}\pi $, $DK$ and $D_{s1}(2460)\pi $ final states. The Section \[sec:PsVDec\] is devoted to decay channels of the pseudoscalar and axial-vector tetraquarks. Here we compute width of their decays to $D_{s}^{\ast }\pi$, $D^{\ast }K$ and $D_{s0}^{\ast}(2317)\pi $ (for $0^{-}$ state), and to $D_{s}^{\ast }\pi$, $D^{\ast }K$ and $%
D_{s1}(2460)\pi $ (for $1^{+}$ state). Section \[sec:Concl\] is reserved for our concluding remarks.
Spectroscopic parameters of the scalar, pseudoscalar and axial vector tetraquarks $Z_{\overline{c}s}$ {#sec:Masses}
=====================================================================================================
In this section we calculate the mass and current coupling of the $Z_{%
\overline{c}s}=[sd][\overline{u}\overline{c}]$ tetraquarks with the quantum numbers $J^{P}=0^{+}$, $0^{-}$ and $1^{+}$ by treating them as diquak-antidiquark systems. In order to simplify the expressions we introduce the notations: in what follows the scalar tetraquark $Z_{\overline{%
c}s}$ will be denoted as $Z_{S}$, whereas for the pseudoscalar and axial-vector ones we will utilize $Z_{PS}$ and $Z_{AV}$, respectively.
The scalar tetraquarks within the context of two-point sum rule approach can be explored using interpolating currents of $C\gamma _{5}\otimes \gamma
_{5}C $ or $C\gamma _{\mu }\otimes \gamma ^{\mu }C$ types, where $C$ is the charge conjugation operator. In the present work we restrict ourselves by the simplest case and employ the current$$J=s_{a}^{T}C\gamma _{5}d_{b}\left( \overline{u}_{a}\gamma _{5}C\overline{c}%
_{b}^{T}-\overline{u}_{b}\gamma _{5}C\overline{c}_{a}^{T}\right) .
\label{eq:Curr1}$$To study the pseudoscalar and axial-vector tetraquarks $Z_{PS}$ and $Z_{AV%
\text{ }}$ we utilize $C\gamma _{\mu }\otimes \gamma _{5}C$ type interpolating current $$J_{\mu }=s_{a}^{T}C\gamma _{\mu }d_{b}\left( \overline{u}_{a}\gamma _{5}C%
\overline{c}_{b}^{T}-\overline{u}_{b}\gamma _{5}C\overline{c}_{a}^{T}\right)
. \label{eq:Curr2}$$Then the correlation functions $\Pi (p)$ and $\Pi _{\mu \nu }(p)$ necessary for the sum rule computations take the forms $$\Pi (p)=i\int d^{4}xe^{ipx}\langle 0|\mathcal{T}\{J(x)J^{\dagger
}(0)\}|0\rangle , \label{eq:CF1}$$and $$\Pi _{\mu \nu }(p)=i\int d^{4}xe^{ipx}\langle 0|\mathcal{T}\{J_{\mu
}(x)J_{\nu }^{\dagger }(0)\}|0\rangle . \label{eq:CF2}$$The current $J_{\mu }$ couples to both the pseudoscalar $0^{-}$ and axial-vector $1^{+}$ states, therefore the function $\Pi _{\mu \nu }(p)$ can be used to calculate parameters of the $Z_{PS}$ and $Z_{AV\text{ }}$ tetraquarks.
We start our analysis from calculations of the scalar state’s spectroscopic parameters. In accordance with QCD sum rule method the correlator given by Eq. (\[eq:CF1\]) should be expressed in terms of physical parameters of the $Z_{S}$ state. In the case under analysis $\Pi (p)$ takes simple form and is defined by the equality $$\Pi ^{\mathrm{Phys}}(p)=\frac{\langle 0|J|Z_{S}(p)\rangle \langle
Z_{S}(p)|J^{\dagger }|0\rangle }{m_{Z_{S}}^{2}-p^{2}}+..., \label{eq:Phys1}$$where $m_{Z_{S}}$ is the mass of the $Z_{S}$ state, and dots stand for contributions of the higher resonances and continuum states. In order to simplify $\Pi ^{\mathrm{Phys}}(p)$ we introduce the matrix element $$\langle 0|J|Z_{S}(p)\rangle =f_{Z_{S}}m_{Z_{S}},$$where $m_{Z_{S}}$ and $f_{Z_{S}}$ are the mass and current coupling of $%
Z_{S}(p)$. Then, for the correlation function we obtain $$\Pi ^{\mathrm{Phys}}(p)=\frac{m_{Z_{S}}^{2}f_{Z_{S}}^{2}}{m_{Z_{S}}^{2}-p^{2}%
}+\ldots . \label{eq:Phys1A}$$The Borel transformation applied to $\Pi ^{\mathrm{Phys}}(p)$ yields $$\mathcal{B}\Pi ^{\mathrm{Phys}%
}(p)=m_{Z_{S}}^{2}f_{Z_{S}}^{2}e^{-m_{Z_{S}}^{2}/M^{2}}+\ldots ,
\label{eq:Phys1AB}$$where $M^{2}$ is the Borel parameter.
The same correlation function $\Pi (p)$ calculated in terms of the quark-gluon degrees of freedom reads$$\begin{aligned}
&&\Pi ^{\mathrm{QCD}}(p)=i\int d^{4}xe^{ipx}\left\{ \mathrm{Tr}\left[ \gamma
_{5}\widetilde{S}_{c}^{b^{\prime }b}(-x)\gamma _{5}S_{u}^{a^{\prime }a}(-x)%
\right] \right. \notag \\
&&\times \mathrm{Tr}\left[ S_{s}^{aa^{\prime }}(x)\gamma _{5}\widetilde{S}%
_{d}^{bb^{\prime }}(x)\gamma _{5}\right] -\mathrm{Tr}\left[ \gamma _{5}%
\widetilde{S}_{c}^{a^{\prime }b}(-x)\right. \notag \\
&&\left. \times \gamma _{5}S_{u}^{b^{\prime }a}(-x)\right] \mathrm{Tr}\left[
S_{s}^{aa^{\prime }}(x)\gamma _{5}\widetilde{S}_{d}^{bb^{\prime }}(x)\gamma
_{5}\right] \notag \\
&&-\mathrm{Tr}\left[ \gamma _{5}\widetilde{S}_{c}^{b^{\prime }a}(-x)\gamma
_{5}S_{u}^{a^{\prime }b}(-x)\right] \mathrm{Tr}\left[ S_{s}^{aa^{\prime
}}(x)\gamma _{5}\widetilde{S}_{d}^{bb^{\prime }}(x)\gamma _{5}\right] \notag
\\
&&\left. +\mathrm{Tr}\left[ \gamma _{5}\widetilde{S}_{c}^{a^{\prime
}a}(-x)\gamma _{5}S_{u}^{b^{\prime }b}(-x)\right] \mathrm{Tr}\left[
S_{s}^{aa^{\prime }}(x)\gamma _{5}\widetilde{S}_{d}^{b^{\prime }b}(x)\gamma
_{5}\right] \right\} . \notag \\
&&{} \label{eq:QCD1}\end{aligned}$$Here we use the short-hand notation$$\widetilde{S}_{q(c)}^{ab}(x)=CS_{q(c)}^{Tab}(x)C, \label{eq:Not}$$with $S_{q}(x)$ and $\ S_{c}(x)$ being the $q=u,\ d$ and $c$-quark propagators, respectively.
The QCD sum rules to evaluate $m_{Z_{S}}$ and $f_{Z_{S}}$ can be obtained by choosing the same Lorentz structures in both of $\Pi ^{\mathrm{Phys}}(p)$ and $\Pi ^{\mathrm{QCD}}(p)$, and equating the relevant invariant amplitudes. In the case under investigation the only Lorentz structure which exists in $\Pi (p)$ is one $\sim I$. For calculation of the mass and coupling it is convenient to employ the two-point spectral density $\rho
_{0}^{\mathrm{QCD}}(s)$. In terms of $\rho _{0}^{\mathrm{QCD}}(s)$ the invariant amplitude $\Pi ^{\mathrm{QCD}}(p^{2})$ can be written down as the dispersion integral $$\Pi ^{\mathrm{QCD}}(p^{2})=\int_{(m_{c}+m_{s})^{2}}^{\infty }\frac{\rho
_{0}^{\mathrm{QCD}}(s)}{s-p^{2}}ds+....$$By applying the Borel transformation to $\Pi ^{\mathrm{QCD}}(p^{2})$ , equating the obtained expression with $\mathcal{B}\Pi ^{\mathrm{Phys}%
}(p^{2}) $, and subtracting the contribution due to higher excited and continuum states we find the final sum rules. For the mass of the $Z_{S}$ state it is given by the formula $$m_{Z_{S}}^{2}=\frac{\int_{(m_{c}+m_{s})^{2}}^{s_{0}}dss\rho _{0}^{\mathrm{QCD%
}}(s)e^{-s/M^{2}}}{\int_{(m_{c}+m_{s})^{2}}^{s_{0}}ds\rho _{0}^{\mathrm{QCD}%
}(s)e^{-s/M^{2}}}, \label{eq:SRmass}$$whereas for the current coupling $f_{Z_{S}}$ we get $$f_{Z_{S}}^{2}=\frac{1}{m_{Z_{S}}^{2}}%
\int_{(m_{c}+m_{s})^{2}}^{s_{0}}dse^{(m_{Z_{S}}^{2}-s)/M^{2}}\rho _{0}^{%
\mathrm{QCD}}(s). \label{eq:SRc}$$In Eqs. (\[eq:SRmass\]) and (\[eq:SRc\]) $s_{0}$ is the continuum threshold parameter which separates contributions stemming from the ground-state and ones due to higher resonances and continuum states. The $%
M^{2}$ and $s_{0}$ are two important auxiliary parameters of sum rule computations choices of which should meet some requirements which will be shortly explained below.
In the case of the current $J_{\mu }$ the correlation function $\Pi _{\mu
\nu }^{\mathrm{Phys}}(p)$ derived using the physical parameters of tetraquarks contains two terms. In fact, the current $J_{\mu }$ couples to the pseudoscalar and axial-vector tetraquarks, therefore after inserting into Eq. (\[eq:CF2\]) full set of states and integrating over $x$ we get expression containing contributions of the ground state pseudoscalar and axial-vector particles, i. e. $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{Phys}}(p)=\frac{\langle 0|J_{\mu
}|Z_{AV}(p)\rangle \langle Z_{AV}(p)|J_{\nu }^{\dagger }|0\rangle }{%
m_{Z_{AV}}^{2}-p^{2}} \notag \\
&&+\frac{\langle 0|J_{\mu }|Z_{PS}(p)\rangle \langle Z_{PS}(p)|J_{\nu
}^{\dagger }|0\rangle }{m_{Z_{PS}}^{2}-p^{2}}+..., \label{eq:Phys2}\end{aligned}$$with $m_{Z_{PS}}$ and $m_{Z_{AV}}$ being the mass of the pseudoscalar and axial-vector states, respectively. Here again by dots we denote contributions coming from higher excitations and continuum states in both the pseudoscalar and vector channels. In general one may consider only one of these terms, and compute parameters of the chosen particle with $%
J^{P}=0^{-}$ or $1^{+}$. In the present work we are interested in both of these particles, therefore keep explicitly two terms in $\Pi _{\mu \nu }^{%
\mathrm{Phys}}(p)$, and use different structures to derive two sets of sum rules.
Further simplification of $\Pi _{\mu \nu }^{\mathrm{Phys}}(p)$ can be achieved by expressing the relevant matrix elements in terms of the masses $%
m_{Z_{PS}}$, $m_{Z_{AV}}$ and current couplings $f_{Z_{AV}}$ , $f_{Z_{PS}}$ $$\begin{aligned}
\langle 0|J_{\mu }|Z_{AV}(p)\rangle &=&f_{Z_{AV}}m_{Z_{AV}}\varepsilon _{\mu
}(p), \\
\langle 0|J_{\mu }|Z_{PS}(p)\rangle &=&f_{Z_{PS}}m_{Z_{PS}}p_{\mu }.\end{aligned}$$Then it is easy to show that $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{Phys}}(p)=\frac{m_{Z_{AV}}^{2}f_{Z_{AV}}^{2}}{%
m_{Z_{AV}}^{2}-p^{2}}\left( -g_{\mu \nu }+\frac{p_{\mu }p_{\nu }}{p^{2}}%
\right) \notag \\
&&+\frac{m_{Z_{PS}}^{2}f_{Z_{PS}}^{2}}{m_{Z_{PS}}^{2}-p^{2}}p_{\mu }p_{\nu
}+... \label{eq:Phys2A}\end{aligned}$$In order to obtain the function $\Pi _{\mu \nu }^{\mathrm{QCD}}(p)$ we substitute the interpolating current $J_{\mu }$ from Eq. (\[eq:Curr2\]) into Eq. (\[eq:CF2\]), and contract the quark fields. As a result, for $%
\Pi _{\mu \nu }^{\mathrm{QCD}}(p)$ we get: $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{QCD}}(p)=i\int d^{4}xe^{ipx}\left\{ \mathrm{Tr}%
\left[ \gamma _{5}\widetilde{S}_{c}^{a^{\prime }b}(-x)\gamma
_{5}S_{u}^{b^{\prime }a}(-x)\right] \right. \notag \\
&&\times \mathrm{Tr}\left[ S_{s}^{aa^{\prime }}(x)\gamma _{\nu }\widetilde{S}%
_{d}^{bb^{\prime }}(x)\gamma _{\mu }\right] +\mathrm{Tr}\left[ \gamma _{5}%
\widetilde{S}_{c}^{b^{\prime }a}(-x)\right. \notag \\
&&\left. \times \gamma _{5}S_{u}^{a^{\prime }b}(-x)\right] \mathrm{Tr}\left[
S_{s}^{aa^{\prime }}(x)\gamma _{\nu }\widetilde{S}_{d}^{bb^{\prime
}}(x)\gamma _{\mu }\right] \notag \\
&&-\mathrm{Tr}\left[ \gamma _{5}\widetilde{S}_{c}^{b^{\prime }b}(-x)\gamma
_{5}S_{u}^{a^{\prime }a}(-x)\right] \mathrm{Tr}\left[ S_{s}^{aa^{\prime
}}(x)\gamma _{\nu }\widetilde{S}_{d}^{bb^{\prime }}(x)\gamma _{\mu }\right]
\notag \\
&&\left. -\mathrm{Tr}\left[ \gamma _{5}\widetilde{S}_{c}^{a^{\prime
}a}(-x)\gamma _{5}S_{u}^{b^{\prime }b}(-x)\right] \mathrm{Tr}\left[
S_{s}^{aa^{\prime }}(x)\gamma _{\nu }\widetilde{S}_{d}^{bb^{\prime
}}(x)\gamma _{\mu }\right] \right\} . \notag \\
&&{} \label{eq:CF3}\end{aligned}$$The correlation function $\Pi _{\mu \nu }^{\mathrm{QCD}}(p)$ has the following Lorentz structures $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{QCD}}(p)=\Pi _{AV}^{\mathrm{QCD}}(p^{2})\left(
-g_{\mu \nu }+\frac{p_{\mu }p_{\nu }}{p^{2}}\right) \notag \\
&&+\frac{p_{\mu }p_{\nu }}{p^{2}}\Pi _{PS}^{\mathrm{QCD}}(p^{2}),
\label{eq:QCD2}\end{aligned}$$where $\Pi _{AV}^{\mathrm{QCD}}(p^{2})$ and $\Pi _{PS}^{\mathrm{QCD}}(p^{2})$ are invariant amplitudes corresponding to the axial-vector and pseudoscalar tetraquarks, respectively. Equating the structures $\sim g_{\mu \nu }$ in Eqs. (\[eq:Phys2A\]) and (\[eq:QCD2\]), and performing the Borel transformation it is possible we derive the sum rules for parameters of the axial-vector tetraquark: They are given by Eqs. (\[eq:SRmass\]) and ([eq:SRc]{}) but with $\rho _{0}^{\mathrm{QCD}}(s)$ replaced by $\rho _{V}^{%
\mathrm{QCD}}(s)$.
The sum rules for the pseudoscalar state are found by computing $p^{\mu }\Pi
_{\mu \nu }^{\mathrm{Phys}}(p)$ and $p^{\mu }\Pi _{\mu \nu }^{\mathrm{QCD}%
}(p),$ and matching obtained expressions, which consist of terms with parameters of the pseudoscalar tetraquark. Then for the mass of the pseudoscalar state we again find the sum rule (\[eq:SRmass\]), but $\rho
_{0}^{\mathrm{QCD}}(s)\rightarrow \rho _{S}^{\mathrm{QCD}}(s)$, whereas the coupling $f_{Z_{PS}}$ is determined by the following expression $$f_{Z_{PS}}^{2}=\frac{1}{m_{Z_{PS}}^{4}}%
\int_{(m_{c}+m_{s})^{2}}^{s_{0}}dse^{(m_{Z_{PS}}^{2}-s)/M^{2}}\rho _{S}^{%
\mathrm{QCD}}(s). \label{eq:SRc2}$$In the present work we calculate the two-point spectral densities $\rho
_{0}^{\mathrm{QCD}}(s),$ $\rho _{V}^{\mathrm{QCD}}(s)$ and $\rho _{S}^{%
\mathrm{QCD}}(s)$ by taking into account quark, gluon and mixed vacuum condensates up to eight dimensions.
The sum rules (\[eq:SRmass\]), (\[eq:SRc\]) and (\[eq:SRc2\]) depend on the masses of $c$ and $s$-quarks, and vacuum expectations of quark, gluon and mixed operators, which are presented below: $$\begin{aligned}
&&m_{c}=(1.27\pm 0.03)~\mathrm{GeV},\ \ m_{s}=96_{-4}^{+8}~\mathrm{MeV}
\notag \\
&&\langle \bar{q}q\rangle =-(0.24\pm 0.01)^{3}\ \mathrm{GeV}^{3},\ \langle
\bar{s}s\rangle =0.8\ \langle \bar{q}q\rangle , \notag \\
&&m_{0}^{2}=(0.8\pm 0.1)\ \mathrm{GeV}^{2},\ \langle \overline{q}g_{s}\sigma
Gq\rangle =m_{0}^{2}\langle \overline{q}q\rangle , \notag \\
&&\langle \overline{s}g_{s}\sigma Gs\rangle =m_{0}^{2}\langle \bar{s}%
s\rangle , \notag \\
&&\langle \frac{\alpha _{s}G^{2}}{\pi }\rangle =(0.012\pm 0.004)\,\mathrm{GeV%
}^{4}, \notag \\
&&\langle g_{s}^{3}G^{3}\rangle =(0.57\pm 0.29)\ \mathrm{GeV}^{6}.
\label{eq:Param}\end{aligned}$$For condensates we use their standard values, whereas the masses of the quarks are borrowed from Ref. [@Olive:2016xmw]. In the chiral limit adopted in the present work $m_{u}=m_{d}=0$.
The sum rules contain also, as it has been just noted above, the auxiliary parameters $M^{2}$ and $s_{0}$. It is clear, that physical quantities evaluated from the sum rules should not depend on the Borel parameter and continuum threshold, but in real calculations one can only reduce their effect to a minimum. In fixing of working regions for $M^{2}$ and $s_{0}$ some conditions should be obeyed. Thus, we fix the upper bound $M_{\mathrm{%
max}}^{2}$ of the window $M^{2}\in \lbrack M_{\mathrm{min}}^{2},\ M_{\mathrm{%
max}}^{2}]$ for the Borel parameter by requiring fulfilment of the following constraint $$\mathrm{PC}=\frac{\Pi ^{\mathrm{QCD}}(M_{\mathrm{max}}^{2},\ s_{0})}{\Pi ^{%
\mathrm{QCD}}(M_{\mathrm{max}}^{2},\ \infty )}>0.13, \label{eq:Rest1}$$where $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})=\mathcal{B}\Pi ^{\mathrm{QCD}%
}(p^{2})$ is the Borel transform of the invariant amplitude after the continuum subtraction. Minimal limit for $\mathrm{PC}$ chosen as $\sim 0.1$ is smaller than in the case of the conventional mesons, but is typical for multiquark systems. The lower limit of the same region $M_{\mathrm{min}}^{2}$ is deduced from convergence of the operator product expansion. By quantifying this condition we require that contribution of the last term in OPE should not exceed $5\%$, i. e. $$\frac{\Pi ^{\mathrm{QCD}(\mathrm{Dim}8)}(M_{\mathrm{min}}^{2},\ \infty )}{%
\Pi ^{\mathrm{QCD}}(M_{\mathrm{min}}^{2},\ \infty )}<0.05, \label{eq:Rest2}$$has to be obeyed. Another condition for the lower limit is exceeding of the perturbative contribution the nonperturbative one. In the present work we apply the following criterion: at the lower bound of $M^{2}$ the perturbative contribution has to constitute $\geq 60\%$ part of the full result.
Analysis of the sum rules for the $Z_{S}$ state enable us to fix the Borel and continuum threshold parameters within the limits: $$M^{2}\in \lbrack 2.5-3.5]~\mathrm{GeV}^{2},\ s_{0}~\in \lbrack 8-10]\
\mathrm{GeV}^{2}. \label{eq:BTparam}$$In these regions the pole contribution defined by Eq. (\[eq:Rest1\]) is $%
\mathrm{PC}>0.14$. At the same time, contribution coming from the pole term at $M_{\mathrm{min}}^{2}$ constitutes $\sim 65\%$, and at $M_{\mathrm{max}%
}^{2}$ approximately $60\%$ of the sum rule (\[eq:SRmass\]) used to evaluate the mass of $Z_{S}$ state. The convergence of OPE expansion in these regions is also satisfied. Thus, contribution of the Dim8 term in OPE does not exceed $3\%$ . All these features are seen in Figs. \[fig:PCS\], \[fig:ConvS\] and \[fig:Pert.NpertS\] , where we plot the pole contribution, contributions due to different nonperturbative terms, and the perturbative and total nonperturbative components of $\Pi ^{\mathrm{QCD}%
}(M^{2},\ s_{0})$ to demonstrate that in the regions for $M^{2}$ and $s_{0}$ given by Eq. (\[eq:BTparam\]) constraints imposed on $\Pi ^{\mathrm{QCD}%
}(M^{2},\ s_{0})$ are fulfilled.
![ The pole contribution in the case of the scalar tetraquark vs the Borel parameter $M^{2}$ at different $s_{0}$.[]{data-label="fig:PCS"}](grScalarPCvsMsq.eps){width="8cm"}
![ Contributions due to different nonperturbative terms in the case of $Z_S$ as functions of $M^2$ (left panel), and $s_0$ (right panel). []{data-label="fig:ConvS"}](grScalarNonPertvsMsq.eps "fig:"){width="8cm"} ![ Contributions due to different nonperturbative terms in the case of $Z_S$ as functions of $M^2$ (left panel), and $s_0$ (right panel). []{data-label="fig:ConvS"}](grScalarNonPertvss0.eps "fig:"){width="8cm"}
![ The perturbative and nonperturbative contributions to $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})$ of the scalar particle. Left: as functions of $M^2$ at central value $s_0=9\ \mathrm{GeV}^2$, right: as functions of $s_0$ at $M^2=3\ \mathrm{GeV}^2$.[]{data-label="fig:Pert.NpertS"}](grScalarContvsMsq.eps "fig:"){width="8cm"} ![ The perturbative and nonperturbative contributions to $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})$ of the scalar particle. Left: as functions of $M^2$ at central value $s_0=9\ \mathrm{GeV}^2$, right: as functions of $s_0$ at $M^2=3\ \mathrm{GeV}^2$.[]{data-label="fig:Pert.NpertS"}](grScalarContvss0.eps "fig:"){width="8cm"}
From obtained sum rules for the mass and current coupling of $%
Z_{S}$ state we find $$m_{Z_{S}}=2628_{-153}^{+166}\ \ \mathrm{MeV},\
f_{Z_{S}}=(0.21_{-0.05}^{+0.06})\cdot 10^{-2}\ \mathrm{GeV}^{4}.$$In Figs. \[fig:Mass1\] and \[fig:Coup1\], $m_{Z_{S}}$ and $\ f_{Z_{S}}$ are depicted as functions of $M^{2}$ and $s_{0}$. It is seen that while effects of varying of these parameters on the mass $m_{Z_{S}}$ are small, dependence of the current coupling $f_{Z_{S}}$ on chosen values of the continuum threshold parameter is noticeable. These effects together with uncertainties of the input parameters generate the theoretical errors in the sum rule calculations, which are their unavoidable feature and may reach $%
30\%$ of the central values.
![ The mass of the $Z_S$ state as a function of the Borel parameter $M^2$ at fixed values of $s_0$ (left panel), and as a function of the continuum threshold $s_0$ at fixed $M^2$ (right panel).[]{data-label="fig:Mass1"}](massZSvsMsq.eps "fig:"){width="8cm"} ![ The mass of the $Z_S$ state as a function of the Borel parameter $M^2$ at fixed values of $s_0$ (left panel), and as a function of the continuum threshold $s_0$ at fixed $M^2$ (right panel).[]{data-label="fig:Mass1"}](massZSvss0.eps "fig:"){width="8cm"}
![ The dependence of the current coupling $f_{Z_S}$ of the scalar $Z_S$ tetraquark on the Borel parameter at chosen values of $s_0$ (left panel), and on the continuum threshold parameter $s_0$ at fixed $M^2$ (right panel).[]{data-label="fig:Coup1"}](fZSvsMsq.eps "fig:"){width="8cm"} ![ The dependence of the current coupling $f_{Z_S}$ of the scalar $Z_S$ tetraquark on the Borel parameter at chosen values of $s_0$ (left panel), and on the continuum threshold parameter $s_0$ at fixed $M^2$ (right panel).[]{data-label="fig:Coup1"}](fZSvss0.eps "fig:"){width="8cm"}
The analogous studies can be carried out for the pseudoscalar and axial-vector tetraquarks. From performed analysis we conclude that regions $$M^{2}\in \lbrack 2.5-3.5]~\mathrm{GeV}^{2},\ s_{0}~\in \lbrack 9.5-11.5]\
\mathrm{GeV}^{2} \label{eq:BTAVparam}$$can be used to evaluate the spectroscopic parameters of the pseudoscalar and axial-vector tetraquarks, as well. Results of computations for the axial-vector state are depicted in Figs. \[fig:PCAV\], \[fig:ConvAV\] and \[fig:Pert.NpertAV\], which confirm our conclusions. The similar results are also valid for the pseudoscalar tetraquark $Z_{PS}$.
![ The pole contribution in the case of the axial-vector tetraquark as a function of the Borel parameter $M^2$.[]{data-label="fig:PCAV"}](grAVPCvsMsq.eps){width="8cm"}
![ Nonperturbative contributions to $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})_{AV}$ as functions of $M^2$ (left panel), and $s_0$ (right panel). []{data-label="fig:ConvAV"}](grAVNonPertvsMsq.eps "fig:"){width="8cm"} ![ Nonperturbative contributions to $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})_{AV}$ as functions of $M^2$ (left panel), and $s_0$ (right panel). []{data-label="fig:ConvAV"}](grAVNonPertvss0.eps "fig:"){width="8cm"}
![ The perturbative and nonperturbative components of $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})_{AV}$. Left: as functions of $M^2$ at $s_0=10.5\ \mathrm{GeV}^2$, right: as functions of $s_0$ at the central value of the Borel parameter $M^2=3\ \mathrm{GeV}^2$.[]{data-label="fig:Pert.NpertAV"}](grAVContvsMsq.eps "fig:"){width="8cm"} ![ The perturbative and nonperturbative components of $\Pi ^{\mathrm{QCD}}(M^{2},\ s_{0})_{AV}$. Left: as functions of $M^2$ at $s_0=10.5\ \mathrm{GeV}^2$, right: as functions of $s_0$ at the central value of the Borel parameter $M^2=3\ \mathrm{GeV}^2$.[]{data-label="fig:Pert.NpertAV"}](grAVContvss0.eps "fig:"){width="8cm"}
In Figs. \[fig:Mass2\] and \[fig:Coup2\] we plot dependence of the axial-vector tetraquark’s mass and current coupling on $M^{2}$ and $s_{0}$. As is seen, estimations made for theoretical errors in the case of $Z_{S}$ are valid for the $Z_{AV}$ state, as well.
Our results for the masses and current couplings of $J^{P}=0^{+},\ 0^{-}$ and $J^{P}=1^{+}$ charm-strange tetraquarks are collected in Table [tab:Results1]{}. The working ranges for the parameters $M^{2}$ and $s_{0}$, and errors of the calculations are also presented in Table \[tab:Results1\].
------------------------------------------------------------------------------------------------------------------------------------------------
$Z$ $Z_S$ $Z_{PS}$ $Z_{AV}$
------------------------- --------------------------------------- ----------------------------------------- ------------------------------------
$M^2 ~(\mathrm{GeV}^2$) $2.5-3.5$ $2.5-3.5$ $2.5-3.5$
$s_0 ~(\mathrm{GeV}^2$) $8-10$ $9.5-11.5$ $9.5-11.5$
$m_{Z} ~(\mathrm{MeV})$ $2628^{+166}_{-153}$ $2719^{+144}_{-156}$ $%
2826^{+134}_{-157}$
$f_{Z}\cdot10^{3}$ $2.1^{+0.6}_{-0.5} ~(\mathrm{GeV}^4)$ $% $2.6^{+0.6}_{-0.7} ~(\mathrm{GeV}%
0.83^{+0.09}_{-0.11} ~(\mathrm{GeV}^3)$ ^4)$
------------------------------------------------------------------------------------------------------------------------------------------------
: The masses and current couplings of the $Z_S$, $Z_{PS}$ and $Z_{AV}$ tetraquarks.[]{data-label="tab:Results1"}
![ The mass of the $Z_{AV}$ state vs Borel parameter $M^2$ at fixed values of $s_0$ (left panel), and vs continuum threshold $s_0$ at fixed values of $M^2$ (right panel).[]{data-label="fig:Mass2"}](massZAVvsMsq.eps "fig:"){width="8cm"} ![ The mass of the $Z_{AV}$ state vs Borel parameter $M^2$ at fixed values of $s_0$ (left panel), and vs continuum threshold $s_0$ at fixed values of $M^2$ (right panel).[]{data-label="fig:Mass2"}](massZAVvss0.eps "fig:"){width="8cm"}
![ The current coupling $f_{Z_{AV}}$ of the $Z_{AV}$ state vs Borel parameter $M^2$ at chosen values of $s_0$ (left panel), and vs $s_0$ at fixed values of $M^2$ (right panel).[]{data-label="fig:Coup2"}](fZAVvsMsq.eps "fig:"){width="8cm"} ![ The current coupling $f_{Z_{AV}}$ of the $Z_{AV}$ state vs Borel parameter $M^2$ at chosen values of $s_0$ (left panel), and vs $s_0$ at fixed values of $M^2$ (right panel).[]{data-label="fig:Coup2"}](fZAVvss0.eps "fig:"){width="8cm"}
Decay channels of the scalar tetraquark $Z_{S}$ {#sec:ScalDec}
===============================================
In this section we calculate the width of processes $Z_{S}\rightarrow
D_{s}\pi $, $Z_{S}\rightarrow DK$ and $Z_{S}\rightarrow D_{s1}(2460)\pi $, which in the light of the result obtained for $m_{Z_{S}}$ are kinematically allowed decay channels of the scalar tetraquark. It is evident that in all these channels the final mesons are particles with negative charges. For simplicity of expressions throughout this paper we do not show explicitly charges of the final mesons. Let us note that first two processes are $S$-wave decay modes, whereas the last one is $P$-wave decay.
In order to evaluate the width of these decays we have to calculate the strong couplings corresponding to the vertices $Z_{S}D_{s}\pi $, $Z_{S}DK$ and $Z_{S}D_{s1}(2460)\pi $. This task can be fulfilled by analysis of corresponding correlation functions and calculating them using light-cone sum rule method. To calculate the strong coupling $g_{Z_{S}D_{s}\pi }$ and width of the decay $Z_{S}\rightarrow D_{s}\pi $ we consider the correlator $$\Pi (p,q)=i\int d^{4}xe^{ipx}\langle \pi (q)|\mathcal{T}\{J^{D_{s}}(x)J^{%
\dag }(0)\}|0\rangle , \label{eq:CFScalar}$$where $$J^{D_{s}}(x)=\overline{c}_{i}(x)\gamma _{5}s_{i}(x),$$is the interpolating current of the pseudoscalar meson $D_{s}$.
The correlation function $\Pi (p,q)$ in terms of the physical parameters of the involved particles is equal to $$\begin{aligned}
&&\Pi ^{\mathrm{Phys}}(p,q)=\frac{\langle 0|J^{D_{s}}|D_{s}\left( p\right)
\rangle }{p^{2}-m_{D_{s}}^{2}}\langle D_{s}\left( p\right) \pi
(q)|Z_{S}(p^{\prime })\rangle \notag \\
&&\times \frac{\langle Z_{S}(p^{\prime })|J^{\dagger }|0\rangle }{p^{\prime
2}-m_{Z_{S}}^{2}}\ldots . \label{eq:PhysScalar}\end{aligned}$$By introducing the matrix elements $$\begin{aligned}
&&\langle 0|J^{D_{s}}|D_{s}\left( p\right) \rangle =\frac{%
m_{D_{s}}^{2}f_{D_{s}}}{m_{c}+m_{s}}, \notag \\
&&\langle D_{s}\left( p\right) \pi (q)|Z_{S}(p^{\prime })\rangle
=g_{Z_{S}D_{s}\pi }p\cdot p^{\prime }, \label{eq:Vertex1}\end{aligned}$$we can rewrite $\Pi ^{\mathrm{Phys}}(p,q)$ in the form$$\Pi ^{\mathrm{Phys}}(p,q)=\frac{g_{Z_{S}D_{s}\pi
}m_{D_{s}}^{2}m_{Z_{S}}f_{D_{s}}f_{Z_{S}}}{(p^{2}-m_{D_{s}}^{2})(p^{\prime
2}-m_{Z_{S}}^{2})(m_{c}+m_{s})}p\cdot p^{\prime }+\ldots$$Here $m_{D_{s}}$ and $f_{D_{s}}$ are the mass and decay constant of the meson $D_{s}$, respectively.
In terms of the quark-gluon degrees of freedom $\Pi (p,q)$ is given by the expression $$\begin{aligned}
&&\Pi ^{\mathrm{QCD}}(p,q)=i\int d^{4}xe^{ipx}\left\{ \left[ \gamma _{5}%
\widetilde{S}_{s}^{ia}(x)\gamma _{5}\widetilde{S}_{c}^{bi}(-x)\gamma _{5}%
\right] _{\alpha \beta }\right. \notag \\
&&\times \langle \pi (q)|\overline{d}_{\alpha }^{b}(0)u_{\beta
}^{a}(0)|0\rangle -\left[ \gamma _{5}\widetilde{S}_{s}^{ia}(x)\gamma _{5}%
\widetilde{S}_{c}^{ai}(-x)\gamma _{5}\right] \notag \\
&&\left. \times \langle \pi (q)|\overline{d}_{\alpha }^{b}(0)u_{\beta
}^{b}(0)|0\rangle \right\} , \label{eq:QCD3}\end{aligned}$$where $\alpha $ and $\beta $ are the spinor indices. To continue we employ the expansion $$\overline{u}_{\alpha }^{a}d_{\beta }^{d}\rightarrow \frac{1}{4}\Gamma
_{\beta \alpha }^{j}\left( \overline{u}^{a}\Gamma ^{j}d^{d}\right) ,
\label{eq:MatEx}$$with $\Gamma ^{j}$ being the full set of Dirac matrices $$\Gamma ^{j}=\mathbf{1,\ }\gamma _{5},\ \gamma _{\lambda },\ i\gamma
_{5}\gamma _{\lambda },\ \sigma _{\lambda \rho }/\sqrt{2}.$$The operators $\overline{u}^{a}(0)\Gamma ^{j}d^{d}(0)$, as well as three-particle operators that appear due to insertion of $G_{\mu \nu }$ from propagators $\widetilde{S}_{s}(x)$ and $\widetilde{S}_{c}(-x){}$ into $%
\overline{u}^{a}(0)\Gamma ^{j}d^{d}(0)$ give rise to local matrix elements of the pion. In other words, instead of the distribution amplitudes the function $\Pi ^{\mathrm{QCD}}(p,q)$ depends on the pion’s local matrix elements. Then, the conservation of four-momentum in the tetraquark-meson-meson vertex can be obeyed by setting $q=0$. In the limit $%
q\rightarrow 0$ we get $p=p^{\prime }$ and have to carry out Borel transformations over one variable $p^{2}$. This condition has to be implemented in the physical side of the sum rule, as well [Agaev:2016dev,Agaev:2017foq]{}.
After substituting Eq. (\[eq:MatEx\]) into the expression of the correlation function and performing the summation over color indices in accordance with recipes presented in a detailed form in Ref. [Agaev:2016dev]{}, we fix local matrix elements that enter to $\Pi ^{\mathrm{%
QCD}}(p,q)$. It turns out that only the matrix element of the pion $$\langle 0|\overline{d}(0)i\gamma _{5}u(0)|\pi (q)\rangle =f_{\pi }\mu _{\pi
}, \label{eq:MatE2}$$where $\mu _{\pi }=-2\langle \bar{q}q\rangle /f_{\pi }^{2}$ contributes to the correlation function. Other matrix elements including three-particle ones either do not contribute to a final expression of $\Pi ^{\mathrm{QCD}%
}(p,q)$ or vanish in the soft limit $q\rightarrow 0$.
In the soft limit the Borel transformation of relevant invariant function $%
\Pi ^{\mathrm{QCD}}(p^{2})$ can be obtained after the following operations: we find the spectral density $\rho ^{\mathrm{pert.}}(s)$ as imaginary part of $\Pi ^{\mathrm{pert.}}(p^{2})$, which is the perturbative component of the full correlation function. It is calculated using Eq. (\[eq:QCD3\]) and keeping in the quark propagators only their perturbative components. All other terms in $\Pi ^{\mathrm{QCD}}(p^{2})$ constitute the nonperturbative peace of the correlator, i.e. function $\Pi ^{\mathrm{n.-pert.}}(p^{2})$. We calculate Borel transformation of $\Pi ^{\mathrm{n.-pert.}}(p^{2})$ directly from Eq. (\[eq:QCD3\]) in accordance with prescriptions of Ref. [Belyaev:1994zk]{} , and by this way bypass intermediate steps, i. e. computation of $\rho ^{\mathrm{n.-pert.}}(s)$, which becomes unnecessary in this case. This approach considerably simplifies calculations and allows us to find explicitly $\Pi ^{\mathrm{n.-pert.}}(M^{2})\equiv \mathcal{B}\Pi ^{%
\mathrm{n.-pert.}}(p^{2})$. For the spectral density $\rho ^{\mathrm{pert.}%
}(s)$ we obtain$$\rho ^{\mathrm{pert.}}(s)=\frac{f_{\pi }\mu _{\pi }}{16\pi ^{2}s}%
(s-m_{c}^{2})(s+2m_{c}m_{s}-m_{c}^{2}). \label{eq:SD1}$$The Borel transformed $\Pi ^{\mathrm{n.-pert.}}(M^{2})$ contains terms up to nine dimensions and reads$$\Pi ^{\mathrm{n.-pert.}}(M^{2})=\frac{f_{\pi }\mu _{\pi }}{12M^{2}}%
e^{-m_{c}^{2}/M^{2}}\sum_{l=1}^{5}F_{l}(M^{2}), \label{eq:Borel1}$$where$$\begin{aligned}
&&F_{1}(M^{2})=-\langle \bar{s}s\rangle
(m_{c}^{2}m_{s}+2m_{c}M^{2}-m_{s}M^{2}), \notag \\
&&F_{2}(M^{2})=\frac{m_{c}}{12M^{2}}\langle \frac{\alpha _{s}G^{2}}{\pi }%
\rangle \int_{0}^{1}\frac{dze^{m_{c}^{2}/M^{2}-m_{c}^{2}/[M^{2}z(1-z)]}}{%
z(z-1)^{3}} \notag \\
&&\times \left[ m_{c}^{3}z+2m_{s}M^{2}(1-z)^{2}z+m_{c}^{2}m_{s}(1-z)\right] ,
\notag \\
&&F_{3}(M^{2})=\frac{m_{c}^{3}}{6M^{4}}\langle \overline{s}g_{s}\sigma
Gs\rangle \left( m_{c}m_{s}+3M^{2}\right) , \notag \\
&&F_{4}(M^{2})=-\frac{m_{c}\pi ^{2}}{18M^{6}}\langle \frac{\alpha _{s}G^{2}}{%
\pi }\rangle \langle \bar{s}s\rangle \left[ m_{c}^{3}m_{s}\right. \notag \\
&&\left. +2m_{c}M^{2}(m_{c}+m_{s})+6M^{4}\right] , \notag \\
&&F_{5}(M^{2})=\frac{m_{c}\pi ^{2}}{108M^{10}}\langle \frac{\alpha _{s}G^{2}%
}{\pi }\rangle \langle \overline{s}g_{s}\sigma Gs\rangle \left(
m_{c}m_{s}+3M^{2}\right) \notag \\
&&\times \left( m_{c}^{4}+6m_{c}^{2}M^{2}+6M^{4}\right) .
\label{eq:Ffunctions}\end{aligned}$$
Then in the soft limit the Borel transformation of $\Pi ^{\mathrm{QCD}%
}(p^{2})$ takes the form$$\begin{aligned}
&&\mathcal{B}\Pi ^{\mathrm{QCD}}(p^{2})=\int_{(m_{c}+m_{s})^{2}}^{\infty
}\rho ^{\mathrm{pert.}}(s)e^{-s/M^{2}}ds \notag \\
&&+\Pi ^{\mathrm{n.-pert.}}(M^{2}).\end{aligned}$$In the same limit the Borel transformation of $\Pi ^{\mathrm{Phys}}(p^{2})$ is given by the expression$$\mathcal{B}\Pi ^{\mathrm{Phys}}(p^{2})=\frac{g_{Z_{S}D_{s}\pi
}m_{D_{s}}^{2}m_{Z_{S}}f_{D_{s}}f_{Z_{S}}}{m_{c}+m_{s}}m^{2}\frac{%
e^{-m^{2}/M^{2}}}{M^{2}}+..., \label{eq:PhysScalA}$$where $m^{2}=(m_{D_{s}}^{2}+m_{Z_{S}}^{2})/2.$
After equating $\mathcal{B}\Pi ^{\mathrm{Phys}}(p^{2})$ and $\mathcal{B}\Pi
^{\mathrm{QCD}}(p^{2})$ one has to subtract contributions of higher resonances and continuum states. In the case of standard sum rules (i.e. $%
q\neq 0$) this can be carried out quite easily, because Borel transformation suppress all undesired terms in the physical side of the equality. But in the soft limit $\mathcal{B}\Pi ^{\mathrm{Phys}}(p^{2})$ contains terms which are not suppressed even after Borel transformation [@Belyaev:1994zk], therefore additional manipulations are required to remove them from the phenomenological side of the sum rule. Acting by the operator $$\mathcal{P}(M^{2},m^{2})=\left( 1-M^{2}\frac{d}{dM^{2}}\right)
M^{2}e^{m^{2}/M^{2}} \label{eq:softop}$$one can achieve this goal [@Ioffe:1983ju]. Then subtraction can be performed in a standard manner and leads to the sum rule$$\begin{aligned}
&&g_{Z_{S}D_{s}\pi }=\frac{m_{c}+m_{s}}{%
m_{D_{s}}^{2}m_{Z_{S}}m^{2}f_{D_{s}}f_{Z_{S}}}\mathcal{P}(M^{2},m^{2})
\notag \\
&&\times \left[ \int_{(m_{c}+m_{s})^{2}}^{s_{0}}\rho ^{\mathrm{pert.}%
}(s)e^{-s/M^{2}}ds+\Pi ^{\mathrm{n.-pert.}}(M^{2})\right] . \notag \\
&&{} \label{eq:SRcoupl1}\end{aligned}$$It is worth noting that we do not perform continuum subtractionin in nonperturbative terms $\sim (M^{2})^{0}$ and $\sim (M^{2})^{-n},\
n=1,2\ldots $ [@Belyaev:1994zk].
With the coupling $g_{Z_{S}D_{s}\pi }$ at hands it is straightforward to evaluate the width of the decay $Z_{S}\rightarrow D_{s}\pi $$$\begin{aligned}
&&\Gamma (Z_{S}\to D_{s}\pi )=\frac{g_{Z_{S}D_{s}\pi }^{2}}{96\pi
m_{Z_{S}}^{2}}\left( m_{Z_{S}}^{2}+m_{D_{s}}^{2}-m_{\pi }^{2}\right) ^{2}
\notag \\
&&\times f(m_{Z_{S}},m_{D_{s}},m_{\pi }),\end{aligned}$$where the function $f(x,y,z)$ is$$f(x,y,z)=\frac{1}{2x}\sqrt{%
x^{4}+y^{4}+z^{4}-2x^{2}y^{2}-2x^{2}z^{2}-2y^{2}z^{2}}.$$
Another decay channel of the doubly charged charm-strange tetraquark is $%
Z_{S}\rightarrow DK$. Correlation function that should be considered in this case is given by the expression $$\Pi _{K}(p,q)=i\int d^{4}xe^{ipx}\langle K(q)|\mathcal{T}\{J^{D}(x)J^{\dag
}(0)\}|0\rangle ,$$where the interpolating current for $D$ meson is $$J^{D}(x)=\overline{c}_{i}(x)\gamma _{5}d_{i}(x).$$The analysis of the channel $Z_{S}\rightarrow DK$ does not differ considerably from consideration of $Z_{S}\rightarrow D_{s}\pi $ decay. Because particles in final states $D_{s}\pi $ and $DK$ are pseudoscalar mesons, differences between two decay channels are encoded in the matrix element $$\langle 0|J^{D}|D\left( p\right) \rangle =\frac{m_{D}^{2}f_{D}}{m_{c}+m_{d}},$$and local matrix element of $K$ meson $$\langle 0|\overline{u}(0)i\gamma _{5}s(0)|K(q)\rangle =\frac{f_{K}m_{K}^{2}}{%
m_{s}+m_{u}}, \label{eq:MatEl3}$$that contributes to $\Pi _{K}^{\mathrm{QCD}}(p,q)$, where $m_{K}$ and $f_{K}$ are the mass and decay constant of $K$ meson. The strong coupling $%
g_{Z_{S}DK}$ with evident replacements is defined by Eq. (\[eq:Vertex1\]).
In the $P$-wave decay $Z_{S}\rightarrow D_{s1}(2460)\pi $ the interpolating current, matrix element and strong coupling of the axial-vector meson $%
D_{s1}(2460)$ are introduced by means of the formulas $$\begin{aligned}
&&J_{\mu }^{D_{s1}}(x)=\overline{c}_{i}(x)\gamma _{\mu }\gamma _{5}s_{i}(x),
\notag \\
&&\langle 0|J_{\mu }^{D_{s1}}|D_{s1}\left( p\right) \rangle
=f_{D_{s1}}m_{D_{s1}}\varepsilon _{\mu }, \notag \\
&&\langle D_{s1}\left( p\right) \pi (q)|Z_{S}(p^{\prime })\rangle
=g_{Z_{S}D_{s1}\pi }\varepsilon ^{\ast }\cdot p^{\prime },
\label{eq:Vertex2}\end{aligned}$$with $m_{D_{s1}}$ , $f_{D_{s1}}$ and $\varepsilon _{\mu }$ being its mass , decay constant and polarization vector, respectively.
The remaining operations and intermediate steps in both cases are standard ones, therefore we refrain from presenting them here in a detailed form, and write down only formula for the decay width $\Gamma (Z_{S}\rightarrow
D_{s1}(2460)\pi ):$ $$\Gamma (Z_{S}\rightarrow D_{s1}(2460)\pi )=\frac{g_{Z_{S}D_{s1}\pi }^{2}}{%
24\pi m_{D_{s1}}^{2}}f^{3}(m_{Z_{S}},m_{D_{s1}},m_{\pi }).$$
Numerical calculations are carried out using the sum rules derived for strong couplings and expressions for widths of different decay modes of $%
Z_{S}$. The masses and decay constants of $D_{s}$ , $D$ and $D_{s1}(2460)$, as well as $\pi $ and $K$ mesons which we employ in numerical computations are collected in Table \[tab:Param\]. The masses of particles are taken from Ref. [@Olive:2016xmw], for decay constants of $D$ and $D_{s}$ mesons we use information from Ref. [@Rosner:2015wva], decay constant of $D_{s1}(2460)$ is borrowed from [@Sungu:2010zz]. Table [tab:Param]{} contains also parameters of the $D^{\ast }$, $D_{s}^{\ast }$ and $D_{s0}^{\ast }(2317)$ mesons which will be used in the next section.
The Borel parameter $M^{2}$ and continuum threshold $s_{0}$ in coupling calculations are chosen as in Eq. (\[eq:BTparam\]). For the strong couplings of the explored vertices and width of the decay modes we obtain: for the channel $Z_{S}\rightarrow D_{s}\pi $ $$\begin{aligned}
&&g_{Z_{S}D_{s}\pi } =(0.51\pm 0.14)\ \mathrm{GeV}^{-1} \notag \\
&&\Gamma (Z_{S} \rightarrow D_{s}\pi )=(8.27\pm 2.32)\ \mathrm{MeV,}
\label{eq:DW1}\end{aligned}$$for the mode $Z_{S}\rightarrow DK$$$\begin{aligned}
&&g_{Z_{S}DK} =(1.54\pm 0.43)\ \mathrm{GeV}^{-1}, \notag \\
&&\Gamma (Z_{S} \rightarrow DK)=(57.41\pm 14.93)\ \mathrm{MeV,}
\label{eq:DW2}\end{aligned}$$and for $Z_{S}\rightarrow D_{s1}(2460)\pi $$$\begin{aligned}
&&g_{Z_{S}Ds1\pi } =26.16\pm 7.36, \notag \\
&&\Gamma (Z_{S} \rightarrow D_{s1}(2460)\pi )=(1.21\pm 0.38)\ \mathrm{MeV.}
\label{eq:DW3}\end{aligned}$$The full width of the scalar tetraquark $Z_{S}$ on the basis of considered decay modes is equal to$$\Gamma _{Z_{S}}=(66.89\pm 15.11)\ \ \mathrm{MeV,}$$which is typical for a diquark-antidiquark state: The tetraquark $Z_{S}$ belongs neither to a class of broad resonances $\Gamma \sim 200\ \mathrm{MeV}
$ nor to a class of very narrow states $\Gamma \sim 1\ \mathrm{MeV}$.
The charmed particle composed of four different quarks as a partner of the $%
X(5568)$ resonance was previously investigated in our work [Agaev:2016lkl]{}. We analyzed this state using the interpolating currents of both $C\gamma _{5}\otimes \gamma _{5}C$ and $C\gamma _{\mu }\otimes \gamma
^{\mu }C$ types. The diquark-antidiquark composition of $X_{c}=[su][%
\overline{c}\overline{d}]$ means that it is a neutral particle. Nevertheless, it is instructive to compare parameters of $X_{c}$ with results for $Z_{S}$ obtained in the present work. In the case of the interpolating current $C\gamma _{5}\otimes \gamma _{5}C$ we found $%
m_{X_{c}}=(2634\pm 62)\ \mathrm{MeV}$ which is very close to our present result. The processes $X_{c}\rightarrow \overline{D}^{0}\overline{K}^{0}$ and $X_{c}\rightarrow D_{s}^{-}\pi ^{+}$ were also subject of studies in Ref. [@Agaev:2016lkl]. Width of these decay channels $\Gamma
(X_{c}\rightarrow \overline{D}^{0}\overline{K}^{0})=(53.7\pm 11.6)\ \mathrm{%
MeV}$ and $\Gamma (X_{c}\rightarrow D_{s}^{-}\pi ^{+})=(8.2\pm 2.1)\ \mathrm{%
MeV}$ are comparable with ones presented in Eqs. (\[eq:DW1\]) and ([eq:DW2]{}).
Parameters Values
---------------------- -----------------------------------------
$m_{D}$ $(1869.5 \pm 0.4) ~\mathrm{MeV}$
$f_{D}$ $(211.9 \pm 1.1)~\mathrm{MeV}$
$m_{D_s}$ $(1969.0 \pm 1.4) ~\mathrm{MeV}$
$f_{D_s}$ $(249.0 \pm 1.2) ~\mathrm{MeV}$
$m_{D_{s1}} $ $(2459.6 \pm 0.9)~\mathrm{MeV} $
$f_{D_{s1}}$ $(481\pm 164) ~\mathrm{MeV}$
$m_{D_{s}^{\ast }}$ $(2112.1\pm 0.4)~\mathrm{MeV}$
$f_{D_{s}^{\ast }}$ $(308\pm 21) ~\mathrm{MeV}$
$m_{D^{\ast}}$ $(2010.26\pm 0.25)~\mathrm{MeV}$
$f_{D^{\ast }}$ $(252.2\pm 22.66) ~\mathrm{MeV}$
$m_{D_{s0}^{\ast }}$ $(2318.0\pm 1.0)~\mathrm{MeV}$
$f_{D_{s0}^{\ast }}$ $201\ \mathrm{MeV}$
$m_{K}$ $(493.677\pm 0.016) ~\mathrm{MeV}$
$f_{K}$ $156~\mathrm{MeV}$
$m_{\pi}$ $(139.57061 \pm 0.00024 )~\mathrm{MeV}$
$f_{\pi}$ $131~\mathrm{MeV}$
: Parameters of the mesons used in numerical calculations.[]{data-label="tab:Param"}
$Z_{PS}\rightarrow \ D_{s}^{\ast }\protect\pi ,\ D^{\ast }K,\
D_{s0}^{\ast}(2317)\protect\pi $ and $Z_{AV}\rightarrow \ D_{s}^{\ast }%
\protect\pi ,\ D^{\ast }K,\ D_{s1}(2460)\protect\pi $ decays of the pseudoscalar and axial-vector tetraquarks {#sec:PsVDec}
=============================================================================================================
The pseudoscalar $Z_{PS}$ and axial-vector $Z_{AV}$ tetraquarks may decay through different channels. Among kinematically allowed decay channels of $%
Z_{PS}$ state are $S-$wave mode $Z_{PS}\rightarrow D_{s0}(2317)\pi $, and $%
P- $wave modes $Z_{PS}\rightarrow D_{s}^{\ast }\pi$ and $\ D^{\ast }K$. The decays of the tetraquark $Z_{AV}$ include $S-$wave channels $%
Z_{AV}\rightarrow D_{s}^{\ast }\pi ,\ D^{\ast }K$ and $P-$wave mode $%
Z_{AV}\rightarrow D_{s1}(2460)\pi$.
It is seen that both $Z_{PS}$ and $Z_{AV}$ states decay to $D_{s}^{\ast }\pi
$ and $D^{\ast }K$, therefore these channels should be analyzed in a connected form. We start our investigation from analysis of the decays $%
Z_{AV}\rightarrow D_{s}^{\ast }\pi $ and $Z_{PS}\rightarrow D_{s}^{\ast }\pi
$, and construct the following correlation function $$\Pi _{\mu \nu }(p,q)=i\int d^{4}xe^{ipx}\langle \pi (q)|\mathcal{T}\{J_{\mu
}^{D_{s}^{\ast }}(x)J_{\nu }^{\dag }(0)\}|0\rangle , \label{eq:CF4}$$where $J_{\mu }^{D_{s}^{\ast }}(x)$ is the interpolating current of the $%
D_{s}^{\ast }$ meson $$J_{\mu }^{D_{s}^{\ast }}(x)=\overline{c}_{l}(x)\gamma _{\mu }s_{l}(x).
\label{eq:Curr3}$$The function $\Pi _{\mu \nu }(p,q)$ will be computed employing QCD sum rule on the light-cone and using a technique of the soft-meson approximation. Because the current $J_{\nu }(x)$ couples to both the pseudoscalar and axial-vector tetraquarks the correlator $\Pi _{\mu \nu }^{\mathrm{Phys}%
}(p,q) $ expressed in terms of the physical parameters of the involved particles and vertices contains two components: Indeed, for $\Pi _{\mu \nu
}^{\mathrm{Phys}}(p,q)$ we find: $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{Phys}}(p,q)=\frac{\langle 0|J_{\mu }^{D_{s}^{\ast
}}|D_{s}^{\ast }\left( p\right) \rangle }{p^{2}-m_{D_{s}^{\ast }}^{2}}%
\langle D_{s}^{\ast }\left( p\right) \pi (q)|Z_{PS}(p^{\prime })\rangle
\notag \\
&&\times \frac{\langle Z_{PS}(p^{\prime })|J_{\nu }^{\dagger }|0\rangle }{%
p^{\prime 2}-m_{Z_{PS}}^{2}}+\frac{\langle 0|J_{\mu }^{D_{s}^{\ast
}}|D_{s}^{\ast }\left( p\right) \rangle }{p^{2}-m_{D_{s}^{\ast }}^{2}}%
\langle D_{s}^{\ast }\left( p\right) \pi (q)|Z_{AV}(p^{\prime })\rangle
\notag \\
&&\times \frac{\langle Z_{AV}(p^{\prime })|J_{\nu }^{\dagger }|0\rangle }{%
p^{\prime 2}-m_{Z_{AV}}^{2}}\ldots . \label{eq:CF5}\end{aligned}$$The terms in Eq. (\[eq:CF5\]) are contributions of vertices $%
Z_{PS}D_{s}^{\ast }\pi $ and $Z_{AV}D_{s}^{\ast }\pi $, where all particles are on their ground states. The dots stand for effects due to the higher resonances and continuum.
We introduce the $D_{s}^{\ast }$ meson matrix element $$\langle 0|J_{\mu }^{D_{s}^{\ast }}|D_{s}^{\ast }\left( p\right) \rangle
=f_{D_{s}^{\ast }}m_{D_{s}^{\ast }}\varepsilon _{\mu },$$where $m_{D_{s}^{\ast }}$ , $f_{D_{s}^{\ast }}$ and $\varepsilon _{\mu }$ are its mass, decay constant and polarization vector, respectively. We define also the matrix elements corresponding to the vertices in the following manner $$\begin{aligned}
&&\langle D_{s}^{\ast }\left( p\right) \pi (q)|Z_{AV}(p^{\prime })\rangle
=g_{Z_{AV}D_{s}^{\ast }\pi }\left[ \left( p\cdot p^{\prime }\right) \left(
\varepsilon ^{\ast }\cdot \varepsilon ^{\prime }\right) \right. \notag \\
&&\left. -\left( q\cdot \varepsilon ^{\prime }\right) \left( p^{\prime
}\cdot \varepsilon ^{\ast }\right) \right] , \label{eq:ME1}\end{aligned}$$and$$\langle D_{s}^{\ast }\left( p\right) \pi (q)|Z_{PS}(p^{\prime })\rangle
=g_{Z_{PS}D_{s}^{\ast }\pi }p^{\prime }\cdot \varepsilon . \label{eq:ME2}$$After some manipulations the ground state terms in $\Pi _{\mu \nu }^{\mathrm{%
Phys}}(p,q)$ can be easily rewritten as: $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{Phys}}(p,q)=g_{Z_{AV}D_{s}^{\ast }\pi }\frac{%
m_{D_{s}^{\ast }}f_{D_{s}^{\ast }}m_{Z_{AV}}f_{Z_{AV}}}{\left(
p^{2}-m_{D_{s}^{\ast }}^{2}\right) \left( p^{\prime 2}-m_{Z_{AV}}^{2}\right)
} \notag \\
&&\times \left( \frac{m_{Z_{AV}}^{2}+m_{D_{s}^{\ast }}^{2}}{2}g_{\mu \nu
}-p_{\mu }p_{\nu }^{\prime }\right) \notag \\
&&+\frac{g_{Z_{PS}D_{s}^{\ast }\pi }f_{D_{s}^{\ast }}m_{Z_{PS}}f_{Z_{PS}}}{%
\left( p^{2}-m_{D_{s}^{\ast }}^{2}\right) \left( p^{\prime
2}-m_{Z_{PS}}^{2}\right) m_{D_{s}^{\ast }}} \notag \\
&&\times \frac{m_{Z_{PS}}^{2}-m_{D_{s}^{\ast }}^{2}}{2}p_{\mu }p_{\nu
}^{\prime }+\ldots . \label{eq:CF5A}\end{aligned}$$One sees that $\Pi _{\mu \nu }^{\mathrm{Phys}}(p,q)$ contains two structures $\sim g_{\mu \nu }$ and $\sim p_{\mu }p_{\nu }^{\prime }.$ The same structures appear in the second part of the sum rule which is the correlation function Eq. (\[eq:CF4\]) calculated in terms of quark propagators. For $\Pi _{\mu \nu }^{\mathrm{QCD}}(p,q)$ we get $$\begin{aligned}
&&\Pi _{\mu \nu }^{\mathrm{QCD}}(p,q)=i\int d^{4}xe^{ipx}\left\{ \left[
\gamma _{\nu }\widetilde{S}_{s}^{ia}(x){}\gamma _{\mu }\right. \right.
\notag \\
&&\left. \times \widetilde{S}_{c}^{bi}(-x){}\gamma _{5}\right] _{\alpha
\beta }\langle \pi (q)|\overline{d}_{\alpha }^{b}(0)u_{\beta
}^{a}(0)|0\rangle \notag \\
&&\left. -\left[ \gamma _{\nu }\widetilde{S}_{s}^{ia}(x){}\gamma _{\mu }%
\widetilde{S}_{c}^{ai}(-x){}\gamma _{5}\right] \langle \pi (q)|\overline{d}%
_{\alpha }^{b}(0)u_{\beta }^{b}(0)|0\rangle \right\} . \label{eq:CF6}\end{aligned}$$We use invariant amplitudes corresponding to structures $\sim g_{\mu \nu }$ from $\Pi _{\mu \nu }^{\mathrm{Phys}}(p,q)$ and $\Pi _{\mu \nu }^{\mathrm{QCD%
}}(p,q)$ to derive sum rule for the coupling $g_{Z_{AV}D_{s}^{\ast }\pi }.$ To this end, we equate these invariant amplitudes and carry out calculations in accordance with scheme described in rather detailed form in the previous section. Obtained by this way sum rule is employed to evaluate the strong coupling $g_{Z_{AV}D_{s}^{\ast }\pi }$. It is utilized as an input parameter at the second stage of analysis, when we employ invariant amplitudes corresponding to structures $\sim p_{\mu }p_{\nu }^{\prime }$ to derive sum rule for $g_{Z_{PS}D_{s}^{\ast }\pi }$.
The decays $Z_{AV}\rightarrow D^{\ast }K$ and $Z_{PS}\rightarrow D^{\ast }K$ can be investigated in the same way, but one has to start from the correlator $$\Pi _{\mu \nu }(p,q)=i\int d^{4}xe^{ipx}\langle K(q)|\mathcal{T}\{J_{\mu
}^{D^{\ast }}(x)J_{\nu }^{\dag }(0)\}|0\rangle , \label{eq:CF7}$$with $J_{\mu }^{D^{\ast }}(x)$$$J_{\mu }^{D^{\ast }}(x)=\overline{c}_{l}(x)\gamma _{\mu }d_{l}(x)
\label{eq:Curr4}$$The remaining analysis does not differ from calculations of the decays $%
Z_{AV}\rightarrow D_{s}^{\ast }\pi $ and $Z_{PS}\rightarrow D_{s}^{\ast }\pi$, and therefore we do not provide further details.
There are also two processes $Z_{PS}\rightarrow D_{s0}^{\ast }(2317)\pi $ and $Z_{AV}\rightarrow D_{s1}(2460)\pi $ which are not connected with each other, and can be studied separately. Let us consider, for example, decay $%
Z_{PS}\rightarrow D_{s0}^{\ast }(2317)\pi $ that can be explored by means of the correlator $$\Pi _{\nu }(p,q)=i\int d^{4}xe^{ipx}\langle \pi (q)|\mathcal{T}%
\{J^{D_{s0}^{\ast }}(x)J_{\nu }^{\dag }(0)\}|0\rangle , \label{eq:CF8}$$where the interpolating current $J_{\mu }^{D_{s0}^{\ast }}(x)$ is chosen in the form$$J^{D_{s0}^{\ast }}(x)=\overline{c}^{i}(x)s^{i}(x). \label{eq:Curr5}$$The correlation function $\Pi _{\nu }(p,q)$ has the following phenomenological representation $$\begin{aligned}
\Pi _{\nu }^{\mathrm{Phys}}(p,q) &=&\frac{\langle 0|J^{D_{s0}^{\ast
}}|D_{s0}^{\ast }\left( p\right) \rangle }{p^{2}-m_{D_{s0}^{\ast }}^{2}}%
\langle D_{s0}^{\ast }\left( p\right) \pi (q)|Z_{PS}(p^{\prime })\rangle
\notag \\
&&\times \frac{\langle Z_{PS}(p^{\prime })|J_{\nu }^{\dagger }|0\rangle }{%
p^{\prime 2}-m_{Z_{PS}}^{2}}+\ldots .\end{aligned}$$Using of the matrix element$$\langle 0|J^{D_{s0}^{\ast }}|D_{s0}^{\ast }\left( p\right) \rangle
=f_{D_{s0}^{\ast }}m_{D_{s0}^{\ast }},$$and also the vertex$$\langle D_{s0}^{\ast }\left( p\right) \pi (q)|Z_{PS}(p^{\prime })\rangle
=g_{Z_{PS}D_{s0}^{\ast }\pi }p\cdot p^{\prime },$$it can be rewritten as $$\Pi _{\nu }^{\mathrm{Phys}}(p,q)=g_{Z_{PS}D_{s0}^{\ast }\pi }\frac{%
f_{Z_{PS}}m_{Z_{PS}}f_{D_{s0}^{\ast }}m_{D_{s0}^{\ast }}}{(p^{2}-m^{2})^{2}}%
m^{2}p_{\nu }^{\prime }+\ldots , \label{eq:CFDS0}$$where $m^{2}=(m_{Z_{PS}}^{2}+m_{D_{s0}^{\ast }}^{2})/2$. In order to match the obtained expression with the same structure from $\Pi _{\nu }^{\mathrm{%
QCD}}(p,q)$ we keep in Eq. (\[eq:CFDS0\]) dependence on $p_{\nu }^{\prime
}$, whereas in the invariant amplitude, i. e. in the function $\sim p_{\nu
}^{\prime }$ implement the soft limit.
The same correlation function $\Pi _{\nu }(p,q)$ in terms of quark propagators and pion’s matrix elements is given by formula$$\begin{aligned}
&&\Pi _{\nu }^{\mathrm{QCD}}(p,q)=i\int d^{4}xe^{ipx}\left\{ \left[ \gamma
_{\nu }\widetilde{S}_{s}^{ia}(x)\widetilde{S}_{c}^{bi}(-x)\gamma _{5}\right]
_{\alpha \beta }\right. \notag \\
&&\times \langle \pi (q)|\overline{d}_{\alpha }^{b}(0)u_{\beta
}^{a}(0)|0\rangle -\left[ \gamma _{\nu }\widetilde{S}_{s}^{ia}(x){}%
\widetilde{S}_{c}^{ai}(-x){}\gamma _{5}\right] \notag \\
&&\left. \times \langle \pi (q)|\overline{d}_{\alpha }^{b}(0)u_{\beta
}^{b}(0)|0\rangle \right\} .\end{aligned}$$After calculations one finds that in $\Pi _{\nu }^{\mathrm{QCD}}(p,q)$ survives only the structure $\sim p_{\nu }^{\prime }$. By equating invariant amplitudes from both sides and performing all manipulations it is possible to derive the sum rule for the coupling $g_{Z_{PS}D_{s0}^{\ast }\pi }$. The similar analysis has been carried out for the decay $Z_{AV}\rightarrow
D_{s1}(2460)\pi $, as well.
In numerical calculations of the $Z_{PS}$ and $Z_{AV}$ states’ strong couplings the Borel parameter and continuum threshold are chosen within the same ranges as in computations of their masses (see, Table [tab:Results1]{}). As input parameters we employ also mass and decay constant of the mesons $D_{s}^{\ast },\ D^{\ast }$ and $\ D_{s0}^{\ast }(2317)$ from Table \[tab:Param\]. It is worth noting that the decay constants $%
f_{D_{s}^{\ast }}$, $f_{D^{\ast }}$ and $f_{D_{s0}^{\ast }}$ have been taken from Refs. [@Agaev:2015faa; @Lucha:2014xla; @Narison:2003td], respectively.
Results for strong couplings and width of decay modes of $Z_{PS}$ and $%
Z_{AV} $ tetraquarks are presented in Table \[tab:Results2\]. Using these predictions one can evaluate full widths of the pseudoscalar and axial-vector tetraquarks $Z_{PS}$ and $Z_{AV}$: $$\Gamma_{Z_{PS}}=(38.1\pm 7.1)~\mathrm{MeV},$$ and $$\Gamma_{Z_{AV}}=(47.3\pm 11.1)~\mathrm{MeV}.$$ As is seen, the tetraquarks $Z_{PS}$ and $Z_{AV}$ are narrower than the scalar state $Z_S$. Nevertheless, we cannot classify them as narrow resonances.
Decay Strong couplings Decay Width
-------------------------------- -------------------------------------- ----------------------------------
$Z_{AV} \to D_{s}^{\ast }\pi $ $(0.26 \pm 0.07)~\mathrm{GeV}^{-1}$ $(7.94 \pm 2.21)~\mathrm{MeV}$
$Z_{AV} \to D^{\ast } K$ $(0.63 \pm 0.17) ~\mathrm{GeV}^{-1}$ $(37.38 \pm 10.84)~\mathrm{MeV}$
$Z_{AV}\to D_{s1}\pi$ $(1.55 \pm 0.43) ~\mathrm{GeV}^{-1}$ $(2.02 \pm 0.59)~\mathrm{MeV}$
$Z_{PS}\to D_{s}^{\ast }\pi $ $3.18 \pm 0.94$ $(4.37 \pm 1.27)~\mathrm{MeV}$
$Z_{PS}\to D^{\ast }K$ $8.24 \pm 2.39 $ $(19.09 \pm 5.73)~\mathrm{MeV}$
$Z_{PS} \to D_{s0}^{\ast}\pi$ $(0.76 \pm 0.18) ~\mathrm{GeV}^{-1}$ $(14.64 \pm 3.94)~\mathrm{MeV}$
: The strong couplings and decay widths of the $Z_{AV}$ and $Z_{PS}$ tetraquarks. []{data-label="tab:Results2"}
Conclusions {#sec:Concl}
===========
In the present work we have investigated the charm-strange tetraquarks $Z_{%
\overline{c}s}=[sd][\overline{u}\overline{c}]$ by calculating their spectroscopic parameters and decay channels. It is easy to see that these states bear two units of electric charge $-|e|$ and belong to a class of doubly charged tetraquarks. Their counterparts with the structure $Z_{c%
\overline{s}}=[uc][\overline{s}\overline{d}]$ have evidently a charge $+2|e|$. We have considered scalar, pseudoscalar and axial-vector doubly charged states. Their masses have been obtained using QCD two-point sum rule method. Our results have allowed us to fix possible decay channels of these states and found their widths. Investigations confirm that the doubly charged diquark-antidiquarks are neither broad states nor very narrow resonances.
Observation of doubly charged tetraquarks may open new stage in exploration of multiquark systems. In fact, resonances that are interpreted as hidden charm (bottom) tetraquarks may be also considered as excited states of charmonia (bottomonia) or their superpositions. The charged resonances can not be explained by this way, and are serious candidates to genuine tetraquarks. They may have diquark-antidiquark structure or be bound states of conventional mesons. In the last case, charged and neutral conventional mesons create shallow molecular states with large decay width. Therefore, it is reasonable to assume that doubly charged tetraquarks presumably exist only as diquark-antidiquarks, because binding of two mesons with the same electric charge to form a molecular state due to repulsive forces between them seems problematic.
The doubly charged tetraquarks deserve further detailed investigations. These studies should embrace also $Z_{b\overline{c}}$-type states that constitute a subclass of open charm-bottom states. Experimental exploration and discovery of $Z_{c\overline{s}}$ and/or $Z_{b\overline{c}}$ tetraquarks may have far-reaching consequences for hadron spectroscopy.
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
The work of S. S. A. was supported by Grant No. EIF-Mob-8-2017-4(30)-17/01/1 of the Science Development Foundation under the President of the Azerbaijan Republic. K. A. thanks TÜBITAK for the partial financial support provided under Grant No. 115F183.
[99]{} S. K. Choi *et al.* \[Belle Collaboration\], Phys. Rev. Lett. **91**, 262001 (2003). V. M. Abazov *et al.* \[D0 Collaboration\], Phys. Rev. Lett. **93**, 162002 (2004). D. Acosta *et al.* \[CDF Collaboration\], Phys. Rev. Lett. **93**, 072001 (2004). B. Aubert *et al.* \[BaBar Collaboration\], Phys. Rev. D **71**, 071103 (2005). H. X. Chen, W. Chen, X. Liu and S. L. Zhu, Phys. Rept. **639**, 1 (2016). H. X. Chen, W. Chen, X. Liu, Y. R. Liu and S. L. Zhu, Rept. Prog. Phys. **80**, 076201 (2017). A. Esposito, A. Pilloni and A. D. Polosa, Phys. Rept. **668**, 1 (2017).
A. Esposito, A. L. Guerrieri, F. Piccinini, A. Pilloni and A. D. Polosa, Int. J. Mod. Phys. A **30**, 1530002 (2015). C. A. Meyer and E. S. Swanson, Prog. Part. Nucl. Phys. **82**, 21 (2015) S. K. Choi *et al.* \[Belle Collaboration\], Phys. Rev. Lett. **100**, 142001 (2008). R. Mizuk *et al.* \[Belle Collaboration\], Phys. Rev. D **80**, 031104 (2009). K. Chilikin *et al.* \[Belle Collaboration\], Phys. Rev. D **88**, 074026 (2013). V. M. Abazov *et al.* \[D0 Collaboration\], Phys. Rev. Lett. **117**, 022003 (2016). The D0 Collaboration, D0 Note 6488-CONF, (2016).
R. Aaij *et al.* \[LHCb Collaboration\], Phys. Rev. Lett. **117**, 152003 (2016). The CMS Collaboration, CMS PAS BPH-16-002, (2016).
S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **93**, 074024 (2016). W. Chen, H. X. Chen, X. Liu, T. G. Steele and S. L. Zhu, Phys. Rev. Lett. **117**, 022002 (2016). Z. G. Wang, Commun. Theor. Phys. **66**, 335 (2016). W. Wang and R. Zhu, Chin. Phys. C **40**, 093101 (2016). C. M. Zanetti, M. Nielsen and K. P. Khemchandani, Phys. Rev. D **93**, 096011 (2016). S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **93**, 114007 (2016). J. M. Dias, K. P. Khemchandani, A. Martinez Torres, M. Nielsen and C. M. Zanetti, Phys. Lett. B **758**, 235 (2016). Z. G. Wang, Eur. Phys. J. C **76**, 279 (2016). C. J. Xiao and D. Y. Chen, Eur. Phys. J. A **53**, 127 (2017). S. S. Agaev, K. Azizi and H. Sundu, Eur. Phys. J. Plus **131**, 351 (2016). T. J. Burns and E. S. Swanson, Phys. Lett. B **760**, 627 (2016). F. K. Guo, U. G. Meißner and B. S. Zou, Commun. Theor. Phys. **65**, 593 (2016). Q. F. Lu and Y. B. Dong, Phys. Rev. D **94**, 094041 (2016). M. Albaladejo, J. Nieves, E. Oset, Z. F. Sun and X. Liu, Phys. Lett. B **757**, 515 (2016). C. B. Lang, D. Mohler and S. Prelovsek, Phys. Rev. D **94**, 074509 (2016). A. Esposito, A. Pilloni and A. D. Polosa, Phys. Lett. B **758**, 292 (2016).
X. W. Kang and J. A. Oller, Phys. Rev. D **94**, 054010 (2016).
A. Esposito, M. Papinutto, A. Pilloni, A. D. Polosa and N. Tantalo, Phys. Rev. D [**88**]{}, 054029 (2013). A. L. Guerrieri, M. Papinutto, A. Pilloni, A. D. Polosa and N. Tantalo, PoS LATTICE [**2014**]{}, 106 (2015).
W. Chen, H. X. Chen, X. Liu, T. G. Steele and S. L. Zhu, Phys. Rev. D **95**, no. 11, 114005 (2017). D. Ebert, R. N. Faustov, V. O. Galkin, Phys. Lett. B **696**, 241 (2011).
S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D**93**, 094006 (2016). Y. R. Liu, X. Liu and S. L. Zhu, Phys. Rev. D **93**, 074023 (2016). I. I. Balitsky, V. M. Braun and A. V. Kolesnichenko, Nucl. Phys. B **312**, 509 (1989). V. M. Belyaev, V. M. Braun, A. Khodjamirian and R. Ruckl, Phys. Rev. D **51**, 6177 (1995). B. L. Ioffe and A. V. Smilga, Nucl. Phys. B **232**, 109 (1984). S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **93**, 074002 (2016).
S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **95**, 034008 (2017). S. S. Agaev, K. Azizi and H. Sundu, Eur. Phys. J. C **77**, 321 (2017). S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **95**, 114003 (2017). S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **96**, 034026 (2017). C. Patrignani *et al.* \[Particle Data Group\], Chin. Phys. C **40**, 100001 (2016). J. L. Rosner, S. Stone and R. S. Van de Water, arXiv:1509.02220 \[hep-ph\]. J. Y. Sungu, H. Sundu, K. Azizi, N. Yinelek and S. Sahin, PoS FACESQCD , 045 (2010). S. S. Agaev, K. Azizi and H. Sundu, Phys. Rev. D **92**, 116010 (2015). W. Lucha, D. Melikhov and S. Simula, Phys. Lett. B **735**, 12 (2014). S. Narison, Phys. Lett. B **605**, 319 (2005).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a convex disk $K$ and a positive integer $k$, let $\delta_T^k(K)$ and $\delta_L^k(K)$ denote the $k$-fold translative packing density and the $k$-fold lattice packing density of $K$, respectively. Let $T$ be a triangle. In a very recent paper [@sriamorn], I proved that $\delta_L^k(T)=\frac{2k^2}{2k+1}$. In this paper, I will show that $\delta_T^k(T)=\delta_L^k(T)$.'
author:
- 'Kirati Sriamorn[^1]'
title: On the Multiple Packing Densities of Triangles
---
**Keywords** Multiple packing $\cdot$ Packing density $\cdot$ Triangle
**Mathematics Subject Classification** 05B40 $\cdot$ 11H31 $\cdot$ 52C15
Introduction
============
Let $D$ be a connected subset of $\mathbb{R}^2$. A family of bounded sets $\mathcal{F}=\{S_1,S_2,\ldots\}$ is said to be a *k-fold packing* of $D$ if $\bigcup S_i\subset D$ and each point of $D$ belongs to the interiors of at most $k$ sets of the family. In particular, when all $S_i$ are translates of a fixed measurable bounded set $S$, the corresponding family is called a *k-fold translative packing* of $D$ with $S$. When the translation vectors form a lattice, the corresponding family is called a *k-fold lattice packing* of $D$ with $S$. Let $I=[0,1)$, and let $M(S,k,l)$ be the maximum number of bounded sets in a $k$-fold translative packing of $lI^2$ with $S$. Then, we define $$\delta_T^k(S)=\limsup_{l\rightarrow\infty} \frac{M(S,k,l)|S|}{|lI^2|}.$$ Similarly, we can define $\delta_L^k(S)$ for the $k$-fold lattice packings.
A family of bounded sets $\mathcal{F}=\{S_1,S_2,\ldots\}$ is said to be a *k-fold covering* of $D$ if each point of $D$ belongs to at least $k$ sets of the family. In particular, when all $S_i$ are translates of a fixed measurable bounded set $S$ the corresponding family is called a *k-fold translative covering* of $D$ with $S$. When the translation vectors form a lattice, the corresponding family is called a *k-fold lattice covering* of $D$ with $S$. Let $m(S,k,l)$ be the minimal number of translates in a $k$-fold translative covering of $lI^2$ with $S$. Then, we define $$\vartheta_T^k(S)=\liminf_{l\rightarrow\infty} \frac{m(S,k,l)|S|}{|lI^2|}.$$ Similarly, we can define $\vartheta_L^k(S)$ for the $k$-fold lattice coverings.
We usually denote by $\delta_{T}$ and $\delta_{L}$ the $1$-fold packing densities $\delta^1_{T}$ and $\delta^1_{L}$, respectively. It is well known that $\delta_{T}(K)=\delta_{L}(K)$ holds for every convex disk $K$ [@rogers]. In particular, we have $\delta_{T}(T)=\delta_{L}(T)$ for every triangle $T$. For the case that $K=C$ is a *centrally symmetric* convex disk, Fejes Tóth [@fejes] proved that $\delta(C)=\delta_{T}(C)=\delta_{L}(C)$ where $\delta(C)$ is the (congruent) packing density of $C$. In fact, we have the following statements:
- Let $\{C_1,C_2,\ldots,C_N\}$ be a packing of a convex hexagon $H$ (Fig. \[packingofH\]), where $C_i$ is a convex disk. Then we can find *convex polygons* $R_1,R_2,\ldots,R_N$ (Fig. \[polygonal\]) such that
(a) $R_i\supseteq C_i$ for every $i=1,2,\ldots,N$;
(b) $\{R_1,R_2,\ldots,R_N\}$ is also a packing of $H$;
(c) the number $s_i$ of sides of $R_i$ satisfies $$\label{mean_sides_inequality}
\sum_{i=1}^{N}s_i\leq 6N.$$
![A packing $\{C_i\}$ of a convex hexagon $H$[]{data-label="packingofH"}](PackingH.eps)
![The polygons $R_1,R_2,\ldots,R_N$[]{data-label="polygonal"}](PolygonalDecomposition.eps)
- (*Dowker’s Theorem*) Given a convex disk $C$, $n\geq 3$, let $A(n)$ denote the minimum area of an $n$-gon circumscribed about $C$. Then $$\label{dowker_convexity}
A(n)\leq \frac{A(n-1)+A(n+1)}{2}.$$
- Let $C$ be a *centrally symmetric* convex disk and let $n\geq 4$ be an even integer. Then one can find a convex $n$-gon $P_n$ circumscribed about $C$ with minimum area such that it is centrally symmetric and has the same center as $C$. As a consequence, we have that $$\label{one_fold_lattice_density}
\delta_L(C)\geq \frac{|C|}{|P_6|}=\frac{|C|}{A(6)},$$ where $A(6)$ is the minimum area of a hexagon circumscribed about $C$.
By using these results, we can show now that $\delta(C)=\delta_{L}(C)$ where $C$ is a centrally symmetric convex disk. Let $H$ be a convex hexagon, and let $\{C_1,C_2,\ldots,C_N\}$ be a packing of $H$ with congruent copies of $C$. By (\[mean\_sides\_inequality\]) and (\[dowker\_convexity\]), we have that $$\begin{aligned}
\frac{|C_1|+|C_2|+\cdots +|C_N|}{|H|}&=\frac{N|C|}{|H|}\\
&\leq \frac{N|C|}{|R_1|+|R_2|+\cdots+|R_N|}\\
&\leq \frac{N|C|}{A(s_1)+A(s_2)+\cdots+A(s_N)}\\
&\leq \frac{|C|}{A(6)},\end{aligned}$$ and hence $$\label{one_fold_congruent_density}
\delta(C)\leq \frac{|C|}{A(6)}.$$ Since $\delta(C)\geq \delta_L(C)$, by (\[one\_fold\_lattice\_density\]) and (\[one\_fold\_congruent\_density\]), we obtain $$\delta(C)=\delta_L(C).$$
In a very recent paper, Sriamorn [@sriamorn] studied the $k$-fold lattice coverings and packings with triangles $T$. He proved that $$\delta_L^k(T)=\frac{2k^2}{2k+1},$$ and $$\vartheta_L^k(T)=\frac{2k+1}{2}.$$ Furthermore, Sriamorn and Wetayawanich [@sriamorn1] showed that $\vartheta_T^k(T)=\vartheta_L^k(T)$ for every triangle $T$. In this paper, I will prove the following result:
\[main\_thm\] For every triangle $T$, we have $\delta_T^k(T)=\delta_L^k(T)=\frac{2k^2}{2k+1}.$
To prove this result, analogous to the proof of $\delta(C)=\delta_L(C)$ above, I will define an *$r$-stair polygon* (Definition \[def\_stair\_polygon\]) and use it in place of “convex $n$-gon" above. More precisely, I will show the following statements:
- Let $\{T_1,T_2,\ldots,T_N\}$ be a $k$-fold *translative* packing of $lI^2$ (for some positive $l$) with a triangle $T$. Then we can find *stair polygons* $S_1,S_2,\ldots,S_N$ such that
(a) $S_i\supseteq T_i$ for every $i=1,2,\ldots,N$;
(b) $\{S_1,S_2,\ldots,S_N\}$ is a $k$-fold packing of $lI^2$ (Lemma \[S\_i\_exact\_k\_fold\]);
(c) we have $$\sum_{i=1}^{N}r_i\leq (2k-1)N,$$ where $S_i$ is an $r_i$-stair polygon (Lemma \[approximate\_average\_stair\_point\]).
- For $r\geq 0$, let $A^*(r)$ denote the minimum area of an $r$-stair polygon containing $T$ (see (\[A\_star\]) below). Then $$A^*(r)\leq \frac{A^*(r-1)+A^*(r+1)}{2}.$$
- For the case of $k$-fold lattice packings, we have $$\delta_L^k(T)=\frac{k|T|}{A^*(2k-1)}=\frac{2k^2}{2k+1}.$$
In order to construct the desired stair polygons, I will introduce a strict partial order of triangles and a term “press" (Section \[def\_of\_press\]). Furthermore, due to technical reasons, I will introduce a concept of a *normal* $k$-fold translative packing (Section \[def\_of\_normal\]). An advantage of using this concept is that, by Theorem \[normal\_packing\] below, we may assume, without loss of generality, that the $k$-fold translative packing of our concern is normal, i.e., none of translates coincide. This could simplify our proof.
It is worth noting that when we talk about ($1$-fold) packings, we often refer to the concept of shadow cells [@pach]. I will show here a way to extend this concept to $k$-fold packings. For a nonzero vector $v$ and a point $q\in\mathbb{R}^2$, denote by $L(q,v)$ the ray parallel to $v$ and starting at $q$. Suppose that $K$ is a convex disk and $K\cap L(q,v)\neq\emptyset$, then we define $$\partial K(q,v)=\operatorname*{argmin}_{p\in K\cap L(q,v)}{d(p,q)},$$ where $d(p,q)$ is the Euclidean distance between points $p$ and $q$ (Fig. \[raycutk\]). Obviously, if $q\in K$, then $\partial K(q,v)=q$ for every nonzero vector $v$.
![$\partial K(q,v)$[]{data-label="raycutk"}](RayCutK.eps)
(*$k$-fold shadow cell*). Let a family of convex disks $\{K_1,K_2,\ldots\}$ be a $k$-fold packing of the plane and let $v$ be a nonzero vector. For every $i$, let $S_i$ be defined as the set of those points $q\in\mathbb{R}^2$ which either (i) $q\in K_i$ or (ii) $K_i\cap L(q,v)\neq\emptyset$ and there are at most $k-1$ numbers of $j$ such that $j\neq i$ and $d(\partial K_j(q,v),q)\leq d(\partial K_i(q,v),q)$. $S_i$ is called a *$k$-fold shadow cell* of $K_i$ (Fig. \[shadowcell\]).
![An example for $2$-fold shadow cell[]{data-label="shadowcell"}](ShadowCellNew.eps)
When $k=1$, we can give another definition of shadow cells by changing the condition $d(\partial K_j(q,v),q)\leq d(\partial K_i(q,v),q)$ in the above definition to $d(\partial K_j(q,v),q)< d(\partial K_i(q,v),q)$. Noting that the definition obtained this way will be equivalent to the definition of shadow cells described in [@pach]. However, we could not do the same thing for the case $k>1$, otherwise the family of shadow cells $\{S_1,S_2,\ldots\}$ might not be a $k$-fold packing of the plane. As shown in Fig. \[counterexample\], let $k=2$, if we use the condition $d(\partial K_j(q,v),q)< d(\partial K_i(q,v),q)$, then $q$ will lie in $S_1$, $S_2$ and $S_3$ for all $q\in D$, and hence $\{S_1,S_2,\ldots\}$ is not a $2$-fold packing.
![A counter example[]{data-label="counterexample"}](CounterExampleShadow.eps)
Naturally, we could use the concept of $k$-fold shadow cells instead of stair polygons in our proof. However, I found that in general it is difficult to say clearly what shape the shadow cells are. Even for the case of $k$-fold translative packings with a triangle, although we might show that the shadow cells are polygons, but it is still quite hard to say how many sides they have, and hence it is not so easy to estimate their areas or to obtain the desired properties. In contrast, the shape of stair polygons is much more simple. Therefore, when we study a $k$-fold translative packing of a triangle, it seems that using the concept of stair polygons is better than using the concept of shadow cells.
Normal $k$-Fold Translative Packing {#def_of_normal}
===================================
Let $D$ be a connected subset of $\mathbb{R}^2$ and $\mathcal{K}=\{K_1,K_2,\ldots\}$ a family of convex disks. Suppose that $\mathcal{K}$ is a $k$-fold packing of $D$. We say that $\mathcal{K}$ is *normal* provided $K_i\neq K_j$ for all $i\neq j$. When $\mathcal{K}$ is normal and $K_i$ are translates of a fixed convex disk $K$, the corresponding family is called a *normal k-fold translative packing* of $D$ with $K$. Let $\widetilde{M}(K,k,l)$ be the maximum number of convex disks in a normal $k$-fold translative packing of $lI^2$ with $K$. Then, we define $$\widetilde{\delta}_T^{k}(K)=\limsup_{l\rightarrow\infty}\frac{\widetilde{M}(K,k,l)|K|}{|lI^2|}.$$
\[normal\_packing\] For every convex disk $K$, we have $$\widetilde{\delta}_T^{k}(K)=\delta_T^k(K).$$
Trivially, we have that $\widetilde{\delta}_T^{k}(K)\leq\delta_T^k(K)$. Let $\{K_1,\ldots K_M\}$ be a $k$-fold translative packing of $lI^2$ with $K$. For any $K_i$, one can see that for every $0<\varepsilon<1$, there exist infinitely many points $(x,y)$ in the plane such that $K_i\supset (1-\varepsilon)K_i+(x,y)$. Hence, for every $0<\varepsilon<1$, there exist $M$ points $(x_1,y_1),\ldots,(x_M,y_M)$ in the plane such that $\{(1-\varepsilon)K_1+(x_1,y_1),\ldots,(1-\varepsilon)K_M+(x_M,y_M)\}$ is a normal $k$-fold translative packing of $lI^2$ with $(1-\varepsilon)K$. Therefore, $M\leq \widetilde{M}((1-\varepsilon)K,k,l)$. This implies that $M(K,k,l)\leq \widetilde{M}((1-\varepsilon)K,k,l)$, and hence $$\begin{aligned}
\delta_T^{k}(K)&=\limsup_{l\rightarrow\infty}\frac{M(K,k,l)|K|}{|lI^2|}\\
&\leq\limsup_{l\rightarrow\infty}\frac{\widetilde{M}((1-\varepsilon)K,k,l)|K|}{|lI^2|}\\
&=\frac{1}{(1-\varepsilon)^2}\widetilde{\delta}_T^{k}((1-\varepsilon)K)\\
&=\frac{1}{(1-\varepsilon)^2}\widetilde{\delta}_T^{k}(K).\end{aligned}$$ By letting $\varepsilon$ tend to zero, one obtains the result.
Some Notations {#def_of_press}
==============
In this paper, we denote by $T$ the triangle with vertices $(0,0)$, $(1,0)$ and $(0,1)$. If $T'=T+(x,y)$ where $(x,y)\in\mathbb{R}^2$, then we denote by $I^2(T')$ the square $I^2+(x,y)$, and denote by $H(T')$ the hypothenuse of $T'$.
For $(x_1,y_1),(x_2,y_2)\in\mathbb{R}^2$, we define the relation $\prec$ by $(x_1,y_1)\prec (x_2,y_2)$ if and only if either $$x_1+y_1<x_2+y_2$$ or $$x_1+y_1=x_2+y_2 \text{~and~} x_1<x_2.$$ One can easily show that $\prec$ is a strict partial ordering over $\mathbb{R}^2$.
Let $K$ be a nonempty bounded set. We define $$V(K)=\{u\in\mathbb{R}^2: u\prec u' \text{~or~} u=u', \text{~for all~} u'\in K\}.$$ Denote by $v(K)$ the point $u$ in $V(K)$ such that for all $u'\in V(K)$, $u'\prec u$ or $u'=u$. For example, $v(T+(x,y))=(x,y)$ and $v(I^2+(x,y))=(x,y)$.
Suppose that $T_1$ and $T_2$ are two distinct translates of $T$ and $I^2(T_1)\cap I^2(T_2)\neq \emptyset$. We say that $T_1$ *presses* $T_2$ provided $v(T_2)\prec v(T_1)$ (Fig. \[typeofpress\]). As immediate consequence of the definition, one can see that for every two translates $T_1,T_2$ of $T$, if $I^2(T_1)\cap I^2(T_2)\neq\emptyset$ and $T_1\neq T_2$, then either $T_1$ presses $T_2$ or $T_2$ presses $T_1$.
![$T_1$ presses $T_2$[]{data-label="typeofpress"}](TypeOfPress.eps)
Suppose that $T_1,~T_2$ and $T_3$ are three distinct translates of $T$ and $I^2(T_1)\cap I^2(T_3)\neq\emptyset$. If $T_1$ presses $T_2$ and $T_2$ presses $T_3$, then $T_1$ presses $T_3$.
Since $T_1$ presses $T_2$ and $T_2$ presses $T_3$, we have that $v(T_2)\prec v(T_1)$ and $v(T_3)\prec v(T_2)$. Hence $v(T_3)\prec v(T_1)$. This implies immediately that $T_1$ presses $T_3$.
\[cut\_by\_the\_others\] Suppose that $T_1,\ldots,T_n$ are $n$ distinct translates of $T$ and $I^2(T_1)\cap\cdots\cap I^2(T_n)\neq\emptyset$. Then, there exists $i\in\{1,\ldots,n\}$ such that $T_j$ presses $T_i$ for all $j\neq i$.
It is easy to see that there exists $i\in\{1,\ldots,n\}$ such that $v(T_i)\prec v(T_j)$ for all $j\neq i$. We note that $I^2(T_i)\cap I^2(T_j)\neq\emptyset$. By the definition, we have that $T_j$ presses $T_i$ for all $j\neq i$.
\[hypothenuse\] Let $T'$ be a translate of $T$ and $u$ be a point in $I^2(T')$. We have that $u\in T'$ if and only if $u\prec u'$ for some $u'\in H(T')$.
Suppose that $T'=T+(x,y)$. If $u\in T'\cap I^2(T')$, then let $u'=(x+1,y)$. It is clear that $u\prec u'$ and $u'\in H(T')$. Conversely, if $u\prec u'$ for some $u'\in H(T')$, then it is obvious that $u\in T'$.
\[I\_cut\_T\_must\_cut\] Suppose that $T_1,\ldots,T_{n+1}$ are $n+1$ distinct translates of $T$. If $T_i$ presses $T_{n+1}$ for all $i=1,\ldots,n$, then $(I^2(T_1)\cap\cdots\cap I^2(T_n))\cap T_{n+1}\subset T_1\cap\cdots\cap T_n\cap T_{n+1}$.
Suppose that $u\in (I^2(T_1)\cap\cdots\cap I^2(T_n))\cap T_{n+1}$. Since $T_i$ presses $T_{n+1}$ and $u\in T_{n+1}$, by Lemma \[hypothenuse\], it is not hard to see that $u\prec u'_i$ for some $u'_i\in H(T_i)$. Again, by Lemma \[hypothenuse\], we have that $u\in T_i$ for all $i=1,\ldots,n$, and hence $u\in T_1\cap\cdots\cap T_n\cap T_{n+1}$.
\[def\_stair\_polygon\] For a non-negative integer $r$, we call a planar set $S$ a *half-open r-stair polygon* (Fig. \[stairpolygon\]) if there are $x_0<x_1<\cdots< x_{r+1}$ and $y_0>y_1>\cdots > y_r>y_{r+1}$ such that $$S=\bigcup_{i=0}^r[x_i,x_{i+1})\times[y_{r+1},y_i).$$
![A half-open $5$-stair polygon[]{data-label="stairpolygon"}](StairPolygon.eps)
Let $A^*(r)$ denote the minimum area of a half-open $r$-stair polygon containing $Int(T)$. Clearly, $A^*$ is a decreasing function. By elementary calculations, one can obtain $$\label{A_star}
A^*(r)=\frac{r+2}{2(r+1)},$$ where $r=0,1,2,\ldots$. Let $B^*$ be the function on $[0,+\infty)$ defined by $$B^*(x)=\frac{x+2}{2(x+1)},$$ It is obvious that $B^*$ is a decreasing convex function and $B^*(r)=A^*(r)$ , for all $r=0,1,2,\ldots$. For convenience, we also denote the function $B^*$ by $A^*$. In [@sriamorn], Sriamorn showed that $$\label{lattice_density}
\delta_L^k(T)=\frac{k|T|}{A^*(2k-1)}=\frac{2k^2}{2k+1}.$$
The Construction of Stair Polygons $S_i$ {#key_section}
========================================
In this section, we suppose that $\mathcal{T}=\{T_1,T_2,\ldots,T_N\}$ is a *normal* $k$-fold translative packing of $lI^2$ with $T$. We will use the terminologies given above to construct the desired stair polygons $S_1,S_2,\ldots,S_N$. In fact, due to technical reasons (but not essential), we will construct *half-open* stair polygons instead of (closed) stair polygons. This could make it easier to prove our desired results.
Denote by $\mathcal{C}_i$ the collection of triangles $T_j$ that press $T_i$. Let $$U_i=\bigcup_{\substack{T_{i_1},\ldots,T_{i_k}\in\mathcal{C}_i \text{~are distinct}\\ I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k})\neq\emptyset}} R(v(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k}))),$$ and $$S_i=I^2(T_i)\setminus U_i,$$ where $R(x_0,y_0)$ denotes the set $\{(x,y): x\geq x_0, y\geq y_0\}$ (for example, see Fig. \[onefold\] and Fig. \[twofold\] ).
![Two examples to illustrate the construction of stair polygons $S_i$ in a $1$-fold packing. Only triangles which press $T_i$ are shown.[]{data-label="onefold"}](OneFold.eps)
![Three examples to illustrate the construction of stair polygons $S_i$ in a $2$-fold packing. Only triangles which press $T_i$ are shown.[]{data-label="twofold"}](TwoFold.eps)
We have the following lemmas.
\[int\_T\_intersect\_U\] $Int(T_i)\cap U_i=\emptyset$.
Assume that $Int(T_i)\cap U_i\neq\emptyset$. By the definition of $U_i$, it can be deduced that there exist $T_{i_1},\ldots,T_{i_k}\in\mathcal{C}_i$ such that $Int(T_i)\cap(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k}))\neq \emptyset$. By Lemma \[I\_cut\_T\_must\_cut\], we know that $Int(T_i)\cap(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k}))\subset Int(T_i)\cap T_{i_1}\cap\cdots\cap T_{i_k}$, and hence $Int(T_i)\cap T_{i_1}\cap\cdots\cap T_{i_k}\neq \emptyset$. Since $v(T_i)\prec v(T_{i_j})$ for all $j=1,\ldots,k$, one can see that $H(T_i\cap T_{i_1}\cap\cdots\cap T_{i_k})\subset H(T_i)$. From $Int(T_i)\cap T_{i_1}\cap\cdots\cap T_{i_k}\neq \emptyset$, we know that $Int(T_i)\cap Int(T_{i_1})\cap\cdots\cap Int(T_{i_k})\neq \emptyset$. This is impossible, since $\mathcal{T}$ is a $k$-fold packing of $\mathbb{R}^2$.
\[S\_i\_stair\_polygon\] $S_i$ is a half-open stair polygon containing $Int(T_i)$.
We note that $\mathcal{C}_i$ is finite, hence it is obvious that $S_i$ is a half-open stair polygon. By Lemma \[int\_T\_intersect\_U\], we have $Int(T_i)\subset I^2(T_i)\setminus U_i=S_i$.
We may assume, without loss of generality, that $S_i$ is a half-open $r_i$-stair polygon and $$S_i=\bigcup_{j=0}^{r_i}[x_j^{(i)},x_{j+1}^{(i)})\times[y_{r_i+1}^{(i)},y_j^{(i)}),$$ where $x_0^{(i)}<x_1^{(i)}<\cdots<x_{r_i+1}^{(i)}$ and $y_0^{(i)}>y_1^{(i)}>\cdots>y_{r_i+1}^{(i)}$ (Fig. \[Si\]). Let $$Z(S_i)=\{(x_j^{(i)},y_j^{(i)}): j=1,\ldots,r_i\}.$$
![$S_i$[]{data-label="Si"}](Si.eps)
\[x\_coordinate\_on\_stair\_point\] For every $(x',y')\in Z(S_i)$, there exists $j\in\{1,\ldots,N\}\setminus\{i\}$ such that $(x',y')\in S_j$ and $x'=x_0^{(j)}$ where $x_0^{(j)}$ is the $x$-coordinate of $v(S_j)$ (Fig. \[SiSj\]).
By the definitions of $S_i$ and $Z(S_i)$, it is not hard to see that there exist $T_{i_1},\ldots,T_{i_k}\in\mathcal{C}_i$ such that $I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k})\neq\emptyset$ and $v(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k}))=(x',y')$. This implies that there is a $j\in\{i_1,\ldots,i_k\}$ such that the $x$-coordinate of $v(I^2(T_{j}))$ is $x'$. Clearly, $(x',y')\in T_j\cap I^2(T_j)\subset S_j$ and $v(S_j)=v(I^2(T_{j}))$.
![$S_i$ and $S_j$[]{data-label="SiSj"}](SiSj.eps)
\[not\_exceed\_k\_1\] Suppose that $i_1,\ldots,i_{k+1}$ are $k+1$ distinct positive integers in $\{1,2,\ldots,N\}$. Then $$\bigcap_{j=1}^{k+1}S_{i_j}=\emptyset.$$
If $I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_{k+1}})=\emptyset$, then it is obvious that $S_{i_1}\cap\cdots\cap S_{i_{k+1}}= \emptyset$. Assume that $I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_{k+1}})\neq \emptyset$. By Lemma \[cut\_by\_the\_others\], we may assume, without loss of generality, that $T_{i_j}$ presses $T_{i_{k+1}}$ for all $j=1,\ldots,k$. Therefore $I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_{k}})\subset R(v(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k})))\subset U_{i_{k+1}}$. Hence $$\bigcap_{j=1}^{k+1}S_{i_j}=(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_{k+1}}))\setminus(U_{i_1}\cup\cdots\cup U_{i_{k+1}})= \emptyset.$$
\[S\_i\_exact\_k\_fold\] $\{S_1,S_2,\ldots,S_N\}$ is a $k$-fold packing of $lI^2$.
Since $T_i\subset lI^2$, it is obvious that $S_i\subset lI^2$. Hence, the result follows immediately from Lemma \[not\_exceed\_k\_1\].
\[to\_prove\_no\_R\_cut\_by\_other\] Let $L_i=(\overline{S_i}\setminus S_i)\cap I^2(T_i)$. If $T_i$ presses $T_j$, then $L_i\cap S_j=\emptyset$.
Assume that there is some point $(x,y)\in L_i\cap S_j$. We have that $(x,y)\in U_i$, and hence there exist $T_{i_1},\ldots,T_{i_k}\in \mathcal{C}_i$ such that $(x,y)\in R(v(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k})))$. Let $v(I^2(T_{i_1})\cap\cdots\cap I^2(T_{i_k}))=(x',y')$. It is obvious that $x'\leq x$ and $y'\leq y$. Since $(x,y)\in S_j\subset I^2(T_j)$ and $(x',y')\in I^2(T_i)\setminus Int(T_i)$, one can deduce that $(x',y')\in I^2(T_j)$, and hence $I^2(T_j)\cap I^2(T_{i_s})\neq\emptyset$ for all $s=1,\ldots,k$. For $s=1,\ldots,k$, since $T_i$ presses $T_j$ and $T_{i_s}$ presses $T_i$, we have that $T_{i_s}$ presses $T_j$, i.e., $T_{i_s}\in\mathcal{C}_j$. Therefore $(x,y)\in U_j$, which is a contradiction.
![$L_i\cap S_j=\emptyset$[]{data-label="LS"}](LS.eps)
\[condition\_for\_cut\] For every $i,j\in\{1,2,\ldots,N\}$, we have $L_i\cap S_j=\emptyset$ or $L_j\cap S_i=\emptyset$.
If $I^2(T_i) \cap I^2(T_j)=\emptyset$ or $i=j$, then the result is trivial. When $I^2(T_i)\cap I^2(T_j)\neq \emptyset$ and $i\neq j$, we have that either $T_i$ presses $T_j$ or $T_j$ presses $T_i$. The result follows directly from Lemma \[to\_prove\_no\_R\_cut\_by\_other\].
\[more\_than\_r\_k\_1\] For $i=1,2,\ldots,N$, let $$n_i=card\{S_j : v(S_j)\in Int(S_i)\cup Z(S_i), j=1,\ldots,N\}.$$ Then, we have $$n_i\geq r_i-k+1.$$
Suppose that $Z(S_i)=\{(x_1^{(i)},y_1^{(i)}),\ldots,(x_{r_i}^{(i)},y_{r_i}^{(i)})\}$. By Lemma \[x\_coordinate\_on\_stair\_point\], we know that for every $j=1,\ldots,r_i$, there exists an $i_j\in\{1,\ldots,N\}\setminus\{i\}$ such that $(x_j^{(i)},y_j^{(i)})\in S_{i_j}$ and $x_j^{(i)}=x_{i_j}$ where $x_{i_j}$ is the $x$-coordinate of $v(S_{i_j})$ (see Figure \[Sij\]). Let $y_i$ and $y_{i_j}$ be the $y$-coordinates of $v(S_i)$ and $v(S_{i_j})$, respectively. Let $$\mathcal{F}=\{S_{i_j} :y_{i_j}\leq y_i, j=1,\ldots,r_i\}.$$ By Lemma \[condition\_for\_cut\], we know that $L_{i_j}\cap S_i=\emptyset$ for all $S_{i_j}\in\mathcal{F}$. We note that $S_i\notin \mathcal{F}$. Since $\{S_1,\ldots,S_N\}$ is a $k$-fold packing of $lI^2$, one can deduce that $card\{\mathcal{F}\}\leq k-1$. It is not hard to see that for every $S\in\{S_{i_1},S_{i_2},\ldots,S_{i_{r_i}}\}\setminus \mathcal{F}$, we have $v(S)\in Int(S_i)\cup Z(S_i)$. Hence $$n_i\geq card\{Z(S_i)\}-card\{\mathcal{F}\}\geq r_i-k+1.$$
![$S_{i_j}$[]{data-label="Sij"}](Sij.eps)
\[less\_than\_kN\] $$\sum_{i=1}^{N}n_i\leq kN.$$
For $i=1,\ldots,N$, let $\mathcal{F}_i=\{S_j : v(S_j)\in Int(S_i)\cup Z(S_i), j=1,\ldots,N\}$ and $\mathcal{F}^*_i=\{S_j : v(S_i)\in Int(S_j)\cup Z(S_j), j=1,\ldots,N\}$. Clearly, we have $n_i=card\{\mathcal{F}_i\}$. Let $n_i^*=card\{\mathcal{F}^*_i\}$. It is not hard to show that $\sum_{i=1}^{N}n_i=\sum_{i=1}^{N}n_i^*$. On the other hand, since $\{S_1,\ldots,S_N\}$ is a $k$-fold packing of $lI^2$, it is obvious that $n_i^*\leq k$. Hence $$\sum_{i=1}^{N}n_i=\sum_{i=1}^{N}n_i^*\leq kN.$$
The following lemma follows immediately from Lemmas \[more\_than\_r\_k\_1\] and \[less\_than\_kN\].
\[approximate\_average\_stair\_point\] $$\sum_{i=1}^{N}r_i\leq (2k-1)N.$$
Proof of Main Theorem
=====================
Let $\mathcal{T}=\{T_1,T_2,\ldots,T_N\}$ be a *normal* $k$-fold translative packing of $lI^2$ with $T$. Let $S_i$ be the half-open $r_i$-stair polygon defined by $\mathcal{T}$ as shown in Section \[key\_section\]. By Lemma \[S\_i\_exact\_k\_fold\], Lemma \[approximate\_average\_stair\_point\], the convexity of $A^*$ and (\[lattice\_density\]), one obtains $$\begin{aligned}
\frac{|T_1|+|T_2|+\cdots+|T_N|}{|lI^2|} &= \frac{N|T|}{|lI^2|}\\
&\leq \frac{kN|T|}{|S_1|+|S_2|+\cdots+|S_N|}\\
&\leq \frac{kN|T|}{A^*(r_1)+A^*(r_2)+\cdots+A^*(r_N)}\\
&\leq \frac{k|T|}{A^*(2k-1)}=\delta_L^k(T),\end{aligned}$$ and hence $$\delta_T^k(T)\leq \delta_L^k(T).$$ This completes the proof.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by 973 Programs 2013CB834201 and 2011CB302401.
[10]{} Sriamorn, K., Wetayawanich, A.: On the multiple covering densities of triangles. (2015). doi:10.1007/s00454-015-9711-0
Sriamorn, K.: Multiple lattice packings and coverings of the plane with triangles. arXiv:1412.5096 (2014)
Sriamorn, K., Xue, F.: On the covering densities of quarter-convex disks. Discrete Comput. Geom.**54**, 246–258(2015). doi:10.1007/s00454-015-9696-8
Rogers, C.A.: The closest packing of convex two-dimensional domains. Acta Math.**86**, 309–321 (1951)
Fejes Tóth, L.: Some packing and covering theorems. Acta Math. Acad. Sci. Hungar.**12**, 62–67 (1950)
Pach, J., Agarwal, P.K.: Combinatorial Geometry, John Wiley $\&$ Sons, pp.47–53 (2011)
[^1]: School of Mathematical Sciences, Peking University, Beijing, China
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The radio access network (RAN) is regarded as one of the potential proposals for massive Internet of Things (mIoT), where the random access channel (RACH) procedure should be exploited for IoT devices to access to the RAN. However, modelling of the dynamic process of RACH of mIoT devices is challenging. To address this challenge, we first revisit the frame and minislot structure of the RAN. Then, we correlate the RACH request of an IoT device with its queue status and analyze the queue evolution process. Based on the analysis result, we derive the closed-form expression of the RA success probability of the device. Besides, considering the agreement on converging different services onto a shared infrastructure, we further investigate the RAN slicing for mIoT and bursty ultra-reliable and low latency communications (URLLC) service multiplexing. Specifically, we formulate the RAN slicing problem as an optimization one aiming at optimally orchestrating RAN resources for mIoT slices and bursty URLLC slices to maximize the RA success probability and energy-efficiently satisfy bursty URLLC slices’ quality-of-service (QoS) requirements. A slice resource optimization (SRO) algorithm exploiting relaxation and approximation with provable tightness and error bound is then proposed to mitigate the optimization problem.'
author:
- 'Peng Yang, Xing Xi, Tony Q. S. Quek, , Jingxuan Chen, Xianbin Cao, , Dapeng Wu, [^1]'
bibliography:
- 'Network\_slicing.bib'
title: 'RAN Slicing for Massive IoT and Bursty URLLC Service Multiplexing: Analysis and Optimization'
---
Massive IoT, random access channel, bursty URLLC, RAN slicing
Introduction
============
the explosive growth of the Internet of Things (IoT), massive IoT (mIoT) devices, the number of which is predicted to reach 20.8 billion by 2020, will access to the wireless networks for implementing advanced applications, such as e-health, public safety, smart traffic, virtual navigation/management, remote maintenance and control, and environment monitoring. To address the IoT market, the third-generation partnership project (3GPP) has identified mIoT as one of the three main use cases of 5G and has already initiated several task groups to standardize several solutions including extended coverage GSM (EC-GSM), LTE for machine-type communication (LTE-M), and narrowband IoT (NB-IoT) [@Cellular-Ericsson; @LTE-Nokia].
For establishing massive connections among the wireless networks and mIoT devices, the investigation of reliable and efficient access mechanisms should be prioritized. In accomplishing the massive connections, when an active IoT device wants to transmit signal in the uplink, it randomly chooses a random access (RA) preamble from an RA preamble pool and transmits it through an RA channel (RACH). If more than one device tries to access to a base station (BS) simultaneously, then interference occurs at the RRH. During the past few years, a rich body of works on RA mechanisms has been developed [@xing2019novel; @jang2016non; @jiang2019reinforcement; @zhang2019dnn; @jiang2018analyzing; @grau2019preamble; @jang2019recursive; @jin2017recursive; @jang2019versatile] to mitigate interference and improve the RA success probability or reduce the access delay of an IoT device.
Most of the studies [@xing2019novel; @jang2016non; @jiang2019reinforcement; @zhang2019dnn; @jiang2018analyzing; @grau2019preamble; @jang2019recursive; @jin2017recursive; @jang2019versatile], however, assumed that the whole network resources were reserved for the IoT service and did not investigate the case of the coexistence of IoT service and many other services such as enhanced mobile broadband (eMBB) and ultra-reliable and low latency communications (URLLC). The research of the coexistence of IoT service and other services is essential as future networks are convinced to converge variety of services with different latency, reliability, and throughput requirements onto a shared physical infrastructure rather than deploying individual network solution for each service [@alliance20155g]. What is more, owing to the shared characteristic of network resources, some conclusions obtained in the case of providing sole IoT service may become inapplicable if multiple types of services are required to be supported by the networks.
Network slicing is considered as a promising technology in future networks for providing scalability and flexibility in allocating network resources to various services. Recently, many network slicing frameworks have been developed to provide performance guarantees to IoT or massive machine-type communications (mMTC) service, eMBB service, and URLLC service [@popovski20185g; @budhiraja2019tactile; @ksentini2017toward; @rost2017network; @wen2018robustness; @buyakar2018poster; @guo2019enabling].
Different from previous works, this paper investigates the mIoT and bursty URLLC service multiplexing via slicing the radio access network (RAN). This study is highly challenging because i) performance requirements of a massive number of IoT devices should be satisfied. Yet, the typical 5G cellular IoT, NB-IoT can admit only 50,000 devices per cell [@3GPP15Cellular]; ii) RAN slicing operation (e.g., creating, activating, and releasing slices) has to be conducted in a timescale of minutes to hours to keep pace with the upper layer slicing. However, the wireless channel generally changes in a timescale of millisecond to seconds. Results of the RAN slicing operation are desired to be achieved based on the time-varying channel. Thus, the RAN slicing should tackle a two timescale issue [@tang2019service].
These challenges motivate us to investigate the RAN slicing for mIoT and bursty URLLC service provision to maximize the utility of mIoT slices and that of bursty URLLC slices. The main contributions of this paper can be summarized as the following:
- We revisit the frame and minislot structure for mIoT transmission to accommodate more RA requests from a massive number of IoT devices.
- We adopt a queueing model to track the IoT packet arrival, accumulate and departure processes and analyze the queue evolution process by employing probability and stochastic geometry theories. Based on the analysis result, we derive the closed-form expression of the RA success probability of a randomly chosen IoT device.
- We define mIoT slice utility and bursty URLLC slice utility and formulate the RAN slicing for mIoT and bursty URLLC service multiplexing as a resource optimization problem. The objective of the optimization problem is to maximize the total mIoT and URLLC slice utilities, subject to limited physical resource constraints. The mitigation of this problem is difficult due to the existence of indeterministic objective function and thorny non-convex constraints and the requirement of tackling a two timescale issue as well.
- To mitigate this thorny optimization problem, we propose a slice resource optimization (SRO) algorithm. In this algorithm, we first exploit a sample average approximate (SAA) technique and an alternating direction method of multipliers (ADMM) to tackle the indeterministic objective function and the two timescale issue. Then, a semidefinite relaxation (SDR) scheme joint with a Taylor expansion scheme are leveraged to approximate the non-convex problem as a convex one. The tightness of the SDR scheme and the error bound of the Taylor expansion are also analyzed.
The remaining of this paper is organized as follows. We review the prior arts in Section II. In Section III, we describe our system model and formulate the service multiplexing problem in Section IV. The problem-mitigating algorithm is presented in Sections V and VI. In Section VII, we give the simulation results and conclude this paper in Section VIII.
Prior arts
==========
Recently, many researches have been conducted to increase RA success probabilities and/or reduce the access delay of mIoT devices. They can be generally classified into two groups: traffic detection and estimation based algorithms and algorithms without traffic detection and estimation.
The fundamental idea of the traffic detection and estimation based algorithms is to design an RA algorithm based on the detected and/or estimated users’ activity and traffic congestion situation and so on. For example, to reduce the access delay, a grant-free non-orthogonal RA system relying on the accurate user activity detection and channel estimation was proposed in [@zhang2019dnn]. A traffic-aware spatiotemporal model for the contention-based RA analysis is conducted for mIoT networks in [@jiang2018analyzing]. With the spatiotemporal model, a hybrid power ramping and back-off RA scheme was then developed to improve the RA success probability. Besides, an extended pseudo-Bayesian backlog estimation scheme was exploited in [@jang2019versatile] to estimate the number of backlogged nodes to attempt access. A versatile access control mechanism was then designed to reduce the access delay based on the estimation results.
For algorithms without detection and estimation, they design RA schemes without detecting users’ activity or estimating the statistical characteristic of traffic. For instance, the work in [@xing2019novel] proposed to improve the RA success probability of an IoT device by exploiting a distributed queue mechanism and then proposed an access resource grouping mechanism to reduce the access delay caused by the queuing process of the distributed queue mechanism. To increase RA success probability, the work in [@jang2016non] proposed to increase the number of preambles at the first step of the RA procedure by utilizing a spatial group mechanism and improve resource utilization through non-orthogonally allocating uplink channel resources at the second step of the RA procedure. Additionally, without knowing the statistical characteristic of traffic, a reinforcement learning-based algorithm was proposed in [@jiang2019reinforcement] to determine the uplink resource configuration for RA such that the average number of served IoT devices was maximized while ensuring a high RA success probability.
Except for the IoT service, future networks are envisioned to simultaneously support different services and applications with significantly different requirements on reliability, latency and bandwidth. As a result, researchers are now paying more attention to the service multiplexing of IoT/mMTC and many other services such as eMBB and URLLC. For example, instead of slicing the RAN via orthogonal resource allocation among different services, the work in [@popovski20185g; @budhiraja2019tactile] studied the potential advantages of allowing for non-orthogonal RAN resources sharing in uplink communications from a collection of mMTC, eMBB, and URLLC devices to the same BS. The work in [@ksentini2017toward] developed a two-level scheduling process to allocate dynamically dedicated bandwidth to each network slice according to workload demand and slices’ quality of service (QoS) requirement such that flexible resource allocation could be implemented. The work in [@rost2017network] proposed to maintain slice-specific radio resource control elements with which the RAN protocol stacks and different slices were configured. Besides, the work in [@wen2018robustness] aimed to optimize the virtual network functions and infrastructure resources such as the system bandwidth to implement slice recovery and reconfiguration for mMTC and eMBB service provision. The work in [@buyakar2018poster] proposed to maintain slice isolation between mMTC and eMBB slices and meet the performance requirements of these slices through limiting and dynamically updating the amount of resources allocated to each slice and monitoring the resource usage of each slice. After representing the slice performance requirements as the required amount of resources per deadline interval, an idea of earliest-deadline and first-scheduling was exploited in [@guo2019enabling] to allocate radio resources to mMTC, eMBB, and URLLC slices effectively.
System model
============
We consider a coordinated-multipoint-enabled RAN slicing system for mIoT and bursty URLLC multiplexing service provision. From the viewpoint of the infrastructure composition, the system mainly includes one baseband unit (BBU) pool and multiple remote radio heads (RRHs) that connect to the BBU via fronthaul links. From the perspective of network slicing, two types of inter-slices, i.e., mIoT slices and URLLC slices, are exploited in this system with $\mathcal{S}^I$ and $\mathcal{S}^u$ representing the sets of mIoT slices and URLLC slices, respectively. We focus on the modelling of uplink IoT data transmission in mIoT slices and the modelling of downlink URLLC data transmission in URLLC slices. The IoT devices are spatially distributed in ${\mathbb R}^2$ according to an independent homogeneous Poisson point process (PPP) $\Phi_s = \{u_{i,s}; s \in \mathcal{S}^I, i = 1, 2, \ldots\}$ with intensity $\lambda_s^I$, where $u_{i,s}$ denotes the location of the $i$-th IoT device in the $s$-th mIoT slice. There are also $N^u$ URLLC devices that are randomly and evenly distributed in ${\mathbb R}^2$. The RRHs are spatially distributed in ${\mathbb R}^2$ according to an independent PPP $\Phi_R = \{v_j; j = 1, 2, \ldots\}$ with intensity $\lambda_R$, where $v_j$ represents the location of the $j$-th RRH. The number and locations of IoT devices and RRHs will be fixed once deployed. Besides, each RRH is equipped with $K$ antennas, and each device is equipped with a single antenna. In IoT network slices, each IoT device is assumed to connect to its geographically closest RRH [@jiang2018analyzing]; thus, the cell area of each RRH constitutes a Voronoi tessellation. In URLLC network slices, RRHs cooperate to transmit signals to a URLLC device to improve its signal-to-noise ratio (SNR). A flexible frequency division multiple access (FDMA) technique is utilized to achieve the inter-slice and intra-slice interference isolation [@tang2019service].
The system time is discretized and partitioned into time slots and minislots with a time slot consisting of $T$ minislots. On the one hand, at the beginning of each time slot, a RAN slicing coordinator [@How2019yang] will decide whether to accept or reject received network slice requests which will be defined in the following subsections. Once a slice request is accepted, a network slice management will be responsible for activating or creating a virtual slice that is well resource-configured to satisfied the QoS requirements of devices in the slice [@How2019yang]. The slice configuration process is time costly and will generally be conducted in a timescale of minutes to hours [@tang2019service]. On the other hand, at the beginning of each minislot, each active IoT device may try to connect to its associated RRH, and RRHs will generate cooperated beamformers based on sensed channel coefficients.
mIoT slice model
----------------
By referring to the concept of a network slice [@rost2017network], especially from the viewpoint of the QoS requirement of a slice, we can define a mIoT slice request as follows.
\[def:IoT\_slice\_definition\] A mIoT slice request is defined as a tuple $\{\lambda_s^I, \gamma_s^{th}, N_{a,s}\}$ for any slice $s \in {\mathcal S}^I$, where $\gamma_s^{th}$ is the requirement of data transfer rate from an IoT device in $s$ to its associated RRH, $N_{a,s}$ denotes the number of accumulated packets in a queue of an IoT device in $s$.
In this paper, all mIoT slice requests are always accepted by the RAN slicing coordinator. IoT devices with the same data transfer rate are assigned to the similar slice. For an IoT device in $s$, if it has the opportunity to send its endogenous arrival packets to the corresponding RRH, then it will randomly select a preamble (e.g., orthogonal Zadoff¨CChu sequences) from a BBU-maintained preamble pool and transmit the preamble to the RRH at the data rate $\gamma_s^{th}$. Just like the literature [@soorki2017stochastic; @gharbieh2017spatiotemporal], if the RRH can successfully decode the preamble, then a connection between the IoT device and the RRH are considered to be set up although the whole connection establishment process usually follows an RA four-step procedure [@grau2019preamble]. In other word, the RA success probability is regarded as the probability of successfully transmitting a preamble in this paper. Next, we will analyze an IoT device queue evolution model, with the analysis of which the RA success probability of the IoT device will be derived.
### Queue evolution model
The queue evolution process consists of the packet arrival process, packet accumulate process, and packet departure process.
During minislot $t$, a Poisson distribution with intensity (or arrival rate) $\epsilon_{w,s}(t)$ is exploited to model the random, mutually independent endogenous packet arrivals in an IoT device in slice $s$. Then during minislot $t$ with a duration $\tau$, the arrival intensity of new packets can be expressed as $\mu_{w,s}(t) = \epsilon_{w,s}(t)\tau$. Once arrived, new packets will not be sent out immediately in general and will enter a queue, which is modelled as an $M/M/k$ queue with unlimited capacity, to wait for their scheduling. In the $M/M/k$ queue, packets will be scheduled according to the first-come, first-served (FCFS) basis. The unlimited queue capacity indicates that the age of information [@kaul2011minimizing; @kaul2012real] of new arrivals will not be considered, and packets will not be dropped before sending out. Besides, owing to the RA behavior of a slotted-ALOHA protocol, new arrivals during $t$ will only be counted at minislot $t + 1$. Thus, the accumulated number of packets $N_{a,s}(t)$ of a randomly selected IoT device in slice $s$ at $t$ is determined by the accumulated number of packets and the number of new arrivals at $t - 1$ and whether the preamble of the device can be successfully decoded by its associated RRH. Table \[table:queue\_evolution\] shows the evolution of accumulated packets in an IoT device. In this table, $x_s = \gamma_s^{th}/L$ packets at the head of the queue will be popped out if the corresponding RA succeeds, where $L$ denotes the IoT packet length; otherwise, they will be kept in the queue and wait for the opportunity of re-transmission at the next minislot. The operation $[x]^+ = \max(x,0)$.
Value Success Failure
---------------- ------------------------------------------------------ -----------------------------------------
$N_{a,s}(1)$ $0$ $0$
$N_{a,s}(2)$ $[N_{w,s}{(1)}$ - $x_s]^+$ $N_{w,s}{(1)}$
$N_{a,s}{(3)}$ $[N_{a,s}{(2)}$ + $N_{w,s}{(2)}$ - $x_s]^+$ $N_{a,s}{(2)}$ + $N_{w,s}{(2)}$
$\ldots$ $\ldots$ $\ldots$
$N_{a,s}{(t)}$ $ [N_{a,s}(t$ - $1)$ + $N_{w,s}(t$ - $1)$ - $x_s]^+$ $N_{a,s}(t$ - $1)$ + $N_{w,s}(t$ - $1)$
: Accumulated packets evolution in an IoT device[]{data-label="table:queue_evolution"}
With the evolution of accumulated packets, we can define the non-empty probability of the queue of an IoT device in $s$ as the following.
\[def:non\_empty\_prob\] At minislot $t$, for a randomly selected IoT device in slice $s \in \mathcal{S}^I$, the probability that its queue is not empty can be defined as $$\label{eq:non_empty_prob}
P_{ne,s}{(t)} = {\mathbb P}\{N_{a,s}(t) > 0\}, \forall s \in {\mathcal{S}^I}$$
(\[eq:non\_empty\_prob\]) implicitly reflects that new arrival packets at $t$ will not be sent out immediately. According to the evolution of $N_{a,s}(t)$, it can be observed that $P_{ne,s}(t)$ is determined by the probability distribution of $N_{a,s}(t-1)$ and the RA success probability. Since these probabilities and their correlations are unknown, the derivation of the explicit expression of $P_{ne,s}(t)$ is difficult.
Next, we describe the packet departure process combined with a frame and minislot structure for mIoT packets transmission. As mentioned above, partly because of the limitation on the frame and minislot structure, NB-IoT and LTE-M can only admit 50,000 devices. For NB-IoT, only one physical resource block (PRB) with a bandwidth of $180$ KHz in the frequency domain is allocated for IoT transmission, and each physical channel occupies the whole PRB. For LTE-M, although the physical channels are time and frequency multiplexed, it only reserves six in-band PRBs with a total bandwidth of $1.08$ MHz in the frequency domain for IoT data transmission. Therefore, the frame and minislot structure for mIoT transmission should be revisited if more RA requests from IoT devices want to be accepted.
Fig. \[fig:fig\_frame\_minislot\_structure\] depicts a frame and minislot structure for mIoT transmission in each mIoT slice[^2]. In this structure, both the frequency division multiplexing scheme and code division multiplexing scheme are leveraged to admit more IoT devices in the way of alleviating the mutual device interference. Particularly, the frequency division multiplexing scheme alleviates signal interference through orthogonal frequency allocation, and the code division multiplexing scheme mitigates the co-channel signal interference via reducing the cross-correlation of simultaneous transmissions. The combination of the two schemes may significantly mitigate interference experienced at an RRH. In this way, the QoS requirements of more IoT devices may be satisfied, and the RAN slicing system may support more IoT devices. For a mIoT slice $s \in \mathcal{S}^I$, each subframe includes $F_s$ orthogonal uplink physical RA channels (PRACHs). A single tone mode with a tone spacing of size of $a$ MHz is adopted for each uplink PRACH, which indicates that each PRACH occupies a PRB. At the beginning of each minislot, an active IoT device, i.e., the device’s queue is non-empty, will randomly choose a preamble from a set of non-dedicated RA preambles of size $\xi$ and transmit the preamble through a randomly selected PRACH. For each preamble, it has an equal probability ${\frac{1}{\xi}}$ to be chosen by each IoT device. Similarly, each PRACH has an equal probability $\frac{1}{F_s}$ to be selected. Thus, the average number of IoT devices in mIoT slice $s \in \mathcal{S}^I$ choosing the same PRACH and the same preamble is $\frac{\lambda_s^I}{\xi F_s}$. Notably, a greater $\xi F_s$ may significantly reduce signal interference experienced at each RRH.
Then, the following question should be tackled: *how many PRBs should be reserved for mIoT transmission?* To improve the resource utilization, the resource allocated to mIoT should be determined according to the requirements of mIoT and other coexistence services. It motivates us to optimize the resources orchestrated for the mIoT service except for analyzing the RACH procedure of IoT devices. The optimization procedure will be discussed in detail in the next section.
![The frame and minislot structure. ’R’ and ’D’ denote the resource block reserved for preamble and IoT data transmission. PBCH, PSS and SSS represent the PRBs for physical broadcast channel, primary synchronization signal and secondary synchronization signal transmission, respectively.[]{data-label="fig:fig_frame_minislot_structure"}](frame_slot_structure.eps){width="2.9in"}
### Access control scheme
In a mIoT network slice, as the slotted-ALOHA protocol allows all active IoT devices to request for RA at the beginning of each minislot without checking the status of channels, IoT devices may simultaneously transmit preambles. It may incur severe slice congestion that may lower the RA success probabilities of IoT devices and degrade the system performance. Access control has been considered as an efficient proposal of alleviating congestion, and many access control schemes have been proposed [@jiang2018analyzing; @Study163GPP]. In this paper, we aim at illustrating the performance difference between a network slicing system without access control and with access control. Therefore, we adopt the following two schemes [@jiang2018analyzing]:
- **Unrestricted scheme:** each active IoT device requests the RACH at the beginning of minislot $t$ without access restriction. If mIoT slices are not crowded or in a light-crowded condition, then this scheme may quickly flush queues of IoT devices. However, if a heavy-crowded condition is encountered, then this scheme may result in a high packet queueing delay.
- **Access class barring (ACB) scheme:** at the beginning of $t$, each active IoT device throws a random number $q \in [0, 1]$ and can request the RACH only if $q < P_{ACB}$, where $P_{ACB}$ is an ACB factor determined by RRHs based on the slice congestion condition. The ACB scheme can relieve slice congestion to some extent by reducing RACH requests of active IoT devices.
With the introduced access control schemes, we can define the non-restriction probability of a randomly selected IoT device in $s$ as follows.
\[mydef:unrestricted\_prob\] At minislot $t$, for a randomly selected IoT device in slice $s \in \mathcal{S}^I$, the probability that its RACH request is not restricted is defined as $$\label{eq:non_restriction_prob}
P_{nr,s}{(t)} = {\mathbb P}\{{\rm Unrestricted} \text{ } {\rm RACH} \text{ } {\rm requests}\}, \forall s \in {\mathcal{S}^I}$$
For all $s \in \mathcal{S}^I$ at any minislot $t$, we have $P_{nr,s}(t) = 1$ for the unrestricted scheme and $P_{nr,s}(t) = P_{ACB}$ for the ACB scheme.
### Analysis of RA success probability
For an RRH, two significant reasons may lead to an error preamble decoding i) the achieved preamble transfer rate at the RRH is less than a preset threshold; ii) the RRH simultaneously decodes at least two similar co-channel preambles, and thus preamble collision occurs. The research of the mitigation of preamble collision has been well conducted in [@jang2019versatile; @jiang2018collision]. Just like [@jiang2018random], we focus on the exploration of enabling successful single preamble transmission that is discussed in detail as follows.
We utilize a power-law path-loss model to calculate the path-loss between an IoT device and its RRH in mIoT slices and utilize a truncated channel inversion power control scheme to eliminate the ’near-far’ effect. In the power-law path-loss model, the IoT device transmit power decays at the rate of $r^{-{\varphi}}$ with $r$ representing the propagation distance and $\varphi$ denoting the path-loss exponent. In the power control scheme, IoT devices associated with the same RRH compensate for the path-loss to maintain the average received signal power at the RRH equal to a threshold $\rho_o$. Without loss of generality, the cutoff threshold $\rho_o$ is set to be the same for all RRHs. Owing to the channel deep fading, severe co-channel interference, and insufficient transmit power, an IoT device may experience uplink preamble transmission outage. The following definition describes the definition of the probability that a randomly selected IoT device can successfully transmit a chosen preamble to its corresponding RRH.
\[mydef:preamble\_trans\_suc\_prob\] At minislot $t$, for a randomly selected active IoT device in slice $s \in \mathcal{S}^I$, its RA success probability is defined as $$\label{eq:preamble_trans_suc_prob}
P_{s}{(t)} = {\mathbb P}\{r_s(t) \ge \gamma_s^{th}\}, \forall s \in {\mathcal{S}^I}$$ where $r_s(t) = a{\log _2}(1 + SIN{R_s}(t))$ denotes the achieved preamble transfer rate at the IoT device’s associated RRH and $SIN{R_s}(t)$ is the signal-to-interference-plus-noise ratio (SINR).
Then, for any active IoT device in $s$, its QoS requirement is given by $$\label{eq:mMTC_QoS}
P_s(t) \ge {\pi _s}, \forall s \in {\mathcal{S}^I}$$ where $\pi_s$ denotes a threshold of the required RA success probability.
This definition shows that the QoS requirement of each active IoT device in $s$ should be satisfied if the slice request of $s$ is accepted. The definition also states that $P_{s}{(t)}$ is correlated with the non-empty probability $P_{ne,s}{(t)}$. Recall that the RA success probability of an IoT device impacts its non-empty probability, we can know that the RA success probability and the non-empty probability are intertwined. Additionally, $SIN{R_s}(t)$ is a function of complicated co-channel interference. Thus, it is hard to obtain the closed-form expression of $P_{s}{(t)}$.
Without any loss in generality, we perform the analysis of RA success probability on an RRH located at the origin. According to Slivnyak’s theorem [@haenggi2012stochastic], the analysis holds for a generic RRH located at a generic location. For a randomly selected IoT device with non-empty queue in $s \in \mathcal{S}^I$, the theoretical preamble transfer rate experienced at the RRH located at the origin can take the form $$\label{eq:preamble_trans_suc_prob}
r_s(t) = a \log_2 \left (1 + \frac{{\rho_o {h_o}}}{{{\sigma ^2} + {{\cal I}_{s}}(t)}} \right ), \forall s \in {\mathcal{S}^I}$$ where $\sigma^2$ represents the noise power, ${{\cal I}_{s}}(t)$ denotes signal interference received at the RRH, the useful signal power equals to $\rho_o h_o$ due to the truncated channel inversion power control[^3] [@elsawy2014stochastic] with $h_o$ denoting the channel power gain between the IoT device and the RRH. It is noteworthy that the channel power gain experienced at a generic RRH is related to the spatial locations of both the RRH and its associated IoT devices. Nevertheless, we drop the spatial indices for notation lightening. Besides, just like [@elsawy2014stochastic], all channel gains are assumed to be known and be independent of each other, independent of the spatial locations, symmetric and are identically distributed (i.i.d.). Considering both the particular IoT device deployment environment and the convenience of theoretical analysis, the Rayleigh fading is assumed, and the channel power gain $h_o$ is assumed to be exponentially distributed with unit mean.
Based on the following five facts, we next present the analytical expression of signal interference
- **Fact 1**: the average signal received from any single IoT device belonging to inter-cells is strictly less than $\rho_o$.
- **Fact 2**: the average interference signal received from any single interfering IoT device associated with the origin RRH strictly equals to $\rho_o$.
- **Fact 3**: IoT devices choosing the same co-channel preamble as the randomly selected IoT device may become an interfering IoT device.
- **Fact 4**: at each minislot, IoT devices with non-empty queue may become interfering IoT devices.
- **Fact 5**: IoT devices in difference slices may not mutually interfere.
Note that Fact 1 and Fact 2 are direct consequences of the device-RRH association policy and power control scheme. Fact 5 holds due to the exploration of intra-slice isolation. Therefore, the aggregate interference received at the origin RRH can take the following form $$\label{eq:I_intra}
\begin{array}{l}
{\cal I}_{s}(t) = \sum\limits_{{u_{m,s}} \in \Phi_s \backslash \{ o\} } {{\mathbbm 1}({p_m}||{d_m}|{|^{ - \varphi }} = {\rho _o}){\mathbbm 1}(N_{a,s}{(t)} > 0)} \times \\
\qquad \quad \text{ } {{\mathbbm 1}({f_m} = {f_o}){\rho _o}{h_m}}, \forall s \in {\mathcal{S}^I}
\end{array}$$ where $o$ represents the randomly selected IoT device associated with the RRH at the origin, $p_m$ represents the transmit power of the $m$-th IoT device, $||d_m||$ is the distance between the $m$-th IoT device and the origin RRH, $f_o$ denotes the preamble and channel chosen by the randomly selected IoT device, $f_o = f_m$ indicates that the randomly selected IoT device and the $m$-th IoT device select the same preamble and channel. ${\mathbbm 1}(\cdot)$ is the indicator function that equals to one if the statement ${\mathbbm 1}(\cdot)$ is ture; otherwise, it equals to zero. Just like [@zhang2015resource], in (\[eq:I\_intra\]), co-channel inter-cell interference is assumed as a part of thermal noise mainly because of the severe wall penetration loss.
Then, for the randomly selected IoT device in $s \in {\mathcal{S}^I}$, we can rewrite (\[eq:preamble\_trans\_suc\_prob\]) as the following form with (\[eq:I\_intra\]) $$\label{eq:QoS_analysis}
\begin{array}{l}
P_s(t) = {\mathbb P}\{ SINR_s(t) \ge \theta _s^{th}\} \\
\qquad \text{ } = {\mathbb P}\{ {h_o} \ge \frac{{\theta _s^{th}}}{{{\rho _o}}}({\sigma ^2} + {{\cal I}_{s}(t)})\} \\
\qquad \text{ } \mathop = \limits^{(a)} {{\mathbb E}}\left[ {\exp \left\{ { - \frac{{\theta _s^{th}}}{{{\rho _o}}}({\sigma ^2} + {{\cal I}_{s}(t)})} \right\}} \right]\\
\qquad \text{ } = \exp \left\{ { - \frac{{\theta _s^{th}}}{{{\rho _o}}}{\sigma ^2}} \right\}{\cal L}_{{{\mathcal I}_{s}(t)}}\left(\frac{\theta_s^{th}}{\rho_o}\right), \forall s \in {\mathcal{S}^I}
\end{array}$$ where $\theta_s^{th} = {2^{\gamma _s^{th}/a}} - 1$. (a) follows from the full probability law over ${\cal I}_{s}(t)$, and ${\cal L}_{{{\cal I}_{s}(t)}} (\cdot)$ denotes the Laplace transform (LT) of the probability density function (PDF) of the random variable ${\cal I}_{s}(t)$. Note that the notation ${\cal L}_{{{\cal I}_{s}(t)}} (\cdot)$ is a terminology that is a slight abuse of subscript ${{\cal I}_{s}(t)}$.
The following lemma characterizes the LT of aggregate interference $\mathcal{I}_s(t)$.
\[lem:LT\_interference\_expression\] For the origin RRH, the LT of its received aggregate interference from active IoT devices associated with it is given by $$\label{eq:LT_interference_expression}
{\cal L}_{{{\cal I}_{s}(t)}} \left(\varpi_s\right) = {\frac{{1 + \varpi_s{\rho _o}}}{{{{\left( {1 + \alpha_s \varpi_s{\rho _o}/\left( {1 + \varpi_s{\rho _o}} \right)} \right)}^{3.5}}}} - \frac{{1 + \varpi_s{\rho _o}}}{{{{\left( {1 + \alpha_s } \right)}^{3.5}}}}}$$ where $\varpi_s = \frac{\theta_s^{th}}{\rho_o}$, $\alpha_s = \frac{{{P_{nr,s}}(t){P_{ne,s}}(t){\lambda _{{s}}^I}}}{{3.5{\lambda _R}{\xi F_s}}}$, for all $s \in {\mathcal{S}^I}$.
Please refer to Appendix A.
With the conclusion in Lemma \[lem:LT\_interference\_expression\], we can then obtain the mathematical expression of the RA success probability of a randomly selected IoT device at $t$ in the following corollary.
\[lem:QoS\_prob\_analysis\] For a randomly selected IoT device in a mIoT slice $s \in {\cal S}^I$, its RA success probability at minislot $t$ is given by $$\label{eq:QoS_prob_analysis}
{P_s}(t) = {\frac{{(1 + \varpi_s{\rho _o})}{{e^{-\varpi_s{\sigma ^2}}}}}{{{{\left( {1 + \alpha_s \varpi_s{\rho _o}/\left( {1 + \varpi_s{\rho _o}} \right)} \right)}^{3.5}}}} - \frac{{(1 + \varpi_s{\rho _o})}{{e^{-\varpi_s{\sigma ^2}}}}}{{{{\left( {1 + \alpha_s } \right)}^{3.5}}}}}$$
By substituting (\[eq:LT\_interference\_expression\]) into (\[eq:QoS\_analysis\]), we can obtain (\[eq:QoS\_prob\_analysis\]).
Although Corollary \[lem:QoS\_prob\_analysis\] presents a mathematical expression of $P_s(t)$, the expression is not in the closed-form as it is a function of $P_{ne,s}(t)$ the closed-form expression of which is not obtained. Next, we derive the closed-form expression of $P_{ne,s}(t)$.
### Analysis of non-empty probability
According to the definition of non-empty probability, $P_{ne,s}(t)$ is correlated with the number of accumulated packets $N_{a,s}(t)$ of the randomly selected IoT device in mIoT slice $s$. Thus, we theoretically analyze the non-empty probability of the randomly selected IoT device as the following.
As the number of the accumulated packets in the queue of a randomly selected IoT device in slice $s \in \mathcal{S}^I$ at the $1^{\rm st}$ minislot is empty, its non-empty probability $P_{ne,s}^{1}$ at the $1^{\rm st}$ minislot can take the form $$\label{eq:non_empty_prob_1}
P_{ne,s}^{1} = {\mathbb P}\{N_{a,s}^1 > 0\} = 0, \forall s \in {\mathcal{S}^I}$$ where we write $x^t$ instead of $x(t)$ to lighten the notation.
The following lemma presents the closed-form expression of the non-empty probability of a randomly selected IoT device served by the origin RRH when minislot $t > 1$.
\[lem:non\_empty\_prob\_lemma\] The number of accumulated packets of a randomly selected IoT device served by the origin RRH at minislot $t > 1$ may be approximately Poisson distributed. Therefore, for any mIoT slice $s \in \mathcal{S}^I$, we approximate the number of accumulated packets $N_{a,s}^{t}$ at minislot $t$ as a Poisson distribution with intensity $\mu_{a,s}^{t}$, which is given by $$\label{eq:mu_accumulated_packets}
\mu _{a,s}^t = \left [\mu _{w,s}^{t - 1} + \mu _{a,s}^{t - 1} - P_{s}^{t - 1}\left( {1 - {e^{ - \mu _{w,s}^{t - 1} - \mu _{a,s}^{t - 1}}}} \right) \right ]^+, \forall s \in {\mathcal{S}^I}$$
Then, the non-empty probability of the device at minislot $t$ can be written as $$\label{eq:non_empty_prob_m}
P_{ne,s}^t = 1 - {e^{ - \mu _{a,s}^t}}, \forall s \in {\mathcal{S}^I}$$
Please refer to Appendix B.
Combine with (\[eq:QoS\_prob\_analysis\]) and (\[eq:mu\_accumulated\_packets\]), the closed-form expression of $P_s(t)$ ($s \in \mathcal{S}^I$) can be obtained.
Bursty URLLC slice model
------------------------
Similar to the definition of a mIoT slice request, a bursty URLLC slice request can be defined as the following.
\[mydef:URLLC\_slice\_request\] A bursty URLLC slice request is defined as four tuples $\{I_s^u, D_s, \alpha, \beta\}$ for slice $s \in {\mathcal S}^u$, where $I_s^u$ denotes the number of URLLC devices in $s$, $D_s$ denotes the transmission latency requirement of each URLLC device in $s$, $\alpha$ and $\beta$ represent the packet blocking probability and the codeword error decoding probability of each URLLC device, respectively.
In this definition, URLLC devices are grouped into $|{\mathcal S}^u|$ clusters according to the transmission latency requirement of each device. Owing to the low latency requirement URLLC packets should be immediately scheduled upon arrival; thus, all URLLC slice requests will always be accepted by the RAN coordinator. Except for the low packet error decoding probability that has been emphasized for URLLC transmission in a plenty of works [@tang2019service], this paper attempts to orchestrate slice resources to reduce the packet blocking probability for bursty URLLC transmission. This is because the bursty characteristic of URLLC traffic [@How2019yang] may lead to the packet blocking in URLLC slices, which may significantly reduce the reliability of URLLC transmission. Therefore, the indicators $\alpha$ and $\beta$ are involved to reflect the ultra-reliable requirement of URLLC transmission jointly.
Then, we address the following question: *how to orchestrate slice resources for reducing packet blocking probability and codeword error decoding probability?*
### Reduction of packet blocking probability
As mentioned above, the bursty feature of URLLC traffic is the crucial factor that leads to the URLLC packet blocking for URLLC transmission. Therefore, we next model the bursty URLLC traffic based on which we discuss how to orchestrate slice resources to alleviate the impact of bursty URLLC traffic.
During minislot $t$, an independent homogeneous Poisson distribution with intensity ${\bm \lambda} = \{\lambda_s; s \in {\mathcal{S}^u}\}$ is utilized to model the number of bursty URLLC packets aggregated at RRHs, where $\lambda_s$ denotes the intensity of new arrivals destined to devices belonging to URLLC slice $s$.
Once arrived, new URLLC arrivals will enter a queue maintained by an RRH to be immediately scheduled. An $M/M/W^u$ queueing system with limited bandwidth $W^u$ is exploited to model the queue. Without loss of any generality, we assume that each RRH maintains the same queue due to the exploration of cooperated transmission. In the queue, a packet destined to URLLC device $i \in {\mathcal{I}_s^u}$, $s \in {\mathcal{S}^u}$ will be allocated with a block of system bandwidth $\omega_{i,s}^u(t)$ for a period of time $d_s \le D_s$ at $t$. Owing to stochastic variations in the bursty packet arrival process, the limited bandwidth may not be enough to serve new arrivals occasionally. As such, URLLC packet blocking may happen. To reduce the probability of URLLC packet blocking, the redesign of URLLC frame and minislot structure may be required.
At minislot $t$, let $P_b(\bm \omega^u(t), \bm \lambda, \bm d, W^u(t))$ denote the packet blocking probability experienced at an RRH, where ${{\bm \omega}^u(t) = \{\omega_{1,1}^u(t), \ldots, \omega_{i,s}^u(t), \ldots, \omega_{I_{|{\mathcal S}^u|}^u,|{\mathcal S}^u|}^u(t)\}}$ and ${\bm d} = \{d_1, \ldots, d_s \ldots, d_{|{\mathcal S}^u|}\}$. The Theorem 1 in [@anand2018resource] provides us with a clue of redesigning the URLLC frame and minislot structure in the time-frequency plane for bursty URLLC traffic transmission. This theorem indicates that if we narrow the PRB of the URLLC frame in the frequency domain while widening it in the time domain, then the number of concurrent transmissions will be increased. As a result, the packet blocking probability is reduced.
Therefore, for a URLLC packet destined to device $i \in {\mathcal I}_s^u$, $s \in \mathcal{S}^u$, we should scale up $d_s$ and choose $d_s$ and $\omega_{i,s}^u(t)$ at $t$ using the following equation $$\label{eq:omega_s}
d_{{s}} = D_s \ {\rm and} \ \omega_{{i,s}}^{u}(t) = \frac{b_{i,s}^u(t){r_{{i,s}}^{u}(t)}}{{\kappa D_s}}, \forall i \in {\mathcal I}_s^u, s \in {\mathcal S}^u$$ where $r_{i,s}^u(t)$ denotes channel uses for transmitting a URLLC packet, $\kappa$ is a constant reflecting the number of channel uses per unit time per unit bandwidth of FDMA frame structure and numerology, $b_{i,s}^u(t)$ is an indicator variable reflecting whether the QoS requirement of device $i$ in slice $s$ can be satisfied at $t$. As network resources are limited and shared by all network slices, not all URLLC devices can be guaranteed to be served at every minislot. Certainly, we can adjust the slice priority weight that will be introduced in the following section to guide the resource orchestration for enforcing the entire URLLC devices coverage.
Based on the result in (\[eq:omega\_s\]) and the conclusion of the Lemma 3.2 in our previous work [@How2019yang], we can derive the minimum upper bound of bandwidth orchestrated for URLLC slices in the following lemma.
\[lem:upper\_bound\_urllc\_bandwidth\] At minislot $t$, for a given $M/M/W^u$ queue with packet arrival intensity $\bm \lambda$ and a family of packet transmit rates $\{{\kappa/r_{i,s}^u(t)}\}$, let $W^u({\bm r}(t))$ denote the minimum upper bound of bandwidth orchestrated for URLLC slices such that $P_Q^{M/M/W^u}\le \varsigma$ and $P_b({\bm \omega}^u(t), {\bm \lambda}, {\bm D}$, $W^u({\bm r}(t)))$ is of the order of $\alpha$, where $P_Q^{M/M/W^u}$ represents the queueing probability, and ${\bm D} = \{D_1, \ldots, D_{|{\mathcal S}^u|}\}$. If $\varsigma > \alpha$, then we have $$\label{eq:URLLC_bandwidth}
\begin{array}{l}
W^u(\bm r(t)) \approx \sum\limits_{s \in {\mathcal S}^u } {\sum\limits_{i \in {\mathcal I}_s^u} {{\lambda _{s}}b_{i,s}^u(t)\frac{{{{r_{i,s}^u(t)}}}}{\kappa}} } + \\
\frac{{{\alpha} - \varsigma {\alpha}}}{{\varsigma - {\alpha}}}\sqrt {\frac{\left({\sum\limits_{s \in {{\mathcal S}^u}} {\sum\limits_{i \in {\mathcal I}_s^u}{b_{i,s}^u(t)\lambda _s^2D_s^2}} }\right)\left( {\sum\limits_{s \in {\mathcal S}^u} {\sum\limits_{i \in {\mathcal I}_s^u} {{\lambda _{s}}b_{i,s}^u(t)\frac{{r_{i,s}^u(t)}^2}{{{\kappa ^2}{D_s}}}} }}\right)}{{\mathop {\min }\limits_{s \in {{\mathcal S}^u}} \{ {\lambda _s}{D_s}\} }}}
\end{array}$$
We omit the proof here as the similar proof can be found in the proof section of Lemma 3.2 in [@How2019yang].
### Reduction of codeword error decoding probability
The crucial factor that impacts the codeword error decoding probability is the network capacity. Next, we discuss the relationship between the network capacity and codeword error decoding probability.
For any URLLC slice $s \in {\cal S}^u$, during minislot $t$, let $x_{i,s}^u(t)$ be the original data symbol destined to a URLLC device $i \in {\mathcal{I}_s^u}$ with ${\mathbb E}[|x_{i,s}^u(t)|^2] = 1$, $\bm g_{ij,s}(t) \in {\mathbb C}^K$ be the transmit beamformer pointing at the device $i$ from the $j$-th RRH and $\bm h_{ij,s}(t) \in \mathbb{C}^K$ be the channel coefficient between the $i$-th URLLC device and the $j$-th RRH. The channel coefficient may change over minislots. However, it is assumed to be i.i.d. over each minislot and remain unchanged during each minislot. Then, the received signal at device $i$ in $s$ during minislot $t$ is given by $$\label{eq:URLLC_receiving_signal}
{\hat x_{i,s}^u(t) = }{\sum\limits_{j \in {\cal J}} {{\bm h}_{ij,s}^{\rm{H}}} (t){{\bm g}_{ij,s}}(t)x_{i,s}^u(t) + {\sigma _{i,s}}(t)}, \forall i \in {\mathcal I}_s^u, s \in {\mathcal S}^u$$ where the first term is the useful signal for $i$ and $\sigma_{i,s}(t) \sim {\cal CN}(0, \sigma_{i,s}^2) $ is the additive white Gaussian noise (AWGN) experienced at $i$. Similar to [@tang2019service], interference signal is not involved in (\[eq:URLLC\_receiving\_signal\]) due to the utilization of a flexible FDMA mechanism. Then the SNR received at device $i$ in $s$ at minislot $t$ can be written as $$\label{eq:URLLC_SINR}
SINR_{i,s}^u(t) = \frac{{|\sum\nolimits_{j \in {\cal J}} {{{\bm h}_{ij,s}^{\rm H}}{{(t)}}{{\bm g}_{ij,s}}(t)} {|^2}}}{{\phi \sigma _{i,s}^2}}, \forall i \in {\mathcal I}_s^u, s \in {\mathcal S}^u$$ where $\phi > 1$ is an SNR loss coefficient. The perception of perfect channel status information (CSI) or accurate channel coefficients requires the information exchange between an RRH and its associated device before data transmission, the process of which is generally time consuming. URLLC packets, however, have a stringent latency requirement. As a result, perfect CSI or accurate channel parameters may be unavailable for URLLC transmission, which may incur the SNR loss. The coefficient $\phi$ is then utilized to characterize the SNR loss [@hou2018burstiness].
Shannon capacity formula is created under a crucial assumption of transmitting a block with long enough blocklength. However, URLLC packets are typically very short to satisfy the ultra-low latency requirement. Thus, the famous Shannon capacity formula cannot be utilized to model the URLLC transmission data rate and capture the corresponding codeword error decoding probability. For URLLC transmission, we resort to the capacity analysis for a finite blocklength channel coding regime derived in [@yang2014quasi]. For any device $i \in {\mathcal{I}_s^u}$, $s \in \mathcal{S}^u$, the number of transmitted information bits $L_{i,s}^u(t)$ at minislot $t$ using $r_{i,s}^u(t)$ channel uses in AWGN channel can be accurately correlated with the codeword error decoding probability $\beta$ according to the following equation $$\label{eq:URLLC_bit_length}
\begin{array}{l}
L_{i,s}^u(t) \approx r_{i,s}^u(t)C(SNR_{i,s}^u(t)) - \\
\quad {Q^{ - 1}}(\beta)\sqrt {r_{i,s}^u(t)V(SNR_{i,s}^u(t))}, \forall i \in {\mathcal I}_s^u, s \in {\mathcal S}^u
\end{array}$$ where $C(SNR_{i,s}^u(t)) = \log_2(1 + SNR_{i,s}^u(t) )$ is the AWGN channel capacity under infinite blocklength assumption, $V(SNR_{i,s}^u(t)) = \ln^2 2\left( {1 - \frac{1}{{{{(1 + SNR_{i,s}^u(t))}^2}}}} \right)$ is the channel dispersion, $Q(\cdot)$ is the $Q$-function. It is noteworthy that a URLLC packet will usually be coded before transmission and the generated codeword will be transmitted in the air interface such that the transmission reliability can be improved.
The complicated expression of $V(SNR_{i,s}^u(t))$ in (\[eq:URLLC\_bit\_length\]) significantly hinders the theoretical analysis of network resources orchestrated for URLLC slices. Fortunately, as $V(SNR_{i,s}^u(t))$ is maximized by $\ln^2 2$, the closed-form expression of the minimum upper bound of $r_{i,s}^u(t)$ ($i \in {\mathcal I}_s^u, s \in {\mathcal S}^u $) with a codeword error decoding probability $\beta$ can be given by [@How2019yang] $$\label{eq:URLLC_channel_use}
\begin{array}{l}
r_{i,s}^u(t) = \frac{{L_{i,s}^u(t)}}{{C(SINR_{i,s}^u(t))}} + \frac{{{{({Q^{ - 1}}(\beta))}^2}}}{{2{{(C(SINR_{i,s}^u(t)))}^2}}}\\
\quad + \frac{{{{({Q^{ - 1}}(\beta))}^2}}}{{2{{(C(SINR_{i,s}^u(t)))}^2}}}\sqrt {1 + \frac{{4L_{i,s}^u(t)C(SINR_{i,s}^u(t))}}{{{{({Q^{ - 1}}(\beta))}^2}}}}
\end{array}$$
Problem formulation
===================
This section aims to formulate the problem of RAN slicing for mIoT and bursty URLLC service multiplexing based on the above models.
In mIoT slices, each RRH may transmit feedback signal to its connected IoT devices for the connection establishment according to an RA four-step procedure [@grau2019preamble]. Meanwhile, in URLLC slices, each RRH may transmit URLLC packets to URLLC devices. As the transmit power $E_j$ of each RRH is limited, we have the following transmit power constraint $$\label{eq:RRH_energy}
\begin{array}{l}
\sum\limits_{s \in {\cal S}} {{(1 + \alpha_g)\frac{\lambda_s^I}{\lambda_R}{{\hat E}_{j}^I}}} + \\
\qquad \sum\limits_{s \in {\cal S}^u} {\sum\limits_{i \in {\cal I}_s^u} {b_{i,s}^u(t){{\bm g}_{ij,s}^{\rm H}}{{(t)}}{{\bm g}_{ij,s}}(t)} } \le {E_j}, \forall j \in \mathcal{J}
\end{array}$$ where ${\hat E}_{j}^I$ is assumed to be a constant and denotes the transmit power of the $j$-th RRH for connecting to its associated IoT devices over downlink, $\alpha_g$ represents a safety margin coefficient. As a PPP with intensity $\lambda_s^I$ is utilized to model the distribution of IoT devices, the actual number of IoT devices may be greater than $\lambda_s^I$ once deployed. As a result, the coefficient $\alpha_g$ is introduced to reserve transmit power for exceeded IoT devices.
In the RAN slicing system, as the total limited system bandwidth $W$ will be shared by mIoT slices and URLLC slices, we have the following bandwidth constraint $$\label{eq:total_bandwidth}
\sum\limits_{s \in {\cal S}^I} {(1+\alpha_g)\omega_s(\bar t)} + W^u(\bm r(t)) \le W$$ where $\omega_s(\bar t)$ denotes the bandwidth allocated to mIoT slice $s \in \mathcal{S}^I$ that is correlated with $F_s$ by means of $F_s = \left\lfloor {{{\omega _s}(\bar t)}/{{a}}} \right\rfloor$, and $\alpha_g \omega_s(\bar t)$ denotes a block of reserved bandwidth.
In (\[eq:total\_bandwidth\]), $F_s$ is an integer, and some integer variable recovery schemes [@tang2019systematic] can be leveraged to obtain the suboptimal $F_s$. However, considering the high computational complexity of optimizing an integer variable and the utilization of the scheme of spectrum safety margin, we directly relax the integer variable into a continuous one, i.e., let $F_s = {{{\omega _s}(\bar t)}/{{a}}}$. Without loss of any generality, we regard $\omega_s(\bar t)$ as an independent variable below. Besides, as at least one PRB should be allocated to mIoT slices, we have $$\label{eq:mMTC_bandwidth}
\omega_s(\bar t) \ge a, \forall s \in {\cal S}^I$$
Owing to the exploration of mIoT and bursty URLLC service multiplexing, we should orchestrate network resources for all mIoT slices and URLLC slices to simultaneously maximize the utilities of mIoT slices and URLLC slices.
For a mIoT slice $s \in \mathcal{S}^I$, its primary goal is to offload as many data packets as possible from IoT devices. In this way, the number of accumulated packets in each IoT device should be kept at a low level. Considering that a great RA success probability of an IoT device will lead to a low number of accumulated packets in its queue, we define the utility of a mIoT slice as the following.
\[mydef:IoT\_slice\_utility\] Over a time slot of duration $T$, the mIoT slice utility is defined as the time-average of RA success probabilities of IoT devices in all mIoT slices, which is given by $$\label{eq:IoT_utility}
{{\bar U}^I} = \frac{1}{T}\sum\limits_{t = 1}^T {{U^I}(t)} = \frac{1}{T}\sum\limits_{t = 1}^T {\tilde P(t) }$$ where $\tilde P(t) = \sum\limits_{s \in {{\mathcal S}^I}} {\frac{{{\lambda _s^I}{P_s}(t)}}{{\sum\nolimits_{s \in {{\mathcal S}^I}} {{\lambda _s^I}} }}}$ with the numerator $\lambda_s^I P_s (t)$ representing the expected sum of RA success probabilities of IoT devices in slice $s \in \mathcal{S}^I$ and the denominator $\sum\nolimits_{s\in\mathcal{S}^I}{\lambda_s^I}$ denoting a normalization coefficient.
In (\[eq:IoT\_utility\]), $\frac{\lambda_s^I}{\sum\nolimits_{s \in \mathcal{S}^I} {\lambda_s^I}}$ can be regarded as an intra-slice priority coefficient. A mIoT slice serving more IoT devices will be orchestrated with more network resources.
For a URLLC slice $s \in \mathcal{S}^u$, its primary objective is to maximize the slice gain that is reflected by the parameters in the slice request at a low cost. Therefore, we define an energy-efficient utility for URLLC slices, as presented below.
\[mydef:URLLC\_slice\_utility\] Over a time slot of duration $T$, the bursty URLLC slice utility is defined as the time-average energy efficiency for serving all URLLC devices, which is given by $$\label{URLLC_long_term_utility}
\begin{array}{l}
{{\bar U}^u} = \frac{1}{T}\sum\limits_{t = 1}^T {{U^u}(t)} = \frac{1}{T}\sum\limits_{t = 1}^T {\sum\limits_{s \in {{\mathcal S}^u}} {U_s^u({D_s},{\bm g_{ij,s}}(t))} } \\
\qquad = \frac{1}{T}\sum\limits_{t = 1}^T {\sum\limits_{s \in {{\mathcal S}^u}} {\sum\limits_{i \in {\mathcal I}_s^u} \frac{b_{i,s}^u(t)}{{1 - {e^{ - {D_s}}}}} } } - \\
\qquad \quad \frac{\eta }{T}\sum\limits_{t = 1}^T {\sum\limits_{s \in {{\mathcal S}^u}} {\sum\limits_{j \in {\mathcal J}} {\sum\limits_{i \in {\mathcal I}_s^u} {b_{i,s}^u(t)\bm g_{ij,s}^{\rm{H}}(t){\bm g_{ij,s}}(t)} } } }
\end{array}$$ where $\eta$ is an energy efficiency coefficient reflecting a tradeoff between the URLLC slice gain and the RRH energy consumption.
Then, over a time slot of duration $T$, the RAN slicing problem for mIoT and URLLC service multiplexing can be formulated as follows
\[eq:original\_problem\] $$\begin{aligned}
{2}
& \mathop {\rm maximize}\limits_{{b_{i,s}^u(t)},{{\omega_s}}(\bar t),{{\bm g}_{ij,s}}(t)} {{\bar U}^I} + {\tilde \rho} {{\bar U}^u} \\
& {\rm subject \text{ } to:} \nonumber \\
& b_{i,s}^u(t) \in \{ 0,1\}, \forall s \in {\cal S}^u, i \in {\cal I}_s^u \\
& \rm {constraints \text{ } (\ref{eq:mMTC_QoS}), (\ref{eq:RRH_energy})-(\ref{eq:mMTC_bandwidth}) \text{ } are \text{ } satisfied.}\end{aligned}$$
where $\tilde \rho$ is an inter-slice priority coefficient reflecting the priority of orchestrating network resources for mIoT slices and URLLC slices.
The mitigation of (\[eq:original\_problem\]) is quite challenging mainly because
- **indeterministic objective function**: (\[eq:original\_problem\]) should be optimized at the beginning of the $1^{\rm st}$ minislot. The time-averaged objective function of (\[eq:original\_problem\]) can only be exactly computed according to the future channel information. Therefore, the value of the objective function is indeterministic at the beginning of the $1^{\rm st}$ minislot.
- **two timescale issue**: the creation of a network slice is performed at a timescale of time slot. Thus, the variable $\omega_s(\bar t)$ should be determined at the beginning of the time slot $\bar t$ and kept unchanged over the whole time slot. The channel, however, is time-varying. As a result, the beamformer $\bm g_{ij,s}(t)$ should be optimized at each minislot. In summary, the variables in (\[eq:original\_problem\]) should be optimized at two different timescales.
- **thorny optimization problem**: at each minislot $t$, the constraint (\[eq:mMTC\_QoS\]) is non-convex over $\omega_s(\bar t)$, and the constraints (\[eq:RRH\_energy\]), (\[eq:total\_bandwidth\]) are non-convex over $\bm g_{ij,s}(t)$, which together lead to a non-convex problem.
Problem solution with system generated channel
==============================================
This section aims to tackle these challenges by exploiting of an SAA technique [@kim2015guide], an ADMM method [@boyd2011distributed], a semidefinite relaxation scheme and a Taylor expansion scheme.
Sample average approximation and alternating direction method of multipliers
----------------------------------------------------------------------------
As mIoT slices and URLLC slices share the network resources, both $\bar U^I$ and $\bar U^u$ may be determined by channel coefficients experienced by URLLC slices. At each minislot $t$, due to the i.i.d. assumption on the channel coefficients of URLLC slices, we have $$\label{objfun_approx}
\frac{1}{T}\sum\limits_{t = 1}^T {{U^I}(t)} + \frac{1}{T}\sum\limits_{t = 1}^T {\tilde \rho {U^u}(t)} \approx {E_{\hat {\bm h}}}\left[ {{{\hat U}^I} + \tilde \rho {{\hat U}^u}} \right]$$ where $\hat {\bm h}$ represents the channel samples of URLLC slices collected at the beginning of the time slot $\bar t$.
Given a collection of channel samples $\{\bm h_m\}$ with $\bm h_m = [\bm h_{11,1m};\ldots;\bm h_{1J,sm};\ldots;\bm h_{N^uJ,|\mathcal{S}^u|m}]$ and $m \in \mathcal{M}=\{1,\ldots,M\}$, Just like [@How2019yang], as constraints (\[eq:original\_problem\]b) and (\[eq:original\_problem\]c) construct a nonempty compact set, the conclusion of Proposition 5.1 in [@How2019yang] is applicable to this paper by exploiting the SAA technique. The conclusion indicates that if the number of channel samples $M$ is reasonably large, then $\frac{1}{M}\sum\limits_{m = 1}^M {U_m^I } + \frac{\tilde \rho}{M}\sum\limits_{m = 1}^M {U_m^u }$ converges to ${E_{\bm {\hat h}}}\left[ {{{\hat U}^I} + {\tilde \rho {\hat U}^u}} \right]$ uniformly on the nonempty compact set almost surely. In other words, the SAA technique enables us to use the channel samples collected at the beginning of a time slot to approximate the unknown channel coefficients over the time slot. For notation lightening, we write $x_m$ instead of $x(m)$ that represents a variable corresponding to the channel sample $\bm h_m$.
Recall that the variable $\omega_s(\bar t)$ will be kept unchanged over the time slot $\bar t$ and the beamformer $\bm g_{ij,s}(t)$ should be calculated at each minislot $t$, we can further consider (\[eq:original\_problem\]) as a global consensus problem, which can be effectively mitigated by an ADMM method. In this problem, $\omega_s(\bar t)$ is a global consensus variable that should be maintained in consensus for all $\bm h_m$, and $\bm g_{ij,sm}$ that is calculated based on $\bm h_m$ is a local variable.
The fundamental principle of an ADMM method is to impose augmented penalty terms characterizing global consensus constraints on the objective function of an optimization problem. In this way, the local variables can be driven into the global consensus while still attempting to maximize the objective function. Let ${\bm G}_{i,sm} = {\bm g}_{i,sm} {\bm g}_{i,sm}^{\rm H} \in {\mathbb R}^{JK \times JK}$, ${\bm H}_{i,sm} = {\bm h}_{i,sm} {\bm h}_{i,sm}^{\rm H} \in {\mathbb R}^{JK \times JK}$, where ${\bm g}_{i,sm} = [{\bm g}_{i1,sm};\ldots;{\bm g}_{iJ,sm}] \in {\mathbb C}^{JK \times 1}$ and ${\bm h}_{i,sm} = [{\bm h}_{i_1,sm};\ldots;{\bm h}_{i_J,sm}] \in {\mathbb C}^{JK \times 1}$. By applying the matrix property ${{{\bm G}_{i,sm}} = {{\bm g}_{i,sm}}{\bm g}_{i,sm}^{\rm H} \Leftrightarrow {{\bm G}_{i,sm}} \succeq 0,} \text{ } {{\rm rank}({{\bm G}_{i,sm}}) \le 1}$ and utilizing the conclusions of SAA and ADMM, we can approximate (\[eq:original\_problem\]) as the following problem at the beginning of the time slot $\bar t$
\[eq:SAA\_admm\_problem\] $$\begin{aligned}
{2}
& \mathop {{\rm{minimize}}}\limits_{\{ \omega _{sm},\omega_s(\bar t),{b_{i,sm}^u},{\bm G_{i,sm}}\}\hfill}
\sum\limits_{m = 1}^M {\left[ { - \frac{{U_m^I}}{M} - \frac{{\tilde \rho U_m^u}}{M}} \right]} + \nonumber \\
& \underbrace {\sum\limits_{m = 1}^M {\sum\limits_{s \in {{\mathcal S}^I}} {\left[ {{\psi _{sm}}\left( {{\omega _{sm}} - {\omega _s}(\bar t)} \right) + \frac{\mu }{2}\left\| {{\omega _{sm}} - \omega{_s}(\bar t)} \right\|_2^2} \right]} } }_{{\rm augmented} \text{ } {\rm penalty} \text{ } {\rm terms}} \\
& {\rm subject \text{ } to:} \nonumber \\
& P_{sm} \ge \pi_s, \forall s \in {\cal S}^I, m \in {\cal M} \\
& \sum\limits_{s \in {\cal S}^I} {{(1 + \alpha_g)\frac{\lambda_s^I}{\lambda_R}{\hat E}_{j}^I}} + \sum\limits_{s \in {{\cal S}^u}} {\sum\limits_{i \in {\cal I}_s^u} {b_{i,sm}^u {\rm tr}({{\bm Z}_j}{{\bm G}_{i,sm}})} } \le {E_j} \nonumber \\
& \qquad \qquad \forall j\in {\cal J},m\in {\cal M} \\
& \sum\limits_{s \in {{\cal S}^I}} {(1+\alpha_g) \omega _{s}(\bar t)} + W^u(\bm r_m) \le W, m \in {\cal M} \\
&{{\bm G}_{i,sm}} \succeq 0, \forall s \in {\cal S}^u, i \in {\cal I}_s^u, m \in {\cal M} \\
&{{\rm rank}({{\bm G}_{i,sm}}) \le 1}, \forall s \in {\cal S}^u, i \in {\cal I}_s^u, m \in {\cal M} \\
& b_{i,sm}^u \in \{0,1\}, \forall s \in {\cal S}^u, i \in {\cal I}_s^u, m \in {\cal M}\end{aligned}$$
where $\psi_{sm}$ is the Lagrangian multiplier, $\mu$ is a penalty coefficient, ${\bm Z}_j$ is a square matrix with $J \times J$ blocks, and each block in ${\bm Z}_j$ is a $K \times K$ matrix. In ${\bm Z}_j$, the block in the $j$-th row and $j$-th column is a $K \times K$ identity matrix, and all other blocks are zero matrices.
(\[eq:original\_problem\]) is now reduced to a deterministic single timescale problem (\[eq:SAA\_admm\_problem\]). What is more, (\[eq:SAA\_admm\_problem\]) can be split into $M$ separate problems that can be optimized in parallel as its objective function is separable. Thus, the following ADMM-based framework from (\[eq:arg\_lagarangian\]) to (\[eq:psi\_update\]) can be exploited to mitigate (\[eq:SAA\_admm\_problem\])
\[eq:arg\_lagarangian\] $$\begin{aligned}
{2}
& \left\{ {\begin{array}{*{20}{l}}
{\omega _{sm}^{(k + 1)},b_{i,sm}^{u(k+1)}}\\
{\bm G_{i,sm}^{(k + 1)}}
\end{array}} \right\} = \mathop {{\rm{argmin}}}\limits_{\left\{ {\scriptstyle\omega _{sm},b_{i,sm}^{u},\hfill\atop
\scriptstyle{\bm G_{i,sm}}\hfill} \right\}} \overline {\cal L} (\omega _{sm},{{\bm G}_{i,sm}}) \\
& {\rm subject \text{ } to:} \nonumber \\
& {\rm for} \text{ } {\rm the} \text{ } m-{\rm th} \text{ } {\rm sample}, (\ref{eq:SAA_admm_problem}b)-(\ref{eq:SAA_admm_problem}g) \text{ } \rm{are} \text{ } satisfied.\end{aligned}$$
$$\label{eq:omega_update}
\omega _s^{(k + 1)}(\bar t) = \frac{1}{M}\sum\limits_{m = 1}^M {\left( {\omega _{sm}^{(k + 1)} + \frac{1}{\mu }\psi _{sm}^{(k)}} \right)}, \text{ } \forall s\in {\cal S}^I$$
$$\label{eq:psi_update}
\psi _{sm}^{(k + 1)} = \psi _{sm}^{(k)} + \mu \left( {\omega _{sm}^{(k + 1)} - \omega _s^{(k + 1)}(\bar t)} \right), \text{ } \forall s\in {\cal S}^I$$
where the augmented partial Lagrangian function $$\label{eq:average_Lagrangian_func}
\begin{array}{l}
\bar{\mathcal{L}}(\omega _{sm},\bm {\bm G}_{i,sm}) = {- \frac{{U_m^{I(k)}}}{M} - \frac{{\tilde \rho U_m^{u(k)}}}{M} +} \\
{ \sum\limits_{s \in {{\cal S}^I}} {\left[ {\psi _{sm}^{(k)}\left( {\omega _{sm} - \omega _s^{(k)}(\bar t)} \right) + \frac{\mu }{2}{{\left\| {\omega _{sm} - \omega _s^{(k)}(\bar t)} \right\|}_2^2}} \right]} }
\end{array}$$
This ADMM-based framework can be executed on multiple processors. Each processor is responsible for optimizing (\[eq:arg\_lagarangian\]) and calculating (\[eq:psi\_update\]) with a global value as an input. (\[eq:omega\_update\]) is centrally updated in such a way that local variables converge to the global value, which is the solution of (\[eq:SAA\_admm\_problem\]). Unfortunately, (\[eq:arg\_lagarangian\]) is a mixed-integer non-convex optimization problem as there are zero-one variables, continuous variables and non-convex constraints in (\[eq:arg\_lagarangian\]). As a result, the optimization of (\[eq:arg\_lagarangian\]) is quite difficult. We next discuss how to handle this hard problem.
Alternative optimization
------------------------
In this subsection, we exploit a widely applied scheme, i.e., an alternative optimization scheme, to handle the mixed-integer non-convex optimization problem. Specifically, we first assume that continuous variables are known and attempt to mitigate a zero-one optimization problem. Given the zero-one variables, we then try to optimize a non-convex optimization problem. The process is alternatively conducted until convergence.
### URLLC device associations
Given continuous variables $\{{\bm G}_{i,sm}^{(k)}, \omega_{sm}^{(k)}\}$ at the $k$-th iteration, the association problem of URLLC devices in URLLC slices can take the following form
\[eq:arg\_lagarangian\_b\_isu\] $$\begin{aligned}
{2}
& \{ b_{i,sm}^{u(k+1)} \} = \mathop {{\rm{argmin}}}\limits_{\{b_{i,sm}^u\}} - \frac{{\tilde \rho U_m^{u(k)}}}{M} \\
& {\rm subject \text{ } to:} \nonumber \\
& {\rm for} \text{ } m, \text{ } (\ref{eq:SAA_admm_problem}c), (\ref{eq:SAA_admm_problem}d),(\ref{eq:SAA_admm_problem}g) \text{ } \rm{are} \text{ } satisfied.\end{aligned}$$
This problem is non-linear and hard to be handled. In theory, an exhaustive algorithm can obtain the optimal solution of (\[eq:arg\_lagarangian\_b\_isu\]). The computation complexity of this algorithm is $O(2^{N^u})$ that may be impractical in implementation. Therefore, a greedy scheme of the computational complexity $O(N^u)$, which is summarized as the following, is proposed to obtain $\{b_{i,sm}^{u(k+1)}\}$
a) initialize two device sets, i.e., candidate device set ${\mathcal{I}}^{u-} = {\mathcal I}^u$, association device set ${\mathcal I}^{u+} = \emptyset$.
b) select the device that maximizes $\frac{{\tilde \rho U_m^{u(k)}}}{M}$ from ${\mathcal{I}}^{u-}$, remove it from ${\mathcal{I}}^{u-}$, and add it to ${\mathcal{I}}^{u+}$. Given ${\mathcal{I}}^{u+}$, check the feasibility of (\[eq:arg\_lagarangian\_b\_isu\]). If (\[eq:arg\_lagarangian\_b\_isu\]) is feasible, then accept the device; otherwise, remove the device from ${\mathcal{I}}^{u+}$. Continue till ${\mathcal{I}}^{u-} = \emptyset$.
### joint bandwidth and beamforming optimization
Given the obtained $b_{i,sm}^{u(k+1)}$, (\[eq:SAA\_admm\_problem\]) will be reduced to the following joint bandwidth and beamforming problem.
\[eq:bandwidth\_beamforming\_problem\] $$\begin{aligned}
{2}
& \left\{
{\omega _{sm}^{(k + 1)}},
{\bm G_{i,sm}^{(k + 1)}}
\right\} = \mathop {{\rm{argmin}}}\limits_{\{\omega _{sm},{\bm G_{i,sm}}\}} \overline {\cal L} (\omega _{sm},{{\bm G}_{i,sm}}) \\
& {\rm subject \text{ } to:} \nonumber \\
& {\rm for} \text{ } m, (\ref{eq:SAA_admm_problem}b)-(\ref{eq:SAA_admm_problem}f) \text{ } \rm{are} \text{ } satisfied.\end{aligned}$$
In (\[eq:bandwidth\_beamforming\_problem\]), the low-rank constraint (\[eq:SAA\_admm\_problem\]f) is non-convex, and its objective function is not convex and even not quasi-convex w.r.t. ${\omega_{sm}}$, the tackling of which is quite tricky.
To tackle the non-convex low-rank constraint (\[eq:SAA\_admm\_problem\]e), we resort to the semidefinite relaxation technique. The primary procedures of SDR are i) directly drop the low-rank constraint; ii) solve the optimization problem without the low-rank constraint to obtain the solution; iii) owing to the relaxation, the obtained solution cannot satisfy the low-rank constraint in general. If it is, its principal component is the optimal solution to the problem. If not, then some manipulations such as randomization/scale [@ma2010semidefinite] are needed to perform on the solution to impose the low-rank constraint.
For the tricky objective function, we are reminded of the art of dealing with a non-convex function, i.e., study the structure of the function if it is non-convex. A crucial observation is that $P_{sm}$ is quasi-concave w.r.t. $\omega_{sm}$ although the objective function is not quasi-convex w.r.t. $\omega_{sm}$. Therefore, we resort to the Taylor expansion to approximate the tricky objection function.
The following analysis is built under the following two facts
- the value of the objective function of (\[eq:bandwidth\_beamforming\_problem\]) is mainly determined by that of ${\tilde P_{m}^{(k)}}$ (or $U_m^{I(k)}$);
- for all $s \in \mathcal{S}^u$, $m \in {\mathcal{M}}$, the solution $\omega_{sm}$ maximizing ${\tilde P_{m}^{(k)}}$ must locate in the range of $[\omega_{sm}^{lb}, S_{sm}^{\star}]$ shown in Fig. \[fig:fig\_quasi\_convex\_structure\], where $\omega_{sm}^{lb}$ denotes the lower bound of $\omega_{sm}$ satisfying the constraint (\[eq:SAA\_admm\_problem\]b), $S_{sm}^{\star}$ is the $\omega_{sm}$ maximizing $P_{sm}$, and the notation $P_{sm}|_{\omega_{sm}}$ is utilized to explicitly indicate that $P_{sm}$ is a function of $\omega_{sm}$.
Fact 1 holds because the linear terms w.r.t. $\omega_{sm}$ will donate little to the objective function as the consensus constraint is active. Besides, the quadratic terms pull local values towards the consensus; thus, they will also donate little to the objective function. Fact 2 holds because the total bandwidth is limited and shared. For example, given a value $\omega_{sm, 2} \in [S_{sm}^{\star}+\delta_{\omega}, W]$ with $\delta_{\omega}$ being a small positive constant, there must exist a value $\omega_{sm,1} \in [\omega_{sm}^{lb}, S_{sm}^{\star}]$ such that $P_{sm}|_{\omega_{sm,1}} = P_{sm}|_{\omega_{sm,2}}$. Thus, a small $\omega_{sm}$ will be preferred as it indicates that more bandwidth can be allocated to URLLC slices to further improve the objective function.
For all $s \in \mathcal{S}^I$, it can be proved that $P_{sm}$ is concave in the interval $(a_1, a_2]$ by evaluating the second-order derivative of $P_{sm}$. Therefore, we can leverage the $2^{\rm nd}$ degree Taylor expansion to approximate $P_{sm}$ in this interval. Considering that $P_{sm}$ is convex in the interval $[\omega_{sm}^{lb}, a_1]$, the $1^{\rm st}$ degree Taylor expansion is always leveraged to obtain the lower bound of $P_{sm}$. However, this interval is usually rather narrow, and the value of $P_{sm}$ in this interval is much lower than the value of that in the interval $(a_1, a_2]$. What is more, the error bound of the $1^{\rm st}$ degree Taylor expansion is greater than that of the $2^{\rm nd}$ expansion. Therefore, we explore the $2^{\rm nd}$ degree Taylor expansion to approximate $P_{sm}$ in the interval $[\omega_{sm}^{lb}, a_2]$. Fig. \[fig\_Taylor\_expansion\] shows an example of the $2^{\rm nd}$ degree Taylor expansion of $P_{sm}$. Given a local point $\bm \omega_{m}^{(k,q)}$ at the $q$-th iteration, the Taylor expansion of $-\tilde P_{m}^{(k)}$ at the local point can be given by $$\label{eq:P_sm_Taylor_expansion}
\begin{array}{*{20}{l}}
{ - {{\tilde P}_{m}^{(k)}} \approx - \tilde P_{m}^{(k,q)} - \nabla \tilde P_m^{(k,q)}{{({{\bm \omega} _m} - {\bm \omega} _m^{(k,q)})}^{\rm T}} - }\\
{\frac{1}{2}({{\bm \omega} _m} - {\bm \omega} _m^{(k,q)})H({\bm \omega} _{m}^{(k,q)}){{({\bm \omega _m} - {\bm \omega} _m^{(k,q)})}^{\rm T}} - R_2(\bm \omega_{m})}
\end{array}$$ where ${\bm \omega}_m = [\omega_{1m}, \ldots, \omega_{|{\mathcal S}^I|m}]$, $\nabla \tilde P_m^{(k,q)}$ is the gradient of $\tilde P_m^{(k)}$ over ${\bm \omega_m}$ at the local point ${\bm \omega_m^{(k,q)}}$ with $$\label{eq:first_order_Taylor_expansion}
\begin{array}{l}
\frac{{\partial P_{sm}^{(k)}}}{{\partial \omega _{sm}^{(k,q)}}} = \frac{{{\lambda _s^I}(1 + {\varpi _s}{\rho _o}){e^{ - {\varpi _s}{\sigma ^2}}}}}{{\sum\nolimits_{s \in {{\cal S}^I}} {{\lambda _s^I}} }} \times \\
\qquad \qquad \left[ {\frac{{3.5{y_{sm}}{z_s}\omega _{sm}^{2.5(k,q)}}}{{{{({y_{sm}}{z_s} + \omega _{sm}^{(k,q)})}^{4.5}}}} - \frac{{3.5{y_{sm}}\omega _{sm}^{2.5(k,q)}}}{{{{({y_{sm}} + \omega _{sm}^{(k,q)})}^{4.5}}}}} \right]
\end{array}$$ and $H({\bm \omega}_m^{(k,q)})$ is a Hessian matrix with $$\label{eq:second_order_Taylor_expansion}
\begin{array}{l}
\frac{{{\partial ^2}P_{sm}^{(k)}}}{{\partial \omega _{sm}^{2(k,q)}}} = \frac{{\lambda _s^I(1 + {\varpi _s}{\rho _o}){e^{ - {\varpi _s}{\sigma ^2}}}}}{{\sum\nolimits_{s \in {{\mathcal S}^I}} {\lambda _s^I} }}\left[ {\frac{{15.75y_{sm}^2z_s^2\omega _{sm}^{1.5(k,q)}}}{{{{({y_{sm}}{z_s} + \omega _{sm}^{(k,q)})}^{5.5}}}} - } \right.\\
\frac{{7{y_{sm}}{z_s}\omega _{sm}^{1.5(k,q)}}}{{{{({y_{sm}}{z_s} + \omega _{sm}^{(k,q)})}^{4.5}}}} + \frac{{7{y_{sm}}\omega _{sm}^{1.5(k,q)}}}{{{{({y_{sm}} + \omega _{sm}^{(k,q)})}^{4.5}}}} - \left. {\frac{{15.75y_{sm}^2\omega _{sm}^{1.5(k,q)}}}{{{{({y_{sm}} + \omega _{sm}^{(k,q)})}^{5.5}}}}} \right]
\end{array}$$ $$\label{eq:second_order_intersection_term}
\frac{{{\partial ^2}P_{sm}^{(k)}}}{{\partial \omega _{sm}^{(k,q)}\partial \omega _{s'm}^{(k,q)}}} = 0, \forall s \ne s'$$ $y_{sm} = \frac{{a{P_{nr,sm}}{P_{ne,sm}}{\lambda_s^I}}}{{3.5{\lambda _R}}}$, ${z_s} = \frac{{\theta_s^{th}}}{{ 1 + \theta_s^{th} }}$. Besides, we write $\omega_{sm}^{2.5(k,q)}$ rather than ${\left( {\omega _{sm}^{(k,q)}} \right)^{2.5}}$ for lightening the notation.
\[lemma:error\_bound\] Let the function $\tilde P_{m}^{(k)}: \mathbb{R}^{|{\mathcal S}^I|} \to \mathbb{R}$ be three times differentiable in a given interval $[\omega_{sm}^{lb}, S_{sm}^{\star}]$ for all $s \in {\mathcal{S}^I}$, then the error bound of the $2^{\rm nd}$ degree Taylor expansion of $\tilde P_{m}^{(k)}$ at the local point $\bm \omega_{sm}^{(k,q)}$ with $\omega_{sm}^{(k,q)} \in [\omega_{sm}^{lb}, S_{sm}^{\star}]$ is given by $$\label{eq:error_bound_lemma}
{R_2}({\bm \omega _m}) = \frac{1}{{3!}}{\left[ {\sum\limits_{s \in {{\cal S}^I}} {\left( {{\omega _{sm}} - \omega _{sm}^{(k,q)}} \right)\frac{\partial }{{\partial \omega _{sm}^{(k,q)}}}} } \right]^3} \tilde P_m^{(k)}{{|_{\bm \omega _m^{lb}}^{{\bm S_m^ \star }}}}$$ where $\tilde P_m^{(k)}{{|_{\bm \omega _m^{lb}}^{{\bm S_m^ \star }}}} = \max \left\{ {\tilde P_m^{(k)}{|_{\bm \omega _m^{lb}}},\tilde P_m^{(k)}{|_{\bm S_m^ \star }}} \right\}$, $\bm \omega_{m}^{lb} = [\omega_{1m}^{lb}, \ldots, \omega_{|\mathcal{S}^I|m}^{lb}]$ and $\bm S_{m}^{\star} = [S_{1m}^{\star}, \ldots, S_{|\mathcal{S}^I|m}^{\star}]$.
Please refer to Appendix C.
After conducting the $2^{\rm nd}$ degree Taylor approximation, the objective function becomes a convex function. Although the constraint (\[eq:SAA\_admm\_problem\]b) is $P_{sm}$ related, we need not to conduct the Taylor approximation on (\[eq:SAA\_admm\_problem\]b) as $P_{sm}$ is quasi-concave and unimodal. In fact, the probability constraint (\[eq:SAA\_admm\_problem\]b) is equivalent to the following inequality $$\label{eq:variable_range}
\omega_{sm}^{lb} \le \omega_{sm} \le \omega_{sm}^{ub}, \forall s \in \mathcal{S}^I$$ where $\omega_{sm}^{ub} \le W$ represents the upper bound of $\omega_{sm}$ satisfying (\[eq:SAA\_admm\_problem\]b).
Next, a low-complexity bisection-search-based scheme, the main procedures of which are described below, is developed to obtain $\omega_{sm}^{lb}$, $S_{sm}^{\star}$, and $\omega_{sm}^{ub}$
a) let the function $Q_{sm} = P_{sm} - \pi_s$. Perform the bisection search method [@yang2018three] on $Q_{sm} = 0$ to obtain $\omega_{sm}^{lb}$ and $\omega_{sm}^{ub}$ that are the two zero points of $Q_{sm}$.
b) with the obtained $\omega_{sm}^{lb}$ and $\omega_{sm}^{ub}$, find the maximum value $S_{sm}^{\star}$ of $P_{sm}$ using the bisection search method again.
According to the above analysis, at the $q$-th iteration, we can rewrite (\[eq:bandwidth\_beamforming\_problem\]) as
\[eq:SCA\_bandwidth\_beamforming\_problem\] $$\begin{aligned}
{2}
& \left\{
{\omega _{sm}^{(k + 1,q+1)}},{\bm G_{i,sm}^{(k + 1,q+1)}}
\right\} = \mathop {{\rm{argmin}}}\limits_{\{\omega _{sm},{\bm G_{i,sm}}\}} \bar {\cal L}^{(q)} (\omega _{sm},{\bm G_{i,sm}}) \\
& {\rm subject \text{ } to:} \nonumber \\
& {\rm for} \text{ } m, (\ref{eq:SAA_admm_problem}c)-(\ref{eq:SAA_admm_problem}e),(\ref{eq:variable_range}) \text{ } \rm{are} \text{ } satisfied.\end{aligned}$$
where $$\label{eq:average_Lagrangian_func}
\begin{array}{l}
\bar {\mathcal{L}}^{(q)}(\omega _{sm},\bm G_{i,sm}) = -{ \frac{1}{M}{\tilde P_{m}^{(k)}} - \frac{{\tilde \rho U_m^{u(k)}}}{M} +} \\
{ \sum\limits_{s \in {{\cal S}^I}} {\left[ {\psi _{sm}^{(k,q)}\left( {\omega _{sm} - \omega _s^{(k,q)}(\bar t)} \right) + \frac{\mu }{2}{{\left\| {\omega _{sm} - \omega _s^{(k,q)}(\bar t)} \right\|}_2^2}} \right]} } \nonumber
\end{array}$$
In (\[eq:SCA\_bandwidth\_beamforming\_problem\]), the objective function is convex, (\[eq:SAA\_admm\_problem\]c) is affine, and the constraint (\[eq:SAA\_admm\_problem\]d) can be proved to be convex w.r.t. both $\omega_{sm}$ and $\bm G_{i,sm}$ [@How2019yang]. Therefore, (\[eq:SCA\_bandwidth\_beamforming\_problem\]) is a convex problem that can be effectively mitigated by some standard convex optimization tools such as CVX [@CVX] and MOSEK [@MOSEK].
Then we can summarize the main steps of mitigating the problem (\[eq:SAA\_admm\_problem\]) in Algorithm \[alg1\].
**Initialization:** Randomly initialize $\bm G_{i,s}^{(0,0)}$, $\{\omega_{s}^{(0,0)}\}$, let $k_{\rm max}=250$, $q_{\rm max}=250$, $q = 0$, $k = 0$, and generate channel samples {$\bm H_{i,sm}$}. Given $\bm G_{i,sm}^{(k,q)}$, $\omega_{sm}^{(k,q)}$, call the greedy scheme to obtain $b_{i,sm}^{u(k,q+1)}$. Optimize (\[eq:SCA\_bandwidth\_beamforming\_problem\]) with obtained $b_{i,sm}^{u(k,q+1)}$ to achieve $\bm G_{i,sm}^{(k,q+1)}$ and $\omega_{sm}^{(k,q+1)}$. Update $q = q + 1$. Let $\omega_{sm}^{(k+1,q+1)} = \omega_{sm}^{(k,q+1)}$, update $\psi _{sm}^{(k + 1)}$ using (\[eq:psi\_update\]). Call (\[eq:omega\_update\]) to update $\omega_{s}^{(k+1)}(\bar t)$. Update $k = k + 1$.
\[lemma\_5\] For all $i\in\mathcal{I}_s^u$, $s \in {\mathcal{S}^u}$, and $m \in {\mathcal{M}}$, the obtained power matrix $\bm G_{i,sm}^{(k,q)}$ by Algorithm 1 at the $(k,q)$-th iteration satisfies the low-rank constraint, i.e., the SDR for the power matrix utilized in Algorithm 1 is tight.
Please refer to Appendix D.
Optimization of beamforming with system sensed channels
=======================================================
In Section V, we obtained a family of global consensus variables $\{\omega_{s}(\bar t)\}$ with the system generated channel samples. The time-varying actual channels may require the re-optimization of beamformers and device associations at each minislot. According to system sensed channels at each minislot, we next discuss how to calculate beamforms and device associations.
At each minislot $t$, given the global consensus variables $\{\omega_s(\bar t)\}$, the original problem (\[eq:original\_problem\]) will be reduced to the following problem
\[eq:mini\_time\_scale\_transformed\_problem\] $$\begin{aligned}
{2}
& \mathop {{\rm{maximize}}}\limits_{\{b_{i,s}^u(t), {\bm G}_{i,s}(t)\}} \text{ } {\tilde \rho} {{U}^u(t)} \\
& {\rm subject \text{ } to:} \nonumber \\
& \rm {constraints \text{ } (\ref{eq:RRH_energy}), (\ref{eq:total_bandwidth}),(\ref{eq:original_problem}b) \text{ } are \text{ } satisfied.}\end{aligned}$$
In (\[eq:mini\_time\_scale\_transformed\_problem\]), the channels are system sensed ones at $t$. According to the convexity analysis in Section V, (\[eq:mini\_time\_scale\_transformed\_problem\]) is a mixed-integer non-convex programming problem with positive semidefinite matrices, which is hard to be mitigated. Therefore, the alternative optimization scheme presented in subsection V-B can be leveraged to achieve the solutions $b_{i,s}^u(t)$ and $\bm G_{i,s}(t)$ of (\[eq:mini\_time\_scale\_transformed\_problem\]). Lemma \[lemma\_5\] indicates that the achieved ${\rm rank}(\bm G_{i,s}(t)) \le 1$. Thus, we can obtain the beamformers $\bm g_{i,s}(t)$ by performing the eigendecomposition on $\bm G_{i,s}(t)$. To sum up, over a time slot $\bar t$, the slice resource optimization algorithm designed for the RAN slicing system can be summarized as follows.
**Initialization:** $\{{\bm H}_{i,s}(t)\}$, $\forall i \in {{\mathcal I}^u}$, $s \in {{\mathcal S}^u}$, and let $P_{s}^{1} \in [0,1]$, $\mu_{a,s}^{1}=0$, $\forall s \in {{\mathcal S}^I}$. Call Algorithm \[alg1\] to obtain $\{\omega_s(\bar t)\}$, for all $s\in {\mathcal S}^I$. Given $\{\omega_s(\bar t)\}$, mitigate (\[eq:mini\_time\_scale\_transformed\_problem\]) by exploiting the alternative optimization scheme to obtain beamformers $\{{\bm g}_{i,s}(t)\}$ and URLLC device associations $b_{i,s}^u(t)$ for all $i \in {\mathcal I}_s^u$, $s \in {\mathcal S}^u$.
Simulation results
==================
In this section, we aim at evaluating the proposed algorithm via extensive simulations.
Comparison algorithms and parameter setting
-------------------------------------------
We compare the following three algorithms to verify the effectiveness of the proposed algorithm and to explain the impact of access control schemes on the RAN system performance intuitively i) SRO algorithm that adopts the unrestricted access control scheme; ii) SRO-ACB$_{\rm I}$ algorithm that utilizes the ACB access control scheme with $P_{ACB} = 0.9$; iii) SRO-ACB$_{\rm II}$ algorithm that adopts the ACB access control scheme with $P_{ACB} = 0.5$.
The parameter setting is as follows: RRHs and IoT devices are deployed following independent PPPs in a one km$^2$ area. URLLC devices are randomly and uniformly distributed in this area. There are three mIoT slices and two URLLC slices in the RAN slicing system. For the mIoT slices, set the new endogenous packet arrivals rate $\epsilon_{w,s}(t) = [1.5, 1.0, 0.5]$ packets/minislot, $\pi_s = 0.5$, $\forall s,t$. Let the path-loss component $\varphi = 4$, $L = 2000$ bits, $\tau = 1$ unit, $\sigma^2 = -90$ dBm, $\rho_o = -90$ dBm, ${\hat E}_{j}^I = 0.03$ mW, $\lambda_R = 3$ RRHs/km$^2$, $\{\lambda_{s}^I\} = [18000, 18000, 18000]$ IoT devices/km$^2$, $\{\gamma_s^{th}\} = \{5.8, 4.35, 2.9\}$ Kbits/minislot. For the URLLC slices, the transmit antenna gain at each RRH is set to be $5$ dB, and a log-normal shadowing path-loss model is leveraged to simulate the path-loss between an RRH and a URLLC device with the log-normal shadowing parameter being $10$ dB. A path-loss is computed by $h({\rm dB}) = 128.1 + 37.6\log_{10}d$, where $d$ (in km) represents the distance between a device and an RRH. Let $L_{i,s}^u = 160$ bits, $\sigma_{i,s}^2 = -100$ dBm, $\lambda_{s} = \lambda = 0.1$ packets/minislot, $\forall i,s$, $\{I_s^u\} = \{3, 5\}$ devices, and $\{D_s\} = \{1, 2\}$ milliseconds, $E_j = 3$ W, $\forall j$ [@tang2019service]. Other simulation parameters are shown in Table \[table\_simulation\_parameters\].
Para. Value Para. Value Para. Value
---------- ----------- ------------ -------------------- --------------- -----------------------
$J$ $3$ $K$ $2$ $\tilde \rho$ $1$
$\eta$ $100$ $T$ $60$ $W$ $60$ MHz
$M$ $100$ $\phi$ $1.5$ $a$ $0.18$ MHz
$\xi$ $54$ $\alpha_g$ $0.05$ $\kappa$ $5.12 \times 10^{-4}$
$\alpha$ $10^{-5}$ $\beta$ $2 \times 10^{-8}$ $\varsigma$ $2 \times 10^{-5}$
: Simulation parameters[]{data-label="table_simulation_parameters"}
Performance evaluation
----------------------
To evaluate the comparison algorithms, the following performance indicators are utilized i) RA success probability $P_s(t)$ that is computed using (\[eq:QoS\_prob\_analysis\]); ii) expected queue length per IoT device at minislot $t$, $E[Q_s(t)] = \mu_{a,s}(t)$; iii) total slice utility $\bar U$ that is the objective function of (\[eq:original\_problem\]).
We first evaluate the convergence of the proposed SRO algorithm. Fig. \[fig:fig\_convergence\] illustrates the convergence of SRO with $\Delta_{\omega} = {\rm{ }}\sum\nolimits_{s \in {{\mathcal S}^I}} {\left| {\omega _s^{(k + 1)}(\bar t) - \omega _s^{(k)}(\bar t)} \right|} $. It shows that SRO can converge after several iterations.
![The convergence curve of the proposed SRO algorithm.[]{data-label="fig:fig_convergence"}](Convergence_curve.eps){width="2.9in"}
We next plot the tendency of the RA success probability $P_s(t)$ and the corresponding expected queue length $E[Q_s(t)]$ during a time slot in Fig. \[fig:fig\_queueLen\_Pst\]. Fig. \[fig:fig\_queueLen\_Pst\](a) and \[fig:fig\_queueLen\_Pst\](c) show the tendency of $P_s(t)$ and $E[Q_s(t)]$ in the case of $\{\gamma_s^{th}\} = \{1.8, 1.35, 0.9\}$. Fig. \[fig:fig\_queueLen\_Pst\](b) and \[fig:fig\_queueLen\_Pst\](d) depict the tendency of $P_s(t)$ and $E[Q_s(t)]$ in the case $\{\gamma_s^{th}\} = \{5.8, 4.35, 2.9\}$.
From Fig. \[fig:fig\_queueLen\_Pst\], we obtain the following interesting conclusions: the queue of each IoT device is not stable when the queue serving rate $\gamma_s^{th}$ is small. In this case, the average queue length monotonously increases over minislot $t$. On the contrary, the queue of each IoT device is periodically flushed when a great queue serving rate is configured.
![Trends of $P_s(t)$ and $E[Q_s(t)]$. (a) and (c) are results of the parameter setting $\{\gamma_s^{th}\} = \{1.8, 1.35, 0.9\}$ Kbits/minislot; (b) and (d) correspond to the parameter setting $\{\gamma_s^{th}\} = \{5.8, 4.35, 2.9\}$ Kbits/minislot.[]{data-label="fig:fig_queueLen_Pst"}](Ps_EQs_vs_t.eps){width="3.1in"}
Let the IoT device intensity $\bm \lambda^I = [900 n, 900 n, 900 n]$ with $n \in \{6, 8, \ldots, 26\}$. Under the existence of both mIoT and URLLC slices, we plot trends of the total slice utility $\bar U$ and bursty URLLC slice utility $\bar U^u$ w.r.t. $n$ in Fig. \[fig:fig\_utility\_vs\_lamdba\_IoT\] to understand the impact of the mIoT slices on the performance of all comparison algorithms. In this figure, $B = [b_{11}^u, \ldots, b_{31}^u, b_{12}^u, \ldots, b_{52}^u]$, $\omega^I=[\omega_{SRO}^I, \omega_{ACB_{\rm I}}^I, \omega_{ACB_{\rm II}}^I]$ MHz with $\omega_{SRO}^I$, $\omega_{ACB_{\rm I}}^I$ and $\omega_{ACB_{\rm II}}^I$ representing the bandwidth allocated to mIoT slices by executing SRO, SRO-ACB$_{\rm I}$, and SRO-ACB$_{\rm II}$ algorithms, respectively, and $\bar U^I = [\bar U_{SRO}^I, \bar U_{ACB_{\rm I}}^I, \bar U_{ACB_{\rm II}}^I]$ with $\bar U_{SRO}^I$ denoting the achieved mIoT slice utility of SRO.
\
The following observations can be obtained from Fig. \[fig:fig\_utility\_vs\_lamdba\_IoT\]: i) when $n < 16$, all algorithms almost obtain the same total slice utility, and the obtained utilities are robust to the average number of IoT devices; ii) when $16 \le n \le 26$, the conclusion changes. For the SRO algorithm, its achieved $\bar U$ decreases with an increasing $n$ due to increasing interference. A great $n$, however, does not cause a significant decrease in the total slice utilities obtained by SRO-ACB$_{\rm I}$ and SRO-ACB$_{\rm II}$. Thanks to the exploration of an access control scheme, both SRO-ACB$_{\rm I}$ and SRO-ACB$_{\rm II}$ can achieve greater $\bar U$ than SRO. For example, compared with SRO, SRO-ACB$_{\rm II}$ improves $\bar U$ by $6.65\%$ when $n = 24$; iii) when $n = 26$, which means that the total average number of IoT devices reaches $70,200$ devices, the RAN slicing system fails to create and manage mIoT slices as the QoS requirements of mIoT slices serving such a massive average number of devices cannot be simultaneously satisfied. In this case, all system resources are allocated to URLLC slices, and the maximum bursty URLLC slice utility is obtained; iv) as mIoT slices and URLLC slices share the system resources, an increasing $n$ results in a decreasing bursty URLLC slice utility $\bar U^u$; Besides, it is interesting to find that the two access-control-based algorithms may not outperform SRO in terms of obtaining $\bar U^u$. It indicates that URLLC slices do not benefit from access control schemes of mIoT slices when changing $n$; v) the RAN slicing system can always accommodate the QoS requirements of all URLLC devices.
Next, to understand the impact of URLLC slices on the performance of all comparison algorithms, we plot the trends of the total slice utilities and the mIoT slice utilities obtained by all comparison algorithms w.r.t. URLLC packet arrival rate $\lambda$ with $\lambda = \{0.1,0.5,1.0,\ldots,4.5,5.0\}$ packets per unit time in Fig. \[fig:fig\_utility\_vs\_lamdba\_URLLC\]. Similarly, the following notations are involved in this figure: $\omega^u = [\omega_{SRO}^u, \omega_{ACB_{\rm I}}^u, \omega_{ACB_{\rm II}}^u]$, $\bar U^u = [\bar U_{SRO}^u, \bar U_{ACB_{\rm I}}^u, \bar U_{ACB_{\rm II}}^u]$ with $\omega_{SRO}^u$ and $\bar U_{SRO}^u$ denoting the bandwidth allocated to URLLC slices and the URLLC slice utility obtained by running the SRO algorithm, respectively.
\
From Fig. \[fig:fig\_utility\_vs\_lamdba\_URLLC\], we can observe that: i) the obtained $\bar U$ of all algorithms decrease with $\lambda$ mainly due to the decrease of the bursty URLLC slice utility. Two algorithms adopting the access control scheme always achieve greater utilities $\bar U$ than SRO. For example, when $\lambda = 5$, compared with the SRO algorithm, the obtained $\bar U$ of SRO-ACB$_{\rm II}$ is increased by $29.41\%$; ii) for all algorithms, the computed bandwidth for URLLC slices increases with an increasing $\lambda$. However, their obtained URLLC slice utilities $\bar U^u$ are reduced owing to the increase of energy consumption; iii) SRO-ACB$_{\rm II}$ may achieve greater $\bar U$ than SRO-ACB$_{\rm I}$ as a greater $\bar U^I$ is obtained by reducing the number of interfering IoT devices; iv) the obtained mIoT slice utilities $\bar U^I$ of SRO-ACB$_{\rm I}$ and SRO-ACB$_{\rm II}$ are robust to the URLLC packet arrival rate. The obtained $\bar U^I$ of SRO decreases with an increasing $\lambda$; v) an important observation is that the $\bar U^I$ of the access-control-based SRO-ACB$_{\rm I}$ algorithm is $1.65$ times that of the SRO algorithm when $\lambda = 5$. It explicitly reflects that mIoT slices can still benefit from access control schemes even though $\lambda$ is changed.
Figs. \[fig:fig\_utility\_vs\_lamdba\_IoT\] and \[fig:fig\_utility\_vs\_lamdba\_URLLC\] illustrate the situation of a given total system bandwidth. We next change the total bandwidth $W$ and plot its impact on the obtained total slice utilities of all algorithms in Fig. \[fig:fig\_total\_utility\_vs\_bandwidth\].
![Trend of achieved total slice utility vs. system bandwidth.[]{data-label="fig:fig_total_utility_vs_bandwidth"}](total_slice_utility_vs_bandwidth.eps){width="3.1in"}
The following conclusions can be obtained from this figure i) when $W=45$ MHz, the QoS requirements of all IoT devices cannot be simultaneously satisfied. As a result, the total bandwidth is allocated to URLLC slices; ii) when $W$ locates in the range of $(45, 55]$ MHz, the achieved total slice utilities $\bar U$ of SRO and SRO-ACB$_{\rm I}$ increase with $W$. Owing to the utilization of the access control scheme, SRO-ACB$_{\rm I}$ and SRO-ACB$_{\rm II}$ obtain higher $\bar U$ than SRO. For example, compared with the SRO algorithm, the SRO-ACB$_{\rm II}$ algorithm improves the achieved total slice utility by $6.66\%$ when $W = 50$ MHz. iii) when $W > 55$ MHz, all algorithms cannot remarkably improve $\bar U$.
At last, we discuss other crucial parameters’ impact on the performance of the comparison algorithms. We reconfigure $\{\gamma_s^{th}\}$ of mIoT slices as $\gamma_1^{th} = 3.6 m$, $\gamma_2^{th} = 2.7 m$ and $\gamma_3^{th} = 1.8 m$ Kbits/minislot with $m \in \{1.5, 1.6, \ldots, 2.1\}$ and $\{D_s\}$ of URLLC slices as $D_1 = 0.00025 d$ second and $D_2 = 0.0005 d$ second with $d \in \{2, 3, \ldots, 10\}$. The impact of QoS requirements of network slices on the total slice utility is plotted in Figs. \[fig:fig\_total\_utility\_vs\_gammath\] and \[fig:fig\_total\_utility\_vs\_Ds\]. The impact of energy efficiency coefficient $\eta$ is plotted in Fig. \[fig:fig\_total\_utility\_vs\_eta\]. In this figure, we denote the energy consumption of RRHs of all algorithms by $E^u = [E_{SRO}^u, E_{ACB_{\rm I}}^u, E_{ACB_{\rm II}}^u]$ with $E_{SRO}^u = \sum\nolimits_{t = 1}^T {\sum\limits_{s \in {{\mathcal S}^u}} {\sum\limits_{i \in {\mathcal I}_s^u} {b_{i,s}^u{\rm tr}({\bm G_{i,s}})} } } $.
![Trend of achieved total slice utility vs. $m$.[]{data-label="fig:fig_total_utility_vs_gammath"}](total_slice_utility_vs_Csth.eps){width="3.1in"}
![Trend of achieved total slice utility vs. $d$.[]{data-label="fig:fig_total_utility_vs_Ds"}](total_slice_utility_vs_Ds.eps){width="3.1in"}
![Trend of achieved total slice utility vs. $\eta$.[]{data-label="fig:fig_total_utility_vs_eta"}](total_slice_utility_vs_eta.eps){width="3.1in"}
From these figures, the following observations can be achieved: i) the obtained utilities $\bar U$ of all algorithms decrease with an increasing $m$. This is because a great $m$ indicates that the accumulated IoT packets in the queue of each IoT device can be quickly emptied, and then a small $P_s(t)$ is obtained; ii) a great $D_s$ will reduce RRHs’ energy consumption. However, it also reduces the URLLC slice gain. Then, it may be hard to conclude the trend of $\bar U^u$ w.r.t. $D_s$ as the energy efficiency coefficient $\eta$ significantly affects the value of $\bar U^u$; iii) it is also uneasy to conclude the trend of $\bar U^u$ w.r.t. $\eta$. An increasing $\eta$ causes a decrease of RRHs’ energy consumption. Yet, the value of $\bar U^u$ is determined by the multiplier of $\eta$ and $E^u$; iv) the SRO-ACB$_{\rm II}$ algorithm may perform better than the SRO algorithm. However, the performance of the other access-control-based algorithm, SRO-ACB$_{\rm I}$, is slightly worse than the SRO algorithm. Besides, it cannot ensure that the $\bar U^I$ obtained by the access-control-based algorithms are always higher than that of SRO. At sometimes, access control schemes may drag down the utility of the mIoT service.
To sum up, in the case of service multiplexing, RA control schemes for alleviating signal interference and enhancing mIoT slice utility may be preferred for mIoT slices. However, considering both the CAPEX and the improvement of slice utility, RA control schemes should be carefully designed and employed because some RA control schemes may worsen the mIoT and even the total slice utilities.
Conclusion
==========
In this paper, we revisited the frame and minislot structure of a RAN slicing system to admit more IoT devices and proposed a queue evolution model to analyze the RACH of a randomly chosen IoT device. Based on the analysis result, we derived the closed-form expression of the RA success probabilities of the device with unrestricted access control scheme and ACB access control scheme. Next, we formulated the RAN slicing for mIoT and bursty URLLC service multiplexing as an optimization problem to optimally orchestrating RAN resources for mIoT slices and URLLC slices, and efficient mechanisms such as SAA and ADMM were exploited to mitigate the optimization problem. Simulation results showed that RA control schemes should be carefully designed and employed in the case of service multiplexing.
Proof of Lemma 1
----------------
For the origin RRH, the LT of its aggregate interference from interfering IoT devices in $s \in {\mathcal{S}^I}$ can be derived as $$\label{eq:LT_intra_interference}
\begin{array}{*{20}{l}}
{{{\mathcal L}_{{{\mathcal I}_s}(t)}}({\varpi _s}) = {E_{{{\mathcal I}_s}(t)}}\left[ {{e^{ - {\varpi _s}{{\mathcal I}_s}(t)}}} \right]}\\
= {E_{{{\mathcal I}_s}(t)}}\left[ {\exp ( { - {\varpi _s}\sum\limits_{{u_{m,s}} \in {\Phi _s}\backslash \{ o\} } {{\mathbbm 1}({p_m}||{d_m}|{|^{ - \varphi }} = {\rho _o}) \times } } } \right.\\
\text{ } \left. {\left. {{\mathbbm 1}({N_{a,s}}(t) > 0){\mathbbm 1}({f_m} = {f_o}){\rho _o}{h_m}} \right)} \right]
\\
\mathop = \limits^{(a)} {E_{{\Phi _s}}}\left[ {{\prod _{{u_{m,s}} \in {\Phi _s}\backslash \{ o\} }}{E_{{h_m}}}\left[ {\exp \left( { - {\varpi _s} \times } \right.} \right.} \right.\\
\text{ } \left. {\left. {{\mathbbm 1}({p_m}||{u_{m,s}}|{|^{ - \varphi }} = {\rho _o}){\mathbbm 1}({N_{a,s}}(t) > 0){\mathbbm 1}({f_m} = {f_o}){\rho _o}{h_m}} \right)} \right] \\
{\mathop = \limits^{(b)} \sum\limits_{n = 0}^\infty {P\{ |{Z_s}| = n\} \prod\limits_{{u_{m,s}} \in {Z_s}} {{E_{{h_m}}}\left[ {{e^{ - {\varpi _s}{\rho _o}{h_m}}}} \right]} } }\\
{\mathop = \limits^{(c)} P\{ |{Z_s}| = 0\} + \sum\limits_{n = 1}^\infty {P\{ |{Z_s}| = n\} {{\left( {\frac{1}{{1 + {\varpi _s}{\rho _o}}}} \right)}^n}} }\\
{\mathop = \limits^{(d)} {{\tilde P}_{{X_s}}}\{ {X_s} = 1\} + \left\{ {\sum\limits_{n' = 0}^\infty {{{\tilde P}_{{X_s}}}\{ {X_s} = n'\} {{\left( {\frac{1}{{1 + {\varpi _s}{\rho _o}}}} \right)}^{n'}}} - } \right.}\\
\text{ } {\left. {\sum\limits_{n' = 0}^1 {{{\tilde P}_{{X_s}}}\{ {X_s} = n'\} {{\left( {\frac{1}{{1 + {\varpi _s}{\rho _o}}}} \right)}^{n'}}} } \right\}(1 + {\varpi _s}{\rho _o})}
\end{array}$$ where $\varpi_s = \frac{\theta_s^{th}}{\rho_o}$, $Z_s$ denotes the set of interfering IoT devices in mIoT slice $s$, $X_s$ represents the number of active IoT devices associated with the origin RRH in $s$. According to the conclusion of Lemma 1 in [@yu2013downlink], the probability mass function (PMF) ${\tilde P}_{X_s} \{ X_s = n'\}$ can be written as $$\label{eq:intra_interference_proba}
{\tilde P}_{X_s}\{ X_s = n'\} = \frac{{{{3.5}^{3.5}}\Gamma (n' + 3.5){{(\frac{{{P_{nr,s}}(t)P_{ne,s}(t){\lambda _{s}^I}}}{{{\lambda _R}\xi F_s}})}^{n' }}}}{{\Gamma (3.5)(n')!{{(\frac{{P_{nr,s}}(t){P_{ne,s}(t){\lambda _{s}^I}}}{{{\lambda _R} \xi F_s}} + 3.5)}^{n' + 3.5}}}}$$ with $\Gamma(\cdot)$ being the gamma function. Besides, in (\[eq:LT\_intra\_interference\]), (a) follows from the i.i.d. distribution of $h_m$ and its further independence from the Poisson point process $\Phi_s$; (b) follows from the expectation of a discrete random variable; (c) follows from the LT over $h_m$; (d) follows from the fact that the number of active IoT device in a specific Voronoi cell is one more than the number of active interfering IoT devices in this cell.
From (\[eq:intra\_interference\_proba\]), we can deduce that $X_s$ ($s\in \mathcal{S}^I$) is a gamma¨CPoisson random variable with $X_s \sim {\rm gamma-Poisson}(\alpha_s, 3.5)$ and $\alpha_s = \frac{{{P_{nr,s}}(t){P_{ne,s}}(t){\lambda _{s}^I}}}{{3.5{\lambda _R}\xi F_s}}$.
For a gamma¨CPoisson random variable $X_s \sim {\rm gamma-Poisson}(\alpha, \beta)$, the following expression holds: $E[e^{X_s}] = (1+\alpha-\alpha e)^{-\beta}$. Thus, we can rewrite (\[eq:LT\_intra\_interference\]) as (\[eq:LT\_interference\_expression\]). This completes the proof.
Proof of Lemma 2
----------------
As new endogenous packet arrivals in any IoT device at each minislot $t$ is modelled as a Poisson distribution, the departure process of packets can be regarded as an approximated thinning process of new arrivals, where the thinning factor is related to the RA success probability. The number of accumulated packets in the queue of any IoT device can then be approximated as a Poisson distribution with intensity $\mu_{a,s}^{t}$ ($s\in \mathcal
S^I$) after the thinning process in a specific minislot $t$ ($t > 1$) [@jiang2018random]. Thus, we can derive the expression of $\mu_{a,s}^{t}$ ($t > 1$) via combining with the following facts
- [**Fact 1:**]{} the accumulated packets during the $t-1$-th minislot will contribute to the accumulated packets at the $m$-th minislot.
- [**Fact 2:**]{} the arrival packets during the $t-1$-th minislot will also contribute to the accumulated packets in the queue of an IoT device at the $m$-th minislot.
- [**Fact 3:**]{} an IoT device can send packets only if its preamble is successfully transmitted.
- [**Fact 4:**]{} at the same minislot, the new packet arrival process and the packet accumulated process are independent.
Similar as the Theorem 2 in [@jiang2018random], we can infer that at the $2^{\rm nd}$ minislot, for all $s \in \mathcal{S}^I$, $\mu_{a,s}^{2}$ depends on the intensity of new packet arrivals $\mu_{w,s}^{1}$ and the probability $P_{s}^{1}$ of a randomly selected IoT device at the $1^{\rm st}$ minislot, which is given by $$\label{eq:mu_acc_2_1}
\begin{array}{l}
\mu _{a,s}^2 = \mu _{w,s}^1 - x_s P_s^1\left( {1 - {e^{ - \mu _{w,s}^1}}} \right)
\end{array}$$
The detailed proof of (\[eq:mu\_acc\_2\_1\]) is omitted for brevity, and a similar proof can be found in the proof section of Theorem 2 in [@jiang2018random].
Considering that $\mu_{a,s}(t)$ is non-negative at each minislot $t$, we have $$\label{eq:mu_acc_2}
\begin{array}{*{20}{l}}
{\mu _{a,s}^2 = }
\end{array}{\left[ {\mu _{w,s}^1 - {x_s}P_s^1\left( {1 - {e^{ - \mu _{w,s}^1}}} \right)} \right]^ + }$$
Then, according to the definition of non-empty probability and the Poisson approximation, the non-empty probability of a randomly selected IoT device in mIoT slice $s \in \mathcal{S}^I$ at the $2^{\rm nd}$ minislot can be approximated as $$\label{eq:non_empty_prob_2_expression}
P_{ne,s}^2 = 1 - {e^{ - \mu _{a,s}^2}}$$
At the $3^{\rm rd}$ minislot, the intensity of accumulated data packets in the queue of a randomly selected IoT device can be derived as the following $$\label{eq:mu_acc_3}
\begin{array}{l}
\mu _{a,s}^3 = P_s^2\left( {\sum\nolimits_{n = 1}^\infty {( {{{[n - x_s]}^ + } \sum\nolimits_{z = 0}^n {{P_{N_{w,s}^2}}(z){P_{N_{a,s}^2}}(n - z)} } )} } \right)\\
\quad + (1 - P_s^2)\left( {\sum\nolimits_{n = 1}^\infty {n \sum\nolimits_{z = 0}^n {{P_{N_{w,s}^2}}(z){P_{N_{a,s}^2}}(n - z)} } } \right)\\
\mathop = \limits^{(a)} P_s^2\left [ {\sum\nolimits_{n = 1}^\infty {\sum\nolimits_{z = 0}^n {\frac{{{{\left( {\mu _{w,s}^2} \right)}^z}{e^{ - \mu _{w,s}^2}}}}{{z!}}\frac{{{{\left( {\mu _{a,s}^2} \right)}^{n - z}}{e^{ - \mu _{a,s}^2}}}}{{(n - z)!}}} \times n} - } \right.\\
\quad \left. {x_s\sum\nolimits_{n = 1}^\infty {\sum\nolimits_{z = 0}^n {\frac{{{{\left( {\mu _{w,s}^2} \right)}^z}{e^{ - \mu _{w,s}^2}}}}{{z!}}\frac{{{{\left( {\mu _{a,s}^2} \right)}^{n - z}}{e^{ - \mu _{a,s}^2}}}}{{(n - z)!}}} } } \right ]^+ + \\
\quad (1 - P_s^2)\sum\nolimits_{n = 1}^\infty {\sum\nolimits_{z = 0}^n {\frac{{{{\left( {\mu _{w,s}^2} \right)}^z}{e^{ - \mu _{w,s}^2}}}}{{z!}}\frac{{{{\left( {\mu _{a,s}^2} \right)}^{n - z}}{e^{ - \mu _{a,s}^2}}}}{{(n - z)!}}} \times n} \\
\mathop = \limits^{(b)} \left [ \mu _{w,s}^2 + \mu _{a,s}^2 - x_sP_s^2\left( {1 - {e^{ - \mu _{w,s}^2 - \mu _{a,s}^2}}} \right) \right ]^+
\end{array}$$ where $P_{N_{w,s}^2}$ and $P_{N_{a,s}^2}$ represent the PMFs of new arrival packets and accumulated packets at the $2^{\rm nd}$ minislot, respectively. Besides, (a) follows from the fact: for any two independent Poisson distribution $\Phi_{X_1}$ and $\Phi_{X_2}$, $P_{X_1,X_2}({X_1} + {X_2} = x) = \sum\nolimits_{y = 0}^x {P_{X_1}({X_1} = y)P_{X_2}(X_2 = x - y)} $; (b) holds as $\Phi_{X_1,X_2}$ is a two dimensional Poisson distribution with an intensity $\lambda_{X_1} + \lambda_{X_2}$, and $\sum\nolimits_{x = 1}^\infty {P_{X_1, X_2}({X_1} + {X_2} = x)} = 1 - P_{X_1,X_2}({X_1} + {X_2} = 0)$.
Similarly, we have $$\label{eq:non_empty_prob_3}
P_{ne,s}^3 = 1 - {e^{ - \mu _{a,s}^3}}$$
When $t > 3$, since the accumulated packets evolution model of the queue of any IoT device is the similar as that at $t = 3$, we can directly extend the conclusion obtained at $t = 3$ to that at $t > 3$. Therefore, we can obtain the closed-form expression of $\mu_{a,s}^t$ for all $s \in \mathcal{S}^I$ at $t > 1$ with $$\label{eq:mu_acc_m_expression_proof}
\mu _{a,s}^t = \left [ \mu _{w,s}^{t - 1} + \mu _{a,s}^{t - 1} - x_sP_s^{t - 1}\left( {1 - {e^{ - \mu _{w,s}^{t - 1} - \mu _{a,s}^{t - 1}}}} \right) \right ]^+$$ and $$\label{eq:non_empty_prob_m_expression_proof}
P_{ne,s}^t = 1 - {e^{ - \mu _{a,s}^t}}$$ This completes the proof.
Proof of Lemma 4
----------------
The $2^{\rm nd}$ degree Taylor expansion of $\tilde P_{m}^{(k)}$ at the local point $\bm \omega_{m}^{(k,q)}$ is $$\tilde P_{2,m}^{(k)} = \sum\limits_{j = 0}^2 {\frac{1}{{j!}}{{\left[ {\sum\limits_{s \in {\mathcal S}^I} {\left( {{\omega _{sm}} - \omega _{sm}^{(k,q)}} \right)\frac{\partial }{{\partial \omega _{sm}^{(k,q)}}}} } \right]}^j}{\tilde P_m^{(k)}}|_{\bm \omega_m^{(k,q)}}}$$
The $3^{\rm rd}$ degree Taylor expansion of $\tilde P_{m}^{(k)}$ at $\bm \omega_{m}^{(k,q)}$ must be more accurate than $\tilde P_{2,m}^{(k)}$ with $$\tilde P_{3,m}^{(k)} = \tilde P_{2,m}^{(k)} + {\frac{1}{{3!}}{{\left[ {\sum\limits_{s \in {\mathcal S}^I} {\left( {{\omega _{sm}} - \omega _{sm}^{(k,q)}} \right)\frac{\partial }{{\partial \omega _{sm}^{(k,q)}}}} } \right]}^3}{\tilde P_m^{(k)}}|_{\bm \omega_m^{(k,q)}}}$$
Since the error of $\tilde P_{2,m}^{(k)}$ is no greater than the maximum difference between $\tilde P_{3,m}^{(k)}$ and $\tilde P_{2,m}^{(k)}$, we have $$\label{eq:error_R2}
R_{2}(\bm \omega_m) = \max \{ {\frac{1}{{3!}}{{[ {\sum\limits_{s \in {\mathcal S}^I} {( {{\omega _{sm}} - \omega _{sm}^{(k,q)}} )\frac{\partial }{{\partial \omega _{sm}^{(k,q)}}}} } ]}^3}{\tilde P_m^{(k)}}|_{\bm \omega_m^{(k,q)}}} \}$$
In (\[eq:error\_R2\]), $\bm \omega_{m}^{(k,q)}$ is a constant vector, the $\max$ operation will not affect the constant vector and the vector $\bm \omega_{m}$. For any $s\in \mathcal{S}^I$, the maximum value obtainable by $\frac{{{\partial ^3}{\tilde P_m^{(k)}}|_{\bm \omega_m^{(k,q)}}}}{{\partial \omega _{sm}^{3(k,q)}}}$ will not exceed the greatest value of that derivative in the interval $[\omega_{sm}^{lb}, S_{sm}^{\star}]$. Additionally, the maximum value of $\frac{{{\partial ^3}{P_m}|_{\bm \omega_m^{(k,q)}}}}{{\partial \omega _{sm}^{3(k,q)}}}$ will generally occur at one of the endpoints of the interval $[\omega_{sm}^{lb}, S_{sm}^{\star}]$. Therefore, we obtain (\[eq:error\_bound\_lemma\]). This completes the proof.
Proof of Lemma 5
----------------
For all $i \in {\mathcal{I}^u}$, $s \in {\mathcal{S}^u}$, $m \in {\mathcal{M}}$, a feasible way of proving that ${\rm rank}(\bm G_{i,sm}) \le 1$ is to utilize the Lagrange method. However, owing to the complicated expression of $W^{u}(\bm r_m)$ w.r.t. $\bm G_{i,sm}$ it will be uneasy to do that. Fortunately, we find that the proof can be conducted if a family of auxiliary variables is introduced.
For the constraint (\[eq:SAA\_admm\_problem\]d), if we introduce the auxiliary variables $\{\nu_{i,sm}\}$ and let $$\label{eq:linear_snr_tau_ism}
\frac{{\rm{tr}}({\bm H_{i,sm}}{{\bm G}_{i,sm}})}{\phi \sigma _{i,s}^2} \ge \nu _{i,sm},\forall i \in {\mathcal I}_s^u,s \in {{\mathcal S}^u},m \in {\mathcal{M}}$$ then (\[eq:SAA\_admm\_problem\]d) is equivalent to $$\label{eq:}
\sum\limits_{s \in {{\cal S}^I}} {(1+\alpha_g)\omega _{sm}(\bar t)} + W^u(\bm f_m) \le W, \text{ } {\rm and} \text{ } {\rm (\ref{eq:linear_snr_tau_ism})}$$ where $\bm f_m = \{f_{i,sm};i \in {\mathcal{I}_s^u}, s \in {\mathcal{S}^u}\}$ and $$\begin{array}{*{20}{l}}
{{f_{i,sm}} = \frac{{L_{i,s}^u}}{{{{\log }_2}(1 + {\nu _{i,sm}})}} + \frac{{{{({Q^{ - 1}}(\beta))}^2}}}{{2\log _2^2(1 + {\nu _{i,sm}})}}}\\
{\qquad \quad + \frac{{{{({Q^{ - 1}}(\beta))}^2}}}{{2\log _2^2(1 + {\nu _{i,sm}})}}\sqrt {1 + \frac{{4L_{i,s}^u{{\log }_2}(1 + {\nu _{i,sm}})}}{{{{({Q^{ - 1}}(\beta))}^2}}}} }
\end{array}$$
We omit the proof of the equivalence as a similar proof can be found in the proof section of constraints’ equivalence in [@How2019yang].
The partial Lagrangian function of (\[eq:SCA\_bandwidth\_beamforming\_problem\]) can be written as $$\label{eq:lagrangian_func}
\begin{array}{l}
L( \ldots ) = \sum\limits_{s \in {{\mathcal S}^u}} {\sum\limits_{i \in {\mathcal I}_s^u} {\left[ {\frac{{\tilde \rho \eta }}{M}{\rm{tr}}({\bm G_{i,sm}}) - {{\bar \mu }_{i,sm}}\frac{{{\rm{tr}}({\bm H_{i,sm}}{\bm G_{i,sm}})}}{{\phi \sigma _{i,s}^2}} + } \right.} } \\
\qquad \qquad \left. {\sum\limits_{j \in {\mathcal J}} {{{\bar \lambda }_{jm}}{\rm{tr}}(b_{i,sm}^{u(k,q)}{\bm Z_j}{\bm G_{i,sm}})} - {{\bm {\bar X}}_{i,sm}}{\bm G_{i,sm}}} \right]
\end{array}$$ where $\bar \lambda_{jm}$, $\bar \mu_{i,sm}$, and $\bar {\bm X}_{i,sm}$ are Lagrangian multipliers corresponding to constraints (\[eq:SAA\_admm\_problem\]c), (\[eq:linear\_snr\_tau\_ism\]) and (\[eq:SAA\_admm\_problem\]e). Besides, only terms related to ${\bm G}_{i,sm}$ are included in this function for brevity. According to the Karush-Kuhn-Tucker (KKT) conditions, the necessary condition for obtaining the optimal matrix power at the $(k,q)$-th iteration ${\bm G}_{i,sm}^{(k,q)\star}$ is given by $$\label{eq:KKT_condition_2}
\begin{array}{l}
\frac{{\partial L( \ldots )}}{{\partial {\bm G}_{i,sm}^{(k,q)\star}}} = \frac{{\tilde \rho \eta }}{M}{\bm I_{i,sm}} + \frac{\bar \mu_{i,sm}{{\bm H_{i,sm}}}}{{\phi \sigma _{i,s}^2}} - \\
\qquad \qquad \text{ } \sum\limits_{j \in {\mathcal J}} {{{\bar \lambda }_{jm}}{b_{i,sm}^{u(k,q)}\bm Z_j}} - {{\bm X}_{i,sm}} = 0
\end{array}$$ where ${\bm I_{i,sm}} \in \mathbb{R}^{JK \times JK}$ is an identity matrix.
Then, we can conclude that ${\rm rank}({\bm X}_{i,sm}) \ge JK - 1$. The reasons are i) ${{\bar \lambda }_{jm}}$, $b_{i,sm}^{u(k,q)}$, and $\bar \mu _{i,sm}$ are nonnegative and the matrix $\bm I_{i,sm}$ is full rank; ii) ${\rm rank}(\bm H_{i,sm}) \le 1$.
Next, according to the complementary slackness condition, we have $$\label{eq:complementary_slack_condition_2}
{{\bm X}_{i,sm}}{\bm G}_{i,sm}^{(k,q)\star} = 0$$
Based on (\[eq:complementary\_slack\_condition\_2\]) and the rank result of $\bm X_{i,sm}$, we can conclude that ${\rm rank}(\bm G_{i,sm}^{(k,q)\star}) \le 1$. This completes the proof.
[^1]: P. Yang and T. Q. S. Quek are with the Information Systems Technology and Design, Singapore University of Technology and Design, 487372 Singapore.
X. Xi, J. Chen, and X. Cao are with the School of Electronic and Information Engineering, Beihang University, Beijing 100083, China, and also with the Key Laboratory of Advanced Technology, Near Space Information System (Beihang University), Ministry of Industry and Information Technology of China, Beijing 100083, China.
D. Wu is with the Department of Electrical and Computer Engineering, University of Florida, Gainesville FL 32611 USA.
[^2]: We do not show all channels in this figure as the detailed research of the physical layer supporting the mIoT service is out of the scope of this paper.
[^3]: Owing to the truncated channel inversion power control, not all of the IoT devices in mIoT slices can communicate in the uplink when the cutoff threshold is relatively high [@elsawy2014stochastic]. However, this paper assumes that the transmit power of each IoT device is large enough such that the IoT device will not experience preamble outage resulting from the insufficient power.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
F. Krauss$^a$, R. Kuhn$^{b,c}$ and G. Soff$^c$\
$^a$Cavendish Laboratory, University of Cambridge, Cambridge CB3 0HE, U.K.\
$^b$Max Planck Institut für Physik Komplexer Systeme, 01187 Dresden, Germany\
$^c$Institut für Theoretische Physik, TU Dresden, 01062 Dresden, Germany\
E-mail:
title: |
AMEGIC++ 1.0\
A Matrix Element Generator In C++
---
Program Summary {#program-summary .unnumbered}
===============
[*Title of the program :*]{}\
[*Program obtainable from :*]{} authors upon request\
[*Licensing provisions :*]{} none\
[*Operating systems under which the program has been tested :*]{} UNIX, LINUX, VMS\
[*Programming language :*]{} C++\
[*Separate documentation available :*]{} no\
[*Nature of the physical problem:* ]{}\
The theoretical description of multi particle production, even at the tree–level, suffers from two problems :
1. The rapidly increasing number of amplitudes forbids the traditional method of summing and squaring individual Feynman amplitudes by means of the completeness relations for spinors and polarization vectors. Instead, the helicity method is employed, translating the amplitudes into complex numbers. Still, the helicity amplitudes for a large number of diagrams have to be constructed which itself is a formidable task.
2. The complex structure of the high–dimensional phase space imperatively requires using Monte Carlo methods for its evaluation. Here, efficiency is of paramount importance, and one has to employ non–flat phase space measures which must be optimized for the process under consideration and its specific singularity structure in phase space.\
[*Method of solution:*]{}\
Automatic generation of helicity amplitudes related to Feynman diagrams describing the process at the tree–level. Translating of the amplitudes into character strings and storing in libraries for increased efficiency during evaluation, i.e. phase space integration. Integration by means of multichannel methods with specific channels which are constructed from the Feynman diagrams.\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the collision of two particles with equal masses moving in the equatorial plane near horizon of the rotating regular Ayón-Beato-García (ABG) black hole (BH) and calculate the center-of-mass (CM) energy for the colliding particles for both extremal and non-extremal cases. It turns out that CM energy depends not only on rotation parameter $a$ but also on charge $Q$. Particularly for the extremal rotating regular ABG BH, CM energy of two colliding particles could be arbitrarily high for critical angular momentum of particles. Furthermore, we also show that, for a non-extremal BH, there exist a finite upper limit of CM energy, which changes with charge $Q$. A comparison, with Kerr and Kerr-Newman black holes, is included.'
author:
- 'Sushant G. Ghosh'
- Pankaj Sheoran
- Muhammed Amir
title: 'Rotating Ayón-Beato-García black hole as a particle accelerator'
---
Introduction
============
Recently, Ban$\tilde{a}$dos-Silk-West (BSW) [@Banados:2009pr] studied collision of two particles (e.g. dark matter particles) in the vicinity of the horizon of the Kerr black hole (BH) and found that center-of-mass (CM) energy of the colliding particles in the equatorial plane can be arbitrarily high in the limiting case of extremal BH. This imply that the extremal rotating BH may be considered as a Planck energy scale particle accelerator, which might allow us to explore ultra high energy collisions and astrophysical phenomena, such as the gamma ray bursts and the active galactic nuclei. In the work of BSW [@Banados:2009pr], two particles of equal mass and equal energy falling freely from rest at infinity and approaching the extremal Kerr BH on the equatorial plane were considered. The energy in the CM frame was computed and the critical value of the particle angular momentum at which the energy blows up were determined. However, Jacobson and Sotiriou [@Jacobson:2009zg] elucidate the mechanism for this result, and point out its practical limitations given that extremal BHs do not exist in nature, e.g., the spin $a$ of astrophysical BHs should not exceed $a/M= 0.998$ [@thorne] ($ M $ is the mass of the BH). In any case, the “third law” of BH thermodynamics (Bardeen et al. 1973) asserts that a BH cannot be spun-up in a finite time to the extreme spin value $a/M = 1$. In particular, Jacobson and Sotiriou [@Jacobson:2009zg] demonstrated that infinite collision energy can only be attained at the horizon, and with a maximally spinning BH. Lake [@Lake:2010bq] also demonstrated that the CM energy of colliding particle near the inner horizon of a non-extremal Kerr BH is finite. It is known that the motion of a particle traveling in the background of a charged spinning BH depends not only on the spin but also on the charge of the BH. Therefore, the CM energy of collision expected to depend both on spin and charge. The BSW scenario is generalized to charged BHs [@Zaslavskii:2010aw], the Kerr-Newman family of BHs [@Wei:2010vca] and general rotating BHs [@Igata:2012js]. It is shown that, for a near-extremal charged BH, there always exists a finite upper bound for CM energy, which decreases with the increase of the charge $ Q $.
The BSW effect near the event horizon of the Kerr-Taub-NUT BH was investigated by [@Liu:2010ja], around the Kaluza-Klein (extremal) BH in [@Mao:2010di] and found the infinitely large CM energy near the horizon in both rotating and non-rotating cases. The BSW effect for the Einstein-Maxwell-dilaton-axion BH [@Hussain:2012zza], and the BTZ BH was discussed in [@Yang:2012we]. In [@Zhu:2011ae], the CM energy was generalized for charged particles moving in an electromagnetic field and braneworld BHs were discussed. Even for nonmaximally rotating BHs, Grib and Pavlov [@Grib:2010dz; @Grib:2010xj; @Grib:2011ph; @Grib:2013vc] argued that the CM energy for two particles collision can be unlimited for the non-maximal rotation as well when one considers the multiple scattering. A general explanation for the arbitrarily high CM energy is presented in terms of the Killing vectors and the Killing horizon by Zaslavskii [@Zaslavskii:2010jd]. Further, the author [@Zaslavskii:2012fh; @Zaslavskii:2010pw; @Zaslavskii:2012qy] clarified that the universal property of acceleration of particles near rotating BHs and give a general argument of this BSW mechanism for the general rotating BHs. Further, Joshi and Patil [@Patil:2011ya] found that the CM energy to be high for the naked singularity of the Kerr BH and other BHs [@Patil:2011uf].
More recently, BSW mechanism is used to calculate the CM energy of collision for two particles freely falling, from rest at infinity, in the horizon of a Ayón-Beato-García (ABG) BH [@Pradhan:2014oaa]. It turns out that the CM energy for ABG BH can be infinitely high for the extremal case [@Pradhan:2014oaa]. The rotating counterpart of ABG BH is another interesting and important spacetime [@Toshmatov:2014nya], which is a solution of Einstein equations coupled to nonlinear electrodynamics. Besides the mass M and the rotation parameter $a$, the rotating ABG spacetime carries with the charge, $Q$. It is widely believed that spacetime singularities do not exist in Nature, but that they represent a limitation of the classical theory. While we do not yet have any solid theory of quantum gravity models of BH solutions without singularities have been proposed [@bardeen; @AyonBeato:1998ub; @Hayward:2005gi]. These spacetimes have an event horizon and no pathological features like singularities or regions with closed timelike curves. The rotating ABG metrics are more important as they can be tested by astrophysical observations, as the BH spin plays a fundamental role in any astrophysical process [@Bambi:2013ufa]. The rotating ABG black hole are axisymmetric, asymptotically flat and depend on the mass and spin of the black hole as well as on a parameter $Q$ that measure potential deviations from the Kerr metric. The rotating ABG metric includes the Kerr metric as the special case if this deviation parameter vanishes.
The main purpose of this paper is to study the collision of two particles in the background of the rotating ABG spacetime and to see what the effects of the charge $Q$ on the CM energy for the particles in the near-horizon collision. It turns out that our results can be reduced to the ones of BSW [@Banados:2009pr] as the charge parameter turns to zero and nonrotating ABG BH when the rotation parameter $a=0$. Besides, we may be more interest to discuss CM energy of the rotating ABG BH because it does not have a simple horizon structure as the previous studied BHs.
Rotating Ayón-Beato-García black hole
=====================================
In this section, we would like to study the rotating ABG BHs. The gravitational field of rotating ABG spacetime [@Toshmatov:2014nya] is described by metric which in the Boyer-Lindquist coordinates with units $c=G=1$ is given by $$\begin{aligned}
\label{abg}
d{s}^2 &=&-f(r,\theta) dt^2+\frac{\Sigma}{\Delta} dr^2
\nonumber \\ &&
-2a\sin^2\theta(1-f(r,\theta))d\phi dt+\Sigma d\theta^2 \nonumber \\ &&
+ \sin^2\theta[\Sigma-a^2(f(r,\theta)-2)\sin^2\theta]d\phi^2,\end{aligned}$$ where the function $f(r,\theta)$ and $\Delta$ are given by $$\begin{aligned}
\label{fr}
f(r,\theta) &=& 1-\frac{2M r \sqrt{\Sigma}}{(\Sigma+Q^2)^{3/2}}+\frac{Q^2\Sigma}{(\Sigma+Q^2)^2},
\nonumber \\
\Delta &=& \Sigma f(r,\theta) + a^2\sin^2\theta\end{aligned}$$ and $\Sigma=r^2+a^2\cos^2\theta$. The parameters $a$, $M$ and $Q$ are respectively correspond to rotation, mass and the electric charge. We shall demonstrate that, for certain range of values of $M$ and $Q$, the metric (\[abg\]) is a BH. The curvature invariant $R$, $R_{ab}R^{ab}$ and $R_{abcd}R^{abcd}$ for the metric (\[abg\]) reveals that the rotating ABG BH is regular everywhere for $ a,M,Q \neq 0 $. The metric (\[abg\]) is a non-singular rotating charged BH which generalizes the standard Kerr BH. In addition, if $Q=0$ the metric (\[abg\]) reduce to Kerr BH [@kerr]. Further when both $a, Q = 0$ the metric (\[abg\]) is Schwarzchild BH [@schw]. If $a\rightarrow 0$ we have an exact ABG BH [@Eloy] which is a solution of Einstein field equations coupled to the nonlinear electrodynamics. We know that the Kerr BH has two horizons like surfaces: static limit surface and event horizon. We are interested in these two surfaces for the rotating ABG spacetime metric (\[abg\]). We start with calculating location and structure of static limit surface which requires the prefactor of $d{t}^2$ to vanish. Then it follows from (\[abg\]) that the static limit surface will satisfy $$\begin{aligned}
\label{sls}
f(r,\theta)= (\Sigma+Q^2)^2-2Mr\sqrt{\Sigma(\Sigma+Q^2)}+{Q^2\Sigma} = 0.\end{aligned}$$ The Eq. (\[sls\]) for $Q=0$ becomes exactly same as Kerr [@kerr]. The surface of no return is known as the event horizon. Beyond event horizon the gravitational force becomes so high that it is impossible for any object, even light to escape from the gravity of the BH. The spacetime (\[abg\]) generically must have two horizons, viz., the inner or Cauchy horizon and the outer or event horizon. The horizons of the BH (\[abg\]) are calculated by equating the $g^{rr}$ component of the metric (\[abg\]) to zero, i.e.,
------------------------ ------------------------ --
 
 
------------------------ ------------------------ --
------------------------ ------------------------
 
------------------------ ------------------------
------------------------ ------------------------
 
------------------------ ------------------------
------------------------ ------------------------
 
------------------------ ------------------------
$$\label{eh}
\Delta = \Sigma f(r,\theta)+a^2 \sin^2 \theta =0,$$
which leads to $$\begin{aligned}
\label{eh1}
\Sigma(\Sigma+Q^2)^2-2Mr\Sigma^{3/2}(\Sigma+Q^2)^{1/2}+Q^2\Sigma^2
\nonumber && \\+a^2(\Sigma+Q^2)^2 \sin^2 \theta = 0.\end{aligned}$$ Clearly, the radii of the horizons depends on $\theta$, which are different from that in the usual Kerr case. The behavior of the static limit surface is shown in Fig. \[fig1\] and that of event horizon in Fig. \[fig2\]-\[fig4\] for different values of mass $M$, charge $Q$ and spinning parameter $a$. The two figures reveals that there exists set of values of parameters for which we have two horizons, i.e., a regular BH with both inner and outer horizons. One can also find values of parameters for which one get extreme BHs where the two horizons are coincides. The region between the static limit surface and the event horizon is termed as quantum ergosphere. The ergosphere is the region which lies outside the BH. In ergosphere it is possible to enter and leave again, and the object moves in the direction of the spin of BH.
Particle orbits in the rotating Ayón-Beato-García black hole
=============================================================
Astrophysical BH candidates are supposed to be the Kerr BHs as predicted in general relativity, but the actual nature of these objects has still to be substantiated [@Bambi:2011mj; @Bambi:2013qj; @Zhang:1997dy]. If one wish to test the nature of an astrophysical BH candidate, it is better to consider a more general spacetime, like rotating ABG BH, than a Kerr, in which the central object is described by a mass $ M $, spin parameter $ a $, and an additional deviation parameter $Q$. The rotating ABG metric can be seen as the prototype of a large class of non-Kerr BH metrics and the Kerr solution must be recovered when this deformation parameter $Q$ vanishes (cf. Fig. \[figure5\]). If the observations require vanishing deformation parameter, the compact object is a Kerr BH or non-Kerr BH otherwise. In general, observations allows both the possibility of a Kerr BH with a certain spin parameter and non-Kerr BH with different spin parameters. Further, we are far from a reliable candidate for a quantum theory of gravity, hence recently more attention is given for phenomenological approaches to somehow solve these singularity problem in classical general relativity and study possible implications. In this context, an important line of research is represented by the work on the regular BH solutions Hence, it pertinent to consider BSW mechanism for the rotating ABG black holes.
![Plot showing spin deformation parameter plane of rotating ABG metric. The curve separates BH from no BH or configuration without an event horizon (i.e., no real root of $\Delta = 0$).[]{data-label="figure5"}](figure5.eps)
Let us consider a motion for a particle with mass $m_{0}$ falling in the background of a rotating ABG BH. The geodesic motion for this BH is determined by the following Hamilton-Jacobi equation $$\label{hje}
\frac{\partial S}{\partial \tau} = -\frac{1}{2} g^{\mu\nu} \frac{\partial S}{\partial x^{\mu}} \frac{\partial S}{\partial x^{\nu}},$$ where $\tau$ is an affine parameter along the geodesics, and $S$ is the Jacobi action. For this BH background the Jacobi action $ S $ can be separated as $$\label{hja}
S = \frac{1}{2} m_0^2 \tau -Et + L \phi + S_{r}(r) + S_{\theta}(\theta),$$ where $ S_{r} $ and $ S_{\theta} $ are function of $ r $ and $ \theta $ respectively. The constants $ m_0 $, $ E $, and $ L $ correspond to rest mass, conserved energy and angular momentum through $m_{0}^2= -p_{\mu}p^{\mu}$, $E=-p_{t}$, and $L=p_{\phi}$. For equatorial plane ($\theta=\pi/2$) in the ABG metric (\[abg\]), we obtain the null geodesics in the form of the first-order differential equations $$\label{u^t}
\frac{d t}{d \tau} = \frac{1}{r^2} \Big[-a (aE - L) + \left(r^2 + a^2\right) \frac{T}{\Delta}\Big],$$ $$\label{u^Phi}
\frac{d \phi}{d \tau} = \frac{1}{r^2} \Big[-(aE - L ) + \frac{a T}{\Delta}\Big],$$ $$\label{u^r}
\frac{d r}{d \tau} = \pm \frac{1}{r^2} \sqrt{T^2 -\Delta \left[m_0^2 r^2 + (L-a E)^2 + K\right]},$$ where $ T = E (r^2 + a^2) -La $ and $K$ is a constant.
------------------------ ------------------------
 
------------------------ ------------------------
These equations determine the propagation of light in the spacetime of the rotating ABG BH. Obviously, for $ Q = 0 $, it is just the null geodesic for the Kerr BH. The constant $K=0$ is the necessary and sufficient condition for particles motion initially in the equatorial plane to remain in the equatorial plane. Any particle which crosses the equatorial plane has $K>0$ [@bard].
The study of effective potential is a very useful tool for describing the motion of test particles. The radial equation for timelike particles moving along geodesic in the equatorial plane ($\theta=\pi/2$) is described by $$\frac{1}{2} \dot{r}^2 + V_{eff} = 0,$$ with the effective potential $$V_{eff} = -\frac{[E (r^2 + a^2) -La]^2 -\Delta [m_0^2 r^2 + (L-a E)^2]}{2 r^4},$$ The circular orbit for the particle is defined as $$\label{lim}
V_{eff}=0 \;\;\;\text{and}\;\;\; \frac{dV_{eff}}{dr}=0,$$ Eq. (\[lim\]) leads to the limitation on the possible values of angular momentum of free falling particle and to achieve the horizon of the BH the angular momentum $L$ must be lying within the range $-4.80898 \lesssim L\lesssim 2.02893 $ for extremal case. For extremal case $a \approx 0.980186499171$ and $r_{e} \approx 1.01388$ is event horizon. Whereas for the case of Kerr-Newman BH the extremal value of spin parameters is $a \approx 0.9$, range of angular momentum is $-4.6864 \lesssim L \lesssim 2.0111$ and the $r_{e} \approx 1$.
Since, $u^{t}= dt/d\tau \geq 0$, then from Eq. (\[u\^t\]) the condition $$\begin{aligned}
&& E \Big[(r^2+a^2)(r^2+Q^2)^2+2Mr^2a^2\sqrt{r^2+Q^2}-Q^2r^2a^2\Big] \nonumber \\
&& \;\;\;\;\;\;\;\;\;\;\;\;\;\geq Lar^2\Big[2M\sqrt{r^2+Q^2}-Q^2\Big],\end{aligned}$$ must be satisfied, as $r \rightarrow r_{+}$, this condition reduces to $$E \geq \frac{aL}{r_{+}^2+a^2} = \Omega_{+}L.$$
------------------------ ------------------------
 
------------------------ ------------------------
Center of mass energy of two colliding particles
================================================
We have dealt with range of the angular momentum, for which the particle can reach the horizon, i.e., if the angular momentum of particles are in the desired range, the collision can take place near horizon of the BH. We are now in the position to study the properties of the CM energy of two colliding particles with same rest mass $m_{0}$ near horizon of regular ABG BH.
------------------------ ------------------------
 
------------------------ ------------------------
------------------------ ------------------------
 
------------------------ ------------------------
------------------------- -------------------------
 
------------------------- -------------------------
------------------------- -------------------------
 
------------------------- -------------------------
------------------------- -------------------------
 
------------------------- -------------------------
------------------------- -------------------------
 
------------------------- -------------------------
Next, We study CM energy of the collision of two particles moving in equatorial plane of rotating ABG BH. Let us consider two colliding particles with the same rest mass $m_{1}= m_{2}=m_{0}$ ($E=m_{0}$) are at rest at infinity. They are coming from rest at infinity and approaching the BH with different angular momenta $ L_{1}$ and $ L_{2}$ and collide at some radius $ r $. Our aim is to compute the energy in CM frame for this collision. We assume that two particles 1 and 2 are at same point with four momentum $$p^{\mu}_{i} = m_{0} u^{\mu}_{i},$$ where $p^{\mu}_{i}$ and $ u^{\mu}_{i} $ are the four momentum and four velocity of the $i-th$ particle ($i=1,2$). The sum of 2-momenta is given as $$p^{\mu}_{tot} = p^{\mu}_{1} + p^{\mu}_{2},$$ The CM energy of two-particles is given by $$\label{formula}
E_{CM}^2 = m_{1}^2+m_{2}^2- 2g_{\mu \nu} p_{1}^{\mu} p_{2}^{\nu},$$ Clearly, $E_{CM}$ is scalar. The formula (\[formula\]) is valid for both massless and massive particles, and independent of coordinate choice. In the CM energy case of equal mass $m_{1}= m_{2}=m_{0}$ the formula (\[formula\]) reduces to $$\label{eqlm}
E_{CM} = m_{0} \sqrt{2} \sqrt{1-g_{\mu \nu} u^{\mu}_{(1)} u^{\nu}_{(2)}},$$ Here we assume that the 2-particles coming from infinity with $E_{1}/m_{0}=E_{2}/m_{0}=1$, for simplicity. The CM energy of two equal mass particles in the rotating ABG spacetime is given by
$$\begin{aligned}
\label{ecm}
\frac{E_{CM}^2}{2 m_{0}^2} &=& \frac{1}{r^2(r^2 f(r) + a^2)} \Big[ a (f(r)-1)(L_{1} + L_{2})r^2 -a^2 (f(r)-3)r^2 - L_{1} L_{2} f(r)r^2 + (1 + f(r)) r^4
\nonumber \\
&-&\sqrt{-r^2\bigg((f(r)-1)a^2 -2a(f(r)-1)L_{1}-r^2+f(r)(L_{1}^2+r^2)\bigg)}
\nonumber \\
&&
\sqrt{-r^2\bigg((f(r)-1)a^2 -2a(f(r)-1)L_{2}-r^2+f(r)(L_{2}^2+r^2)\bigg)}\Big].\end{aligned}$$
Near horizon collision in extremal rotating ABG black hole
----------------------------------------------------------
Note that for an extremal BH, the two horizons always coincides and located at $r=r_{e}$, where $r_{e}$ is the double root of Eq. (\[eh\]), i.e., $\Delta(r_{e})=r_{e}^2f(r_{e})+a^2=0$. We now study the properties of CM energy Eq. (\[ecm\]) as radius $r \rightarrow r_{e}$ for the extremal rotating ABG spacetime. Firstly we find the range of angular momentum of particles with which particles can reach the horizon by solving Eq. (\[lim\]) numerically. The maximum/minimum angular momentum we are looking for extremal BH are listed in Table (\[table1\]). We can see from the Table (\[table1\]) that when charge $Q$ is increases, the extremal horizon $r_{e}$ is also increases.
We can see the behavior of CM energy for extremal BH from Fig. \[fig8\] and Fig. \[fig10\]. The Fig. \[fig8\] shows the variation of $E_{CM}$ vs $r$ for different values of $L_{1}$ and $L_{2}$ with fixed values of spin parameter $a$ and charge $Q$. We can see that $E_{CM}$ is infinite at the event horizon when $L_{1}$ is critical while for any other values of $L_{1}$, $E_{CM}$ is finite. In Fig. \[fig10\] we show the variation of $E_{CM}$ vs $a$ at the horizon for different values of $L_{1}$ and $L_{2}$ with fixed value of charge parameter $Q$.
S.No. $Q$ $a$ $r_{+}$ $L_{min}$ $L_{max}$
------- ----- ----- --------- ----------- -----------
1 0.1 0.9 1.39848 -4.75125 2.58660
2 0.2 0.8 1.49272 -4.66002 2.77108
3 0.3 0.7 1.50591 -4.63034 2.52549
4 0.4 0.6 1.44629 -4.42525 2.90270
5 0.5 0.5 1.25836 -4.27280 2.84983
: Numerical values of max/min angular momentum for non-extremal rotating ABG BH $M=1$, $m_{0}=1$, $E=1$.[]{data-label="table2"}
Near horizon collision in non-extremal rotating ABG black hole
--------------------------------------------------------------
A BH is called non-extremal when the outer and inner horizons are not coincide, i.e. $r \neq r_{e}$. We find the range of angular momentum of particles by solving Eq. (\[lim\]) numerically. The maximum/minimum angular momentum we are looking for non-extremal BH are listed in Table (\[table2\]).
We can see the behavior of CM energy for non-extremal BH from Fig. \[fig9\] and Fig. \[fig11\]. In Fig. \[fig9\] we show the variation of $E_{CM}$ vs $r$ for different values of $L_{1}$ and $L_{2}$ with fixed values of spin parameter $a$ and charge $Q$. Here, we can see that $E_{CM}$ is finite at the event horizon. In Fig. \[fig11\] we show the variation of $E_{CM}$ vs $a$ at the outer horizon for different values of $L_{1}$ and $L_{2}$ with fixed values of charge $Q$. The variation of $E_{CM}$ vs $r$ for different values of $Q$ can be seen in panel (a) of Fig. \[fig12\] and $E_{CM}$ vs $a$ in panel (a) of Fig. \[fig13\]. We have also plot the variation of $E_{CM}$ for Kerr BH in panel (b) of Fig. \[fig12\] and in panel (b) of Fig. \[fig13\].
Eq. (\[ecm\]) confirms that the non-linear charge $Q$ indeed has influence on the CM energy. In the limit $a\rightarrow 0$, Eq. (\[ecm\]) reduces to expression for $E_{CM}$ of ABG BH,
$$\begin{aligned}
\label{ecm1}
\frac{E_{CM}^2}{2 m_{0}^2 }(a\rightarrow 0) &=& \Bigg[1+\frac{1}{f(r)}-\frac{L_{1} L_{2}}{r^2} -\frac{1}{f(r)}\sqrt{1-f(r)\left(1+\frac{L_{1}^2}{r^2}\right)} \sqrt{1-f(r)\left(1+\frac{L_{2}^2}{r^2}\right)}\Bigg].\end{aligned}$$
which is exactly same as obtained in [@Pradhan:2014oaa]. Further, $E_{CM}$ for the Kerr BH [@Banados:2009pr] can be obtained by taking limit $Q \rightarrow 0$ in Eq. (\[ecm\]), which reads
$$\begin{aligned}
\label{ecm2}
\frac{E_{CM}^2}{2 m_{0}^2 }(Q\rightarrow 0) &=& \frac{1}{r(r^2-2r+a^2)} \Big[2a^2 (1+r) - 2a (L_{1}+L_{2})- L_{1} L_{2}(-2+r)
+2(-1+r)r^2
\nonumber \\
&-&\sqrt{2(a-L_{1})^2 - L_{1}^2r + 2r^2}\sqrt{2(a-L_{2})^2 - L_{2}^2r + 2r^2}\Big].\end{aligned}$$
It turns out that Eq. (\[ecm2\]) the $ E_{CM}^2 $ blows up at the horizon in the extremal case [@Banados:2009pr].
Conclusion
==========
We are far from a reliable candidate for a quantum theory of gravity, hence recently more attention is given for phenomenological approaches to somehow solve these singularity problem in classical general relativity and study possible implications. In this context, an important line of research is represented by the work on the regular BH solutions. Hence, we have investigated the collision of two particles falling freely from rest at infinity in the background of a rotating regular ABG BH. It was demonstrated by BSW [@Banados:2009pr] that CM energy of two colliding particles near the horizon of an extremal Kerr BH can be arbitrarily high value under some restriction on the angular momentum of particles. The analysis of BSW when extended to rotating ABG BH to investigate the CM energy $E_{CM}$ for two particles, an unlimited $E_{CM}$ that constraints on the charge $Q$ and rotation parameter $a$. From the point of view of BSW mechanism the CM energy for two colliding particles for extremal BH should be arbitrarily high. So that we have found critical value of the angular momentum for which the CM energy of colliding particles is arbitrarily high near the horizon (extremal BH). We have also seen that when the angular momentum of the colliding particles is not critical (non-extremal BH) the CM energy is finite. For the non-extremal case, we find that the CM energy $E_{CM}$ increases with increase in the charge $Q$. The range of the angular momentum $L$ and the spin parameter $a$ increases for the rotating ABG BH as compared with the Kerr-Newman BH. The CM energy $E_{CM}$ for the extremal rotating ABG BH case is infinite at $r_{e}=1.01388$ while for the extremal Kerr BH case the CM energy $E_{CM}$ is infinite at $r_{e}=1$.
Indeed, to obtain an arbitrary high $E_{CM}$, besides the conditions that the BH is an extremal BH, and one of the particle has critical angular momentum, one still has restriction on the charge $Q$. It may be pointed out that horizon structure of the rotating ABG BHs is significantly complicated to analyze analytically and hence we turn to numerically methods to analyze $E_{CM}$, the range of the angular momentum, with which particle can reach horizon and fall into BH.
It turns out that if the angular momentum of a particle lies outside the range, it will not fall into the BH and no collision will take place. We have also estimated the $E_{CM}$ for non-extremal rotating ABG BH case, and found that the $E_{CM}$ is always finite and significantly affected by charge $Q$ and rotation parameter $a$. It will be of interest to extend the analysis discussed to other regular BH, which will be determine whether these properties of regular BHs are generic. These and related work are subject of forthcoming papers.
M.A. acknowledges University Grant Commission, India for financial support through Maulana Azad National Fellowship scheme (File No.: F1-17.1/2012-13/MANF-2012-13-MUS-RAJ-8679).
[99]{} M. Banados, J. Silk and S. M. West, Phys. Rev. Lett. [**103**]{}, 111102 (2009).
T. Jacobson and T. P. Sotiriou, Phys. Rev. Lett. [**104**]{}, 021101 (2010).
K. S. Thorne, Astrophys. J. 191, 507 (1974).
K. Lake, Phys. Rev. Lett. [**104**]{}, 211102 (2010).
O. B. Zaslavskii, JETP Lett. [**92**]{}, 571 (2010) \[Pisma Zh. Eksp. Teor. Fiz. [**92**]{}, 635 (2010)\].
S. W. Wei, Y. X. Liu, H. Guo and C. EFu, Phys. Rev. D [**82**]{}, 103005 (2010).
T. Igata, T. Harada and M. Kimura, Phys. Rev. D [**85**]{}, 104028 (2012)
C. Liu, S. Chen, C. Ding and J. Jing, Phys. Lett. B [**701**]{}, 285 (2011).
P. J. Mao, L. Y. Jia, J. R. Ren and R. Li, arXiv:1008.2660 \[hep-th\].
I. Hussain, Mod. Phys. Lett. A [**27**]{}, 1250017 (2012).
J. Yang, Y. L. Li, Y. Li, S. W. Wei and Y. X. Liu, arXiv:1202.4159 \[hep-th\].
Y. Zhu, S. -F. Wu, Y. -X. Liu and Y. Jiang, Phys. Rev. D [**84**]{}, 043006 (2011).
A. A. Grib and Y. .V. Pavlov, Grav. Cosmol. [**17**]{}, 42 (2011).
A. A. Grib, Y. .V. Pavlov, O. F. Piattella and O. F. Piattella, Int. J. Mod. Phys. A [**26**]{}, 3856 (2011).
A. A. Grib and Y. .V. Pavlov, Astropart. Phys. [**34**]{}, 581 (2011).
A. A. Grib and Y. .V. Pavlov, Europhys. Lett. [**101**]{}, 20004 (2013).
O. B. Zaslavskii, Phys. Rev. D [**82**]{}, 083004 (2010).
O. B. Zaslavskii, Phys. Lett. B [**712**]{}, 161 (2012).
O. B. Zaslavskii, Class. Quant. Grav. [**28**]{}, 105010 (2011).
O. B. Zaslavskii, Int. J. Mod. Phys. D [**22**]{}, 1350044 (2013).
M. Patil and P. S. Joshi, Class. Quant. Grav. [**28**]{}, 235012 (2011).
M. Patil, P. S. Joshi, M. Kimura and K. i. Nakao, Phys. Rev. D [**86**]{}, 084023 (2012).
P. Pradhan, arXiv:1402.2748 \[gr-qc\].
B. Toshmatov, B. Ahmedov, A. Abdujabbarov and Z. Stuchlik, arXiv:1404.6443 \[gr-qc\].
J. Bardeen, in Proceedings of GR5, Tiflis, U.S.S.R., (1968).
E. Ayon-Beato and A. Garcia, Phys. Rev. Lett. [**80**]{}, 5056 (1998).
S. A. Hayward, Phys. Rev. Lett. [**96**]{}, 031103 (2006).
C. Bambi and L. Modesto, Phys. Lett. B [**721**]{}, 329 (2013).
R.P. Kerr, Phys. Rev. Letters, **11**,237 (1963).
K. Schwarzschild, Uber das Gravitationsfeld eines Maßpunktes nach der Einsteinschen Theorie, Sitzber. Deutsch. Akad. Wiss. Berlin, Kl. Math. Phys. Tech., 189, (1916).
E. Ayón-Beato, A. García, Phys. Lett. B [**464**]{}, 25 (1999).
C. Bambi, Mod. Phys. Lett. A [**26**]{}, 2453 (2011). C. Bambi, Astron. Rev. [**8**]{}, 4 (2013). S. N. Zhang, W. Cui and W. Chen, Astrophys. J. [**482**]{}, L155 (1997). J. M. Bardeen, W. H. Press and S. A. Teukolsky, Astrophys. J. [**178**]{}, 347 (1972).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We use N-body simulations to model the tidal disruption of a star cluster in a Milky-Way-sized dark matter halo, which results in a narrow stream comparable to (but slightly wider than) Pal-5 or GD-1. The mean Galactic dark matter halo is modeled by a spherical Navarro-Frenk-White (NFW) potential with subhalos predicted by the [$\Lambda$CDM]{} cosmological model. The distribution and mass function of the subhalos follow the results from the Aquarius simulation. We use a matched filter approach to look for “gaps” in tidal streams at 12 length scales from $0.1\,\mathrm{kpc}$ to $5\,\mathrm{kpc}$, which appear as characteristic dips in the linear densities along the streams. We find that, in addition to the subhalos’ perturbations, the epicyclic overdensities (EOs) due to the coherent epicyclic motions of particles in a stream also produce gap-like signals near the progenitor. We measure the gap spectra – the gap formation rates as functions of gap length – due to both subhalo perturbations and EOs, which have not been accounted for together by previous studies. [Finally, we project the simulated streams onto the sky to investigate issues when interpreting gap spectra in observations. In particular, we find that gap spectra from low signal-to-noise observations can be biased by the orbital phase of the stream. This indicates that the study of stream gaps will benefit greatly from high-quality data from future missions. ]{}'
author:
- 'W. H. W. Ngan and R. G. Carlberg'
title: 'Using gaps in N-body tidal streams to probe missing satellites'
---
Introduction
============
The [$\Lambda$CDM]{} model – a universe dominated by cold dark matter and a cosmological constant – is a successful model of the universe at large scales [@blumenthal84; @davis85; @bardeen86; @riess98; @perlmutter99; @planck13]. However, as more observations are made and computational power becomes available, discrepancies in the model have been found below galactic scales. Cosmological simulations show that the dark matter halo of a Milky-Way-sized galaxy, in addition to having a smooth component with well-known density profiles [@nfw; @navarro04; @navarro10], should be populated by substructures, or subhalos [@madau08; @stadel09; @zemp09; @aquarius; @gao11]. This theoretical prediction has not been supported by observational evidence, as many efforts over the past two decades have failed to find enough satellite systems in the Milky Way to account for the predicted abundance [@klypin99; @moore99; @strigari07]. This discrepancy is called the [*Missing Satellite Problem*]{}.
Competing solutions to the Missing Satellite Problem can be roughly classified into two types – astrophysical and physical. Astrophysical solutions, on one hand, postulate that the predicted subhalo abundance is correct, but the subhalos have too little stellar content to be observable directly. For example, reionizing radiation or stellar feedback [@koposov09; @maccio09] can suppress star formation in a subhalo. This class of solutions [is]{} founded on the absence of baryonic processes in those cosmological simulations which predicted the subhalo abundances. Physical solutions, on the other hand, postulate that the predicted subhalo abundance is incorrect, as our understanding of CDM may be incomplete. Alternative dark matter solutions, such as warm dark matter [@barkana01; @bode01; @benson13; @schneider13], self-interacting dark matter [@spergel00], or inflationary models with non-scale invariance [@kamionkowski00] offer mechanisms to suppress structure formation at small scales. Clearly, the predicted difference between these two types of solutions is the number of subhalos. In order to test the [$\Lambda$CDM]{} model at sub-galactic scales, measuring the true abundance of subhalos is an important step.
Tidal streams – or simply “streams” – are remnants of stellar systems such as globular clusters (GCs) or dwarf galaxies (DGs) as they are tidally disrupted by a massive host. When the stars become unbound from the progenitor, the stars trace an elongated tail which wraps around the massive host. Tidal remnants have long been useful probes for studying the gravitational potential of the Milky Way [e.g., @johnston98; @helmi03; @law05]. In particular, @ibata02 first used simulations to show that the encounters between the stream stars and subhalos can dynamically heat up the stream, which can be used to probe the presence of subhalos. Moreover, a key influence that subhalos have on streams is that the stream stars near the point of the encounter get scattered into different orbits by the perturbation, causing an abrupt decrease in stellar density in that region of an otherwise smooth stream. Using the abundance of subhalos obtained from high-resolution simulations [@madau08; @aquarius], it has been predicted that streams in the Milky Way described by the [$\Lambda$CDM]{} model should contain many “gaps” [@yoon11; @carlberg12; @pal5_gaps; @CG13].
In the past decade, many streams in the Milky Way have been found (see @G10 for a list). Two streams of particular interest to us are Pal-5 [@pal5] and GD-1 [@gd1]. Both streams are detected as long, narrow tidal tails with length-to-width ratio of $\sim100$. These two streams are interesting as they show varying densities longitudinally along the streams. It is not clear whether those density variations correspond to subhalo perturbations. Other possible origins of those density variations include clumping due to the coherence in the epicyclic orbits of stream stars [@kupper08; @kupper10] and Jeans instabilities [@qc11]. The ultimate goal of this study is to test whether the gaps observed in streams are consistent with the prediction by the [$\Lambda$CDM]{} model. @yoon11 and @carlberg12 independently made the first predictions by simulating ideal streams with massless particles in the presence of orbiting subhalos. In particular, @carlberg12 derived analytical expressions of the gap formation rate as a function of a stream’s intrinsic properties, which are readily comparable to observations. However, neither of the aforementioned studies self-consistently modeled the realistic disruption of the progenitor system.
In this study, we measure the gap formation rate by modeling a stream’s formation and its interaction with subhalos using N-body simulations. This paper is organized as follows. Section 2 describes the details of our simulations, including the subhalo abundances and density profiles that we adapt, and the details of the star cluster and the resulting model stream. Section 3 focuses on the method of detecting gaps in a simulated stream. The method of using match filters is inspired by analyses for observations, but modified here to analyze simulations. Section 4 contains detailed discussion of our key results, including the phenomenology of gaps and comparisons with previous analytical predictions and observations. Section 5 is a summary of our results.
Simulations
===========
Models {#sec:models}
------
The host galaxy is modeled as a dark matter halo, as well as a set of subhalos which orbit around the halo’s potential. A Milky-Way-sized halo is modeled with a static spherical Navarro-Frenk-White (NFW) profile [@nfw] with $v_{max}=210$ [kms$^{-1}$]{} located at $r_{max}=30\,\mathrm{kpc}$. Each individual subhalo is modeled by a spherical Hernquist profile [@hernquist90] $$\Phi_i(r) = \frac{GM_i}{h_i+r}$$ for simplicity, compared to Einasto profiles which produce better fits in simulations but are more complicated to compute [@aquarius]. We use the [formula found in @carlberg09, which approximates the results from both @aquarius and @neto07,]{} where $h_i$ is independent of galactocentric position, and is related to $M_i$ by $$h(M) = 6\,\mathrm{kpc} \times \left( \frac{M}{10^{10}{M_\odot}} \right)^{0.43}.
\label{eqn:subhalo_size}$$ We use the mass and spatial distributions of the subhalos from the results of the Aquarius simulations [@aquarius], where the mass function is independent of the spatial distribution. The mass function is a power law $$\frac{dN}{dM} = 3.26{\times10^{-5}}{M_\odot}^{-1} \left( \frac{M}{2.52{\times10^{7}}{M_\odot}} \right)^{-1.9},
\label{eqn:subhalo_massfunction}$$ and the spatial distribution follows an Einasto profile $$n(r) \propto \exp \left\{ -2.95 \left[ \left(\frac{r}{199\,\mathrm{kpc}} \right)^{0.678} - 1 \right] \right\}.$$ The subhalos’ velocities are initialized with a Gaussian distribution where the velocity dispersion is the solution to the isotropic Jean’s equation [@bt08] using the halo’s potential. The subhalos orbit around this potential as test masses.
The progenitor of the stream, which is an approximation to a globular star cluster, is initialized using $10^6$ particles of equal mass as a King model with parameters $w=4.91$, total mass $4.29{\times10^{4}}{M_\odot}$, and a core radius of $0.01$ kpc. This results in a zero-density radius of $0.103$ kpc. Each N-body particle in the system interacts with the [dark matter halo’s and subhalos’ potentials.]{} With the Galactic center at the origin, the satellite is initially put at $(x,y,z)=(30, 0, 0)$ kpc and velocity $(v_x, v_y, v_z) = (0, 140, 0)$ [kms$^{-1}$]{}. The resulting orbit is confined on the $xy$-plane with eccentricity $0.33$, peri- and apogalacticon at $r_p=15\,\mathrm{kpc}$ and $r_a=30\,\mathrm{kpc}$, respectively. The azimuthal and radial periods are about 0.70 Gyr and 0.47 Gyr, respectively.
Software and Parameters
-----------------------
We use [Gadget-2]{} [@gadget2] for our N-body simulations. Since the public distribution[^1] does not have functionality for external potentials, we modify the code such that in every time step, an external acceleration term which accounts for the potentials of the halo and all the subhalos is added to the accelerations of all the particles after their N-body interactions are computed.
Each of our simulations lasts 10 Gyr, and we impose a maximum time step of 1 Myr. The particle softening is 5 pc. Each simulation produces 500 snapshots, one every 20 Myr, and each consists of the positions and velocities of the N-body particles and subhalos.
Stream Properties
-----------------
The star cluster is modeled as an N-body system which forms a stream as the cluster is disrupted by the tidal field of the massive host. When the cluster is isolated, the energies of the individual particles are conserved to a few percent over 10 Gyr. Using the softening as minimal impact distance, the relaxation timescale in the core is $\gtrsim 110\, \mathrm{Gyr}$, which is much greater than the orbital period at $\lesssim1\,\mathrm{Gyr}$. [Figure \[fig:massloss\]]{} shows the mass enclosed inside 0.103 kpc of the [star cluster’s]{} center [it is orbiting in the absence of subhalos]{}. Because the stream is repeatedly stretched and compressed longitudinally along the eccentric orbit, the mass enclosed in a fixed radius is not always decreasing in time. The bottom panel of [Figure \[fig:massloss\]]{} shows that the mass loss is driven purely by bulge shocking [@bt08], as the periodic bursts have exactly the same period as the radial period of the orbit.
![Top: fraction of mass enclosed in 0.103 kpc of the star cluster while orbiting the dark matter halo without any subhalos. Middle: rate of change of mass enclosed in 0.103 kpc of the star cluster (solid black) and radial position of the star cluster’s orbit (red). This shows that the star cluster experiences bursts of mass loss almost immediately after each pericentric passage. [Bottom: length of the stream. This shows that the stream gets stretched and compressed depending on its orbital phase. The red and blue vertical lines correspond to the snapshots where the star cluster is at the pericenters and apocenters of its orbit, respectively. These times are selected to demonstrate the bias discussed in Section \[sec:case\_study\].]{}[]{data-label="fig:massloss"}](figs/massloss/massloss2.eps){width="3.5in"}
[Figure \[fig:stream\_sigma\_width\]]{} shows the velocity dispersion and traverse FWHM of our simulated stream without any subhalos. The stream is chosen so that its properties are on the same orders of magnitude as Pal-5 [@dehnen04; @odenkirchen09] and GD-1 [@koposov10; @willett09]. In the derivation in @carlberg12, the gap formation rate is expressed as a function of galactocentric distance of the orbit and width of the stream. For our simulated stream, we adapt average values of 22 kpc and 0.3 kpc, respectively, over the entire stream.
![Tangential velocity dispersion and FWHM width along the stream at 8 Gyr. The progenitor is centered at 0 kpc, where positive and negative positions represent the leading and trailing branches of the stream, respectively. The structures in velocity dispersion and width are due to the eccentricity of the stream, as traced by the orbital distance along the stream.[]{data-label="fig:stream_sigma_width"}](figs/stream_8Gyr/sigma_width_8Gyr_multiplot.eps){width="3.5in"}
In a time independent and spherical potential $\Phi(r)$, three interesting conserved quantities are the radial, azimuthal, and latitudinal actions where $$\begin{aligned}
J_r = \frac{1}{\pi}\int_{r_p}^{r_a} dr \sqrt{2E-2\Phi(r)-\frac{L^2}{r}} \\
J_\phi = L_z \\
J_\theta = L-|L_z|\end{aligned}$$ respectively [@bt08], where $r_a$ and $r_p$ are the apo- and pericentric distances of the orbit, respectively. Since our stream’s orbital plane is the $xy$-plane, $J_\theta\approx 0$ (though not exactly 0 because the stream has a finite thickness), so we only consider $J_r$ and $L_z$. For simplicity, we ignore the progenitor and only consider the particles which have already escaped from the cluster, so when computing $J_r$ we assume that the potential due to the progenitor’s potential is negligible.
$J_r$ and $L_z$ are useful since their dispersions are the origins of the stream’s average width. For example, in the epicyclic approximation [@bt08] where $\kappa$ and $a$ are the epicyclic frequency and amplitude, respectively, the radial motion can be written as $$r(t)=a \cos(\kappa t + \psi)$$ where $\psi$ is an arbitrary phase angle. Then it can be shown that $$\begin{aligned}
J_r = \frac{1}{2\pi} \oint p_r dr \propto \int \dot{r}\,dr \propto \kappa a^2 \\
L_z = R_g^2\Omega\end{aligned}$$ where $\Omega$ and $R_g$ are the orbital frequency and radius of the guiding center, respectively. Clearly, dispersions in both $a$ and $R_g$ can affect the width of the stream. Therefore, conserved quantities $J_r$ and $L_z$ are especially valuable in understanding the width of the stream. [Figure \[fig:Jr\_Lz\_distribution\]]{} shows the distributions in $(J_r,L_z)$ when the stream is 8 Gyr old. The two lobes at higher and lower $L_z$ are the trailing and leading branches of the stream, respectively. The absolute dispersions in $J_r$ and $L_z$ are on the same order of magnitude, in rough agreement with the formula $\Delta J_r/\Delta L\sim (r_a-r_p)/\pi r_p$ [@EB11]. In Section \[sec:gap\_morphology\], we will show how the spread in actions affects the morphology of stream gaps.
Plots similar to [Figure \[fig:Jr\_Lz\_distribution\]]{} can be found in @EB11 where they used angle-action variables extensively to study the relation between the stream and the orbit. A similar plot can also be found in @yoon11, but in scaled energy and angular momentum which were first used by @johnston98 to describe the dynamics of tidal streams.
![Distribution of orbital action variables $(J_r,L_z)$ for our stream at 8 Gyr without any subhalos. Along the entire stream with the progenitor masked out, 50,000 unbound particles were randomly selected to be placed on this map. Each pixel in this map represents the density of particles in that bin. The two symmetric lobes are the two branches of the stream. Similar plots can be found in @EB11, and a plot in scaled energy and angular momentum in @yoon11.[]{data-label="fig:Jr_Lz_distribution"}](figs/Jr_Lz_distribution/Jr_Lz_distribution_8Gyr.eps){width="3.5in"}
Method
======
Subhalo Mass Ranges {#sec:subhalo_mass_ranges}
-------------------
@yoon11 divided their subhalo mass spectrum into separate mass ranges in order to resolve the contributions from each mass range. Using the same approach, we divide the subhalos from $6.5{\times10^{4}}$ to $10^8{M_\odot}$ into 13 mass ranges ([Table \[table:subhalo\_massrange\]]{}). Each mass range contains incrementally more subhalos, starting from the higher mass end toward the lower mass end. In the higher mass end, the mass ranges are chosen such that the increase in subhalo masses are roughly the same. In the lower mass end, the mass ranges are chosen such that the increase in subhalo numbers are roughly the same.
In each set of subhalos, we reduce their numbers by eliminating those whose orbits are always inside the perigalacticon and outside the apogalacticon of the [progenitor’s orbit]{}. The largest subhalo in our simulations has a length scale of $\sim 1\,\mathrm{kpc}$ ([Equation (\[eqn:subhalo\_size\])]{}), so [all the subhalos with perigalacticon (apogalacticon) larger (smaller) than that of the progenitor’s orbit by $\sim2\,\mathrm{kpc}$ will interact minimally with the stream. This allows us to safely eliminate the subhalos with perigalacticon larger than $32\,\mathrm{kpc}$, and apogalacticon smaller than $13\,\mathrm{kpc}$.]{}
We run 14 simulations – 1 “smooth stream” without any subhalos, and 13 “[$\Lambda$CDM]{} streams” containing the subhalos in the mass ranges in [Table \[table:subhalo\_massrange\]]{} – with identical initial conditions and dark matter halo potential. This allows us to resolve the effects of the lower mass subhalos whose existence are in question.
[cccccc]{} $5.3{\times10^{7}}$ & 203 & $1.5{\times10^{10}}$ & 30 & $2.2{\times10^{9}}$\
$2.7{\times10^{7}}$ & 593 & $2.9{\times10^{10}}$ & 98 & $4.7{\times10^{9}}$\
$1.3{\times10^{7}}$ & 1,392 & $4.4{\times10^{10}}$ & 220 & $6.9{\times10^{9}}$\
$5.8{\times10^{6}}$ & 3,160 & $5.9{\times10^{10}}$ & 476 & $9.1{\times10^{9}}$\
$2.5{\times10^{6}}$ & 7,038 & $7.3{\times10^{10}}$ & 1,101 & $1.1{\times10^{10}}$\
$1.0{\times10^{6}}$ & 16,394 & $8.8{\times10^{10}}$ & 2,576 & $1.4{\times10^{10}}$\
$3.6{\times10^{5}}$ & 41,515 & $1.0{\times10^{11}}$ & 6,539 & $1.6{\times10^{10}}$\
$2.1{\times10^{5}}$ & 67,599 & $1.1{\times10^{11}}$ & 10,563 & $1.7{\times10^{10}}$\
$1.5{\times10^{5}}$ & 91,601 & $1.1{\times10^{11}}$ & 14,337 & $1.8{\times10^{10}}$\
$1.1{\times10^{5}}$ & 121,181 & $1.2{\times10^{11}}$ & 18,872 & $1.8{\times10^{10}}$\
$9.0{\times10^{4}}$ & 145,220 & $1.2{\times10^{11}}$ & 22,578 & $1.9{\times10^{10}}$\
$7.5{\times10^{4}}$ & 171,163 & $1.2{\times10^{11}}$ & 26,586 & $1.9{\times10^{10}}$\
$6.5{\times10^{4}}$ & 194,726 & $1.2{\times10^{11}}$ & 30,253 & $1.9{\times10^{10}}$
Gap Finding {#sec:gapfinding}
-----------
Gaps are manifested as local minima in the linear density along the stream. To obtain the linear density along a stream in an eccentric orbit, we first fit the stream with two degree-6 polynomials – one each for the leading and trailing streams – in polar coordinates centered at the Galactic center. The points along each line are spaced at 0.002 radians apart. Between each pair of adjacent points a cylinder of radius 1 kpc is drawn which lies lengthwise along the pair of points. The linear density is then the number of particles inside this cylinder divided by the length of the cylinder. This spacing is chosen so that the gaps as wide as the stream are well resolved.
The method used to find gaps in stream densities is inspired by the technique first used by @pal5_gaps to find gaps in observations. They used matched filters of the estimated shape of a density gap at various length scales to look for positions in the stream which potentially contain gap signals. The filter consists of a local minimum which is the underdensity of stars, and two local maxima on both sides of the minimum due to conservation of mass ([Figure \[fig:filter\_scales\]]{}). This method is similar to the wavelets approach, where the integral of the filter function is constructed to vanish inside a certain domain. The potential gap signals are then easily identified as local maxima in the convolution between the filter and the signal.
To obtain the significance of each potential gap signal against noise, @pal5_gaps produced bootstrap samples from the sky background. With the simulations in this study, we can estimate noise levels using the smooth stream. Note that the “smooth stream” itself is not totally smooth. As we will show in Section \[sec:gap\_counting\], there are large density fluctuations near the progenitor due to the coherent in epicyclic motion of the particles, as first explained by @kupper08. When the particles become unbound from the progenitor, they pile up near the base of their cycloid trajectories, creating epicyclic overdensities (hereafter EO) along the stream. Although this intrinsic process to mimic gaps can be confused with gaps caused by subhalos, EOs are only [apparent within $\lesssim5\,\mathrm{kpc}$ away from the progenitor in our streams ([Figures \[fig:density\_gaps\_whole\_massrange0\] and \[fig:density\_gaps\_central\_massrange0\]]{}). The details of the dynamics of EOs is beyond the scope of this study, but this effect can be understood in terms of orbital actions. EOs occur due to coherent epicyclic motions of the particles, which nevertheless have finite dispersions in orbital actions ([Figure \[fig:Jr\_Lz\_distribution\]]{}) and are not perfectly coherent. Therefore, although the escaping particles’ orbits stay roughly coherent in the first few clumps, their orbits eventually drift out of phase as they travel along the stream. This explains why the density peaks of EOs further downstream are not as apparent as the peaks closer to the progenitor [@just09; @kupper10].]{}
After masking $10\,\mathrm{kpc}$ of the smooth stream centered at the progenitor, the rest of the smooth stream is simply noise. Our method to find gaps in a given stream can be summarized as follows.
1. Compute $$C_s(x) = \frac{1}{s} \int_{x-1.5s}^{x+1.5s} \left[ \rho(x^\prime)-\bar\rho_s(x) \right]\, f\left( \frac{2(x-x^\prime)}{s} \right) dx^\prime
\label{eqn:cs}
$$ where $f(t)=(t^6-1)\exp(-1.2321t^2)$ is a matched filter function [@pal5_gaps; @CG13], and $s$ is the filter scale. $\bar\rho_s(x)$ is the mean of $\rho(x)$ inside $[x-1.5s,x+1.5s]$, the domain in which the integral of $f(2x/s)$ itself vanishes. Each potential gap signal would appear as a local maximum in $C_s(x)$. This convolution is computed at 12 logarithmically spaced filter scales from 0.1 to $5\,\mathrm{kpc}$ ([Figure \[fig:filter\_scales\]]{}), and then all the local maxima of each stream are sorted by $C_s$.
2. Repeat the above step using the smooth stream, but with $\pm 5$ kpc from the progenitor masked along the stream. The set of local maxima in $\tilde C_s(x)$ from this convolution is the noise, which are also sorted by $\tilde C_s$.
3. Each local maximum in the signal set are compared against the noise set. A signal element at any position along either branch of the stream that ranks higher than 99% in the noise set is identified as a gap.
Inevitably, this method may detect the same gap at 99% confidence at different scales but in very close proximity. To avoid over-counting, we employ the following scheme to eliminate overlapping gaps. First, we define an overlap as two gaps whose $C_s$ local maxima are identified at $C_{s1}(x_1)$ and $C_{s2}(x_2)$ that are within $s_1$ away from each other along the stream, where $s_2<s_1$. When this occurs, the gap with higher $C_s$ eliminates the lower.
[Our gap detection method requires no prior knowledge whether a given gap is an EO or a subhalo perturbation, both of which can be identified as a series of over- and under-densities. When we count the number of gaps in the end, EOs will be included. One key result of our study is that gaps due to EOs are distributed very differently in lengths compared to gaps due to subhalo perturbation.]{}
![Five example filter scales for the match filter used in @pal5_gaps, which has the functional form $f(t)=(t^6-1)\exp(-1.2321t^2)$. With a physical scale $s$, the integral of $f(2x/s)$ vanishes inside $-1.5s<x<1.5s$. The roots of each filter are located at $\pm s/2$ so that the gap length is simply the distance between the roots. We search for gaps at 12 logarithmically filter scales from 0.1 to $5\,\mathrm{kpc}$ in order to minimize the chance of detecting the same gap at multiple scales.[]{data-label="fig:filter_scales"}](figs/filter_scales/filter_scales.eps){width="3.5in"}
Results and discussion
======================
Gap Morphology {#sec:gap_morphology}
--------------
According to @yoon11, gaps in general are diagonal and not perpendicular to the stream due to a gradient in angular momentum (hence a gradient in orbital velocities) across the width of the stream, which can shear a gap longitudinally. [Figure \[fig:Jr\_Lz\_distribution\]]{} allows us to estimate the shearing effect in our streams using the distributions in angular momenta. For each branch of the stream, the FWHM spread in angular momentum is about $\Delta L \sim \Delta L_z \sim10\,\mathrm{kpc}\,{\mathrm{km}\,\mathrm{s}^{-1}}$. For a narrow stream at $r=22\,\mathrm{kpc}$, the spread in velocity is $\Delta v = \Delta L/r \sim 0.5\,{\mathrm{km}\,\mathrm{s}^{-1}}$. Therefore, a gap that spans the width of the stream will be sheared by less than $1\,\mathrm{kpc}$ per Gyr.
[Figures \[fig:stream\_maps\_smooth\] and \[fig:stream\_maps\_lcdm\]]{} show the time evolution of the smooth and a [$\Lambda$CDM]{} stream, respectively, from 7 to $8\,\mathrm{Gyr}$. The EOs near the progenitor appear to shear by different amounts at different times, but this is due to the radial oscillation in the orbit where the radial period is $\sim0.5\,\mathrm{Gyr}$. Upon closer inspection of [Figure \[fig:stream\_maps\_lcdm\]]{}, we also note that not only do subhalo gaps have complicated morphologies, but their orientations flip back and forth in a radial period due to the spread in $J_r$. Nevertheless, comparing panels of the same radial phase at one radial period apart, the end points of each gap across the width of the stream do not shift by any appreciable amount. Rather, the morphologies of the subhalo gaps are already apparent as each gap first appears.
If the linear density of a stream is calculated by integrating the entire thickness of the stream in traversing slices along the stream, then the contrast of the gap will be reduced. This is because the edges of the gaps are not perfectly straight across the width of the stream, so dividing the stream into slices will smear out the density contrast. To investigate how much the smearing will affect gap detection, we calculate the linear densities in two ways. (1) Integrating cylindrical slices of radius 1 kpc along the stream, hereafter the “whole width,” where 1 kpc was chosen to cover the entire thickness of the whole stream. (2) Integrating only the cylindrical slices of radius 0.04 kpc centered along the best fit line of each branch of the stream, hereafter the “central width,” where 0.04 kpc is chosen to mimic the GD-1’s observed width of 0.08 kpc [@CG13]. This central width then encloses about 30%–40% of the mass of the whole width, depending on its orbital phase where, for example, the stream is radially compressed during pericentric passage.
![[Surface density of the smooth stream from 7 to 8 Gyr projected onto the $xy$-plane. The stream is then aligned to Cartesian coordinates]{} where the horizontal axis is the offset position along the stream from the progenitor, and the vertical axis is the radial offset from the galactocentric distance of the progenitor. [This is done by tracing a best fit line along the stream.]{} For each segment in the line, the particles in between the end points of the segment are rotated such that the Galactic center points toward the $+y$ direction in this plot. Note that the vertical axis has been scaled 30 times the larger than the horizontal axis.[]{data-label="fig:stream_maps_smooth"}](figs/diagonal_gap/linearmap_7-8Gyr.eps){width="3.5in"}
![Stream with $5.8{\times10^{6}}<M/{M_\odot}<10^8$ subhalos from 7 to $8\,\mathrm{Gyr}$, [aligned to]{} Cartesian coordinates similar to the smooth stream in [Figure \[fig:stream\_maps\_smooth\]]{}. Compared to the smooth stream, this [$\Lambda$CDM]{} stream shows much more structures at various scales. In general, whether by EOs (mostly inside $\pm5\,\mathrm{kpc}$) or by subhalo perturbations, gaps have complicated morphologies and do not even have straight edges across the width of the stream.[]{data-label="fig:stream_maps_lcdm"}](figs/diagonal_gap/linearmap_7-8Gyr_mass5.8e6_1e8.eps){width="3.5in"}
Gap Counting {#sec:gap_counting}
------------
[Figures \[fig:density\_gaps\_whole\_massrange0\]–\[fig:density\_gaps\_central\_massrange13\]]{} show the densities along the whole and central widths of the smooth stream and a [$\Lambda$CDM]{} stream from 3 to $10\,\mathrm{Gyr}$. The streams younger than $3\,\mathrm{Gyr}$ are not shown as the stream is $\lesssim 10\,\mathrm{kpc}$ long at those ages, so the gaps are dominated by very prominent EOs. Moreover, the stream itself does not yet have a large enough cross section to produce enough gaps for meaningful statistics. In each panel, the shaded columns represent the gaps that are found on the scale of the columns’ widths. Although these gaps are identified as being 99% significant, the density contrasts of the gaps have not been quantified. [For the rest of this paper, we assume that all gaps identified at 99% significance can be observed.]{} Note that because the gap finding process is applied independently to each snapshot, the shaded columns do not necessarily represent the time evolutions of individual gaps. Instead, the shaded columns show the general distributions of gaps – both in space and in gap lengths.
Our gap finding method is has a number of problems. In the smooth stream ([Figures \[fig:density\_gaps\_whole\_massrange0\] and \[fig:density\_gaps\_central\_massrange0\]]{}), our method by construction identifies 1% of the noise as gaps. This is why there can be spurious gaps detected well beyond 5 kpc away from the progenitor, even though EOs tend to form very close to the progenitor. Also, the overall profile of the stream density can sometimes be confused as a gap as well. One example is a gap at $4\,\mathrm{kpc}$ at $5\,\mathrm{Gyr}$ in [Figure \[fig:density\_gaps\_whole\_massrange13\]]{}, where a smooth density gradient from 3 to $8\,\mathrm{kpc}$ is mistaken as the right half of a long gap. At the 95% confidence limit, both kinds of false positives are quite common and can often be identified by eye. For the results below, we show gaps that are 99% significant, which minimizes the occurrence false positives.
![Linear densities along the smooth stream, integrated for the whole thickness of the stream, from 3 to $8\,\mathrm{Gyr}$. The progenitor is centered at $0\,\mathrm{kpc}$ and is masked out. Shaded columns are gaps identified at 99% confidence at the scale depicted by the columns’ widths.[]{data-label="fig:density_gaps_whole_massrange0"}](figs/gap_finding/king12_1M_halo30_v1.5_1_gaps_cl0.99.eps){width="3.5in"}
![Linear densities and gaps identified at 99% in the smooth stream, but integrated for only a cylinder of diameter $0.08\,\mathrm{kpc}$ along the central line of the stream.[]{data-label="fig:density_gaps_central_massrange0"}](figs/gap_finding/king12_1M_halo30_v1.5_1_central0.04_gaps_cl0.99.eps){width="3.5in"}
![Linear densities and gaps identified at 99% in the [$\Lambda$CDM]{} stream with subhalo masses $6.5{\times10^{4}}<M/{M_\odot}<10^8$, integrated for the whole thickness of the stream.[]{data-label="fig:density_gaps_whole_massrange13"}](figs/gap_finding/king12_1M_halo30_v1.5_mass6.5e4_1e8_1_gaps_cl0.99.eps){width="3.5in"}
![Linear densities and gaps identified at 99% in the [$\Lambda$CDM]{} stream with subhalo masses $6.5{\times10^{4}}<M/{M_\odot}<10^8$, integrate for only a cylinder of diameter $0.08\,\mathrm{kpc}$ along the central line of the stream.[]{data-label="fig:density_gaps_central_massrange13"}](figs/gap_finding/king12_1M_halo30_v1.5_mass6.5e4_1e8_1_central0.04_gaps_cl0.99.eps){width="3.5in"}
Gap Spectrum {#sec:gap_size_distribution}
------------
Following the idealized experiments in @carlberg12, @CG13 derived an analytical relation between gap formation rate $R_\cup$ which is the cumulative number of gaps longer than length $l$ per unit stream length per unit time as a function of gap length (hereafter the “gap spectrum”) such that $$R_\cup^{\mathrm{ideal}} = 0.060\, \hat{r}^{0.44}\, l^{-1.16}\, \mathrm{kpc}^{-1}\,\mathrm{Gyr}^{-1}
\label{eqn:gaprate_gapsize_formula}$$ where $\hat{r}\equiv r/30\,\mathrm{kpc}$, and we adapt $r=22\,\mathrm{kpc}$ for the average galactocentric radius of the stream. In this section we aim to study the validity of [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} in our self-consistent stream. We set the “length” of each gap as the scale $s$ of the matched filter which identified the gap (Section \[sec:gapfinding\])
### Smooth Stream without Subhalos
[Figures \[fig:density\_gaps\_whole\_massrange0\] and \[fig:density\_gaps\_central\_massrange0\]]{} show the gaps identified in the smooth stream integrated using the whole thickness and central thickness, respectively. Clearly, the gaps due to EOs are clustered at $\lesssim5\,\mathrm{kpc}$ on both sides of the progenitor, and all the gaps have very similar sizes. The measured $R_\cup$ would peak at short gaps and quickly drops off beyond $l\gtrsim 1\,\mathrm{kpc}$. [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} is meant to describe an idealized gap spectrum produced by subhalos, and not by EOs.
A key result of this study, as discussed below, is that the gap spectrum for subhalo gaps looks very different than the gap spectrum for EOs. The existence of gaps longer than $\sim 1\,\mathrm{kpc}$ would be an indication that processes other than EOs are responsible for the gaps. Furthermore, subhalo gaps can be found everywhere along the whole stream. EOs can only be observed in the immediate vicinity of the progenitor.
### [$\Lambda$CDM]{} Stream with Independent Sets of Subhalos
As an ideal case, [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} ignores the visibility of gaps when the same position of a stream suffers impacts by multiple subhalos at different times. For instance, after one major impact by a massive subhalo which results in a long and high contrast density gap at an early time, subsequent impacts by less massive subhalos in that same region at a later time may not be visible.
Gap overlapping can be minimized by the following experiment. We run 13 separate simulations with the same initial conditions as the star cluster, but the subhalo masses are selected differentially from [Table \[table:subhalo\_massrange\]]{}. This allows each stream to interact with an independent set of subhalos of a very [small]{} range of masses. Overlapping can still occur within the same simulation for each set of subhalos (hence a small number of gaps can still be eliminated), but to a much lesser extent than using integrated mass ranges.
[Figure \[fig:RU\_summed\]]{} shows the measured gap spectrum from the gaps collected from all 13 simulations using independent sets of subhalos. In the top panel which includes all gaps, the measured gap spectrum matches [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} reasonably well. However, this is a coincidence as the gaps contain EOs which are not described by [Equation (\[eqn:gaprate\_gapsize\_formula\])]{}. In an attempt to eliminate EOs, in the bottom panel of [Figure \[fig:RU\_summed\]]{}, the gaps that are located within $5\,\mathrm{kpc}$ away from the progenitor are eliminated. When computing the gap formation rates in these cases, the number of gaps are divided by a stream length which is reduced by $10\,\mathrm{kpc}$ and a stream age which reduced by $2\,\mathrm{Gyr}$ (i.e., the age of the stream when it is $10\,\mathrm{kpc}$ long). This allows us to facilitate a fair comparison of gap spectra against the cases which include all gaps in the entire stream. Comparing the two panels in [Figure \[fig:RU\_summed\]]{}, we can see the masking of the $10\,\mathrm{kpc}$ around the progenitor reduces the abundance of shorter gaps. This is expected since that region of the stream contains mostly EOs which occur at scales $\lesssim 1\,\mathrm{kpc}$. In general it is difficult to tell whether a given gap within $5\,\mathrm{kpc}$ is due to EOs or subhalos, so in the process some subhalo gaps near the progenitor may have been eliminated as well.
Whether [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} is a good description of the gap spectrum in a general stream likely requires more simulations with varying orbital parameters. Nevertheless, it is apparent that the gap spectrum does not depend strongly on the age or the integrating width for the linear density of the stream. However, it is still an ideal case since a stream realistically interacts with all the subhalos at the same time. As we show in the following section, gap overlapping can significantly alter the gap spectrum.
![Analytical [@carlberg12] and simulated gap spectra. The gaps are collected from 13 simulations with independent sets of subhalos selected differentially from the mass ranges in [Table \[table:subhalo\_massrange\]]{}. In the top panel, all gaps identified (including EOs) [are]{} included in the measured gap spectra. In the bottom panel, the gaps located within $5\,\mathrm{kpc}$ away from the progenitor [are]{} eliminated, and the stream lengths and ages [are]{} adjusted. [Since the gaps are mostly independent, the gap spectra follow the analytical prediction reasonably well.]{} []{data-label="fig:RU_summed"}](figs/gaprate_gapsize/gaprate_gapsize_summed_cl0.99_lim0.eps "fig:"){width="3.5in"} ![Analytical [@carlberg12] and simulated gap spectra. The gaps are collected from 13 simulations with independent sets of subhalos selected differentially from the mass ranges in [Table \[table:subhalo\_massrange\]]{}. In the top panel, all gaps identified (including EOs) [are]{} included in the measured gap spectra. In the bottom panel, the gaps located within $5\,\mathrm{kpc}$ away from the progenitor [are]{} eliminated, and the stream lengths and ages [are]{} adjusted. [Since the gaps are mostly independent, the gap spectra follow the analytical prediction reasonably well.]{} []{data-label="fig:RU_summed"}](figs/gaprate_gapsize/gaprate_gapsize_summed_cl0.99_lim5.eps "fig:"){width="3.5in"}
### [$\Lambda$CDM]{} Stream with all Subhalos {#sec:lcdm_stream_with_all_subhalos}
We now consider the validity of [Equation (\[eqn:gaprate\_gapsize\_formula\])]{} for stream gaps in the presence of all subhalos in each cumulative mass range in [Table \[table:subhalo\_massrange\]]{}. [Figure \[fig:RU\_all\]]{} compares the measured gap spectra of both the whole and central streams for three mass ranges, with and without the gaps within $5\,\mathrm{kpc}$ away from the progenitor. Clearly in all cases, the ideal gap spectrum over estimates the measured spectrum by nearly an order of magnitude due to gap overlapping.
Similar to the simulations with independent sets of subhalos ([Figure \[fig:RU\_summed\]]{}), the gap spectra produced by full sets of subhalos do not have strong dependence on stream age and integrating width. The only exception is the youngest stream shown at 5 Gyr which consistently has higher $R_\cup$ than the older streams. However, when the gaps near the progenitor are eliminated, the numbers of gaps at $5\,\mathrm{Gyr}$ in all cases decrease significantly, where the gap spectra are dominated by a single gap at $2-3\,\mathrm{kpc}$, and a number of extremely short gaps. This is likely because the stream is still young, and the effective length of the stream (after masking $10\,\mathrm{kpc}$ centered at the progenitor) is only $\sim15\,\mathrm{kpc}$. This eliminates a significant part of the stream, making its stream statistics unreliable.
The weak dependence of the gap spectrum on the integrating width for linear density is also worth noting. [Figure \[fig:stream\_maps\_lcdm\]]{} shows that gaps in general have much more complicated morphologies than straight edges across the width of the stream. The explanation for these morphologies requires detailed understanding of how subhalo perturbations manifest in a self-consistent stream, which is beyond the scope of this study. While @carlberg13 studied the dynamics of subhalo perturbations for an idealized stream, we defer the self-consistent case to a future study.
Perhaps the most surprising result is that the gap spectra do not show obvious dependence on subhalo masses. The spectra are difficult to distinguish between the mass range of subhalos which causes the gaps. This is in disagreement with @carlberg12 which derived a relation between the length of a gap and the mass of the subhalo that caused it such that $$l(M) = 8.3 \left( \frac{r}{30\,\mathrm{kpc}} \right)^{0.37} \left( \frac{M}{10^8{M_\odot}} \right)^{0.41}\,\mathrm{kpc}.
\label{eqn:length_mass}$$ From this formula, it is reasonable to expect the inclusion of lower-mass subhalos to show more gaps at the shorter end. In their ideal simulations, however, @carlberg12 did not account for the time evolution of each gap. An example can be seen in [Figures \[fig:density\_gaps\_whole\_massrange13\] and \[fig:density\_gaps\_central\_massrange13\]]{}. The gap located at about $-7\,\mathrm{kpc}$ at $4\,\mathrm{Gyr}$ evolves into a much longer gap centered at about $-9\,\mathrm{kpc}$ at $10\,\mathrm{Gyr}$. Evidently [Equation (\[eqn:length\_mass\])]{} requires revision for self-consistent streams before it can be used to understand the relation between gap spectra and subhalo masses.
![Analytical [@carlberg12] and simulated gap spectra. The gaps are collected from single simulations which include the full ranges of subhalos. The lower limit of the mass range is labeled in each panel, and the upper limit is $10^8{M_\odot}$ in all cases. In the top three panels, all gaps identified (including EOs) [are]{} included in gap spectrum. In the bottom three panels, the gaps located within $5\,\mathrm{kpc}$ away from the progenitor [are]{} eliminated, and the stream lengths and ages [are]{} adjusted. [Compared to [Figure \[fig:RU\_summed\]]{}, the simulated gap spectra here no longer follow the analytical prediction because of overlapping between the gaps.]{}[]{data-label="fig:RU_all"}](figs/gaprate_gapsize/gaprate_gapsize_3panels_cl0.99_lim0.eps "fig:"){width="3.5in"} ![Analytical [@carlberg12] and simulated gap spectra. The gaps are collected from single simulations which include the full ranges of subhalos. The lower limit of the mass range is labeled in each panel, and the upper limit is $10^8{M_\odot}$ in all cases. In the top three panels, all gaps identified (including EOs) [are]{} included in gap spectrum. In the bottom three panels, the gaps located within $5\,\mathrm{kpc}$ away from the progenitor [are]{} eliminated, and the stream lengths and ages [are]{} adjusted. [Compared to [Figure \[fig:RU\_summed\]]{}, the simulated gap spectra here no longer follow the analytical prediction because of overlapping between the gaps.]{}[]{data-label="fig:RU_all"}](figs/gaprate_gapsize/gaprate_gapsize_3panels_cl0.99_lim5.eps "fig:"){width="3.5in"}
Observational Considerations {#sec:case_study}
----------------------------
[We now consider the issues when interpreting gap spectra from observations. A gap spectrum for GD-1 has been observed by @CG13, but we emphasize that the gap spectra from our simulated streams in this study should not be directly compared to the one in @CG13 because our models for both the star cluster and the galaxy halo are chosen in favor of a simple interpretation, and may be missing some complications discussed in Section \[sec:other\_effects\].]{}
[We first project each stream onto sky coordinates. For simplicity, we put the hypothetical observer at the center of the galaxy, and then project each particle onto the azimuthal and altitudinal plane in galactocentric coordinates. Since the our stream progenitor is orbiting along the $xy$-plane in a spherical potential, the smooth stream appears as a straight line along the azimuthal direction, and each [$\Lambda$CDM]{} stream appears only a few degrees off the azimuthal plane due to subhalo perturbations.]{}
[The density along the stream is simply the number of particles in bins of $0.1{^\circ}$ in the azimuthal direction. The match filter approach to detect gaps remain the same as the analysis above, but the 12 filter scales ([Figure \[fig:filter\_scales\]]{}) are now logarithmically spaced in angular units from $0.34{^\circ}$ to $14{^\circ}$, and the noise levels are obtained from the regions at $>10{^\circ}$, rather than $>5\,\mathrm{kpc}$, away from the progenitor. The choice of bin size and filter scales are on the same orders of magnitude as @CG13, but putting the hypothetical observer at the Galactic center may affect angular sizes by factors of $\sim2$.]{}
[To ensure that the behaviors of the simulated streams are typical, we simulate each [$\Lambda$CDM]{} stream 10 times with the same initial conditions for the star cluster, but different realizations of the same subhalo distributions. At the end we take the median numbers of gaps of the 10 streams to avoid outliers.]{}
### Orbital Phase
[One surprising result from Section \[sec:lcdm\_stream\_with\_all\_subhalos\] is that the gap spectrum has little to no dependence on age and subhalo masses. To investigate what this means when interpreting observations, the top panel of [Figure \[fig:numgaps\_time\]]{} shows the cumulative numbers of gaps longer than $0.34{^\circ}$ (i.e., all gaps detected in the entire stream) as functions of time. At $t>5\,\mathrm{Gyr}$, the numbers of gaps due to subhalos vary according to the orbital phase of the stream progenitor. The “bursts” in numbers of gaps in the [$\Lambda$CDM]{} streams occur when the streams are stretched as they passes through the pericenters of their orbits. At $t<5\,\mathrm{Gyr}$, on the other hand, this correlation does not exist for two reasons. First, our detection method (Section \[sec:gapfinding\]) uses the parts of the stream that are $>5\,\mathrm{kpc}$ (before sky projection) or $>10{^\circ}$ (after sky projection) away from the progenitor in order to estimate noise. At $t<5\,\mathrm{Gyr}$, the length of the stream varies between a few to $20\,\mathrm{kpc}$, which may not be long enough to estimate noise. Second, in only $5\,\mathrm{Gyr}$ the stream does not yet have enough time and to interact with subhalos.]{}
[Both the total number of gaps and the dynamical age of the stream are difficult to measure, as the some parts of a stream may not be observable. We define a more useful quantity $S_\cup$ which is the cumulative number of gaps longer than a given gap length per unit stream length. In other words, leaving the age of the stream as an unknown, $S_\cup$ differs from $R_\cup$ in Section \[sec:gap\_size\_distribution\] by a normalization by age, and that $S_\cup$ is after sky projection. In the next section we show that gaps due to EOs and subhalos have very different $S_\cup(l)$ distributions which are directly observable. ]{}
![[Top: time evolution of the number of gaps longer than $0.34{^\circ}$. Each colored line represents the median of 10 realizations of an identical subhalo distribution (with $10^8{M_\odot}$ as upper mass limit), and the black thick line represents the smooth stream. Bottom: the vertical dashed lines are visual guides which show that the variation of the number of gaps are correlated with the orbital phase of the stream progenitor. These plots show that the cumulative number of gaps in a stream on average is increasing in time, but instantaneously the number of gaps observable can have an even stronger dependence on the orbital phase of the stream than on the stream’s dynamical age or subhalo masses.]{}[]{data-label="fig:numgaps_time"}](figs/numgaps_time/numgaps_time.eps){width="3.5in"}
### Signal-to-noise Ratios
[Another important distinction between simulations and observations is the signal-to-noise ratio (S/N). At 5 and $8\,\mathrm{Gyr}$ our smooth stream is represented by about 60,000 and 80,000 particles in total, respectively ([Figure \[fig:massloss\]]{}). @koposov10 estimated that the $60{^\circ}$ visible segment of GD-1 consists of 3000 stars. At an average distance of $\sim10\,\mathrm{kpc}$, the visible segment is $\sim10\,\mathrm{kpc}$ long. In our simulations, after 5 and $8\,\mathrm{Gyr}$, the average stream lengths are about 20 and $40\,\mathrm{kpc}$, respectively ([Figure \[fig:massloss\]]{}). This means that our simulated stream should be represented about eight times fewer particles in order to be comparable to observations. With the progenitor masked, we reduce the number of particles in the stream by randomly sampling the stream using two, four, and eight times fewer particles than the original stream. The particle reduction applies to both the stream of interest and the smooth stream which is the source for estimating noise. This allows us to investigate the importance of high S/N.]{}
[Note that in our simulation each particle is equivalent to about $0.043{M_\odot}$, which is less massive than the typical stars that are detected in observations. Our simulations are not meant to be physical models of the real stream. In this section, we are only concerned about matching the numbers of particles in the simulations to the numbers of stars in the observation. As the stars escape from the progenitor, the stream’s self-gravity becomes negligible [@johnston98], and the particles’ masses are no longer important.]{}
[[Figure \[fig:reduction\_gaps\]]{} shows the density profiles of the [$\Lambda$CDM]{} stream with subhalo masses $1.5{\times10^{5}}<M/{M_\odot}<10^8$ at $8\,\mathrm{Gyr}$ projected onto the sky. The panels show the gaps detected in the same stream after three levels of particle reduction. Even after reduction by a factor of eight, the stream appears to have retained most of its gaps despite a lower S/N.]{}
[In [Figure \[fig:GD1\_reduction\]]{}, each line shows the median of ten gap spectra from the same stream but with 10 realizations of the same subhalo distribution. In each panel, the solid (dashed) lines represent the times when the progenitor is at the pericenter (apocenter) of its orbit. When the stream is compressed and stretched as it oscillates radially (see [Figure \[fig:massloss\]]{}), its length can differ by up to a factor of two. Careful inspection of [Figures \[fig:numgaps\_time\] and \[fig:GD1\_reduction\]]{} shows that during pericentric passages, the numbers of gaps are at maximum, but $S_\cup$ is at minimum because the stream length is also at maximum. For a [$\Lambda$CDM]{} stream at high S/N (upper left panel in [Figure \[fig:GD1\_reduction\]]{}), the gap spectra are not sensitive to this oscillation, except with an excess of shorter gaps and fewer longer gaps, which are expected as the stream, including its longitudinal structure, is compressed during apocentric passage. At low S/N (lower left panel in [Figure \[fig:GD1\_reduction\]]{}), however, the gap spectrum during apocentric passage is consistently higher than that during pericentric passage. This is because the length of the stream is insensitive to the S/N, but the number of gaps is not. Therefore, high S/N data for the stream is important when studying stream gaps. Otherwise, the spectrum may be over- or underestimated depending on the orbital phase.]{}
[The right panels in [Figure \[fig:GD1\_reduction\]]{} show the gap spectra of the smooth stream. They are also somewhat sensitive to S/N, but the most obvious difference from the spectra of [$\Lambda$CDM]{} stream is the shape of the spectra. This is especially obvious during pericentric passage where the gap spectra rapidly drop off to zero for gaps longer than $\sim1{^\circ}$. Even during apocentric passage, the gap spectra remain flat at gap lengths $\gtrsim1{^\circ}$. If the gaps originated from subhalo perturbations, then the gap spectrum should be steep and extend well beyond $1{^\circ}$.]{}
![[Density profiles and gaps detected at 99% confidence for the [$\Lambda$CDM]{} stream with subhalos of masses $1.5{\times10^{5}}<M/{M_\odot}<10^8$ at 8 Gyr, projected onto the sky. Each panel represents a stream whose number of particles have been reduced by the factor indicated and binned at $0.1{^\circ}$ throughout. Without reduction (top panel), the stream contains about 80,000 particles. The stream retains most of its gaps even after particle reduction by a factor of eight.]{}[]{data-label="fig:reduction_gaps"}](figs/reduction_gaps/reduction_gaps.eps){width="3.5in"}
{width="6.5in"}
### The Case of GD-1
[ An interesting confusion for GD-1 in particular is that GD-1’s progenitor has not been identified. If GD-1’s progenitor has evaded observation, and the observation corresponds to a segment of the stream which is close enough to the progenitor such that EOs can be observed, then GD-1 may be a poor choice as a probe for missing satellites. However, this is unlikely because EOs are observable only in a small segment of the stream, and subhalo gaps are observable everywhere along the stream. ]{}
[ Another possibility is that GD-1’s progenitor may have been completely disrupted. In this case, the tidal radius of the progenitor approaches zero. Since the spacing between EOs are proportional to the tidal radius of the progenitor, this means that the EOs should also fade away as the progenitor is disrupted [@kupper10]. Therefore, despite its lack of progenitor GD-1 should be a viable probe for missing satellites. ]{}
[ The gap spectrum of GD-1 has been measured by @CG13. The spectrum show presence of gaps at all lengths between $0.2{^\circ}$ and $10{^\circ}$, which is sufficient to rule out a smooth and spherical potential. To facilitate a conclusive analysis on the origin of the gaps in GD-1, we need to consider a much more realistic model which include the effects discussed in Section \[sec:other\_effects\]. ]{}
Subhalo Mass Limits
-------------------
We consider the effects of subhalos as massive as $10^8{M_\odot}$, since the effects by more massive subhalos are not relevant to us. Gaps caused by subhalos at these masses produce long gaps with high-density contrasts. For example, an obvious gap located at $-7\,\mathrm{kpc}$ at 4 Gyr shown in [Figures \[fig:density\_gaps\_whole\_massrange13\] and \[fig:density\_gaps\_central\_massrange13\]]{} are caused by a $4.5{\times10^{7}}{M_\odot}$ subhalo. In fact, the perturbation by $M\gtrsim10^8{M_\odot}$ ($h\gtrsim 1\,\mathrm{kpc}$) subhalos can be so catastrophic that the stream is warped and divided into segments. As a result, a stream which originated from one progenitor can be observed as a few separate streams. Observations of Pal-5’s and GD-1’s gaps, on the other hand, show small-scale density fluctuations in a long, narrow stream, so these two streams are not sensitive to subhalos above $10^8{M_\odot}$. By coincidence, this upper limit approximately coincides with the upper limit beyond which the models of warm dark matter can be no longer be distinguished from CDM. Therefore, $10^8{M_\odot}$ is a reasonable upper limit where our simulations can be useful.
In the low-mass end, we only consider the effects of subhalos down to $\sim 6{\times10^{4}}{M_\odot}$. From a separate simulation of the same stream but with only the subhalos with masses $6.5{\times10^{4}}<M/{M_\odot}<7.5{\times10^{4}}$, the density profile is indistinguishable from the smooth stream, and the gap statistics are identical. Furthermore, Section \[sec:lcdm\_stream\_with\_all\_subhalos\] shows that the gap spectra have very little dependence on mass. Changing the mass lower limit from $7.5{\times10^{4}}{M_\odot}$ to $6.5{\times10^{4}}{M_\odot}$ produced indistinguishable gap spectra. This means that subhalos less massive than $\sim10^5{M_\odot}$, even though they are much more abundant than those of higher masses ([Equation (\[eqn:subhalo\_massfunction\])]{}), have minimal effects on our stream.
[The Milky Way has about 160 known GCs [@harris], and a few hundred DGs brighter than $L\gtrsim10^3 L_\odot$ after bias corrections (see @bullock10 for a review). It is interesting to ask whether these known satellite systems, rather than the truly “missing” satellites, can contribute to the observed stream gaps. Typical GCs have masses $\sim10^5{M_\odot}$, which correspond to the low end of our mass spectrum of subhalos. In the same mass range, though, there are orders of magnitudes more subhalos (e.g., $\sim10^5$ subhalos at $10^5<M/{M_\odot}<10^6$) than GCs, so GCs are unlikely to contribute significantly to observed stream gaps. On the other hand, DGs are commonly found at $\gtrsim10^7{M_\odot}$ [@strigari08] which is the high end of our mass spectrum of subhalos. At that mass range ($\sim2000$ subhalos at $M>10^7{M_\odot}$), the number of known DGs are only 1 order of magnitude below the number of subhalos, so DGs may contribute to some observed gaps. However, a common limitation in understanding the contributions from both GCs and DGs is their orbits, especially when the kinematics of these satellites are not well constrained. As done in our simulations (Section \[sec:subhalo\_mass\_ranges\]), subhalos that do not approach the stream’s orbit will interact minimally with the stream. [Table \[table:subhalo\_massrange\]]{} shows that in our realizations of subhalos, only $\sim15\%$ of them would approach to within $2\,\mathrm{kpc}$ of GD-1’s orbit. This means that most known satellites may never interact with a GD-1-like stream, and that stream gaps, if they were indeed due to satellites and were not EOs, are more likely due to satellites that have never been observed. ]{}
Other Effects {#sec:other_effects}
-------------
In order to keep our results simple, the galaxy is modeled as stationary, spherical NFW potential, the subhalos as static, test masses, and the satellite as a collisionless King model. These models ignore a number of known dynamical complications.
[*Two-body Relaxation.*]{} The star cluster is modeled as a collisionless system with relaxation timescale of $\sim 110\,\mathrm{Gyr}$. Globular clusters typically have relaxation timescales of $\lesssim10\,\mathrm{Gyr}$ [@harris 2010 Edition], so mass loss should originate from dynamical evaporation, in addition to tidal disruption. As a result, the star cluster should be disrupted even faster than we measured in [Figure \[fig:massloss\]]{}. This may have an important effect on the formation of gaps, which depends on the details of the dynamics of a stream [@carlberg13]. The relation between gaps and mass loss mechanism will be investigated in a future study.
[*Dynamical Friction (DF).*]{} Both the star cluster and subhalos should suffer from DF as they orbit around the [dark matter halo]{}. Comparing the magnitudes of the accelerations due to DF and due to the orbit, $a_{DF}/a_{orbit}\sim10^{-8}\ln\Lambda$ for both the star cluster at 22 kpc and a $10^6{M_\odot}$ subhalo at 100 kpc, where $\ln\Lambda\equiv \ln(b_{max}/b_{min})\approx 10$ is the log of the ratio of the maximum and minimum impact distances [@bt08]. Therefore, DF is negligible throughout our model.
[*Disk Shocking:*]{} @dehnen04 found that the evolution of Pal-5 is driven by the tidal shocks when crossing the Galactic disk, which is not modeled in our simulations. The orbit of Pal-5 in @dehnen04 has peri- and apogalacticon at 5.5 kpc and 19 kpc, respectively, whereas our smooth stream has peri- and apogalacticon at 15 kpc and 30 kpc, respectively. Being farther away from the the Galactic center, if our simulations contained a disk, its effect should be less severe for our simulated stream than Pal-5. Moreover, @dehnen04 concluded that disk shocking is not responsible for the observed structure in Pal-5, while @kupper10 concluded that EOs persist even under the influence of disk shocks, so the absence of a disk should not significantly change our conclusion. The gap formation rate with and without subhalos in the presence of a disk is beyond the scope of this study.
[*Halo Shape and Collapse History:*]{} @SV08 found that [the shape of the halo potential can have a larger effect than subhalos have on the over all structure of a stream. However, their simulations focused on streams which originated from DGs at $\sim10^9{M_\odot}$, as well as subhalos at $\gtrsim10^7{M_\odot}$ which is the high end of our mass spectrum of subhalos. The small gaps from a stream originating from a GC in a non-spherical halo has yet to be studied.]{} In fact, since the initial collapse of the entire halo, the potential cannot be stationary throughout a Hubble time, which is the timescale of our simulations. In the future, [we aim to repeat a similar study using potentials which resulted directly from high resolution simulations such as @madau08 and @aquarius. The self-consistent halo and subhalo potentials from those simulations can eliminate the idealized models in Section \[sec:models\].]{}
Conclusions
===========
For the first time, we used N-body simulations to model the disruption of a collisionless star cluster which formed a narrow stream similar to Pal-5 and GD-1, and we investigated the phenomenology of gaps that originated from the perturbations by subhalos predicted in the [$\Lambda$CDM]{} cosmological model. Analytical predictions of stream gap statistics in previous studies were all based on massless particles distributed to mock realistic streams, but the dynamics of gaps have never been studied in self-consistent models. With a stream from a self-consistent model, we characterized the gap length distribution which can be used as a tool to understand the origin of stream gaps seen in observations.
The properties of the subhalos in our simulations were approximations to those in the Aquarius simulation [@aquarius]. We ran 14 simulations of the same stream – 1 without any subhalos (the smooth stream), and 1 for each cumulative mass range in [Table \[table:subhalo\_massrange\]]{} (the [$\Lambda$CDM]{} streams). In each stream we looked for gaps using a matched filter approach previously used by @pal5_gaps and @CG13. We found that, in addition to subhalo perturbations, the overdensities of particles due to their epicyclic motions as the progenitor loses mass [@kupper08] can also produce gaps. Therefore, even without subhalos, “gaps” can appear within $\sim5\,\mathrm{kpc}$ away from the progenitor. [For the first time, our match filter approach accounted for these EOs together with the gaps due to subhalo perturbations.]{}
@yoon11 first noted that subhalos gaps were typically diagonal and not perpendicular to the stream due to the range in angular momenta across the width of the stream. We investigated whether this could be a hindrance to gap detection. By measuring the distribution of angular momenta in our simulated stream, we estimated that the two ends of a gap across the width of a stream were sheared by no more than a $1\,\mathrm{kpc}$ per Gyr. Rather, subhalo gaps show complicated morphologies which were already imprinted into the stream as soon as the [gaps first]{} occurred. In addition to integrating the entire thickness of the stream, we also considered the case where the linear density are integrated using only the central 0.08 kpc of the stream in order to minimize the impact of gaps morphologies. We found that the resulting gap rate spectra the two cases were similar. Therefore, gap morphology does not affect our conclusion.
We tested the validity of the idealized gap spectrum $R_\cup$, or the cumulative number of gaps per unit stream length per unit stream age as a function of gap length $l$, derived by @carlberg12. We found that overlapping gaps in the stream can significantly reduce $R_\cup$, [and that the dependences of $R_\cup$ on subhalo masses and stream age are smaller than its dependence on the stream’s orbital phase. Therefore, the stream’s orbital phase must be known when interpreting gap formation rates in observations.]{}
[We considered how to interpret gap spectra from observations by projecting the stream onto the sky, and for each [$\Lambda$CDM]{} stream we also simulated them using ten realizations of the same subhalo distributions. One observational concern is the S/N of the stream’s detection. We down-sampled our simulated streams with less particles in order to match the S/N which is similar to the GD-1 detection [@CG13]. Our result indicated that at GD-1’s S/N, the gap spectrum can be biased by the orbital phase of the stream. In addition, we compared gap spectra produced purely by EOs and by EOs and subhalos together in a [$\Lambda$CDM]{} halo. We showed that the gap spectra of the former are limited in gap lengths, and that the latter have a much larger variety of gap lengths. This can be a powerful method to identify the origin of gaps in streams. Therefore, high S/N data such as those from [*Gaia*]{} will be very useful for using stream gaps to constraint the abundance of subhalos. ]{}
The dynamics of stream gaps depend on the details of the dynamics of the stream itself. We adapted a few tools such as match filter and scaling relations which were derived from idealized simulations. In a future study, we aim to use self-consistent streams to repeat experiments akin to @yoon11, @carlberg12, and @carlberg13, where the details of individual gaps can be studied in controlled experiments, in order to revise the aforementioned tools that is applicable quantitatively to realistic streams.
[We thank the anonymous referee for valuable comments, which inspired us to expand the content of this paper. Computations were performed on the GPC supercomputer at the SciNet HPC Consortium. SciNet is funded by: the Canada Foundation for Innovation under the auspices of Compute Canada, the Government of Ontario; Ontario Research Fund - Research Excellence, and the University of Toronto.]{}
Bardeen, J. M. Bond, J. R., Kaiser, N., et al. 1986, , 304, 15
Barkana, R., Haiman, Z., & Ostriker, J. P. 2001, , 558, 482
Benson, A. J., Farahi, A., Cole, S. 2013, , 428, 1774
Binney, J. & Tremaine, S. 2008, Galactic Dynamics (2nd ed.; Princeton, NJ: Princeton Univ. Press)
Blumenthal, G. R., Faber, S. M., Primack, J. R., et al. 1984, , 311, 517
Bode, P., Ostriker, J. P., & Turok, N. 2001, , 556, 93
Bullock, J. S. 2010, arXiv:1009.4505
Carlberg, R. G. 2009, , 705, L223
Carlberg, R. G. 2012, , 748, 20
Carlberg, R. G. 2013, , 775, 95
Carlberg, R. G., & Grillmair C. J. 2013, , 768, 171
Carlberg, R. G., Grillmair C. J., & Hetherington N. 2012, , 760, 75
Comparetta, J., & Quillen, A. C. 2011, , 414, 810
Davis, M., Efstathiou, G., Frenk, C. S., et al. 1985, , 292, 371
Dehnen, W., Odenkirchen, M., Grebel, E. K., et al. 2004, , 127, 2753
Eyre, A., & Binney, J. 2011, , 413, 1852
Gao, L., Frenk, C. S., Boylan-Kolchin, M., et al. 2011, , 410, 2309
Grillmair, C. J. 2010, in Galaxies and Their Masks, ed. D. Block, K. C. Freeman, & I. Puerari (New York: Springer), 247
Grillmair C. J., & Dionatos O. 2006, , 643, L17
Harris, W. E. 1996, , 112, 1487
Helmi A., White S. D. M., & Springel, V. 2003, , 339, 834
Hernquist, L. 1990, , 356, 359
Ibata, R. A., Lewis, G. F., Irwin, M. J., et al. 2002, , 332, 915
Johnston, K. V. 1998, , 495, 297
Just, A., Berczik, P., Petrov, M. I., et al. 2009, , 392, 969
Kamionkowski, M., & Liddle, A. 2000, PhRvL, 84, 4525
Klypin, A., Kravtsov, A. V., Valenzuela, O., et al. 1999, , 522, 82
Koposov, S. E., Rix H. W., & Hogg D. W. 2010, , 712, 260
Koposov, S. E., Yoo, J., Rix, H.-W., et al. 2009, , 696, 2179
Küpper, A. H. W., MacLeod, A., & Heggie D. C. 2008, , 387, 1248
Küpper, A.H.W., Kroupa, P., Baumgardt, H., et al. 2010, , 401, 105
Law D. R., Johnston K. V., & Majewski S. R. 2005, , 619, 807
Macciò A. V., Kang, X., Fontanot, F., et al. 2010, , 402, 1995
Madau, P., Diemand, J., & Kuhlen, M. 2008, , 679, 1260
Moore, B., Ghigna, S., Governato, F., et al. 1999, 524, L19
Navarro, J. F., Frenk, C. S., & White S. D. M. 1997, , 490, 493
Navarro, J. F., Hayashi, E., Power, C., et al. 2004, , 349, 1039
Navarro, J. F., Ludlow, A., Springel, V., et al. 2010 , 402, 21
Neto, A. F., Gao, L., Bett, P., et al. 2007, MNRAS, 381, 1450
Odenkirchen, M., Grebel, E. K., Rockosi, C. M., et al. 2001, , 548, L165
Odenkirchen, M., Grebel, E. K., Kayser, A., et al. 2009, , 137, 3378
Perlmutter, S., Aldering, G., Goldhaber, G., et al. 1999, , 517, 565
Planck Collaboration, et al. 2013, arXiv:1303.5062
Riess, A. G., Filippenko, A. V., Challis, P., et al. 1998, , 116, 1009
Schneider, A., Smith, R. E., & Reed, D. 2013, , 433, 1573
Siegal-Gaskins, J. M., & Valluri, M. 2008, , 681, 40
Spergel D. N., & Steinhardt, P. J. 2000, PhRvL, 84, 3760
Springel, V. 2005, MNRAS, 364, 1105
Springel, V., Wang, J., Vogelsberger, M., et al. 2008, , 391, 1685
Stadel, J., Potter, D., Moore, B., et al. 2009, , 398, L21
Strigari, L. E., Bullock, J. S., Kaplinghat, M. 2007, , 669, 676
Strigari, L. E., Bullock, J. S., Kaplinghat, M., et al. 2008, Nature, 454, 1096
Willett, B. A., Newberg, H. Jo, Zhang, H., et al. 2009, , 697, 207
Yoon, J. H., Johnston, K. V., & Hogg, D. W. 2011, , 731, 58
Zemp, M., Diemand, J., Kuhlen, M., et al. 2009, , 394, 641
[^1]: <http://www.mpa-garching.mpg.de/gadget/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A [*CIS*]{} graph is a graph in which every maximal stable set and every maximal clique intersect. A graph is [*well-covered*]{} if all its maximal stable sets are of the same size, [*co-well-covered*]{} if its complement is well-covered, and [*vertex-transitive*]{} if, for every pair of vertices, there exists an automorphism of the graph mapping one to the other. We show that a vertex-transitive graph is CIS if and only if it is well-covered, co-well-covered, and the product of its clique and stability numbers equals its order. A graph is [*irreducible*]{} if no two distinct vertices have the same neighborhood. We classify irreducible well-covered CIS graphs with clique number at most $3$ and vertex-transitive CIS graphs of valency at most $7$, which include an infinite family. We also exhibit an infinite family of vertex-transitive CIS graphs which are not Cayley.
[**Keywords:**]{} CIS graph; well-covered graph; vertex-transitive graph; Cayley graph; maximal stable set; maximum stable set; maximal clique; maximum clique.
(2010): 05C69, 05C25
author:
- Edward Dobson
- Ademir Hujdurović
- Martin Milanič
- Gabriel Verret
bibliography:
- 'CIS-VT-bib.bib'
title: 'Vertex-transitive CIS graphs[^1]'
---
Introduction {#sec:intro}
============
A [*CIS*]{} graph is a graph in which every maximal stable set and every maximal clique intersect (CIS stands for “Cliques Intersect Stable sets”). The study of CIS graphs is rooted in observations of Berge [@Berge] and Grillet [@MR0244112] (see [@MR1344757]). CIS graphs were studied in a series of papers [@AndBorGur; @BorosGM2014; @MR2657704; @MR2489416; @MR2496915; @MR2064873; @MR1344757; @MR2145519; @BorosGM2014+]. The problem of recognizing CIS graphs is believed to be co-NP-complete [@MR1344757], conjectured to be co-NP-complete [@MR2234986], and conjectured to be polynomial [@AndBorGur]. For further background on CIS graphs, see, e.g., [@BorosGM2014].
A graph is called [*vertex-transitive*]{} if, for every pair of vertices, there exists an automorphism of the graph mapping one to the other. Our goal in this paper is to study vertex-transitive CIS graphs.
Our first main result is that a vertex-transitive graph is CIS if and only if it is well-covered, co-well-covered, and the product of its clique and stability numbers equals its order (Theorem \[thm:VT-CIS\]). (A graph is [*well-covered*]{} if all its maximal stable sets are of the same size and [*co-well-covered*]{} if its complement is well-covered.) We then exhibit several infinite families of vertex-transitive CIS graphs (see Section \[sec:examples\]), including some non-Cayley ones (see Proposition \[prop:Gamma\_r\]).
In view of Theorem \[thm:VT-CIS\], we spend some time studying well-covered graphs. (Well-covered graphs were defined by Plummer [@MR0289347] and are well studied in the literature; see, for example, the survey [@MR1254158] and [@MR2739910; @MR2602814; @MR2568844; @MR2537505; @MR2340625; @MR2206373] for some more recent references.) In particular, we give a full classification of irreducible well-covered CIS graphs with clique number at most $3$ (Theorem \[Omega=3\]).
These results are then used to prove our main result: a classification of connected vertex-transitive CIS graphs of valency at most $7$ (Corollary \[cor:VT-CIS-valency-7\]). In particular, we show that there are only finitely many such graphs for valency at most $6$ but there exists an interesting infinite family of examples with valency $7$. In fact, we prove Corollary \[cor:VT-CIS-valency-7\] under an hypothesis slightly weaker than vertex-transitive (see Theorem \[thm:valency-7\]). We conclude the paper with a few open problems (Section \[sec:questions\]).
Preliminaries {#sec:prelim}
=============
All graphs considered are finite, simple and undirected. Let $\Gamma=(V,E)$ be a graph. We call $V$ the vertex set of $\Gamma$ and write $V= V(\Gamma)$. Similarly, we call $E$ the edge set of $\Gamma$ and write $E = E(\Gamma)$. The [*complement*]{} ${\overline{\Gamma}}$ of $\Gamma$ is the graph with the same vertex set and the complementary edge set $\overline E = \{\{x,y\}\,\mid \,x,y\in V,~x\neq y \textrm{ and } \{x,y\}\not\in E\}$.
For a vertex $v\in V$, let $N(v)$ denote the [*neighborhood*]{} of $v$, that is, the set of vertices of $\Gamma$ that are adjacent to $v$. The [*closed neighborhood*]{} of $v$ is the set $N(v)\cup\{v\}$, denoted by $N[v]$. The [*valency*]{} of $v$ is $|N(v)|$ and $\Gamma$ is said to be [*$k$-regular*]{} (or we say that it has *valency* $k$) if all its vertices have valency $k$. A [*universal*]{} vertex of $\Gamma$ is a vertex adjacent to all other vertices of $V(\Gamma)$. If $v$ is a non-isolated vertex of $\Gamma$ then the *local graph of $\Gamma$ at $v$* is the subgraph of $\Gamma$ induced by $N(v)$.
A [*clique*]{} (respectively, a [*stable set*]{}) of $\Gamma$ is [a set]{} of pairwise adjacent (respectively, non-adjacent) vertices. The inclusion maximal cliques and stable sets of $\Gamma$ are called [*maximal*]{}. The maximal cardinality of a clique (respectively a stable set) of $\Gamma$ is called the *clique* (respectively *stability*) number and denoted $\omega(\Gamma)$ (respectively $\alpha(\Gamma)$). We say that $\Gamma$ is [*triangle-free*]{} if $\omega(\Gamma)\leq 2$.
A [*matching*]{} in $\Gamma$ is a set of pairwise disjoint edges. A matching is *perfect* if every vertex is incident with some edge of the matching. The [*line graph*]{} of $\Gamma$, denoted by $L(\Gamma)$, is the graph with vertex set $E(\Gamma)$ and two edges of $\Gamma$ adjacent in $L(\Gamma)$ if and only if they intersect. For positive integers $m$ and $n$, we denote by $n\Gamma$ the disjoint union of $n$ copies of $\Gamma$, by $C_n$ the cycle of order $n$, by $K_n$ the complete graph of order $n$ and by $K_{m,n}$ the complete bipartite graph with parts of size $m$ and $n$. The [*lexicographic product*]{} of graphs $\Gamma_1$ and $\Gamma_2$ is the graph $\Gamma_1[\Gamma_2]$ with vertex set $V(\Gamma_1)\times V(\Gamma_2)$, where two vertices $(u,x)$ and $(v,y)$ are adjacent if and only if either $\{u,v\}\in E(\Gamma_1)$ or $u = v$ and $\{x,y\}\in E(\Gamma_2)$.
Let $G$ be a group and let $S$ be an inverse-closed subset of $G$ such that $1\not\in S$. The [*Cayley graph of $G$ with connection set $S$*]{} has vertex set $G$ with two vertices $g$ and $h$ adjacent if and only if $g^{-1}h\in S$. We say that a graph is *Cayley* if it isomorphic to some Cayley graph. It is well known that a graph is Cayley if and only if its automorphism group contains a subgroup acting regularly (that is, sharply transitively) on vertices (see, e.g., [@Sabidussi1958]). In particular, Cayley graphs are vertex-transitive.
CIS graphs {#sec:prelim-CIS-well-covered}
----------
We first recall some basic properties of CIS graphs (see, e.g., [@BorosGM2014]).
\[lemma:CIS-complements\]
1. A graph is CIS if and only if its complement is CIS.
2. A disconnected graph is CIS if and only if each of its connected component is CIS. \[CIS:disconnected\]
3. For every two graphs $\Gamma_1$ and $\Gamma_2$, the lexicographic product $\Gamma_1[\Gamma_2]$ is CIS if and only if $\Gamma_1$ and $\Gamma_2$ are CIS. \[CIS:lexproduct\]
In view of Proposition \[lemma:CIS-complements\] (\[CIS:disconnected\]), we often restrict ourselves to the study of connected graphs.
Let $\Gamma$ be a graph and let $R$ be the equivalence relation “having the same neighborhood” on $V(\Gamma)$. If $R$ is the identity relation then we say that $\Gamma$ is *irreducible*. It is not hard to see that the quotient graph of $\Gamma$ with respect to $R$ is irreducible. It is called the *irreducible quotient* of $\Gamma$.
It is easy to see that every equivalence class of $R$ is a stable set[^2], and thus, given an irreducible graph $X$ one can easily recover all graphs which have $X$ as an irreducible quotient.
\[irreducibleQuotient\] A graph is CIS if and only if its irreducible quotient is CIS.
Let $\Gamma$ be a graph, let $X$ be its irreducible quotient and let $\pi$ be the natural projection from $\Gamma$ to $X$. It is not hard to see that a maximal clique of $\Gamma$ consists in choosing exactly one representative of each $\pi$-fiber of a maximal clique of $X$. On the other hand, a maximal stable set of $\Gamma$ consists in choosing all elements of each fiber of a maximal stable set of $X$. The result easily follows.
By Proposition \[irreducibleQuotient\] and the comment preceding it, the study of CIS graphs is reduced to the study of irreducible CIS graphs. We now develop some properties of irreducible CIS graphs.
\[coollemma\] Let $\Gamma$ be an irreducible CIS graph. If $\omega(\Gamma)=t$ then no two $t$-cliques of $\Gamma$ intersect in a $(t-1)$-clique.
Suppose, on the contrary, that $C_1$ and $C_2$ are two $t$-cliques of $\Gamma$ such that $|C_1\cap C_2|=t-1$. Let $v_1$ be the unique vertex contained in $C_1$ but not in $C_2$ and let $v_2$ be the unique vertex contained in $C_2$ but not in $C_1$.
Let $x$ be a neighbor of $v_1$. If $x$ is not adjacent to $v_2$ then $\{v_2,x\}$ is contained in some maximal stable set $S$ which must intersect the maximal clique $C_1$, but $x$ is adjacent to $v_1$ and $v_2$ is adjacent to every vertex of $C_1\setminus\{v_1\}$, which is a contradiction. It follows that $x$ is adjacent to $v_2$. We have shown that $N(v_1)\subseteq N(v_2)$ and, by symmetry, we obtain $N(v_1)= N(v_2)$, contradicting the fact that $\Gamma$ is irreducible.
The following are immediate consequences.
\[corollary:localCIS\] Let $\Gamma$ be an irreducible co-well-covered CIS graph with $\omega(\Gamma)=t$, let $v$ be a non-isolated vertex of $\Gamma$ and let $Y$ be the local graph of $\Gamma$ at $v$. Then $Y$ is a co-well-covered graph with $\omega(Y)=t-1$ and with the property that no two $(t-1)$-cliques intersect in a $(t-2)$-clique.
Clearly, $C_Y$ is a clique of $Y$ if and only if $\{v\}\cup C_Y$ is a clique of $\Gamma$ containing $v$. The result then follows from Lemma \[coollemma\].
\[Omega=2\] Let $\Gamma$ be a connected irreducible graph with $\omega(\Gamma)=2$. Then $\Gamma$ is CIS if and only if $\Gamma\cong K_2$.
\[YOYO\] Let $\Gamma$ be a connected graph with $\omega(\Gamma)=2$. Then $\Gamma$ is CIS if and only if $\Gamma\cong K_{n,m}$ for some $n,m\geq 1$.
We remark that Corollary \[YOYO\] can also be easily derived from the following easy observation (see, e.g., [@AndBorGur]).
\[obs:bull\] Let $\Gamma$ be a CIS graph. If $P$ is an induced path of length three in $\Gamma$ then there exists a vertex in $\Gamma$ adjacent to the two midpoints of $P$ and non-adjacent to its endpoints.
Our last result of this section shows that, among connected CIS graphs, the graphs $L(K_{n,n})$ are characterized by their local graphs. This will be needed in later sections.
\[Local TwiceComple\] Let $\Gamma$ be a connected CIS graph such that every local graph of $\Gamma$ is the disjoint union of two complete graphs. Then $\Gamma\cong L(K_{n,n})$ for some $n\geq 1$.
If $\Gamma$ contains an isolated vertex then, by connectedness, we have $\Gamma\cong K_1\cong L(K_{1,1})$. We thus assume that $\Gamma$ has no isolated vertex. Since every local graph of $\Gamma$ consists of two disjoint complete graphs, every vertex is contained in exactly two maximal cliques, and these two cliques intersect in only that vertex. Let $u$ and $v$ be distinct vertices. By the previous observation, there exists a maximal clique containing $u$ but not $v$ and thus $N[u]\neq N[v]$.
It follows that distinct vertices have distinct closed neighborhoods and, by [@KloksKM1995 Theorem 16], $\Gamma$ is the line graph of a triangle-free graph $Z$. Clearly, $Z$ is connected. A stable set in $\Gamma$ corresponds to a matching in $Z$ and a maximal stable set in $\Gamma$ corresponds to a maximal matching in $Z$. Since $Z$ is triangle-free, a clique in $\Gamma$ corresponds to a star in $Z$ (that is, to a set of edges incident with a fixed vertex), and a maximal clique in $\Gamma$ corresponds to an maximal star in $Z$ (that is, to the set of all edges incident with a fixed vertex).
Since $\Gamma$ is CIS, every maximal stable set of $\Gamma$ intersects every maximal clique of $\Gamma$. Consequently, every maximal matching of $Z$ intersects every maximal star of $Z$, which means that every maximal matching of $Z$ is actually a perfect matching. By [@Sumner1979 Theorem 1], it follows that $Z$ is isomorphic to $K_{2n}$ for some $n\ge 2$ or to $K_{n,n}$ for some $n\ge 1$. Since $Z$ is triangle-free, $Z\cong K_{n,n}$ and the result follows.
Vertex-transitive CIS graphs {#sec:VT}
============================
Our first important result is the following characterization of CIS vertex-transitive graphs, which generalizes [@BorosGM2014 Theorem 3].
\[thm:VT-CIS\] Let $\Gamma$ be a vertex-transitive graph of order $n$. Then $\Gamma$ is CIS if and only if all maximal stable sets are of size $\alpha(\Gamma)$, all maximal cliques are of size $\omega(\Gamma)$, and $\alpha(\Gamma)\omega(\Gamma)= n$.
Let $G$ be the automorphism group of $\Gamma$, let $C$ be a maximal clique, and let $S$ be a maximal stable set of $\Gamma$. Denote by $\mathcal{I}(C,S)$ the set of triples $(x,y,g)$ such that $x\in C$, $y\in S$, $g\in G$ and $x^g=y$. Given $(x,y)\in C\times S$, the set of elements of $G$ which map $x$ to $y$ is a coset of $G_x$ and hence has cardinality $|G_x|$. This shows that $|\mathcal{I}(C,S)|=|C||S||G_x|$. For every $g\in G$, $C^g$ is a maximal clique and hence $|C^g\cap S|\leq 1$.
If $\Gamma$ is CIS then $|C^g\cap S|=1$ for every $g\in G$. In particular, for every $g\in G$, there is a unique choice of $(x,y)\in C\times S$ such that $(x,y,g)\in \mathcal{I}(C,S)$. It follows that $|G|=|\mathcal{I}(C,S)|$. By the orbit-stabilizer theorem, we have $|G|=n|G_x|$ and hence $n=|C||S|$. Since the choice of $C$ and $S$ was arbitrary, all maximal stable sets are of size $|S|$ and all maximal cliques are of size $|C|$.
Conversely, if $|C||S|=n$ then $|\mathcal{I}(C,S)|=|C||S||G_x|=n|G_x|=|G|$. Since $|C^g\cap S|\leq 1$ for every $g\in G$, it follows that in fact $|C^g\cap S|=1$ for every $g\in G$. Setting $g=1$, we obtain $|C\cap S|=1$. Since $C$ and $S$ were arbitrary, it follows that $\Gamma$ is CIS.
\[cor:regular-wc-co-wc\] Every vertex-transitive CIS graph is regular, well-covered, and co-well-covered.
Corollary \[cor:regular-wc-co-wc\] suggests that, in order to study vertex-transitive CIS graphs, it can be useful to first study well-covered CIS graphs. This is what we do in Section \[sec:omega-3\] but, first, we construct some infinite families of connected irreducible vertex-transitive CIS graphs.
Examples {#sec:examples}
--------
The proof of the following result is straightforward.
\[prop:LKnn\] If $n\geq 1$ then $L(K_{n,n})$ is a connected vertex-transitive CIS graph of order $n^2$ and valency $2(n-1)$ with $\alpha(L(K_{n,n}))=\omega(L(K_{n,n}))=n$.
For $n\geq 3$, let $PX(n)$ be the graph with vertex-set $\ZZ_n\times\ZZ_2\times\ZZ_2$ and edge-set $\{(i,x,y),(i + 1,y,z)\mid i\in \ZZ_n,~x,y,z\in\ZZ_2\}$. To $PX(n)$, we add the following set of edges $\{(i,x,y),(i,u,v)\mid i\in \ZZ_n,~x,y,u,v\in\ZZ_2,~(x,y)\neq (u,v)\}$ to obtain the graph $Q_n$.
The graphs $PX(n)$ were first studied in [@PraegerX1989] (where they were denoted $C(2,n,2)$). Proposition \[prop:Gamma\_r\] shows that the graphs $Q_n$ are vertex-transitive CIS graphs of valency $7$. It turns out that there are only finitely many other vertex-transitive CIS graphs of valency at most $7$ (see Corollary \[cor:VT-CIS-valency-7\]). Proposition \[prop:Gamma\_r\] also shows that $Q_n$ is not Cayley when $n$ is prime.
\[prop:Gamma\_r\] Let $n\geq 4$. The graph $Q_n$ is a connected vertex-transitive CIS graph of order $4n$ and valency $7$ with $\alpha(Q_n)=n$ and $\omega(Q_n)=4$. Moreover, if $n$ is prime then $Q_n$ is not a Cayley graph.
Clearly, $Q_n$ has order $4n$, valency $7$, and is connected and irreducible. Let $C$ be a clique of $Q_n$. Since vertices that are adjacent in $Q_n$ have first coordinate differing by at most one, it follows that $C$ is contained in the graph induced on vertices having first coordinate either $i$ or $i+1$, for some $i\in \ZZ_n$. It is easy to check that every maximal clique of this graph (which is isomorphic to $\overline{2C_4}$) has size $4$. It follows that $Q_n$ is co-well-covered with $\omega(Q_n)=4$.
Let $S$ be a maximal stable set of $Q_n$. Clearly, for every $i\in\ZZ_n$, $S$ contains at most one vertex with first coordinate $i$. Suppose that, for some $i\in\ZZ_n$, $S$ does not contain a vertex with first coordinate $i$. Without loss of generality, we may assume that $(i-1,a,b),(i+1,c,d)\in S$ for some $a,b,c,d\in \ZZ_2$. Note that $(i,b+1,c+1)$ is adjacent to neither $(i-1,a,b)$ nor $(i+1,c,d)$ and thus to no element of $S$. This contradicts the maximality of $S$. It follows that $Q_n$ is well-covered with $\alpha(Q_n)=n$.
We say that two vertices of $Q_n$ are *related* if they have the same first coordinate. This is clearly an equivalence relation. We denote the corresponding partition of $V(Q_n)$ by $\mathcal{B}$. If $n\geq 5$, it follows by [@PraegerX1989 Lemma 2.8] that every automorphism of $PX(n)$ preserves $\mathcal{B}$. Since the graph obtained from $PX(n)$ by adding an edge between every related pair of vertices is $Q_n$, it follows that every automorphism of $PX(n)$ is an automorphism of $Q_n$. By [@PraegerX1989 Theorem 2.10], $PX(n)$ is vertex-transitive and thus so is $Q_n$. If $n=4$ then, by [@PraegerX1989 Theorem 2.10], $PX(n)$ contains a vertex-transitive group of automorphisms that preserves $\mathcal{B}$ and we argue as before. We have shown that $Q_n$ is vertex-transitive and thus by Theorem \[thm:VT-CIS\], it is CIS.
Call an edge of $Q_n$ *red* if it is contained in two $4$-cliques. It is easy to check that two vertices of $Q_n$ are connected by a red path if and only if they have the same first coordinate. Since red edges are mapped to red edges by automorphisms of $Q_n$, this shows that $\mathcal{B}$ is preserved under the group of automorphism of $Q_n$. On the other hand, if we remove all edges between related vertices, we obtain $PX(n)$ and thus every automorphism of $Q_n$ is also an automorphism of $PX(n)$. Together with the previous paragraph, this implies that if $n\geq 5$ then $Q_n$ and $PX(n)$ have the same automorphism group.
Recall that a graph is Cayley if and only its automorphism group contains a regular subgroup and thus $Q_n$ is Cayley if and only if $PX(n)$ is. On other hand, it was shown in [@MR1250993 Theorem 3] that $PX(n)$ is not Cayley when $n$ is prime. This concludes the proof.
\[prop:R\_n\] Let $n\geq 2$ and let $R_n$ be the Cayley graph on $\ZZ_{2n}\times \ZZ_4$ with connection set [$S=\{(0,1),(0,3),(n,0),(n,2),(2i,2),(2i+1,0)\mid 0\leq i\leq n-1\}$.]{} Then $R_n$ is a connected CIS graph of order $8n$ and valency $2n+3$ with $\alpha(R_n)=2n$ and $\omega(R_n)=4$.
Clearly $R_n$ is of valency $2n+3$ and, since $S$ generates $\ZZ_{2n}\times \ZZ_4$, $R_n$ is connected. Since $R_n$ is vertex-transitive, it suffices by Theorem \[thm:VT-CIS\] to prove that all maximal cliques containing $(0,0)$ are of size $4$ and that all maximal stable sets containing $(0,0)$ are of size $2n$. We will assume that $n$ is even. The argument in the case when $n$ is odd is analogous.
It is not difficult to see that every maximal clique containing $(0,0)$ is one of the following: $\{(0,0),(0,1),(0,2),(0,3)\}$, $\{(0,0),(2i+1,0),(n+2i+1,0),(n,0)\}$ or $\{(0,0),(2i,2),(n+2i,2),(n,0)\}$, for some $i\in \{0,1,\ldots,n-1\}$. Therefore, all maximal cliques containing $(0,0)$ are of size $4$.
Let $M$ be a maximal stable set containing $(0,0)$. Note that $M\cap S=\emptyset$. Suppose first that $M$ is contained in $\ZZ_{2n}\times \{0,2\}$. Since $(0,0)\in M$ and $M$ is a stable set, $M$ is contained in $\bigcup_{i\in \{0,1,\ldots, n-1\}}\{(2i,0),(2i+1,2)\}$. If $(2j,0)\in M$ for some $j\in \{1\ldots,n-1\}$ then $(2j+ n,1)$ has no neighbor in $M$, contradicting the fact that $M$ is a maximal stable set. Similarly, if $(2j+1,2)\in M$ for some $j\in \{0,1\ldots,n-1\}$ then $(n+2j+1,1)$ has no neighbor in $M$, again a contradiction.
We may thus assume that $M$ contains some element $(x,y)$ with $y\in\{1,3\}$. We will deal with the case when $y=1$ and $x=2k-1$ is odd. The other cases can be dealt with similarly. Let $X=\ZZ_{2n}\times \ZZ_4\setminus (N[(0,0)] \cup N[(2k-1,1)])$. Note that $M\setminus \{(0,0),(2k-1,1)\}\subseteq X$ and that $$X=
\left(\bigcup_{\substack{i\in \{1,3,\ldots,2n-1\}\setminus\{2k-1\}\\j\in \{2,4,\ldots,2n-2\}}}
\{(j,0),(i,1),(i,2),(j,3)\}\right)\setminus \{(n,0),(n+2k-1,1)\}$$ Thus, the subgraph of $R_n$ induced by $X$ is the disjoint union of two isolated vertices ($(n,3)$ and $(n+2k-1,2)$) and $(n-2)$ cycles of length $4$ (of the form $((i,1),(i,2),(n+i,2),(n+i,1))$ for $i\in \{1,3,\ldots,2n-1\}\setminus\{2k-1\}$ or $((j,0),(j,3),(n+j,3),(n+j,0))$ for $j\in \{2,4,\ldots,2n-2\}$). In particular, a maximal stable set in this induced subgraph has size $2(n-2)+2=2n-2$ and thus $|M|=2n$. This concludes the proof.
\[prop:S\_n\] Let $n\geq 2$ and let $S_n$ be the Cayley graph on $\ZZ_{2n}\times \ZZ_4$ with connection set $S=\{(0,1),(0,3),(2i+1,0),(2i+1,1),(2i+1,3)\mid 0\leq i \leq n-1\}$. Then $S_n$ is a connected CIS graph of order $8n$ and valency $3n+2$ with $\alpha(S_n)=2n$ and $\omega(S_n)=4$.
The proof of Proposition \[prop:S\_n\] is very similar to the proof of Proposition \[prop:R\_n\] and is omitted. Note that $R_2\cong Q_4\cong \overline{S_2}$. Using Gordon Royle’s table of vertex-transitive graphs of order at most $32$ [@Royleweb], we obtain the following with the help of computer.
\[prop:up-to-32\] Let $\mathcal{F}$ be the family containing the following graphs:
1. $K_n$, $n\geq 1$,
2. $L(K_{n,n})$, $n\geq 3$,
3. $Q_n$, $n\geq 4$,
4. $R_n$, $n\geq 3$,
5. $S_n$, $n\geq 3$,
and let $\overline{\mathcal{F}}$ be the closure of $\mathcal{F}$ under the operations of taking complements and lexicographic products. Then, up to isomorphism, every vertex-transitive CIS graph of order at most $32$ is in $\overline{\mathcal{F}}$.
Well-covered CIS graphs with clique number at most $3$ {#sec:omega-3}
======================================================
As we saw in Section \[sec:prelim-CIS-well-covered\], when studying CIS graphs, we may restrict our attention to connected irreducible graphs. The main result of this section is Theorem \[Omega=3\], a classification of connected irreducible well-covered CIS graphs with clique number at most $3$. First, we need a definition and a lemma.
Let $\Gamma$ be a graph and $v$ a non-isolated vertex in $\Gamma$. We denote by $\rho_\Gamma(v)$ the minimum of $\{|S\cap N(v)|\mid S$ is a maximal stable set of $\Gamma$ not containing $v\}$ and by $\rho(\Gamma)$ the minimum of $\rho_\Gamma(v)$ as $v$ ranges over all non-isolated vertices of $\Gamma$.
\[lemma:intersection\] Let $\Gamma$ be a well-covered graph with no isolated vertex. Let $S$ be a maximal stable set in $\Gamma$, let $v\in V(\Gamma)\setminus S$ be such that $|S\cap N(v)|=\rho(\Gamma)$ and let $X=\{v\}\cup (S\setminus N(v))$. Note that $X$ is a stable set. Let $S'$ be a maximal stable set containing $X$ and let $W=S'\setminus (S\cup\{v\})$. Then $|S'\setminus S|=\rho(\Gamma)$, $|W|=\rho(\Gamma)-1$, $W\cap N[v]=\emptyset$ and every vertex of $W$ is adjacent to every vertex of $S\cap N(v)$.
Note that $S\setminus S'=S\cap N(v)$ and, since every maximal stable set has cardinality $|S|$, $|S'\setminus S|=|S\setminus S'|=|S\cap N(v)|=\rho(\Gamma)$ and thus $|W|=\rho(\Gamma)-1$. Since $S'$ is a stable set, we have $W\cap N[v]=\emptyset$. Let $w\in W$. By definition of $\rho(\Gamma)$, $w$ has at least $\rho(\Gamma)$ neighbors in $S$, hence they must be exactly the elements of $S\setminus S'=S\cap N(v)$.
\[Omega=3\] Let $\Gamma$ be a connected irreducible well-covered CIS graph with $\omega(\Gamma)\leq 3$. Then $\Gamma$ is isomorphic to $K_1$, $K_2$, $K_3$ or $L(K_{3,3})$.
If $\omega(\Gamma)\leq 2$ then, by Corollary \[Omega=2\], $\Gamma$ is isomorphic to $K_1$ or $K_2$. We thus assume that $\omega(\Gamma)=3$.
We first show that every edge of $\Gamma$ is contained in a triangle. Since $\omega(\Gamma)=3$ and $\Gamma$ is connected, it suffices to show that every edge that intersects a triangle is itself contained in a triangle. Let $T:=\{v,w,z\}$ be a triangle and let $x$ be a neighbor of $v$ not contained in $T$. Suppose that $\{x,v\}$ is not contained in a triangle. In particular, $x$ is adjacent to neither $w$ or $z$. Suppose that $x$ has a neighbor $y$ different than $v$. Since $\{x,v\}$ is not contained in a triangle, $y$ is not adjacent to $v$ and it follows by Observation \[obs:bull\] that neither $(y,x,v,w)$ nor $(y,x,v,z)$ is an induced path of length three. In particular, $y$ is adjacent to both $w$ and $z$. This implies that the edge $\{w,z\}$ is contained in two distinct triangles, contradicting Lemma \[coollemma\]. We may thus assume that $v$ is the unique neighbor of $x$. Let $S$ be a maximal stable set of $\Gamma$ containing $v$ and let $S' = (S\cup \{x\})\setminus \{v\}$. Clearly, $S'$ is a stable set of $\Gamma$ with $|S'| = |S|$ and, since $\Gamma$ is well-covered, it is a maximal stable set. On the other hand, $S'$ does not intersect the maximal clique $T$, contradicting the fact that $\Gamma$ is CIS. This concludes the proof that that every edge of $\Gamma$ is contained in a triangle and, in fact, in a unique triangle by Lemma \[coollemma\]. In particular, every vertex has even valency.
Let $v$ be a vertex of minimal valency in $\Gamma$ and let $r$ be the number of triangles containing $v$. Since every edge is contained in a unique triangle, $v$ has valency $2r$. Since $\Gamma$ is CIS, it is clear that $\rho_\Gamma(v)=r=\rho(\Gamma)$.
Let $u$ be a neighbor of $v$ and let $S$ be a maximal stable set of $\Gamma$ containing $u$ (and thus not $v$). Note that $|S\cap N(v)|=r$ and thus Lemma \[lemma:intersection\] implies that there exists a stable set $S'$ containing $\{v\}\cup (S\setminus N(v))$ with $|S'\setminus S|=r$. Let $T$ be a triangle containing $u$. Since $\Gamma$ is CIS, $T\cap S'\neq\emptyset$. As $u\in S\setminus S'$, it follows that $(T\setminus\{u\})\cap (S'\setminus S)\neq \emptyset$. In particular, the number of triangles containing $u$ is at most $|S'\setminus S|=r$ and thus $u$ has valency $2r$. In particular, every neighbor of a vertex with minimal valency is also of minimal valency. Connectedness of $\Gamma$ implies that $\Gamma$ is $2r$-regular.
Suppose that $r\geq 3$ and let $N(v)=\{a_1,\ldots,a_r,b_1,\ldots,b_r\}$ such that $a_i$ is adjacent with $b_i$. Let $S$ be a maximal stable set in $\Gamma$ containing $\{a_1,\ldots,a_r\}$. By Lemma \[lemma:intersection\], there exists a set $W_1$ of $r-1$ vertices such that $W_1\cap N[v]=\emptyset$ but every vertex of $W_1$ is adjacent to every vertex of $\{a_1,\ldots,a_r\}$. By the same argument, there exists a set $W_2$ of $r-1$ vertices such that $W_2\cap N[v]=\emptyset$ but every vertex of $W_2$ is adjacent to every vertex of $\{a_1,\ldots,a_{r-1},b_r\}$, and a set $W_3$ of $r-1$ vertices such that $W_3\cap N[v]=\emptyset$ but every vertex of $W_3$ is adjacent to every vertex of $\{a_1,\ldots, a_{r-2},b_{r-1},b_r\}$.
Note that $W_1\cap W_2=\emptyset$ otherwise the edge $\{a_r,b_r\}$ would be contained in two distinct triangles. Note also that every element of $(W_1\cup W_2)\cap W_3$ forms a triangle with $\{a_{r-1},b_{r-1}\}$. Since $v\notin(W_1\cup W_2)\cap W_3$ and every edge is in a unique triangle, it follows that $(W_1\cup W_2)\cap W_3=\emptyset$. This implies that $|W_1\cup W_2 \cup W_3|=3(r-1)$. Finally, note that $W_1,W_2,W_3,\{b_1,v\}\subseteq N(a_1)$, but neither $b_1$ nor $v$ are contained in $W_1\cup W_2 \cup W_3$. It follows that $a_1$ has valency at least $3(r-1)+2=3r-1$. This implies $k=2r\geq 3r-1$, contradicting the fact that $r\geq 3$.
We may thus assume that $r\leq 2$. If $r=1$ then clearly $\Gamma\cong K_3$. If $r=2$ then it follows from Theorem \[Local TwiceComple\] that $\Gamma\cong L(K_{3,3})$.
Together with Theorem \[thm:VT-CIS\], Theorem \[Omega=3\] immediately implies that a connected irreducible vertex-transitive CIS graph with clique number at most $3$ is isomorphic to $K_1$, $K_2$, $K_3$ or $L(K_{3,3})$. In contrast, if the clique number is $4$ then there are many infinite families of examples, for example $Q_n$, $R_n$ and $S_n$ (see Section \[sec:examples\]).
Vertex-transitive CIS graphs of valency at most $7$ {#sec:valency}
===================================================
In this section, we classify vertex-transitive CIS graphs of valency at most $7$. We will need the following preliminary results.
\[UniversalVertex\] Let $\Gamma$ be a $k$-regular graph such that every local graph of $\Gamma$ has exactly $n$ universal vertices. Then $\Gamma=Z[K_{n+1}]$ for some graph $Z$ of valency $\frac{k-n}{n+1}$.
Let $R$ be the equivalence relation “having the same closed neighborhood” on $V(\Gamma)$ and let $Z$ be the quotient graph of $\Gamma$ with respect to $R$. It is easy to see that $\Gamma=Z[K_{n+1}]$.
\[lemma:new\] Let $\Gamma$ be a connected, $k$-regular, well-covered, co-well-covered graph. Then either $\omega(\Gamma)\leq \frac{2}{3}(k+1)$ or $\Gamma$ is a complete graph.
Suppose that $\omega(\Gamma)> \frac{2}{3}(k+1)$. Let $\Gamma_{\mathcal{Q}}$ be the graph of maximal cliques of $\Gamma$, that is, the graph with maximal cliques of $\Gamma$ as vertices, and two such cliques adjacent in $\Gamma_{\mathcal{Q}}$ if they intersect in $\Gamma$. Since $\Gamma$ is co-well-covered and connected, it easily follows that $\Gamma_{\mathcal{Q}}$ is connected. By [@CranstonRarXiv Lemma 2.2], this implies that $\Gamma$ has a vertex that is contained in every maximal clique. Since $\Gamma$ is co-well-covered, this vertex is a universal vertex. In particular, there exists a maximal stable set of cardinality one. As $\Gamma$ is well-covered, every maximal stable set has cardinality one and thus $\Gamma$ is a complete graph.
The graphs in Figure \[fig:localgraphs\] will play an important role in the proof of Theorem \[thm:valency-7\].
[0.45]{} ![Local graphs for Theorem \[thm:valency-7\][]{data-label="fig:localgraphs"}](T_2 "fig:"){height="0.8in"}
[0.45]{} ![Local graphs for Theorem \[thm:valency-7\][]{data-label="fig:localgraphs"}](T_3 "fig:"){height="0.8in"}
[0.45]{} ![Local graphs for Theorem \[thm:valency-7\][]{data-label="fig:localgraphs"}](TT_3 "fig:"){height="1.2in"}
[0.45]{} ![Local graphs for Theorem \[thm:valency-7\][]{data-label="fig:localgraphs"}](U_2 "fig:"){height="1.1in"}
We leave the following lemma as an exercise for the reader.
\[lemma:local\] Let $Y$ be a co-well-covered graph of order $k$ with $\omega(Y)=t$ and with the property that no two $t$-cliques intersect in a $(t-1)$-clique.
1. If $(k,t)=(5,3)$ then $Y\cong T_2$.
2. If $(k,t)=(6,3)$ then $Y\cong 2K_3$.
3. If $(k,t)=(7,3)$ then $Y\cong T_3$ or $T_3'$.
4. If $(k,t)=(7,4)$ then $Y\cong U_2$.
\[thm:valency-7\] Let $\Gamma$ be a connected, $k$-regular, irreducible, well-covered, and co-well-covered CIS graph. If $k\leq 7$ then $\Gamma$ is either a complete graph or isomorphic to one of $L(K_{3,3})$, $L(K_{4,4})$, $C_4[K_2]$, $K_{3,3}[K_2]$ or $Q_n$ for some $n\geq 4$.
By Theorem \[Omega=3\] and Lemma \[lemma:new\], we may assume that $(k,\omega(\Gamma))$ is one of the following pairs : $(5,4)$, $(6,4)$, $(7,4)$, or $(7,5)$ (otherwise $\Gamma$ is either complete or isomorphic to $L(K_{3,3})$). Let $t=\omega(\Gamma)-1$ and let $Y$ be a local graph of $\Gamma$. By Corollary \[corollary:localCIS\], $Y$ is a co-well-covered graph of order $k$ with $\omega(Y)=t$ and with the property that no two $t$-cliques intersect in a $(t-1)$-clique, and we may apply Lemma \[lemma:local\].
If $k=5$ then every local graph is isomorphic to $T_2$. By Lemma \[UniversalVertex\], $\Gamma=Z[K_2]$ for some $2$-regular graph $Z$. Since $\Gamma$ is not complete, $Z\cong C_n$ for some $n\geq 4$. By Proposition \[lemma:CIS-complements\] (\[CIS:lexproduct\]), $Z$ must be CIS hence $n=4$ and $\Gamma\cong C_4[K_2]$. If $k=6$ then every local graph is isomorphic to $2K_3$ and, by Theorem \[Local TwiceComple\], $\Gamma\cong L(K_{4,4})$.
From now on, we assume that $k=7$. If $t=4$ then, by Lemma \[lemma:local\], every local graph is isomorphic to $U_2$. Since $U_2$ contains a unique universal vertex, Lemma \[UniversalVertex\] implies that $\Gamma\cong Z[K_2]$ for some graph $Z$. It follows that $\omega(\Gamma) = 2\cdot\omega(Z)$ is even, contradicting the fact that $\omega(\Gamma) = t+1 = 5$.
From now on, we assume that $t=3$ and thus $\omega(\Gamma)=4$. By Lemma \[lemma:local\], every local graph of $\Gamma$ is isomorphic to $T_3$ or $T_3'$. We will show that $\Gamma$ is either isomorphic to $K_{3,3}[K_2]$ or $Q_n$ for some $n\geq 4$. This case is by far the hardest; we thus break up the proof into a series of claims.
<span style="font-variant:small-caps;">Claim 1.</span> Let $C$, $C_1$, and $C_2$ be three pairwise distinct $4$-cliques of $\Gamma$ such that $C_1\cap C_2=\emptyset$ and $C\subseteq C_1\cup C_2$. Then $|C\cap C_1|=|C\cap C_2|=2$ and every vertex of $C_1\setminus C$ is adjacent to every vertex of $C_2\setminus C$.
[*Proof.*]{} Recall that, by Lemma \[coollemma\], distinct $4$-cliques of $\Gamma$ intersect in at most an edge and thus $|C\cap C_1|=|C\cap C_2|=2$. Let $x_1\in C_1\setminus C$ and let $x_2\in C_2\setminus C$. If $x_1$ and $x_2$ are not adjacent then they are part of a maximal stable set $S$. Since $C=(C_1\cap C)\cup(C_2\cap C)$, every vertex of $C$ is adjacent to one of $x_1$ or $x_2$. It follows that $S$ does not intersect $C$, contradicting the fact that $\Gamma$ is CIS. $_\blacksquare$
We call an edge of $\Gamma$ *red* if it is contained in at least two different $4$-cliques. If $\{u,v\}$ is a red edge, we will say that $u$ is a *red* neighbor of $v$.
<span style="font-variant:small-caps;">Claim 2.</span> Let $C$ be a $4$-clique containing two disjoint red edges $e_1$ and $e_2$. For every $i\in\{1,2\}$, let $C_i$ be a $4$-clique distinct from $C$ containing $e_i$. Then $C_1\cap C_2 = \emptyset$ and every vertex of $C_1\setminus C$ is adjacent to every vertex of $C_2\setminus C$. In particular, $C_i\setminus C$ is a red edge.
[*Proof.*]{} Since distinct $4$-cliques intersect in at most an edge, we have $C_i\cap C=e_i$ and thus $|C_i\cap C|=2$ and $C=e_1\cup e_2\subseteq C_1\cup C_2$. Let $e_1=\{v,w\}$. In the local graph at $v$, $C$ and $C_1$ project to two triangles intersecting at the vertex $w$. In particular, there are no edges between $C_1\setminus C$ and $C\setminus C_1=C_2\cap C$. This implies that $C_1\cap C_2=\emptyset$. By Claim 1, every vertex of $C_1\setminus C$ is adjacent to every vertex of $C_2\setminus C$. It follows that $(C_1\setminus C)\cup (C_2\setminus C)$ is a $4$-clique and thus $C_i\setminus C$ is a red edge, being contained both in $C_i$ and in $(C_1\setminus C)\cup (C_2\setminus C)$. $_\blacksquare$
<span style="font-variant:small-caps;">Claim 3.</span> If $\{u,v\}$ is an edge then the local graph at $u$ is isomorphic to the local graph at $v$.
[*Proof.*]{} Suppose otherwise, that is, without loss of generality, that the local graph at $u$ is isomorphic to $T_3$ and the local graph at $v$ is isomorphic to $T_3'$. Since $T_3'$ has a universal vertex, $v$ has a unique neighbor $v'$ such that $N[v]=N[v']$. Clearly, the local graph at $v'$ is also isomorphic to $T_3'$ and thus $v'\neq u$. Note that $v'$ is the unique red neighbor of $v$ and thus $\{u,v\}$ is not a red edge. By the same reasoning, $\{u,v'\}$ is not red either. Let $C$ be the unique $4$-clique containing $\{u,v\}$. Note that $v'\in C$ and write $C=\{u,v,v',w\}$. Since neither $v$ nor $v'$ are red neighbors of $u$ and the local graph at $u$ is isomorphic to $T_3$, it follows that $\{u,w\}$ is a red edge. Let $e_1=\{v,v'\}$ and $e_2=\{u,w\}$. There exist $C_1$ and $C_1'$, two $4$-cliques distinct from $C$ such that $C_1\cap C_1'=e_1$, and $C_2$ a $4$-clique distinct from $C$ and containing $e_2$. By Claim 2, every vertex of $C_2\setminus C$ is adjacent to every vertex of $(C_1\cup C_1')\setminus e_1$.
Let $z$ be the unique red neighbour of $u$ different from $w$. Note that $z\in C_2\setminus C$ and thus the four vertices of $(C_1\cup C_1')\setminus e_1$ are adjacent to $z$. On the other hand, by examining $T_3$, we see that $|N(z)\cap N(u)|=4$. Since the local graph at $v$ is isomorphic to $T_3'$, $u$ is not adjacent to any vertex in $(C_1\cup C_1')\setminus e_1$ and thus $N(u)$ is disjoint from $(C_1\cup C_1')\setminus e_1$. It follows that $|N(z)|\geq 8$, which is a contradiction. $_\blacksquare$
Since $\Gamma$ is connected, Claim 3 implies that all local graphs of $\Gamma$ are pairwise isomorphic. The case when they are all isomorphic to $T_3'$ is easy to deal with. Indeed, this implies by Lemma \[UniversalVertex\] that $\Gamma=Z[K_2]$ for some connected $3$-regular graph $Z$ and it is easy to check that we must have $\omega(Z)=2$. By Proposition \[lemma:CIS-complements\] (\[CIS:lexproduct\]), $Z$ must be a CIS graph. By Corollary \[YOYO\], $Z\cong K_{3,3}$ and consequently $\Gamma\cong K_{3,3}[K_2]$.
From now on, we assume that all local graphs of $\Gamma$ are isomorphic to $T_3$. In particular, every red edge is part of exactly two $4$-cliques, and every vertex has precisely two red neighbors (the vertices of valency $4$ in its local graph) and thus is incident to exactly two red edges. Let $v$ be a vertex. Starting from the local graph at $v$ and removing the two red neighbors of $v$, we obtain a graph with five vertices and two edges. We call these two edges *extremal (with respect to $v$)*.
<span style="font-variant:small-caps;">Claim 4.</span> Every extremal edge is red.
[*Proof.*]{} Note that each red edge appears in exactly four local graphs. On the other hand, in the local graph at a vertex, red neighbors are incident to at most one red edge and thus the local graph contains at most four red edges. Since the number of red edges is equal to the number of vertices, every local graph contains exactly four red edges, two of which must include the extremal edges. $_\blacksquare$
<span style="font-variant:small-caps;">Claim 5.</span> Every red $2$-path is contained in a unique red cycle, which has length $4$. Moreover, the induced graph on this $4$-cycle is a clique.
[*Proof.*]{} Let $P=(u,v,w)$ be a red $2$-path. Clearly, $u$ and $w$ are the only two red neighbors of $v$. By considering the local graph at $v$, it follows that $u$ and $w$ are adjacent and there is a unique vertex, say $x$, adjacent to each of $u$, $v$ and $w$. Now, $\{u,v,w\}$ is a triangle in the local graph at $x$ therefore it contains a red neighbor of $x$. Since $x\notin V(P)$ and $u$ and $w$ are the only two red neighbors of $v$, we may assume without loss of generality that $w$ is a red neighbor of $x$.
Let $e_1=\{v,w\}$ and let $C_1=\{u,v,w,x\}$. Note that $C_1$ is a $4$-clique. Let $C$ be the unique $4$-clique distinct from $C_1$ and containing the edge $e_1$. Note that $C\setminus C_1$ is an extremal edge with respect to $v$ hence it must be red by Claim 4. By Claim 2, it follows that $C_1\setminus C=\{u,x\}$ must be red, concluding the proof of this claim. $_\blacksquare$
We call a $4$-clique *red* if it contains a red cycle, and *black* otherwise.
<span style="font-variant:small-caps;">Claim 6.</span> Every black clique contains exactly two red edges and these are disjoint.
[*Proof.*]{} Let $C$ be a black clique. By Claim 5, $C$ does not contain a red $2$-path and thus it suffices to show that $C$ contains at least two red edges. Let $v\in C$. Suppose first that $C$ contains no edge that is extremal with respect to $v$. By examining $T_3$, we see that this implies that $C$ contains two red neighbours of $v$ and thus a red $2$-path, which is a contradiction. We may thus assume that $C$ contains an extremal edge with respect to $v$, say $e$. By examining $T_3$, we see that $C=e\cup\{v,w\}$ for some red neighbor $w$ of $v$. By Claim 4, $e$ is red and thus $C$ contains at least two red edges. $_\blacksquare$
Let $C$ be a red clique and let $e_1$, $e_2$, $e_3$, and $e_4$ be the four red edges of $C$, labeled such that $e_1$ and $e_2$ are disjoint (and thus also $e_3$ and $e_4$). For every $i\in\{1,2,3,4\}$, there is a unique $4$-clique containing $e_i$ and distinct from $C$, say $C_i$. Clearly, $C_i$ is black and thus, by Claim $6$, it contains a unique red edge distinct from $e_i$, say $f_i$. Note that $e_i\cap f_i=\emptyset$. Now, $f_i$ is red and thus it is contained in a red clique, say $R_i$. Since every red edge is contained in exactly two $4$ cliques, it follows that $R_i$ is the unique clique distinct from $C_i$ and containing $f_i$. By Claim 2, every vertex of $f_1$ is adjacent to every vertex of $f_2$. In particular, $f_1\cup f_2$ is a $4$-clique and thus $R_1=R_2$. By the same reasoning, we have $R_3=R_4$. Note that $f_1$ and $f_3$ are extremal edges of the local graph at the vertex $e_1\cap e_3$. In particular, $f_1$ and $f_3$ are disjoint and there are no edges between them. It follows that $R_3\neq R_1$.
Note that $R_i$ was uniquely determined by $C$ and $e_i$. We call $R_i$ the red clique *adjacent to $C$ at $e_i$*. Note that, in fact, $C$ is adjacent to $R_i$ at $f_i$, hence the adjacency relation is symmetric. Moreover, we have shown in the paragraph above that every red clique is adjacent to exactly two red cliques.
In particular, there is a set of $n$ red cliques indexed by $\ZZ_n$ with $n\geq 3$ such that $C_i$ is adjacent to $C_{i+1}$. It is not hard to check that $\Gamma$ being connected implies that $\bigcup_{i\in\ZZ_n} C_i=V(\Gamma)$.
Now, for every $i\in \ZZ_n$, there are exactly two edges, say $e_{i,0}$ and $e_{i,1}$ of $C_i$ such that $C_{i-1}$ is adjacent to $C_i$ at $e_{i,0}$ and also at $e_{i,1}$. Note that $C_i=e_{i,0}\cup e_{i,1}$.
We define a mapping $\varphi:\bigcup_{i\in\ZZ_n} C_i\to \ZZ_n\times\ZZ_2\times\ZZ_2$ by $\varphi(v)=(i,x,y)$ such that $v$ is contained in $e_{i,x}$ and adjacent to $e_{i+1,y}$. Now we will show that $\varphi$ is a well-defined bijection. Indeed, if $v\in \bigcup_{i\in\ZZ_n} C_i$ then there is a unique $i\in \ZZ_n$ such that $v\in C_i$. Since $C_i$ is the disjoint union of $e_{i,0}$ and $e_{i,1}$, there is a unique $x\in \ZZ_2$ such that $v\in e_{i,x}$. Let $f$ be the unique red edge containing $v$ distinct from $e_{i,x}$. Note that the red clique adjacent to $C_i$ at $e_{i,x}$ is $C_{i-1}$ and thus the red clique adjacent to $C_i$ at $f$ is $C_{i+1}$. It follows that there is a unique $y\in \ZZ_2$ such that $v$ is adjacent to $e_{i+1,y}$.
We now show that $\varphi$ is an isomorphism between $\Gamma$ and $Q_n$. Let $\{v,w\}\in E(\Gamma)$ and let $\varphi(v)=(i,x,y)$. Note that $N[v] \subseteq C_{i-1}\cup C_i \cup C_{i+1}$. If $w\in C_i$ then $\varphi(v)$ and $\varphi(w)$ have the same first coordinate, hence $\varphi(\{v,w\})\in E(Q_n)$. Suppose now that $w\not\in C_{i}$. Without loss of generality, we may assume that $w\in C_{i+1}$. Note that $v\in e_{i,x}$ and $w\in e_{i+1,y}$ and thus $\varphi(w)=(i+1,y,z)$ for some $z\in \ZZ_2$. This shows that $\varphi(\{v,w\})\in E(Q_n)$. As $\{v,w\}$ was an arbitrary edge of $\Gamma$, it follows that $\varphi(E(\Gamma))=E(Q_n)$. Since $|E(\Gamma)|=|E(Q_n)|$, this implies that $\varphi$ is an isomorphism. It is not hard to check that $Q_3$ is not CIS and thus $n\geq 4$.
\[irreducibleQuotientVT\] Let $\Gamma$ be a graph and let $X$ be its irreducible quotient. Then, $\Gamma$ is vertex-transitive if and only if $X$ is vertex-transitive and $\Gamma \cong X[\overline{K_n}]$ for some $n\ge 1$.
Let $R$ be the equivalence relation “having the same neighborhood” on $V(\Gamma)$. If $\Gamma$ is vertex-transitive then the $R$-equivalence classes have the same size, say $n$. Since these equivalence classes are stable sets, it follows that $\Gamma \cong X[\overline{K_n}]$. Moreover, the $R$-equivalence classes form a partition of $V(\Gamma)$ invariant under the action of the automorphism group of $\Gamma$ and thus the automorphism group of $\Gamma$ has an induced action on the quotient, namely $X$. Since $\Gamma$ is vertex-transitive, it follows that $X$ is vertex-transitive. The converse is clear.
Combining Propositions \[irreducibleQuotient\] and \[irreducibleQuotientVT\] and Theorems \[thm:VT-CIS\] and \[thm:valency-7\] we easily obtain the following.
\[cor:VT-CIS-valency-7\] Let $\Gamma$ be a connected vertex-transitive graph of valency at most $7$. Then, $\Gamma$ is CIS if and only if $\Gamma$ is isomorphic to one of the following graphs $K_n$ ($1\leq n \leq 8$), $K_{n,n}$ ($2\leq n\leq 7$), $L(K_{3,3})$, $L(K_{4,4})$, $C_4[K_2]$, $K_3[\overline{K_2}]$, $K_3[\overline{K_3}]$, $K_4[\overline{K_2}]$, $K_{3,3}[K_2]$ or $Q_n$ for some $n\geq 4$.
Some open questions {#sec:questions}
===================
We conclude the paper with some open questions related to the results of this paper. It is known that every perfect graph $\Gamma$ satisfies the inequality $\alpha(\Gamma)\omega(\Gamma)\ge |V(\Gamma)|$ [@MR0309780], and Theorem \[thm:VT-CIS\] implies that it holds with equality for vertex-transitive CIS graphs. This motivates the following question.
Does every CIS graph $\Gamma$ satisfy $\alpha(\Gamma)\omega(\Gamma)\ge |V(\Gamma)|$?
Given the examples that have appeared in this paper, the following is also a natural question.
\[quest-chi\] Does every vertex-transitive CIS graph $\Gamma$ admit a decomposition of its vertex set into $\omega(\Gamma)$ stable sets?
Question \[quest-chi\] can be stated equivalently as follows: [*Does the chromatic number of every vertex-transitive CIS graph equal its clique number?*]{} Generalizing the famous class of perfect graphs [@MR2233847; @MR2063679], graphs with this property were named [*weakly perfect graphs*]{} in [@MR2679902; @MR2738964]. Finally, in view of Theorem \[thm:valency-7\], the following question seems natural.
[Let $G$ be a regular CIS graph such that both it and its complement are connected, well-covered and irreducible. Is $G$ necessarily vertex-transitive?]{}
We suspect the answer is no, but it does not seem easy to produce a counterexample.
[^1]: Authors’ e-mail addresses: `[email protected]` (Edward Dobson), `[email protected]` (Ademir Hujdurović), `[email protected]` (Martin Milanič), `[email protected]` (Gabriel Verret).
[^2]: The equivalence classes of $R$ have appeared in the literature under various names such as [*maximal independent-set modules*]{} [@HeggernesMP2011], or [*similarity classes*]{} [@LozinM2010].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
A linear time-invariant dissipative Hamiltonian (DH) system $\dot x = (J-R)Q x$, with a skew-Hermitian $J$, an Hermitian positive semi-definite $R$, and an Hermitian positive definite $Q$, is always Lyapunov stable and under weak further conditions even asymptotically stable. In various applications there is uncertainty on the system matrices $J, R, Q$, and it is desirable to know whether the system remains asymptotically stable uniformly against all possible uncertainties within a given perturbation set. Such robust stability considerations motivate the concept of stability radius for DH systems, i.e., what is the maximal perturbation permissible to the coefficients $J, R, Q$, while preserving the asymptotic stability. We consider two stability radii, the unstructured one where $J, R, Q$ are subject to unstructured perturbation, and the structured one where the perturbations preserve the DH structure. We employ characterizations for these radii that have been derived recently in \[*SIAM J. Matrix Anal. Appl., 37, pp. 1625-1654, 2016*\] and propose new algorithms to compute these stability radii for large scale problems by tailoring subspace frameworks that are interpolatory and guaranteed to converge at a super-linear rate in theory. At every iteration, they first solve a reduced problem and then expand the subspaces in order to attain certain Hermite interpolation properties between the full and reduced problems. The reduced problems are solved by means of the adaptations of existing level-set algorithms for ${\mathcal H}_\infty$-norm computation in the unstructured case, while, for the structured radii, we benefit from algorithms that approximate the objective eigenvalue function with a piece-wise quadratic global underestimator. The performance of the new approaches is illustrated with several examples including a system that arises from a finite-element modeling of an industrial disk brake.\
**Key words.** Linear Time-Invariant Dissipative Hamiltonian System, Port-Hamiltonian system, Robust Stability, Stability Radius, Eigenvalue Optimization, Subspace Projection, Structure Preserving Subspace Framework, Hermite Interpolation.\
**AMS subject classifications.** 65F15, 93D09, 93A15, 90C26
author:
- Nicat Aliyev
- Volker Mehrmann
- Emre Mengi
bibliography:
- 'PH\_stabradii.bib'
title: 'Computation of Stability Radii for Large-Scale Dissipative Hamiltonian Systems'
---
Introduction
============
Linear time-invariant *Dissipative Hamiltonian (DH) systems* are dynamical systems of the form $$\label{DH}
\dot x \;\; = \;\; (J-R)Qx.$$ They arise as homogeneous part of *port-Hamiltonian (PH) systems* of the form $$\label{ph}
\begin{split}
\dot x & \;\; = \;\; (J-R)Qx(t) + (B-P)u(t), \\
y(t) & \;\; = \;\; (B+P)^H Qx(t) + D u(t),
\end{split}$$ when the input $u$ is $0$ and the output $y$ is not considered. Here $Q=Q^H\in\mathbb{C}^{n \times n}$ is an Hermitian positive definite matrix (denoted as $Q>0$), $J\in\mathbb{C}^{n \times n}$ is a skew-Hermitian matrix associated with the energy flux of the system, $R\in\mathbb{C}^{n \times n}$ is the Hermitian positive semi-definite (denoted by $R\geq 0$) *dissipation matrix* of the system, $B\pm P\in\mathbb{C}^{n \times m}$ are the *port matrices*, and $D$ describes the *direct feed-through* from input to output. The function $\mathcal H(x)= \frac 12 x^HQx$ (called *Hamiltonian function*) describes the total internal energy of the system. Here and elsewhere $A^H$ denotes the conjugate transpose of a complex matrix $A$.
PH and DH systems play an essential role in most areas of science and engineering, see e.g. [@JacZ12; @SchJ14], due to their very important structural properties; e.g., they allow modularized modeling and easy model reduction via Galerkin projection. An important structural property is that DH systems are automatically *Lyapunov stable*, i.e., all eigenvalues of $A=(J-R)Q$ are in the closed left half of the complex plane, and those on the imaginary axis are semisimple, see [@MehMS16a]. However, DH systems are not necessarily *asymptotically stable*, since $A$ may have purely imaginary eigenvalues, e.g., when the dissipation matrix $R$ vanishes, then all eigenvalues are purely imaginary. If a DH system is Lyapunov stable but not asymptotically stable, then arbitrarily small unstructured perturbations (such as rounding errors) may cause the system to become unstable. These issues are our motivation to analyse whether a DH system is *robustly asymptotically stable*, i.e., whether small (structured or unstructured) perturbations keep it asymptotically stable.
\[ex:ex1\] [Disk brake squeal is a well-known problem in mechanical engineering. It occurs due to self-excited vibrations caused by instability at the pad-rotor interface [@Akay_2002]. The transition from stability to instability of the brake system is generally examined by finite element (FE) analysis of the system. In [@GraMQSW16] FE models resulting for disk brakes are derived in form of second order differential equations $$\label{brake1}
M \ddot x+D(\Omega) \dot x+ K(\Omega) x=f,$$ with large and sparse coefficient matrixes $M$, $D(\Omega)$, and $K(\Omega) \in \mathbb{R}^{n \times n}$, where $D(\Omega)$ and $K(\Omega)$ depend on the rotational speed $\Omega>0$ of the disk, and have the form $$D(\Omega)
:=
D_M + \frac{1}{\Omega} D_R,\
K(\Omega)
:=
K_E + \Omega^2 K_g$$ with $D_M, D_R$ representing material, friction-induced damping matrices, $K_E, K_g$ corresponding to elastic, geometric stiffness matrices, respectively. Here, $M>0$ and $K(\Omega)>0$, whereas $D(\Omega)\geq 0$. (The function $f$ represents a forcing term or control, but for the stability analysis one may assume that $f=0$, which we assume in the following.) The incorporation of gyroscopic effects, modeled by the term $G(\Omega) \dot x$, with $G(\Omega) := \Omega D_G=-\Omega D_G^H$, and circulatory effects, modeled by an unsymmetric term $N x$ gives rise to a system $$\label{eq:full_brake}
M \ddot x + (D(\Omega) +G(\Omega)) \dot x + (K(\Omega) + N)x=0,$$ or in first order representation $ \widetilde{M} \dot z + \widetilde{K} z=0$, where $$\label{eq:first_order}
\widetilde{M}=\begin{bmatrix} M & 0 \\ 0 & K(\Omega) \end{bmatrix}, \;\:
\widetilde{K}=\begin{bmatrix} D(\Omega) + G(\Omega) & K(\Omega) + N \\ -K(\Omega) & 0 \end{bmatrix}.$$ Straightforward manipulations yields a system $$\label{insta}
\dot {\widetilde{z}} = (J - R)Q\widetilde{z}$$ with $\widetilde{z} = Q^{-1}z \:$, where $$\label{eq:insta_matrices}
\begin{split}
J =\begin{bmatrix} -G(\Omega) & -(K(\Omega) + \frac12N) \\ K(\Omega) + \frac12N^H & 0 \end{bmatrix}, \hskip 22ex \\
R =\begin{bmatrix} D(\Omega) & \frac12N \\ \frac12N^H & 0 \end{bmatrix}, \quad
Q =\begin{bmatrix} M & 0 \\ 0 & K(\Omega) \end{bmatrix}^{-1}.
\end{split}$$ In the absence of the circulatory effects, i.e., when $N=0$, the system in (\[insta\]) is a DH system and as a result it is Lyapunov stable and typically even asymptotically stable. However, small circulatory effects, i.e., perturbations by a non-symmetric $N$ of small norm, may result in instability. ]{}
Asymptotic stability of a general linear dynamical system in the presence of uncertainty can only be guaranteed when the system has a reasonable *distance to instability*, i.e., to systems with purely imaginary eigenvalues. Hence, an estimation of the distance to instability, which is an optimization problem over admissible perturbations, is an important ingredient of a proper stability analysis.
In this paper we focus on the stability analysis of large-scale (and typically sparse) DH systems of the form in the presence of uncertainties in the coefficients. Considering perturbations in one of the coefficient matrices $J$, $R$, $Q$ of , in [@MehMS16a] characterizations for several structured distances to instability were derived under restricted perturbations of the form $B\Delta C$, with restriction matrices $B \in\mathbb{C}^ {n\times m}$ and $C \in {{\mathbb{C}}}^{p\times n}$ of full column rank and full row rank, respectively, allowing selected parts of the matrices $J, R, Q$ to be unperturbed. We will use an adaptation of the subspace framework introduced in [@Aliyev2017], based on model order reduction techniques to compute the stability radii using the characterizations in [@MehMS16a].
The paper is organized as follows. Section \[defns\] provides formal definitions of the structured and unstructured stability radii, and in Section \[char\] we briefly recall the characterizations of these stability radii derived in [@MehMS16a]. Section \[sec:DH\_unstructured\_dist\] proposes subspace frameworks for computing the unstructured stability radii problems exploiting these characterizations. The performance of the proposed frameworks for the unstructured stability radii is illustrated via the disk brake example and several synthetic examples in Section \[sec:DH\_numexps\]. Finally, Section \[sec:structured\] focuses on the structured stability radius when only $R$ is subject to Hermitian perturbations. We first discuss how small-scale problems can be solved in Section \[sec:small\_scale\_st\]. A new structured subspace framework is discussed in Section \[sec:large\_scale\_st\] followed by several numerical examples in Section \[sec:str\_num\_exp\].
Unstructured and Structured Stability Radii {#defns}
===========================================
In [@MehMS16a] computable formulas for DH systems of the form are derived using several notions of unstructured and structured stability radii. In this section we briefly recall the main definitions and results from [@MehMS16a] for restricted perturbations in one of the following forms. $$\label{pert_sys}
\left( \left(J+B\Delta C\right) - R\right)Q , \;
\left(J - \left(R+B\Delta C\right)\right)Q , \; \text{or} \;
\left(J - R \right)\left(Q+B\Delta C\right).$$ In the following ${\rm i} {\mathbb R}$ denotes the imaginary axis in the complex plane, $\Lambda(A)$ the spectrum of a matrix $A$, and $\|A\|_2$ the spectral norm.
\[defn:unst\_stab\_radii\] Consider a DH system of the form and suppose that $B \in \mathbb{C}^{n \times m}$ and $C \in \mathbb{C}^{p \times n}$ are given full rank restriction matrices.
- The *unstructured restricted stability radius* $r(J; B,C)$ with respect to perturbations of $J$ under the restriction matrices $B$, $C$ is defined by $$r(J; B,C) := \inf \{\|\Delta\|_2 : \Delta\in \mathbb{C}^{m\times p}, \Lambda (\left( \left(J+B\Delta C\right) - R\right)Q)\cap {\rm i}\mathbb{R} \neq \emptyset\}.$$
- The *unstructured restricted stability radius* $r(R; B,C)$ with respect to perturbations of $R$ under the restriction matrices $B$, $C$ is defined by $$r(R; B,C) := \inf \{\|\Delta\|_2 : \Delta\in \mathbb{C}^{m\times p}, \Lambda (\left(J - \left(R+B\Delta C\right)\right)Q\cap {\rm i}\mathbb{R} \neq \emptyset\}.$$
- The *unstructured restricted stability radius* $r(Q; B,C)$ with respect to perturbations of $Q$ under the restriction matrices $B$, $C$ is defined by $$r(Q; B,C) := \inf \{\|\Delta\|_2 : \Delta\in \mathbb{C}^{m\times p}, \Lambda \left(J - R \right)\left(Q+B\Delta C\right)\cap {\rm i}\mathbb{R} \neq \emptyset\}.$$
\[ex:ex2\]
Consider again Example \[ex:ex1\]. Here it is of interest to know whether (for given $\Omega$) the norm of the non-symmetric matrix $N$ is tolerable to preserve the asymptotic stability of the DH system in (\[insta\]) without the circulatory effects. The relevant stability radius for a specified $\Omega$ is given by $$\label{eq:BH_stabradii}
\inf
\left\{
\| N \|_2 \; \bigg| \;
\Lambda
\left(
{\mathcal A}(N) \right) \cap {{\mathrm i}}{\mathbb R} \neq \emptyset
\right\},$$ where $$\begin{split}
{\mathcal A}(N)
& :=
\left(
\begin{bmatrix} -G(\Omega) & -(K(\Omega) + \frac12N) \\ K(\Omega) + \frac12N^H & 0 \end{bmatrix}
-
\begin{bmatrix} D(\Omega) & \frac12N \\ \frac12N^H & 0 \end{bmatrix}
\right)
\begin{bmatrix} M & 0 \\ 0 & K(\Omega) \end{bmatrix}^{-1} \\
& =
\left\{
\begin{bmatrix} -G(\Omega) & -K(\Omega) \\ K(\Omega) & 0 \end{bmatrix}
-
\left(
\begin{bmatrix} D(\Omega) & 0 \\ 0 & 0 \end{bmatrix}
+
\begin{bmatrix} 0 & N \\ 0 & 0 \end{bmatrix}
\right)
\right\}
\begin{bmatrix} M & 0 \\ 0 & K(\Omega) \end{bmatrix}^{-1}.
\end{split}$$ Hence, the stability radius in (\[eq:BH\_stabradii\]) corresponds to the unstructured stability radius $r(R; B, C)$ with the restriction matrices $
B =
\begin{bmatrix}
I & 0
\end{bmatrix}^T
$ and $
C =
\begin{bmatrix}
0 & I
\end{bmatrix}
$ with $n\times n$ blocks.
Furthermore, in the definition of ${\mathcal A}(N)$ the skew-Hermitian perturbations are more influential on the imaginary parts of its eigenvalues, whereas the Hermitian perturbations are more effective in moving its eigenvalues towards the imaginary axis. This leads us to the consideration of the stability radius $$\label{eq:BH_stabradii2}
\inf
\left\{
\| N \|_2 \; \bigg| \;
\Lambda
\left(
{\mathcal A}_0(N) \right) \cap {{\mathrm i}}{\mathbb R} \neq \emptyset
\right\}$$ with $${\mathcal A}_0(N)
:=
\left(
\begin{bmatrix} -G(\Omega) & -K(\Omega) \\ K(\Omega) & 0 \end{bmatrix}
-
\begin{bmatrix} D(\Omega) & \frac12N \\ \frac12N^H & 0 \end{bmatrix}
\right)
\begin{bmatrix} M & 0 \\ 0 & K(\Omega) \end{bmatrix}^{-1}.$$
Examples such as Example \[ex:ex2\] motivate the following definition of the structured stability radius in [@MehMS16a].
\[defn:st\_stab\_radii\] Consider a DH system of the form and suppose that $B \in \mathbb{C}^{n \times m}$ is a given restriction matrix. The *structured restricted stability radius* with respect to Hermitian perturbations of $R$ under the restriction $B$ is defined by $$\label{eq:structured_radii_defn}
\begin{split}
r^{\rm Herm}(R; B) := \inf\{ \|\Delta \|_2 \; | \; \Delta = \Delta^H, \;\;\; \hskip 25ex \\
\hskip 23ex
\Lambda \big( (J - R)Q - (B\Delta B^H)Q \big) \cap {\rm i}\mathbb R\neq \emptyset\}.
\end{split}$$
Characterizations for Stability Radii {#char}
=====================================
The numerical techniques that we will derive for the computation of the unstructured and structured stability radii exploit eigenvalue or singular value optimization characterizations derived in [@MehMS16a].
\[thm\_unstr\] For an asymptotically stable DH system of the form and restriction matrices $B\in\mathbb{C}^{n \times m}$, $C \in\mathbb C^{p \times n}$ the following assertions hold:
- The unstructured stability radius $r(R; B,C)$ is finite if and only if $G_R(\omega) := CQ({\rm i}\omega I_n-(J - R)Q)^{-1}B$ is not identically zero if and only if $r(J; B,C)$ is finite. If $r(R; B,C)$ is finite, then we have $$\label{th2}
r(R; B,C) = r(J;B,C) = \inf_{\omega\in\mathbb R}\frac{1}{\|G_{R}(\omega)\|_2}.$$
- The unstructured stability radius $r(Q; B,C)$ is finite if and only if $G_Q(\omega) : = C({\rm i}\omega I_n-(J - R)Q)^{-1}(J-R)B$ is not identically zero for all $\omega\in\mathbb R$. If $r(Q; B,C)$ is finite, then we have $$\label{th1}
r(Q; B,C) = \inf_{\omega\in\mathbb R}\frac{1}{\|G_{Q}(\omega)\|_2}.$$
For the structured stability radius and Hermitian perturbations of $R$ the following result is obtained in [@MehMS16a].
\[thm\_str\_Herm\] For an asymptotically stable DH system of the form , and a restriction matrix $B\in\mathbb{C}^{n\times m}$ of full column rank, let
1. $W(\lambda) := (J-R)Q - \lambda I$ for a given $\lambda \in {\mathbb C}$ such that $W(\lambda)$ is invertible,
2. $L(\lambda)$ be a lower triangular Cholesky factor of
$\widetilde{H}_0(\lambda) := B^H W(\lambda)^{-H} Q B B^H Q W(\lambda)^{-1} B$,
i.e., $L(\lambda)$ is a lower triangular matrix satisfying $\widetilde{H}_0(\lambda) = L(\lambda) L(\lambda)^H$,
3. $H_0(\lambda) := L(\lambda)^{-1} L(\lambda)^{-H}$,
4. $H_1(\lambda) := {{\mathrm i}}(\widetilde{H}_1(\lambda) - \widetilde{H}_1(\lambda)^H)$, where $\widetilde{H}_1(\lambda) := L(\lambda)^{-1} B^H W(\lambda)^{-H} Q B L(\lambda)^{-H}$.
Then $r^{\rm Herm}(R; B)$ is finite, and given by $$r^{\rm Herm}(R; B)
\; = \;
\left\{
\inf_{\omega \in {\mathbb R}} \: \sup_{t \in {\mathbb R}} \: \lambda_{\min} (H_0({{\mathrm i}}\omega) + t H_1 ({{\mathrm i}}\omega))
\right\}^{1/2},$$ where $\lambda_{\min}( \cdot )$ denotes the smallest eigenvalue of its Hermitian matrix argument, and the inner supremum is attained if and only if $H_1({\rm i} \omega)$ is indefinite.
The characterization in [@MehMS16a] is presented in a slightly different form. In particular, it is stated in terms of an orthonormal basis $U(\lambda)$ for the kernel of $\left((I - BB^+) W(\lambda)\right)$. It turns out that $U(\lambda)$ does not have to be orthonormal, rather the theorem can be stated in terms of any basis for the kernel of $\left((I - BB^+) W(\lambda)\right)$. In Theorem \[thm\_str\_Herm\], we have employed a particular basis that simplifies the formulas and facilitates the computation.
Computation of the Unstructured Stability Radii for Large-Scale Problems {#sec:DH_unstructured_dist}
========================================================================
In this section we study the computation of unstructured stability radii for large-scale DH systems using the characterizations of $r(R; B,C)$, $r(Q; B,C)$, $r(J;B,C)$ given in Theorem \[thm\_unstr\]. One easily observes that $$\begin{split}
G_R({\rm i} \omega) & \; := \; CQ({\rm i}\omega I_n-(J - R)Q)^{-1}B, \\
G_Q({\rm i} \omega) & \; := \; C({\rm i}\omega I_n-(J - R)Q)^{-1}(J-R)B
\end{split}$$ can be viewed as restrictions of transfer functions of control systems to the imaginary axis. To be precise, setting $\widetilde{A} = (J - R ) Q$, $\widetilde{B} = B$ and $\widetilde{C} = C Q$, the matrix-valued function $G_R({\rm i} \omega) := CQ({\rm i}\omega I_n-(J - R)Q)^{-1}B \;$ becomes $$\label{def_GR}
\widetilde{G}_R({\rm i} \omega) := \widetilde{C}({\rm i}\omega I_n - \widetilde{A})^{-1} \widetilde{B}$$ which can be considered as the transfer function of the system $$\label{eq:Disc}
\dot x = \widetilde{A} x + \widetilde{B} u, \quad
y = \widetilde{C} x$$ on the imaginary axis. Theorem \[thm\_unstr\] suggests that if $\widetilde{G}_R({\rm i} \omega) := \widetilde{C}({\rm i}\omega I_n - \widetilde{A})^{-1} \widetilde{B}$ is not identically zero, then $r(R; B,C)$ and $r(J; B,C)$ are finite, and characterized by $$\label{char1}
\begin{split}
r(R; B,C) \; = \; r(J; B,C) & \; = \; \inf_{\omega\in\mathbb{R}}\frac{1}{ \| \widetilde{G}_R({\rm i} \omega) \|_2} \\
& \; = \ \frac{1}{\sup_{\omega\in\mathbb{R}} \| \widetilde{G}_R({\rm i} \omega) \|_2}
\; = \; \frac{1}{\;\; \| \widetilde{G}_R \|_{\mathcal H_\infty}},
\end{split}$$ where $\| \widetilde{G}_R \|_{{\mathcal H}_\infty} := \sup_{\omega \in {\mathbb R}} \: \sigma_{\max} (\widetilde{G}_R({\rm i} \omega))$ denotes the ${\mathcal H}_\infty$-norm of $\widetilde{G}_R$, and $\sigma_{\max}(\cdot)$ denotes the maximal singular value.
For the stability radius $r(Q; B,C)$, consideration of $G_Q({\rm i} \omega) : = C({\rm i}\omega I_n-(J - R)Q)^{-1}(J-R)B$, by setting $\widetilde{A} = (J - R)Q$, $\widetilde{B} = (J - R )B$ and $\widetilde{C} = C$, leads us to a similar characterization.
A Subspace Framework {#sec:sf_descriptor}
--------------------
Recently, in [@Aliyev2017], a subspace framework for the computation of the ${\mathcal H}_\infty$-norm of a large-scale system has been proposed, which is inspired from model order reduction techniques, and has made the computation of ${\mathcal H}_\infty$-norms feasible for very large control systems. We will now discuss how to use these techniques for the computation of the unstructured stability radii $r(R; B,C)$, $r(J; B,C)$, $r(Q; B,C)$ in the large-scale setting.
To briefly summarize the iterative procedure in the subspace framework of [@Aliyev2017], let us assume that in iteration $k$, two subspaces ${\mathcal V}_k$ and ${\mathcal W}_k$ of equal dimension have been determined, as well as matrices $V_k$ and $W_k$ whose columns span orthonormal bases for these subspaces. Applying a Petrov-Galerkin projection to system (\[eq:Disc\]), restricts the state $x$ to ${\mathcal V}_k$, i.e., in (\[eq:Disc\]) we replace $x$ by $V_k x_k$, and imposes that the residual after this restriction is orthogonal to ${\mathcal W}_k$. This projection gives rise to a reduced order system $$\label{eq:red_sys1}
\dot x_k = \widetilde{A}_k x_k+ \widetilde{B}_k u, \quad
y_k = \widetilde{C}_k x_k,$$ with $$\label{eq:red_sys2}
\widetilde{A}_k := W_k^H \widetilde{A} V_k, \;\; \widetilde{B}_k := W_k^H \widetilde{B}, \;\;
\widetilde{C}_k = \widetilde{C} V_k.$$ Then the ${\mathcal H}_\infty$-norm of a transfer function $G(s) := \widetilde{C}( s I_n - \widetilde{A})^{-1} \widetilde{B}$ in (\[eq:Disc\]) can be approximated by computing the ${\mathcal H}_\infty$-norm of $$G_k(s) := \widetilde{C}_k( s I_k - \widetilde{A}_k)^{-1} \widetilde{B}_k$$ for instance by employing the method in [@Boyd1990] or [@Bruinsma1990], in particular, by computing $ \omega_{k+1} := \operatorname*{arg\,max}_{\omega \in {\mathbb R}} \: \sigma_{\max} ( G_k( {{\mathrm i}}\omega) )$. This is computationally cheap if the dimensions of ${\mathcal V}_k, {\mathcal W}_k$ are small. Once $\omega_{k+1}$ has been computed, then the subspaces ${\mathcal V}_k$ and ${\mathcal W}_k$ are expanded into larger subspaces ${\mathcal V}_{k+1}$ and ${\mathcal W}_{k+1}$ in such a way that the corresponding reduced transfer function $G_{k+1}(s)$ satisfies the *Hermite interpolation conditions* $$\label{eq:Hermite_interpolate}
\begin{split}
\sigma_{\max} (G({{\mathrm i}}\omega_{k+1})) & = \sigma_{\max} (G_{k+1}({{\mathrm i}}\omega_{k+1})), \\
\sigma_{\max}' (G({{\mathrm i}}\omega_{k+1})) & = \sigma_{\max}' (G_{k+1}({{\mathrm i}}\omega_{k+1})),
\end{split}$$ where $\sigma_{\max}'(G({\rm i}\omega))$ denotes the derivative of $\sigma_{\max} (G({\rm i} \omega))$ with respect to $\omega$. Denoting the image space of a matrix by $A$ by $ {\rm Im} (A)$, it is shown in [@Aliyev2017] that $$\begin{split}
{\mathcal V}_{k+1} & := {\mathcal V}_k \oplus {\rm Im} (( {{\mathrm i}}\omega_{r+1} I_n - \widetilde{A})^{-1} \widetilde{B}), \\
{\mathcal W}_{k+1} & := {\mathcal W}_k \oplus {\rm Im} ( (\widetilde{C} ( {{\mathrm i}}\omega_{k+1} I_n - \widetilde{A})^{-1})^H),
\end{split}$$ more specifically the inclusions $${\rm Im} (( {{\mathrm i}}\omega_{k+1} I_n - \widetilde{A})^{-1} \widetilde{B}) \subseteq {\mathcal V}_{k+1},\
{\rm Im} ( (\widetilde{C} ( {{\mathrm i}}\omega_{k+1} I_n - \widetilde{A})^{-1})^H) \subseteq {\mathcal W}_{k+1},$$ ensure that the Hermite interpolation conditions (\[eq:Hermite\_interpolate\]) are satisfied. The procedure is then repeated with the expanded subspaces ${\mathcal V}_{k+1}$, ${\mathcal W}_{k+1}$. In [@Aliyev2017], it is shown that the sequence $\{ \omega_k \}$ converges at a super-linear rate and satisfies $$\begin{split}
\sigma_{\max} (G({{\mathrm i}}\omega_{j})) & = \sigma_{\max} (G_{k}({{\mathrm i}}\omega_{j})), \\
\sigma_{\max}' (G({{\mathrm i}}\omega_{j})) & = \sigma_{\max}' (G_{k}({{\mathrm i}}\omega_{j}))
\end{split}$$ for $j = 1,\dots,k$. A disadvantage of this general approach is that even if $\widetilde{A} = (J - R)Q$ has DH structure, this is not necessarily true for $\widetilde{A}_k$, so it cannot be guaranteed from the structure that the reduced system is stable. In the next section we modify the procedure of [@Aliyev2017] to preserve the DH structure.
A Structure Preserving Subspace Framework {#sec:sf_PH}
-----------------------------------------
In this subsection we derive an interpolating, DH structure preserving version of the robust subspace projection framework. Structure preserving subspace projection methods in the context of model order reduction of large-scale PH and DH systems have been proposed in [@Gugercin_2012; @Gugercin_2009; @Polyuga_2009; @Polyuga_2011; @Schaft_2009; @Wu_2014]. Our approach is inspired by [@Gugercin_2012] and uses a general interpolation result from [@Gallivan2005].
\[thm:Gal\_interpolate\] Let $G(s)$ be the transfer function for a full order system as in (\[eq:Disc\]), and let $G_k(s)$ be the transfer function for the reduced system defined by (\[eq:red\_sys1\]), (\[eq:red\_sys2\]).
1. *(Right Tangential Interpolation)* For given $\widehat{s} \in {\mathbb C}$ and $\widehat{b} \in {\mathbb C}^m$, if $$\label{eq:subspace_incl}
\left[ (\widehat{s}I - \widetilde{A})^{-1} \right]^\ell \widetilde{B} \widehat{b} \; \in \; {\mathcal V}_k \quad \quad {\rm for} \;\; \ell = 1, \dots, N,$$ and $\: {\mathcal W}_k$ is such that $W_k^H V_k = I$, then we have $$\label{eq:tangent_interpolate}
G^{(\ell)} (\widehat{s}) \widehat{b} = G^{(\ell)}_k (\widehat{s}) \widehat{b} \quad\quad {\rm for} \;\; \ell = 0, \dots, N-1$$ provided that both $\widehat{s} I - \widetilde{A}$ and $\widehat{s} I - \widetilde{A}_k$ are invertible.
2. *(Left Tangential Interpolation)* For a given $\widehat{s} \in {\mathbb C}$ and $\widehat{c} \in {\mathbb C}^p$, if $$\label{eq:subspace_inclusion_left}
\left( \widehat{c}^{\: H} \widetilde{C} \left[ (\widehat{s}I - \widetilde{A})^{-1} \right]^\ell \right)^H \; \in \; {\mathcal W}_k \quad \quad {\rm for} \;\; \ell = 1, \dots, N,$$ and $\: {\mathcal V}_k$ is such that $W_k^H V_k = I$, then we have $$\label{eq:tangent_interpolate_left}
\widehat{c}^{\: H} G^{(\ell)} (\widehat{s}) = \widehat{c}^{\: H} G^{(\ell)}_k (\widehat{s})
\quad\quad {\rm for} \;\; \ell = 0, \dots, N-1$$ provided that both $\widehat{s} I - \widetilde{A}$ and $\widehat{s} I - \widetilde{A}_k$ are invertible.
### Computation of $r(R; B,C)$ and $r(J; B,C)$ {#sec:unstructured_subspace_rR_rJ}
The computation of $r(R; B, C) = r(J; B,C)$ involves the maximization of the largest singular value of the transfer function $G(s) = CQ (sI - (J-R)Q)^{-1} B$ associated with the system $$\label{eq:PHS}
\begin{split}
\dot x &= (J-R) Qx + B u, \\
y &= CQx.
\end{split}$$ on the imaginary axis. We make use of Theorem \[thm:Gal\_interpolate\] to obtain a reduced order system satisfying the interpolation conditions (\[eq:tangent\_interpolate\]) while retaining the structure in (\[eq:PHS\]). We, in particular, employ right tangential interpolation for a given $\widehat{s} \in {\mathbb C}$ and $\widehat{b} \in {\mathbb C}^m$, and choose ${\mathcal V}_k$ as any subspace satisfying (\[eq:subspace\_incl\]). Let us also define $ W_k \;\; := \;\; QV_k (V_k^H Q V_k)^{-1}$, ${\mathcal W}_k := {\rm Im} (W_k)$, so that $$W_k^H V_k \;\; = \;\; I_k
\quad {\rm and} \quad
( W_k V_k^H )^2 \;\; = \;\; W_k V_k^H,$$ i.e., $W_k V_k^H$ is an oblique projector onto ${\rm Im}( Q V_k)$.
The matrices $\widetilde{A}_k, \widetilde{B}_k, \widetilde{C}_k$ of the reduced system (\[eq:red\_sys1\]), (\[eq:red\_sys2\]) for these choices of $V_k$ and $W_k$ then satisfy $$\label{eq:reduced_A}
\begin{split}
\widetilde{A}_k & \;\; = \;\; W_k^H \widetilde{A} V_k = W_k^H (J-R) Q V_k = W_k^H (J-R) W_k V_k^H Q V_k \\
& \;\; = \;\; (J_k - R_k) Q_k,
\end{split}$$ where $J_k := W_k^H J W_k=-J_k^H$, $R_k := W_k^H R W_k=R_k^H\geq 0$, and $Q_k := V_k^H Q V_k=Q_k^H>0$. Additionally, $
\widetilde{C}_k \;\; = \;\; \widetilde{C} V_k = CQ V_k = C W_k V_k^H Q V_k = C_k Q_k$, where $C_k : = C W_k$, and $\widetilde{B}_k \;\; = \;\; W_k^H B \;\; =: \;\; B_k$.
This construction leads to the following result of [@Gugercin_2012].
\[interpol2\] Consider a linear system of the form with transfer function $G(s) := CQ(s I_n - (J - R)Q)^{-1}B$. Furthermore, for a given point $\widehat{s}\in{{\mathbb{C}}}$ and a given tangent direction $\widehat{b}\in{{\mathbb{C}}}^m$, suppose that $V_k$ is a matrix with orthonormal columns such that $$( \widehat{s}I_n - (J - R)Q )^{-(\ell-1)} (\widehat{s}I_n - (J - R)Q )^{-1}B\widehat{b} \in{\rm Im}(V_k) \quad \text{for} \;\; \ell = 1,\ldots,N.$$ Define $\: W_k := QV_k(V_k^HQV_k)^{-1} \:$ and set $$\label{eq:DH_subspaces}
\begin{split}
J_k: = & W_k^H J W_k, \;\; Q_k: = V_k^H Q V_k, \;\; R_k := W_k^H R W_k \\
B_k: =& W_k^H B, \;\; C_k: = CW_k.
\end{split}$$ Then the resulting reduced order model $$\label{eq:RPHS}
\begin{split}
\dot x_k& \;\; = \;\; (J_k-R_k) Q_k x_k + B_k u, \\
y_k& \;\; = \;\; C_k Q_kx_k
\end{split}$$ is a DH system with transfer function $$\label{eq:rph}
G_k(s) \;\; := \;\; C_kQ_k(s I_k - (J_k - R_k)Q_k)^{-1}B_k$$ that satisfies $$G^{(j)}(\hat{s})\hat{b} \;\; = \;\; G^{(j)}_k(\hat{s})\hat{b} \quad \text{for} \quad j = 0\ldots, N-1,$$ where $G^{(j)}(\hat{s})$ denotes the $j$-th derivative of $\; G(s)$ at the point $\hat{s}$.
Based on Theorem \[interpol2\] we obtain Algorithm \[alg:dti\] for the computation of $r(R; B,C) = r(J; B,C)$.
\[algor1\]
Choose initial interpolation points $\omega_{1}, \dots , \omega_{q} \in {\mathbb R}.$ $V_q \gets {\rm orth} \begin{bmatrix} D({{\mathrm i}}\omega_1)^{-1}B & D({{\mathrm i}}\omega_1)^{-2}B
& \dots & D({{\mathrm i}}\omega_q)^{-1}B & D({{\mathrm i}}\omega_q)^{-2}B \end{bmatrix}$ \[defn\_init\_subspaces0\]\
19ex $\quad \text{and}\quad \quad
W_q \gets QV_q(V_q^HQV_q)^{-1} $. \[defn\_init\_subspaces\] Form $G_{k}$ as in for the choices of $J_k, R_k, Q_k, B_k, C_k$ in (\[eq:DH\_subspaces\]). \[reduced\_transfer\] $\; \displaystyle \omega_{k+1} \gets \: \operatorname*{arg\,max}_{\omega \in {\mathbb R}} \sigma_{\max}(G_{k} ({{\mathrm i}}\omega))$. \[solve\_reduced\] $\widehat{V}_{k+1} \gets \begin{bmatrix} D({{\mathrm i}}\omega_{k+1})^{-1}B & D({{\mathrm i}}\omega_{k+1})^{-2}B. \end{bmatrix}$ \[defn\_later\_subspaces\] $V_{k+1} \gets \operatorname{orth}\left(\begin{bmatrix} V_{k} & \widehat{V}_{k+1} \end{bmatrix}\right)
\quad \text{and}\quad W_{k+1} \gets Q V_{k+1}(V_{k+1}^H Q V_{k+1})^{-1}$.
According to Theorem \[interpol2\], for a given $\widehat{s} \in {\mathbb C}$, setting $$V_k := \begin{bmatrix} D(\widehat{s})^{-1}B & D(\widehat{s})^{-2}B \end{bmatrix},\ W_k := QV_k(V_k^HQV_k)^{-1}$$ where $$\label{eq:defn_Ds}
D(\widehat{s}): = (\widehat{s} I_n - (J - R)Q),$$ we obtain $G(\widehat{s}) = G_k(\widehat{s})$ and $G'(\widehat{s}) = G'_k(\widehat{s})$ and thus the Hermite interpolation conditions $$\label{eq:Hermite_inter_smax}
\sigma_{\max}(G(\widehat{s})) = \sigma_{\max}(G_k(\widehat{s})),\ \sigma'_{\max}(G(\widehat{s})) = \sigma'_{\max}(G_k(\widehat{s}))$$ are satisfied, which suggest the use of the reduced system in the greedy subspace framework outlined in Algorithm \[alg:dti\].
In line 5 of every iteration, the subspace framework computes the ${\mathcal H}_\infty$-norm of a reduced system, in particular it computes the point ${\rm i} \omega_\ast$ on the imaginary axis where this ${\mathcal H}_\infty$-norm is attained. Then the current left and right subspaces are expanded in a way so that the resulting reduced system still has DH structure and its transfer function Hermite interpolates the original transfer function at ${\rm i} \omega_\ast$. Since the Hermite interpolation conditions (\[eq:Hermite\_inter\_smax\]) are satisfied at $\widehat{s} = {\rm i} \omega_1, \dots, {\rm i} \omega_{k}$ at the end of iteration $k$, the rate-of-convergence analysis in [@Aliyev2017] applies to deduce a superlinear rate-of-convergence for the sequence $\{ \omega_k \}$.
The computationally most expensive part of Algorithm \[alg:dti\] is in lines \[defn\_init\_subspaces\] and \[defn\_later\_subspaces\], where many linear systems with possibly many right hand sides have to be solved. If this is done with a direct solver, then for each value $\widehat{\omega} \in {\mathbb R}$ one $LU$ factorization of the matrix $D({{\mathrm i}}\widehat{\omega})$ has to be performed. For large values of $n$, the computation time is usually dominated by these $LU$ factorizations. In contrast to this, the solution of the reduced problem in line \[solve\_reduced\] can be achieved (for small systems) by means of the efficient algorithm in [@Boyd1990; @Bruinsma1990].
### Computation of $r(Q; B,C)$
To compute the stability radius $r(Q; B,C)$ in the large-scale setting, we employ left tangential interpolations (i.e., part (ii) of Theorem \[thm:Gal\_interpolate\]). In this case $r(Q; B,C)$ is the reciprocal of the ${\mathcal H}_\infty$-norm of the transfer function $G(s) := C (sI - (J-R)Q)^{-1} (J-R)B$ corresponding to the system $$\label{eq:PHS2}
\dot x = (J-R) Qx + (J-R)B u, \quad
y(t) = Cx.$$ To obtain a reduced system which has the same structure as (\[eq:PHS2\]) and has a transfer function $G_k(s)$ that satisfies $\widehat{c}^{\: H} G(\widehat{s}) = \widehat{c}^{\: H}G_k(\widehat{s})$ for a given point $\widehat{s} \in {\mathbb C}$ and a direction $\widehat{c} \in {\mathbb C}^p$, let us choose $W_k$ so as to satisfy the condition in (\[eq:subspace\_inclusion\_left\]) for $\widetilde{A} := (J-R)Q$, and $\widetilde{C} := C$. Furthermore, we set $$V_k := (J - R)^H W_k (W_k^H (J - R)^H W_k)^{-1},$$ as well as ${\mathcal W}_k := {\rm Im}(W_k)$, ${\mathcal V}_k := {\rm Im}(V_k)$. The matrix $V_k$ is chosen to satisfy $$V_k^H W_k \;\; = \;\; I_k \quad {\rm and} \quad
(V_k W_k^H)^2 \;\; = \;\; V_k W_k^H,$$ so that $V_k W_k^H$ is an oblique projector onto ${\rm Im} ( (J-R)^H W_k )$.
In (\[eq:PHS2\]), setting $\widetilde{A} := (J-R)Q$, $\widetilde{B} := (J-R)B$, $\widetilde{C} = C$, let us investigate the matrices $\widetilde{A}_k, \widetilde{B}_k, \widetilde{C}_k$ of the corresponding reduced system defined by (\[eq:red\_sys1\]), (\[eq:red\_sys2\]). Specifically, we have that $$\widetilde{A}_k \;\; = \;\; W_k^H \widetilde{A} V_k \; = \; W_k^H (J-R) Q V_k \; = \; W_k^H (J-R) W_k V_k^H Q V_k,$$ where, in the third equality, we employ $V_k W_k^H (J - R)^H W_k = (J- R)^H W_k$, or equivalently $W_k^H (J- R) W_k V_k^H = W_k^H (J-R)$. Hence, defining $J_k := W_k^H J W_k=-J_k^H$, $R_k := W_k^H R W_k=R_k^H\geq 0$, $Q_k := V_k^H Q V_k=Q_k^H>0$, we obtain a DH system with $\widetilde{A}_k \;\; = \;\; (J_k - R_k) Q_k$. We also have $$ \widetilde{B}_k \;\; = \;\; W_k^H \widetilde{B} = W_k^H (J-R) B = W_k^H (J - R) W_k V_k^H B = (J_k - R_k) B_k
$$ with $B_k := V_k^H B$, and $\widetilde{C}_k \;\; = \;\; C V_k \;\; =: \;\; C_k$. These constructions lead to the following analogue of Theorem \[interpol2\].
\[interpol3\] Consider a linear system of the form with the transfer function $G(s) := C(s I_n - (J - R)Q)^{-1}(J-R)B$. For a given point $\widehat{s}\in{{\mathbb{C}}}$ and a direction $\widehat{c}\in{{\mathbb{C}}}^m$, suppose that $W_k$ is a matrix with orthonormal columns such that $$\left( \widehat{c}^{\: H} C ( \widehat{s}I_n - (J - R)Q )^{-1} (\widehat{s}I_n - (J - R)Q )^{-(\ell-1)} \right)^H \in{\rm Im}(W_k)$$ for $\ell= 1, \dots, N$. Letting $\: V_k := (J - R)^H W_k (W_k^H (J - R)^H W_k)^{-1} \:$, and $$\label{eq:DH_subspaces2}
\begin{split}
J_k: = & W_k^H J W_k \;\; Q_k: = V_k^H Q V_k, \;\; R_k := W_k^H R W_k \\
B_k: =& V_k^H B, \;\; C_k: = CV_k,
\end{split}$$ the resulting reduced order system $$\label{eq:RPHS2}
\dot x_k \;\; = \;\; (J_k-R_k) Q_k x_k + (J_k - R_k) B_k u, \quad
y_k \;\; = \;\; C_k x_k$$ is such that $\dot x_k = (J_k-R_k) Q_k x_k$ is dissipative Hamiltonian. Furthermore, the transfer function $$\label{eq:rph2}
G_k(s) \;\; := \;\; C_k (s I_k - (J_k - R_k)Q_k)^{-1} (J_k - R_k) B_k$$ of (\[eq:RPHS2\]) satisfies $$\widehat{c}^{\: H} G^{(\ell)}(\hat{s})\hat{b} \;\; = \;\; \widehat{c}^{\: H} G^{(\ell)}_k(\hat{s}) \quad \text{for} \quad \ell = 0\ldots, N-1.$$
Theorem \[interpol3\] shows that at a given $\widehat{s} \in {\mathbb C}$, the Hermite interpolation properties $G(\widehat{s}) = G_k(\widehat{s})$ and $G'(\widehat{s}) = G_k'(\widehat{s})$ and, in particular, $\sigma_{\max} (G(\widehat{s}))$ $=$ $\sigma_{\max}(G_k(\widehat{s}))$ and $\sigma_{\max}'(G(\widehat{s})) = \sigma_{\max}'(G_k(\widehat{s}))$) can be achieved, while preserving the structure, with the choices $$\begin{aligned}
W_k \; &=& \;
\left[
\begin{array}{cc}
(C D(\widehat{s})^{-1})^H & (C D(\widehat{s})^{-2})^H
\end{array}
\right],\\
V_k \; &=& \; (J - R)^H W_k (W_k^H (J - R)^H W_k)^{-1},\end{aligned}$$ where $D(\widehat{s})$ is as in (\[eq:defn\_Ds\]). This in turn gives rise to Algorithm \[alg:dti2\].
Choose initial interpolation points $\omega_{1}, \dots, \omega_{j} \in {\mathbb R}$. $W_j \gets
{\rm orth}
\left[
\left( C D({{\mathrm i}}\omega_1)^{-1} \right)^H
\;\; \left( C D({{\mathrm i}}\omega_1)^{-2} \right)^H
\;\; \dots \right.$\
35ex $ \left. \;\; \left( C D({{\mathrm i}}\omega_j)^{-1} \right)^H
\;\; \left( C D({{\mathrm i}}\omega_j)^{-2} \right)^H
\right]$,\
$\quad\quad\quad\quad V_j\gets (J - R)^H W_j (W_j^H (J - R)^H W_j)^{-1} $. \[defn\_init\_subspaces2\] Form $G_{k}$ as in for the choices of $J_k$, $R_k$, $Q_k$, $B_k$, $C_k$ in (\[eq:DH\_subspaces2\]). \[reduced\_transfer2\] $\; \displaystyle \omega_{k+1} \gets \operatorname*{arg\,max}_{\omega \in {\mathbb R}} \sigma_{\max}(G_{k} ({{\mathrm i}}\omega))$. \[solve\_reduced2\] $\widehat{W}_{k+1} \gets
\begin{bmatrix} \left( C D({{\mathrm i}}\omega_{k+1})^{-1} \right)^H & \left( C D({{\mathrm i}}\omega_{k+1})^{-2} \right)^H \end{bmatrix}$. \[defn\_later\_subspaces2\] $W_{k+1} \gets \operatorname{orth}\left(\begin{bmatrix} W_{k} & \widehat{W}_{k+1} \end{bmatrix}\right)
\quad \text{and} $\
15ex $V_{
k+1} \gets (J - R)^H W_{k+1} (W_{k+1}^H (J - R)^H W_{k+1})^{-1}$.
At every iteration of this algorithm, the ${\mathcal H}_\infty$-norm is computed for a reduced problem of the form $(\ref{eq:RPHS2})$, in particular, the optimal frequency where this ${\mathcal H}_\infty$-norm is attained is retrieved. Then the subspaces are updated so that the Hermite interpolation properties hold also at this optimal frequency at the largest singular values of the full and reduced problem, respectively. Once again the sequence $\{ \omega_k \}$ by Algorithm \[alg:dti2\] is guaranteed to converge at a super-linear rate, which can be attributed to the Hermite interpolation properties holding between the largest singular values of the full and reduced transfer functions.
Numerical Experiments {#sec:DH_numexps}
---------------------
In this subsection we illustrate the performance of MATLAB implementations of Algorithms \[alg:dti\] and \[alg:dti2\] via some numerical examples. We first discuss some implementation details and then present numerical results on two sets of random synthetic examples in Section \[sec:numexp\_syn\], and data from a FE model of a brake disk in Section \[sec:numexp\_brake\].
### Implementation Details and Test Setup {#sec:test_setup}
Algorithms \[alg:dti\] and \[alg:dti2\] are terminated when at least one of the following three conditions is fulfilled:
1. The relative distance between $\omega_{k}$ and $\omega_{k-1}$ is less than a prescribed tolerance for some $k> j$, i.e., $$\left| \omega_k- \omega_{k-1} \right| < \varepsilon \cdot \frac{1}{2} (\omega_k + \omega_{k-1}).$$
2. Letting $f_k := \max_{\omega \in {\mathbb R} \cup \infty} \sigma_{\max}(G_k({\rm i} \omega))$, two consecutive iterates $f_k, f_{k-1}$ are close enough in a relative sense, i.e., $$\left| f_k - f_{k-1} \right| < \varepsilon \cdot \frac{1}{2} (f_k + f_{k-1}).$$
3. The number of iterations exceeds a specified integer, i.e., $k> k_{\max}$.
In all numerical examples that we present, we set $\varepsilon = 10^{-6}$ and $k_{\max} = 100$.
In general, Algorithms \[alg:dti\] and \[alg:dti2\] converge only locally. The choice of the initial interpolation points affects the maximizers that the subspace frameworks converge to, in particular, whether convergence to a global maximizer occurs. The initial interpolation points are chosen based on the following procedure.
First, we discretize the interval $[ \lambda_{\min}^{\Im} , \lambda_{\max}^{\Im} ]$ into $\rho$ equally spaced points, say $\omega_{0,1}, \dots, \omega_{0,\rho}$, including the end-points $\lambda_{\min}^{\Im}, \lambda_{\max}^{\Im}$, where $\rho$ is specified by the user and $\lambda_{\min}^{\Im}, \lambda_{\max}^{\Im}$ denote the imaginary parts of the eigenvalues of $(J-R)Q$ with the smallest and largest imaginary part, respectively. Then we approximate the eigenvalues $z_1, \dots, z_\rho$ of $(J-R)Q$ closest to ${\rm i} \omega_{0,1}, \dots, {\rm i} \omega_{0,\rho}$, and permute them into $z_{j_1}, \dots, z_{j_\rho}$ where $\{ j_1, \dots, j_\rho \} = \{ 1, \dots, \rho \}$ so that $\sigma_{\max}(G({\rm i} \Im z_{j_1} )) \geq \dots \geq \sigma_{\max}(G({\rm i} \Im z_{j_\rho} ))$. The interpolation points $\omega_1, \dots, \omega_\ell$ employed initially are then chosen as the imaginary parts of $z_{j_1}, \dots, z_{j_\ell}$, where again $\ell\leq \rho$ is specified by the user.
### Results on Synthetic Examples {#sec:numexp_syn}
We now present results for two families of linear DH systems with random coefficient matrices; the first family consists of dense systems of order $800$, whereas the second family consists of sparse systems of order $5000$.
**Dense Random examples.** In the dense family the coefficient matrices $J$, $Q$, $R$ are formed by the MATLAB commands
>> J = randn(800); J = (J - J')/2;
>> Q = randn(800); Q = (Q + Q')/2; mineig = min(eig(Q));
>> if (mineig < 10^-4) Q = Q + (-mineig + 5*rand)*eye(n); end
>> p = round(80*rand);
>> Rp = randn(p); Rp = (Rp + Rp')/2; mineig = min(eig(Rp));
>> if (mineig < 10^-4) Rp = Rp + (-mineig + 5*rand)*eye(p); end
>> R = [Rp zeros(p,800-p); zeros(800-p,p) zeros(800-p,800-p)];
>> X = randn(800); [U,~] = qr(X); R = U'*R*U;
The restriction matrices $B$ and $C$ are chosen as $800\times 2$ and $2\times 800$ random matrices created by the MATLAB command `randn`. To compute $r(R; B,C) = r(J; B,C)$, as well as $r(Q;B,C)$, we ran
1. the Boyd-Balakrishnan (BB) algorithm [@Boyd1990],
2. the subspace framework that does not preserve the DH structure [@Aliyev2017 Algorithm 1] described in Subsection \[sec:sf\_descriptor\], and
3. the subspace frameworks that preserve structure, i.e., Algorithms \[alg:dti\] and \[alg:dti2\], introduced in Subection \[sec:sf\_PH\]
on $100$ such random examples. The spectrum of a typical $(J-R)Q$ of size $800$ generated in this way is depicted in Figure \[fig:spectra\_randomPH\] on the left.
-2.5ex
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![The spectra of $A = (J-R)Q$ for a dense random $J, R, Q \in {\mathbb R}^{800\times 800}$ (left), and a sparse random $J, R, Q \in {\mathbb R}^{5000\times 5000}$ (right). The MATLAB commands yielding these $J, R, Q$ are specified in Section \[sec:numexp\_syn\].[]{data-label="fig:spectra_randomPH"}](dense_spectrum.pdf "fig:"){width="6.cm"} ![The spectra of $A = (J-R)Q$ for a dense random $J, R, Q \in {\mathbb R}^{800\times 800}$ (left), and a sparse random $J, R, Q \in {\mathbb R}^{5000\times 5000}$ (right). The MATLAB commands yielding these $J, R, Q$ are specified in Section \[sec:numexp\_syn\].[]{data-label="fig:spectra_randomPH"}](sparse_spectrum.pdf "fig:"){width="6.cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The progress of Algorithm \[alg:dti2\], as well as Algorithm 1 in [@Aliyev2017], to compute $r(Q; B,C)$ for this example is presented in Figure \[fig:progress\_alg2\], which includes on the top left a plot of $f(\omega) := \sigma_{\max}(C ({\rm i} \omega I - (J-R) Q)^{-1} (J-R)B)$ for $\omega \in [-2000,0]$ along with the converged maximizers by the respective Algorithms. Algorithm \[alg:dti2\] converges to the global maximizer $\omega_{\ast,1} = -731.9774$ with $f(\omega_{\ast,1}) = 32.321399$, while Algorithm 1 in [@Aliyev2017] converges to the local maximizer $\omega_{\ast,2} = -1602.1187$ with $f(\omega_{\ast,2}) = 29.028197$. The globally optimal peak $(\omega_{\ast,1}, f(\omega_{\ast,1}))$ and the locally optimal peak $(\omega_{\ast,2}, f(\omega_{\ast,2}))$ are marked in the plot with a square and a circle, respectively.
The remaining five plots in Figure \[fig:progress\_alg2\] illustrate the progress of Algorithm \[alg:dti2\]. In each one of these plots, the black curve is a plot of the reduced function $f_k(\omega) := \max_{\omega \in {\mathbb R} \cup \infty} \sigma_{\max}(C_k ({\rm i} \omega I - (J_k-R_k) Q_k)^{-1} (J_k-R_k)B_k)$ with respect to $\omega$, and the circle marks the global maximizer of this reduced function. The top right shows the initial reduced function in black interpolating the full function at ten points, and the other four show the reduced function after iterations $1$-$4$ from middle-left to bottom-right. Observe that, at every iteration, the refined reduced function interpolates the full function at the maximizer of the previous reduced function in addition to the earlier interpolation points. We also list the iterates of Algorithm \[alg:dti2\] in Table \[table:dense\_iterates\_alg2\] indicating a quick converge. The algorithm terminates after performing six subspace iterations.
The results of Algorithms \[alg:dti\] and \[alg:dti2\] for the first $10$ random examples are presented in Tables \[table:dense\_comparison\] and \[table:dense\_comparisonQ\], respectively. Results from [@Aliyev2017 Algorithm 1] and the BB Algorithm [@Boyd1990] are also included in these tables for comparison purposes. For the computation of $r(J; B,C) = r(R; B,C)$, the new structure-preserving Algorithm \[alg:dti\] and [@Aliyev2017 Algorithm 1] perform equally well on these first $10$ examples. They both return the globally optimal solutions in $9$ out of $10$ examples, perform similar number of subspace iterations and require similar amount of cpu-time.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4.pdf "fig:"){width="5.6cm"} ![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4_iter1.pdf "fig:"){width="5.6cm"}
![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4_iter2.pdf "fig:"){width="5.6cm"} ![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4_iter3.pdf "fig:"){width="5.6cm"}
![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4_iter4.pdf "fig:"){width="5.6cm"} ![**(Top Left)** The plot of $\sigma_{\max}(C({\rm i} \omega - (J-R)Q)^{-1} (J-R)B)$ as a function of $\omega \in [-2000,0]$ along with the maxima computed by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] marked with the square and circle, respectively, for a dense random example of order $800$. **(Top Right)** The black curve is the initial reduced function for Algorithm \[alg:dti2\] interpolating the full function at $10$ points, whereas the circle is the global maximum of this reduced function. **(Middle Left - Bottom Right)** Plots of the reduced functions after iterations $1$-$4$ of Algorithm \[alg:dti2\] displayed with black curves along with the maximizers of the reduced functions marked with circles.[]{data-label="fig:progress_alg2"}](denseexample4_iter5.pdf "fig:"){width="5.6cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
A more decisive conclusion can be drawn when we consider all of the $100$ random examples. The left-hand columns in Figure \[fig:dense\_functions\] depict the ratios $(f_{\rm BB} - f_{\rm SF}) / ((f_{\rm BB} + f_{\rm SF})/2)$, where $f_{BB}$ are the globally maximal values of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ over $\omega$ returned by the BB algorithm and $f_{\rm SF}$ are the values returned by the subspace frameworks, specifically by Algorithm \[alg:dti\] on the top and by [@Aliyev2017 Algorithm 1] at the bottom. The results by Algorithm \[alg:dti\] match with the ones by the BB algorithm $81$ times out of $100$, while the results by Algorithm [@Aliyev2017 Algorithm 1] match with the ones by the BB algorithm $67$ times out of $100$. (In the examples where the results by the subspace frameworks differ from those by the BB algorithm, the subspace frameworks converge to local maximizers that are not global maximizers.)
On these $100$ random examples Algorithm \[alg:dti\] performs slightly fewer iterations, on average $17.6$, whereas [@Aliyev2017 Algorithm 1] on average performs $20.3$ iterations. On the other hand, the total run-time on average is better for [@Aliyev2017 Algorithm 1] compared with Algorithm \[alg:dti\] here, $21.3\, s$ vs $30.9\, s$. We observe this behavior on various other DH systems; Algorithm \[alg:dti\] seems to be more robust for the computation of $r(J; B,C) = r(R; B,C)$ in converging to the globally maximal value of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ compared with [@Aliyev2017 Algorithm 1], however, this is at the expense of slightly more computation time.
On the other hand, for the computation of $r(Q; B, C)$, Table \[table:dense\_comparisonQ\] indicates that Algorithm \[alg:dti2\] returns exactly the same globally maximal values (up to tolerances) as the BB algorithm for all of the first $10$ examples except one, whereas application of [@Aliyev2017 Algorithm 1] results in locally maximal solutions that are not globally maximal $4$ times. Fewer number of subspace iterations in favor of Algorithm \[alg:dti2\] are also apparent from the table. Once again, the plots of the ratios $(f_{\rm BB} - f_{\rm SF}) / ((f_{\rm BB} + f_{\rm SF})/2)$ are shown in Figure \[fig:dense\_functions\] on the right-hand column for all $100$ examples with $f_{\rm SF}$ now representing the values returned by Algorithm \[alg:dti2\] on the top and by [@Aliyev2017 Algorithm 1] at the bottom. Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1] return locally optimal solutions that are not globally optimal $21$ and $27$ times, respectively. In this case the difference between the number of subspace iterations for these $100$ examples is more pronounced in favor of Algorithm \[alg:dti2\]; indeed the number of subspace iterations is on average $7.2$ for Algorithm \[alg:dti2\] and and $17.0$ for [@Aliyev2017 Algorithm 1]. This difference in the number of iterations is also reflected in the average run-times which are $13.2\, s$ and $19\, s$ for Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1], respectively.
-------------- ---------------- ------------------------------------
$k$ $\omega_{k+1}$ $\sigma_{\max}(G_k(\omega_{k+1}))$
\[0.5ex\] 10 -600.705819 26.182525
11 -674.769938 28.262865
12 -697.139310 34.834307
13 -731.573363 35.133647
14 -731.942586 32.309246
15 -731.977386 32.321399
16 -731.977385 32.321399
-------------- ---------------- ------------------------------------
: Iterates of Algorithm \[alg:dti2\] to compute $r(Q; B, C)$ on a DH system with dense random $J, R, Q \in {\mathbb R}^{800\times 800}$ and random restriction matrices $B \in {\mathbb R}^{800\times 2}, C \in {\mathbb R}^{2\times 800}$. The algorithm is initiated with $10$ interpolation points and terminates after $6$ iterations with $32$ dimensional subspaces. []{data-label="table:dense_iterates_alg2"}
--------------- ------------------ ---------------------- --------------------- ------------------ --------------- ------------------ ---------------
\[0.5ex\] Ex. Alg. \[alg:dti\] [@Aliyev2017 Alg. 1] BB Alg. [@Boyd1990] Alg. \[alg:dti\] [@Aliyev2017] Alg. \[alg:dti\] [@Aliyev2017]
\[0.5ex\] 1 32.559659 32.559659 32.559659 9 9 14 13
2 46.703932 46.703932 46.703932 15 12 16 12.4
3 26.227029 **24.023572** 26.227029 7 12 12.9 15.1
4 **62.748090** 108.030409 108.030409 17 41 17.8 27.2
5 35.974956 35.974957 35.974956 9 14 13.4 13.9
6 53.522033 53.522033 53.522033 6 3 11.4 10.2
7 31.739000 31.739000 31.739000 4 12 11.6 13.8
8 76.958658 76.958658 76.958658 35 8 43.2 11
9 37.007241 37.007241 37.007241 6 2 13 11.5
10 155.642871 155.642871 155.642871 2 2 8.6 8.5
--------------- ------------------ ---------------------- --------------------- ------------------ --------------- ------------------ ---------------
: Run-time (in $s$) comparison of Algorithm \[alg:dti\] and [@Aliyev2017 Algorithm 1] to compute $r(R; B, C) = r(J; B, C)$ for $10$ dense random examples of order $800$. The third column refers to the number of subspace iterations.[]{data-label="table:dense_comparison"}
-1.3ex
[|c||ccc|cc|cc|]{} & & &\
\[0.5ex\] $\#$ & Alg. \[alg:dti\] & [@Aliyev2017 Alg. 1] & BB Alg. [@Boyd1990] & Alg. \[alg:dti\] & [@Aliyev2017] & Alg. \[alg:dti\] & [@Aliyev2017]\
\[0.5ex\]
1 & 9.809182 & 9.809182 & 9.809182 & 3 & 26 & 11.1 & 19.5\
2 & 22.386670 & 22.386670 & 22.386670 & 5 & 26 & 10.2 & 18.1\
3 & 8.364927 & 8.364927 & 8.364927 & 3 & 8 & 11 & 13.3\
4 & 32.321399 & **29.028197** & 32.321399 & 6 & 37 & 10.5 & 25.2\
5 & 15.071678 & 15.071678 & 15.071678 & 7 & 15 & 12.7 & 14\
6 & 21.641484 & 21.641484 & 21.641484 & 4 & 8 & 10.2 & 11.6\
7 & 12.858494 & **12.763161** & 12.858494 & 6 & 4 & 11.8 & 11.2\
8 & 31.901305 & **27.996873** & 31.901305 & 8 & 12 & 11.5 & 12.5\
9 & 9.228945 & 9.228945 & 9.228945 & 3 & 8 & 11.3 & 13\
10 & **47.697528** & **47.697528** & 71.534252 & 10 & 8 & 11.8 & 10.2\
-3ex
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![**(Left Column)** Ratios $(f_{BB} - f_{SF}) / ((f_{BB} + f_{SF})/2)$ for $100$ dense random DH examples of order $800$, where $f_{BB}$ and $f_{SF}$ denote the maximal values of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1}B)$ over $\omega$ computed by the BB algorithm and the subspace framework (i.e., Algorithm \[alg:dti\] for the plot on the top, [@Aliyev2017 Algorithm 1] for the plot at the bottom). **(Right Column)** Same as the left column on the same $100$ dense random examples of order $800$ except now this concerns a comparison of the maximization of $\sigma_{\max}(C ({\rm i} \omega I - (J-R)Q)^{-1}(J-R)B)$ using Algorithm \[alg:dti2\] in the top plot and [@Aliyev2017 Algorithm 1] in the bottom plot.[]{data-label="fig:dense_functions"}](dense_PH_function.pdf "fig:"){width="6cm"} ![**(Left Column)** Ratios $(f_{BB} - f_{SF}) / ((f_{BB} + f_{SF})/2)$ for $100$ dense random DH examples of order $800$, where $f_{BB}$ and $f_{SF}$ denote the maximal values of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1}B)$ over $\omega$ computed by the BB algorithm and the subspace framework (i.e., Algorithm \[alg:dti\] for the plot on the top, [@Aliyev2017 Algorithm 1] for the plot at the bottom). **(Right Column)** Same as the left column on the same $100$ dense random examples of order $800$ except now this concerns a comparison of the maximization of $\sigma_{\max}(C ({\rm i} \omega I - (J-R)Q)^{-1}(J-R)B)$ using Algorithm \[alg:dti2\] in the top plot and [@Aliyev2017 Algorithm 1] in the bottom plot.[]{data-label="fig:dense_functions"}](dense_PH_function_perturbQ.pdf "fig:"){width="6cm"}
![**(Left Column)** Ratios $(f_{BB} - f_{SF}) / ((f_{BB} + f_{SF})/2)$ for $100$ dense random DH examples of order $800$, where $f_{BB}$ and $f_{SF}$ denote the maximal values of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1}B)$ over $\omega$ computed by the BB algorithm and the subspace framework (i.e., Algorithm \[alg:dti\] for the plot on the top, [@Aliyev2017 Algorithm 1] for the plot at the bottom). **(Right Column)** Same as the left column on the same $100$ dense random examples of order $800$ except now this concerns a comparison of the maximization of $\sigma_{\max}(C ({\rm i} \omega I - (J-R)Q)^{-1}(J-R)B)$ using Algorithm \[alg:dti2\] in the top plot and [@Aliyev2017 Algorithm 1] in the bottom plot.[]{data-label="fig:dense_functions"}](dense_descriptor_function.pdf "fig:"){width="6cm"} ![**(Left Column)** Ratios $(f_{BB} - f_{SF}) / ((f_{BB} + f_{SF})/2)$ for $100$ dense random DH examples of order $800$, where $f_{BB}$ and $f_{SF}$ denote the maximal values of $\sigma_{\max}(CQ ({\rm i} \omega I - (J-R)Q)^{-1}B)$ over $\omega$ computed by the BB algorithm and the subspace framework (i.e., Algorithm \[alg:dti\] for the plot on the top, [@Aliyev2017 Algorithm 1] for the plot at the bottom). **(Right Column)** Same as the left column on the same $100$ dense random examples of order $800$ except now this concerns a comparison of the maximization of $\sigma_{\max}(C ({\rm i} \omega I - (J-R)Q)^{-1}(J-R)B)$ using Algorithm \[alg:dti2\] in the top plot and [@Aliyev2017 Algorithm 1] in the bottom plot.[]{data-label="fig:dense_functions"}](dense_descriptor_function_perturbQ2.pdf "fig:"){width="6cm"}
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Sparse Random Examples.** The $5000\times 5000$ sparse matrices $J, Q, R$ are constrained to be banded with bandwidth $10$. The matrix $J$ is generated as in the dense family randomly using the `randn` command, but the entries that fall outside of the bandwidth $10$ are set equal to zero. The matrix $Q>0$ is created using the commands
>> A = sprandn(n,n,1/n); >> Q = (A + A')/2,
followed by setting the entries outside the bandwidth $10$ again to zero. Finally, the following commands ensure that $Q>0$.
>> mineig = eigs(Q,1,'smallestreal');
>> if (mineig<10^-4) Q=Q+(-mineig+5*rand)*speye(n); end
To form $R\geq 0$, first a diagonal matrix $D$ of random rank not exceeding $500$ is generated by the commands
>> p = round(500*rand); D = sparse(5000,5000); h = n/p;
>> for j=1:p k = floor(j*h); D(k,k) = 5*rand; end
Then we set `R = sparse(X’*D*X)` for a square random matrix $X$ with bandwidth $5$. The matrices $B$, $C$ are random, and of size $5000 \times 2$, $2 \times 5000$, respectively. The spectrum of a typical such sparse matrix $(J-R)Q$ is displayed in Figure \[fig:spectra\_randomPH\] on the right-hand side. We again apply Algorithms \[alg:dti\] and \[alg:dti2\] to $100$ such random sparse examples. Since the matrices are too large to apply the BB algorithm, we compare the structure-preserving algorithms directly with the unstructured algorithm [@Aliyev2017 Algorithm 1]. The retrieved estimates of $\max_{\omega \in {\mathbb R} \cup \infty} \sigma_{\max}(G({\rm i} \omega))$ for $G({\rm i}\omega) = CQ ({{\mathrm i}}\omega I - (J-R)Q)^{-1}B$ and $G({\rm i}\omega) = C ({{\mathrm i}}\omega I - (J-R)Q)^{-1}(J-R)B$ are compared on the top and at the bottom, respectively, in Figure \[fig:sparse\_functions\]. Specifically, the ratio $
\frac{2 (f_{\rm ST} - f_{\rm UN})}{f_{\rm ST} + f_{\rm UN}}
$ is plotted for each random example with $f_{\rm UN}$ denoting the estimate by the unstructured algorithm [@Aliyev2017 Algorithm 1], and $f_{\rm ST}$ denoting the estimate by the structured algorithm, i.e., Algorithm \[alg:dti\] for the top plot, Algorithm \[alg:dti2\] for the bottom plot. According to the top plot, which concerns the computation of $r(J; B, C) = R(R; B, C)$, the two algorithms return exactly the same results (up to tolerances) for all but $6$ examples; the structured algorithm returns better estimates for $4$ of these $6$ examples, while the unstructured algorithm returns better estimates for the other two. The structured algorithm appears to be even more robust for the computation of $r(Q; B, C)$ in terms of avoiding locally optimal solutions away from global solutions; as displayed at the bottom, the structured algorithm returns a better estimate for $40$ of the $100$ examples, the unstructured algorithm returns the better estimate for $6$ examples, and the results match exactly up to the tolerances for the remaining $54$ examples.
The structured algorithms perform typically fewer iterations as compared to the unstructured algorithm. Indeed the average value of the number of subspace iterations performed on these $100$ examples is $6.3$ for the structured and $9.8$ for the unstructured algorithm for the computation of $r(J; B,C) = r(R; B, C)$, while these average values are $9.7$ and $12.9$ for the computation of $r(Q; B, C)$. On the other hand, the unstructured algorithm is slightly superior when run-times are taken into account. The average run-times are $14.3\, s$ for the structured and $12.7\, s$ for the unstructured algorithm for the computation of $r(J; B, C) = r(R; B, C)$, whereas these figures are $16.5\, s$ and $14\, s$ for the computation of $r(Q; B, C)$.
We also list the computed maximal values of $\sigma_{\max}(CQ({\rm i} \omega I - (J-R)Q)^{-1}B)$ and $\sigma_{\max}(C({\rm i} \omega I - (J-R)Q)^{-1}(J-R)B)$ over $\omega$ for the first $10$ of these sparse examples in Tables \[table:sparse\_iterates\_alg\] and \[table:sparse\_iterates\_algQ\]. Included in these tables are also the number of subspace iterations, as well as the run-time required by the structured and the unstructured algorithm.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ --
![ **(Top)** Plot of the ratios $(f_{\rm ST} - f_{\rm UN})/((f_{\rm ST} + f_{\rm UN})/2)$ on $100$ sparse random examples of order $5000$, where $f_{\rm ST}, f_{\rm UN}$ represent the computed maximal value of $\sigma_{\max}(CQ({\rm i} \omega I - (J-R)Q)^{-1} B)$ over $\omega$ by Algorithm \[alg:dti\] and [@Aliyev2017 Algorithm 1], respectively. **(Bottom)** Similar to the top plot, only now for the computed maximal values $f_{\rm ST}$, $f_{\rm UN}$ of $\sigma_{\max}(C({\rm i} \omega I - (J-R)Q)^{-1} (J-R)B)$ over $\omega$ by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1], respectively. []{data-label="fig:sparse_functions"}](sparse_function.pdf "fig:"){width="8cm"}
![ **(Top)** Plot of the ratios $(f_{\rm ST} - f_{\rm UN})/((f_{\rm ST} + f_{\rm UN})/2)$ on $100$ sparse random examples of order $5000$, where $f_{\rm ST}, f_{\rm UN}$ represent the computed maximal value of $\sigma_{\max}(CQ({\rm i} \omega I - (J-R)Q)^{-1} B)$ over $\omega$ by Algorithm \[alg:dti\] and [@Aliyev2017 Algorithm 1], respectively. **(Bottom)** Similar to the top plot, only now for the computed maximal values $f_{\rm ST}$, $f_{\rm UN}$ of $\sigma_{\max}(C({\rm i} \omega I - (J-R)Q)^{-1} (J-R)B)$ over $\omega$ by Algorithm \[alg:dti2\] and [@Aliyev2017 Algorithm 1], respectively. []{data-label="fig:sparse_functions"}](sparse_function_perturbQ.pdf "fig:"){width="8cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ --
---------------- ------------------------ ------------------------ ------------------ --------------- ------------------ ---------------
\[0.5ex\] $\#$ Alg. \[alg:dti\] [@Aliyev2017 Alg. 1] Alg. \[alg:dti\] [@Aliyev2017] Alg. \[alg:dti\] [@Aliyev2017]
\[0.5ex\] 1 $1.393086\times 10^3$ $1.393086\times 10^3$ 2 3 10.7 10.1
2 $8.323309\times 10^2$ $8.323309\times 10^2$ 6 11 12.3 11.9
3 $1.289416 \times 10^3$ $1.289416 \times 10^3$ 4 3 11.2 10.6
4 $8.850355 \times 10^2$ $8.850355 \times 10^2$ 5 16 13.2 14
5 $6.891467 \times 10^2$ $6.891467 \times 10^2$ 26 46 30.3 31.4
6 $5.652337\times 10^6$ $5.652337\times 10^6$ 1 1 6.6 6.1
7 $8.834190\times 10^2$ $8.834190\times 10^2$ 1 1 9.7 9.6
8 $3.402375\times 10^3$ $3.402375\times 10^3$ 1 2 9.1 9
9 $8.240097\times 10^2$ $8.240097\times 10^2$ 2 3 12.3 12.1
10 $1.256781\times 10^3$ $1.256781\times 10^3$ 2 10 10.9 11.3
---------------- ------------------------ ------------------------ ------------------ --------------- ------------------ ---------------
: Comparison of Algorithms \[alg:dti\] and [@Aliyev2017 Algorithm 1] to compute $r(R; B,C) = r(J; B, C)$ on sparse random DH systems of order $5000$ of bandwidth $10$. The MATLAB commands to generate these random sparse examples are explained in Section \[sec:numexp\_syn\]. The third column lists the number of subspace iterations, while the run-times (in $s$) are listed in the last column.[]{data-label="table:sparse_iterates_alg"}
-2ex
---------------- ----------------------- --------------------------------- ------------------- --------------- ------------------- ---------------
\[0.5ex\] $\#$ Alg. \[alg:dti2\] [@Aliyev2017 Alg. 1] Alg. \[alg:dti2\] [@Aliyev2017] Alg. \[alg:dti2\] [@Aliyev2017]
\[0.5ex\] 1 $1.320177\times 10^3$ $\mathbf{9.360406\times 10^2}$ 15 39 18.3 26.7
2 $8.284036\times 10^2$ $8.284036\times 10^2$ 2 13 11.1 12.7
3 $9.288427\times 10^2$ $\mathbf{4.863413 \times 10^2}$ 31 11 46.1 11.8
4 $6.583171\times 10^2$ $6.583171 \times 10^2$ 5 17 13.1 14.6
5 $7.722736\times 10^2$ $7.722736\times 10^2$ 3 9 11.3 11.4
6 $2.260647\times 10^6$ $2.260647\times 10^6$ 1 1 6.4 6.1
7 $1.660164\times 10^3$ $1.660164\times 10^3$ 4 12 10.6 11.1
8 $1.515824\times 10^3$ $1.515824\times 10^3$ 8 9 11.8 10
9 $4.610935\times 10^2$ $\mathbf{2.535771\times 10^2}$ 13 4 17.9 12.1
10 $1.011299\times 10^3$ $1.011299\times 10^3$ 3 7 10.8 10.6
---------------- ----------------------- --------------------------------- ------------------- --------------- ------------------- ---------------
: Comparison of Algorithms \[alg:dti2\] and [@Aliyev2017 Algorithm 1] to compute $r(Q; B, C)$. The display is analogous to Table \[table:sparse\_iterates\_alg\], in particular the numerical experiments are carried out exactly on the same $10$ sparse random examples of order $5000$ employed for Table \[table:sparse\_iterates\_alg\]. []{data-label="table:sparse_iterates_algQ"}
### The FE model of a Disk Brake {#sec:numexp_brake}
The only large-scale computation required by Algorithm \[alg:dti\] is the solution of the linear systems $$\label{eq:large_lin_sys}
D({\rm i} \omega) \: X = B \quad {\rm and} \quad D({\rm i} \omega)^2 \: Y = B$$ at a given $\omega \in {\mathbb R}$ in lines \[defn\_init\_subspaces0\] and \[defn\_later\_subspaces\], where $D({\rm i} \omega) = {\rm i} \omega I - (J - R)Q$. For the DH system resulting from a FE model of a disk brake in (\[insta\]) and (\[eq:insta\_matrices\]), the mass matrix $M$ and the stiffness matrix $K(\Omega)$ are available from the FE modeling. In other words, we have the sparse matrix $Q^{-1}$, but not $Q$, which turns out to be dense. Trying to invert $Q^{-1}$ and/or solve a linear system with the coefficient matrix $Q$ is computationally very expensive and would require full matrix storage.
This difficulty can be avoided by exploiting that $$( {\rm i} \omega I - (J - R)Q )^{-1} B \;\;\; = \;\;\; Q^{-1} ( {\rm i} \omega Q^{-1} - (J - R) )^{-1} B.$$ Hence, to compute $X, Y$ as in (\[eq:large\_lin\_sys\]), we proceed as follows.
1. We first solve $( {\rm i} \omega Q^{-1} - (J - R) ) \widehat{X} = B$ for $\widehat{X}$, and set $X = Q^{-1} \widehat{X}$.
2. Then we solve $( {\rm i} \omega Q^{-1} - (J - R) ) \widehat{Y} = X$ for $\widehat{Y}$, and set $Y = Q^{-1} \widehat{Y}$.
A second observation that further speeds up the computation is the particular structure of the coefficient matrix $\{ {\rm i} \omega Q^{-1} - (J - R) \}$ with $N = 0$. Setting $\widetilde{M}({\rm i} \omega; \Omega) := {\rm i} \omega M + D(\Omega) + G(\Omega)$, we have $${\rm i} \omega Q^{-1} - (J - R) \;\; = \;\;
\left[
\begin{array}{cc}
\widetilde{M}({\rm i} \omega; \Omega) & K(\Omega) \\
-K(\Omega) & {\rm i} \omega K(\Omega)
\end{array}
\right].$$ Hence, to solve $\left\{ {\rm i} \omega Q^{-1} - (J - R) \right\} Z = W$, for a given $W
=
\left[
\begin{array}{cc}
W_1^T & W_2^T
\end{array}
\right]^T$ and the unknown $Z
=
\left[
\begin{array}{cc}
Z_1^T & Z_2^T
\end{array}
\right]^T$ with $W_1, W_2, Z_1, Z_2$ having all equal number of rows, we perform a column block permutation and then eliminate the lower left block to obtain $$\left[
\begin{array}{cc}
K(\Omega) & \widetilde{M}({\rm i} \omega; \Omega) \\
0 & -K(\Omega) - {\rm i} \omega \widetilde{M}({\rm i} \omega; \Omega)
\end{array}
\right]
\left[
\begin{array}{c}
Z_2 \\
Z_1
\end{array}
\right]
\;\; = \;\;
\left[
\begin{array}{c}
W_1 \\
W_2 - {\rm i} \omega W_1
\end{array}
\right],$$ which in turn yields $$(-K(\Omega) - {\rm i} \omega \widetilde{M}({\rm i} \omega; \Omega) ) Z_1 = W_2 - {\rm i} \omega W_1,\
K(\Omega) Z_2 = W_1 - \widetilde{M}({\rm i} \omega; \Omega) Z_1.$$ At every subspace iteration, the highest costs arise from the computation of the $LU$ factorizations of the sparse matrices $K(\Omega)$ and $K(\Omega) + {\rm i} \omega \widetilde{M}({\rm i} \omega; \Omega)$. The main cost for Algorithm \[alg:dti2\] is the solution of the linear systems $$D({\rm i} \omega)^H X \;\; = \;\; C^H
\quad {\rm and} \quad
\left[ D({\rm i} \omega)^H \right]^2 Y \;\; = \;\; C^H$$ at a given $\omega \in {\mathbb R}$. This can be treated similarly by exploiting that $$D({\rm i} \omega)^{-H} C^H
\;\; = \;\;
( {\rm i} \omega Q^{-1} - (J-R))^{-H} (CQ^{-1})^H.$$
We have applied Algorithm \[alg:dti\] to compute the unstructured stability radius $r(R; B, B^T)$ for the DH system of the form (\[insta\]), (\[eq:insta\_matrices\]) resulting from the FE brake model with $N = 0$, where $G(\Omega), K(\Omega), D(\Omega), M \in {\mathbb R}^{4669 \times 4669}$ so that $J, R, Q \in {\mathbb R}^{9338\times 9338}$.
The plot of the computed $r(R; B, B^T)$ vs the rotation speed $\Omega$ is presented in Figure \[fig:brake\_squeal\_unstructured\] at lower frequencies (i.e., $\Omega \in [2.5, 100]$) on the top, and at higher frequencies (i.e., $\Omega \in [900, 1700]$) at the bottom. For smaller frequencies, the stability radius initially decreases with respect to $\Omega$, but around $\Omega = 1100$ the stability radius suddenly increases. The non-smooth nature of the stability radius with respect to $\Omega$ is apparent from the figure. One should note, in particular, the sharp turns near $\Omega = 1120$ and $\Omega = 1590$; this non-smoothness is due to the fact that $\sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ has multiple global maximizers. This means that two distinct points on the imaginary axis can be attained with perturbations of minimal norm.
The computed values of $r(R; B, B^T)$ are listed in Table \[tab:brake\_squeal\_unstructured\] for some values of $\Omega$. In this table, for each $\Omega$, the value $\omega_\ast$, where the singular value function $\sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ is maximized globally is displayed, the number of subspace iterations, the run-time (in $s$) and the subspace dimension at termination are included as well. In all cases, $2$ or $3$ subspace iterations are sufficient to achieve the prescribed accuracy tolerance. This leads to considerably smaller reduced systems of size $72 \times 72$ or $78\times 78$ compared with the original problem of size $9338\times 9338$. The $\omega$ value maximizing $\sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ differs substantially, depending on whether the frequency $\Omega$ is small or large.
The resulting reduced problems at termination capture the full problem remarkably well around the global maximizer. This is depicted in Figure \[fig:brake\_squeal\_unstructured\_sval\], where, for $\Omega = 1000$, the singular value function $f(\omega) = \sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ for the full problem (solid curve) and $f_k(\omega) = \sigma_{\max}(B^T_kQ_k ({\rm i} \omega I - (J_k-R_k)Q_k)^{-1} B_k)$ for the reduced problem at termination (dashed curve) are plotted near the global maximizer $\omega_\ast = -178880.9$. It is fairly difficult to distinguish these two curves from each other.
$\Omega$ $r(R; B, B^T)$ $\omega_\ast$ iterations run-time dimension
---------- ---------------- --------------------- ------------ ---------- -----------
2.5 0.01066 $-1.938\times 10^5$ 2 22.0 72
5 0.01038 $-1.938\times 10^5$ 2 21.7 72
10 0.01026 $-1.938\times 10^5$ 3 24.9 78
50 0.00999 $-1.938\times 10^5$ 2 22.0 72
100 0.00988 $-1.938\times 10^5$ 2 22.1 72
1000 0.00809 $-1.789\times 10^5$ 2 21.4 72
1050 0.00789 $-1.789\times 10^5$ 2 21.8 72
1100 0.00834 $-1.789\times 10^5$ 3 24.8 78
1116 0.01091 $-1.789\times 10^5$ 3 25.6 78
1150 0.00344 $-1.742\times 10^5$ 2 21.5 72
1200 0.00407 $-1.742\times 10^5$ 2 21.8 72
1250 0.00471 $-1.742\times 10^5$ 2 21.8 72
1300 0.00516 $-1.742\times 10^5$ 2 21.2 72
: Computed stability radii $r(R; B, B^T)$ by Algorithm \[alg:dti\] for several $\Omega$ values for the DH system of order $9338$ originating from the FE brake model. The other columns display $\omega_\ast$ corresponding to $\: \operatorname*{arg\,max}_{\omega} \sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B) \:$, the number of subspace iterations, the total run-time (in $s$) and the subspace dimension at termination. []{data-label="tab:brake_squeal_unstructured"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Plot of the stability radius $r(R; B, B^T)$ for the DH system (\[insta\]), (\[eq:insta\_matrices\]) resulting from the FE model of a disk-brake as a function of the rotation speed $\Omega$ for $\Omega \in [2.5, 100]$ (top plot), and $\Omega \in [900, 1700]$ (bottom plot). The order of the DH system under consideration in this plot is $9338$. []{data-label="fig:brake_squeal_unstructured"}](brake_squeal_unstructured.pdf "fig:"){width="8cm"}
![ Plot of the stability radius $r(R; B, B^T)$ for the DH system (\[insta\]), (\[eq:insta\_matrices\]) resulting from the FE model of a disk-brake as a function of the rotation speed $\Omega$ for $\Omega \in [2.5, 100]$ (top plot), and $\Omega \in [900, 1700]$ (bottom plot). The order of the DH system under consideration in this plot is $9338$. []{data-label="fig:brake_squeal_unstructured"}](brake_squeal_unstructured2.pdf "fig:"){width="8cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![ Plot of the singular value functions $f(\omega) = \sigma_{\max}(B^TQ ({\rm i} \omega I - (J-R)Q)^{-1} B)$ (solid curve) and $f_k(\omega) = \sigma_{\max}(B^T_kQ_k ({\rm i} \omega I - (J_k-R_k)Q_k)^{-1} B_k)$ (dashed curve) at termination of Algorithm \[alg:dti\] near the global maximizer $\omega_\ast = -178880.9$ for the DH system of order $9338$ arising from the FE disk-brake model with $\Omega = 1000$. The circle marks $(\omega_\ast, f(\omega_\ast))$.[]{data-label="fig:brake_squeal_unstructured_sval"}](brake_squeal_unstructured_sval.pdf "fig:"){width="8cm"}
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Computation of the Structured Stability Radius {#sec:structured}
==============================================
In the last section we have studied stability radii for dissipative Hamiltonian systems where the restriction matrices, however, allowed unstructured perturbations in the system coefficients. In this section we put additional constraints on the perturbations, in particular we require that the perturbations are structured themselves. We discuss only perturbations in the dissipation matrix, since this is usually the most uncertain part of the system, due to the fact that modeling damping or friction very exactly is usually extremely difficult. We deal with the computation of $r^{\rm Herm}(R; B)$ defined as in (\[eq:structured\_radii\_defn\]). We first describe a numerical technique for small-scale problems in Section \[sec:small\_scale\_st\] and then develop a subspace framework that converges superlinearly with respect to the subspace dimension in Section \[sec:large\_scale\_st\]. Both techniques use the eigenvalue optimization characterization of $r^{\rm Herm}(R; B)$ in Theorem \[thm\_str\_Herm\].
Small-Scale Problems {#sec:small_scale_st}
--------------------
### Inner Maximization Problems {#sec:small_scale_inner}
The eigenvalue optimization characterization of $r^{\rm Herm}(R; B)$ is a min-max problem, where the inner maximization problem is concave, indeed it can alternatively be expressed as a semi-definite program (SDP). Formally, for a given $\omega \in {\mathbb R}$, and $H_0({\rm i} \omega), H_1({\rm i} \omega)$ representing the Hermitian matrices defined in Theorem \[thm\_str\_Herm\], we have $$\label{eq:inner_max}
\begin{split}
\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)
& \; = \;
\sup_{t \in {\mathbb R}} \: \lambda_{\min} (H_0({\rm i} \omega) + t H_1 ({\rm i} \omega)) \\
& \; = \;
\sup\{ z \; | \; z,t \in {\mathbb R} \;\; {\rm s.t.} \;\; H_0({\rm i} \omega) + t H_1 ({\rm i} \omega) - zI \geq 0 \},
\end{split}$$ where the characterization in the second line is a linear convex SDP. Here $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ is related to a structured backward error for the eigenvalue ${\rm i} \omega$, specifically it corresponds to the square of the distance (see [@MehMS16a Definition 3.2 and Theorem 4.9]) $$\label{eq:backward_error}
\begin{split}
\eta^{\rm Herm}(R; B, {\rm i}\omega)
\; := \; \hskip 49ex \\
\hskip 13ex
\inf\{ \|\Delta \|_2 \; | \; \Delta = \Delta^H, \;\; {\rm i} \omega \in \Lambda \big( (J - R)Q - (B\Delta B^H)Q \big) \}.
\end{split}$$ We have that $\eta^{\rm Herm}(R; B, {\rm i}\omega)$ is finite if and only if the suprema in (\[eq:inner\_max\]) are attained, which happens if and only if $H_1({\rm i}\omega)$ is indefinite, i.e., $H_1({\rm i}\omega)$ has both negative and positive eigenvalues.
The most widely used techniques to solve a linear convex SDP are different forms of interior-point methods. Implementations of some of these interior-point methods are made available through the package `cvx` [@Grant2008; @cvx]. Hence, one option is to use `cvx` to compute $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ directly. An alternative, and also theoretically well understood approach, is to employ the software package `eigopt` [@Mengi2014] for the eigenvalue optimization problem in the first characterization in (\[eq:inner\_max\]). This second approach forms piece-wise quadratic functions that lie globally above the eigenvalue function, and maximizes these piece-wise quadratic functions instead of the eigenvalue function. Each piece-wise quadratic function is defined as the minimum of several other quadratic functions, all of which have the same curvature $\gamma$ (which must be a global upper bound on the second derivative of the eigenvalue function at all points where the eigenvalue function is differentiable). Any slightly positive real number for the curvature $\gamma$ serves the purpose (e.g., $\gamma = 10^{-6}$), since the smallest eigenvalue function in (\[eq:inner\_max\]) is a concave function of $t$.
In our experience, `eigopt` performs the computation of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ significantly faster than `cvx`. The only downside is that an interval containing the optimal $t$ for (\[eq:inner\_max\]) must be supplied to `eigopt`, whereas such an interval is not needed by `cvx`.
### Outer Minimization Problems {#sec:small_scale_outer}
The minimum of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ with respect to $\omega \in {\mathbb R}$ yields the distance $r^{\rm Herm}(R; B)$, and the minimizing $\omega \in {\mathbb R}$ yields the point ${\rm i} \omega$ that first becomes an eigenvalue on the imaginary axis under the smallest perturbation possible. This is a non-convex optimization problem, indeed the objective $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ may even blow up at some $\omega$. We again resort to `eigopt` for the minimization of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$. For the sake of completeness, a formal description is provided in Algorithm \[eigopt\] below, where we use the abbreviations $$\widetilde{\eta}_j \; := \; \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_j)
\quad {\rm and} \quad
\widetilde{\eta}^{\;'}_j \; := \; \frac{d\, \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_j)}{d \omega}.$$ Introducing $t(\omega) := \operatorname*{arg\,max}_{t \in {\mathbb R}} \: \lambda_{\min}(H_0({\rm i} \omega) + t H_1({\rm i} \omega))$, the algorithm approximates the smallest eigenvalue function $$\lambda_{\min}(H_0({\rm i} \omega) + t(\omega) H_1({\rm i} \omega))
\; = \;
\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$$ with the piece-wise quadratic function $$\begin{aligned}
Q_k(\omega) &:=& \max \{ q_j(\omega) \; | \; j = 0, \dots, k \}, \\
q_j(\omega) \; &:=& \; \widetilde{\eta}_j +
\widetilde{\eta}_j^{\;'} ( \omega - \omega_j ) + (\gamma/2) ( \omega - \omega_j )^2\end{aligned}$$ at iteration $k$. It computes the global minimizer $\omega_{k+1}$ of $Q_k(\omega)$, and refines the piece-wise quadratic function $Q_k(\omega)$ with the addition of one more quadratic piece, namely $q_{k+1}(\omega) \; := \; \widetilde{\eta}_{k+1} +
\widetilde{\eta}_{k+1}^{\;'} ( \omega - \omega_{k+1} ) + (\gamma/2) ( \omega - \omega_{k+1} )^2$. Here, $\gamma$ is supposed to be a lower bound for the second derivative $\lambda_{\min}''(H_0({\rm i} \omega) + t(\omega) H_1({\rm i} \omega))$ for all $\omega$ sufficiently close to the global minimizer of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$. In theory, it can be shown that for all $\gamma$ small enough, every convergent subsequence of the sequence $\{ \omega_k\}$ converges to a global minimizer of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$.
At step $k$ of the algorithm $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ and its derivative need to be computed at $\omega_{k+1}$. We rely on one of the two approaches (`cvx` or `eigopt`) described in Section \[sec:small\_scale\_inner\] for the computation of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_{k+1})$, and employ a finite difference formula to approximate its derivative.
Compute $\widetilde{\eta}_0 := \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_0)$ and $\widetilde{\eta}_0^{\;'} := d\, \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_0) / d\omega$ Compute $$\begin{aligned}
\widetilde{\eta}_{k+1} &:=& \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_{k+1}),\\
\widetilde{\eta}_{k+1}^{\;'} &:=& d\, \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_{k+1}) / d\omega\end{aligned}$$
Large-Scale Problems {#sec:large_scale_st}
--------------------
The characterization via eigenvalue optimization in Theorem \[thm\_str\_Herm\] is in terms of the matrix-valued functions $H_0({\rm i} \omega), H_1({\rm i} \omega)$, which are of small size provided that $B$ has few columns. The large-scale nature of this characterization is hidden in the matrix-valued function $W({\rm i} \omega) := (J-R)Q - {\rm i} \omega I$ defined in Theorem \[thm\_str\_Herm\]. Note that, in particular, both $H_0({\rm i} \omega)$ and $H_1({\rm i} \omega)$ are defined in terms of $W({\rm i} \omega)^{-1} B$. This is also reflected in Algorithm \[eigopt\] when $J, R, Q$ are large; at iteration $k$ of the algorithm, the matrices $H_0({\rm i} \omega_{k+1})$, $H_1({\rm i} \omega_{k+1})$ need to be formed for the computation of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega_{k+1})$, which in turn requires the solution of the linear system $W({\rm i} \omega_{k+1}) Z = B$.
To cope with the large-scale setting, we benefit from structure preserving two-sided projections similar to those described in Section \[sec:sf\_PH\]. In particular, for a given subspace ${\mathcal V}_k$ and a matrix $V_k$ whose columns form an orthonormal basis for ${\mathcal V}_k$, we set $$\label{eq:defn_Wr}
W_k \;\; := \;\; QV_k (V^H_k Q V_k)^{-1}.$$ Furthermore, we define the projected matrices $$\label{eq:projected_matrices}
\begin{split}
J_k \;\; & := \;\; W_k^H J W_k, \;\; R_k \;\; := \;\; W_k^H R W_k, \\
Q_k \;\; & := \;\; V_k^H Q V_k, \;\;\;\: B_k \;\; := \;\; W_k^H B.
\end{split}$$ Recall also the identities $$\label{eq:identities}
W_k^H V_k = I \quad {\rm and} \quad (W_k V_k^H)^2 = W_k V_k^H,$$ the latter of which means that $W_k V_k^H$ is an oblique projector onto ${\rm Im}(Q V_k)$.
Although these identities are still available, however, we no longer have a tool such as Theorem \[thm:Gal\_interpolate\] that we could depend on to establish interpolation results. This is because there is no apparent transfer function, as there is indeed no apparent linear port-Hamiltonian system that can be tied to the eigenvalue optimization characterization. But the following simple observation turns out to be very useful.
\[thm:reduced\_Wlambda\] Consider a DH model (\[DH\]) and a reduced model $\dot x_k= (J_k-R_k)Q_k x_k$ with coefficients as in (\[eq:projected\_matrices\]). With $W(\lambda) = (J - R) Q - \lambda I$ and $W_k(\lambda) := (J_k - R_k) Q_k - \lambda I$, we then have $$W_k(\lambda) \;\; = \;\; W_k^H \: W(\lambda) \: V_k \quad\quad \mbox{\rm for all}\ \lambda \in {\mathbb C}.$$
From the definition of $J_k, R_k, Q_k$ in (\[eq:projected\_matrices\]) we obtain $$\begin{split}
W_k(\lambda) & \;\; = \;\; (W_k^H (J - R) W_k) V_k^H Q V_k - \lambda I \\
& \;\; = \;\; W_k^H (J - R) Q V_k - \lambda W_k^H V_k \\
& \;\; = \;\; W_k^H \left\{ (J-R) Q - \lambda I \right\} V_k \;\; = \;\; W_k^H W(\lambda) V_k,
\end{split}$$ where we have employed the identities in (\[eq:identities\]).
Our reduced problems are expressed in terms of the reduced versions of $H_0(\lambda)$, $L(\lambda)$, and $H_1(\lambda)$ defined via $$H_{k,0}(\lambda) := L_k(\lambda)^{-1} L_k(\lambda)^{-H},$$ with $L_k(\lambda)$ denoting a lower triangular Cholesky factor of $$\widetilde{H}_{k,0}(\lambda) := B^H_k W_k(\lambda)^{-H} Q_k B_k B^H_k Q_k W_k(\lambda)^{-1} B_k,$$ and $H_{k,1}(\lambda) := {{\mathrm i}}(\widetilde{H}_{k,1}(\lambda) - \widetilde{H}_{k,1}(\lambda)^H)$ with $$\widetilde{H}_{k,1}(\lambda) := L_k(\lambda)^{-1} B^H_k W_k(\lambda)^{-H} Q_k B_k L_k(\lambda)^{-H}.$$ Note that to ensure the uniqueness of $L(\lambda)$ and $L_k(\lambda)$, we define them as the Cholesky factors of $\widetilde{H}_{0}(\lambda)$ and $\widetilde{H}_{k,0}(\lambda)$ with real and positive entries along the diagonal. Our goal is to come up with reduced counterparts of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, $\eta^{\rm Herm}(R; B, {\rm i}\omega)$ that Hermite-interpolate the full functions at prescribed points. For a given subspace ${\mathcal V}_k$, a matrix $V_k$ whose columns form an orthonormal basis for ${\mathcal V_k}$, and for $W_k$ as in (\[eq:defn\_Wr\]), we introduce $$\begin{split}
\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)
\; & := \;
\sup_{t \in {\mathbb R}} \: \lambda_{\min} (H_{k,0}({\rm i} \omega) + t H_{k,1} ({\rm i} \omega)) \\
\eta^{\rm Herm}_k(R; B, {\rm i}\omega)
\; & := \;
\inf\{ \|\Delta \|_2 \; | \; \Delta = \Delta^H, \;\; {\rm i} \omega \in \Lambda \big( (J_k - R_k)Q_k - (B_k\Delta B^{H}_k)Q_k \big) \}.
\end{split}$$ Recall that $\eta^{\rm Herm}(R; B, {\rm i}\omega) = \widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)^{1/2}$, and a similar relation holds for the reduced problems, i.e., $\eta^{\rm Herm}_k(R; B, {\rm i}\omega) = \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)^{1/2}$.
We start our analysis by establishing that the quantities $\eta^{\rm Herm}_k(R; B, {\rm i}\omega)$, $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ are independent of the choice of basis for the subspace ${\mathcal V}_k$. For this proof we introduce the notation $\eta^{\rm Herm}_{V_k,W_k}(R; B, {\rm i}\omega))$, $\widetilde{\eta}^{\rm Herm}_{V_k,W_k}(R; B, {\rm i}\omega)$ to emphasize the particular choices of basis $V_k, W_k$ for the subspaces ${\mathcal V_k}, {\mathcal W}_k$ used in the definitions of $\eta^{\rm Herm}_k(R; B, {\rm i}\omega), \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$. Similarly, we indicate the spaces and the bases with the help of the notations $J_{W_k}$, $R_{W_k}$, $Q_{V_k}$, and $ B_{W_k}$.
\[thm:eta\_basis\_independent\] Let the columns of $V_k$ and $\widetilde{V}_k$ form orthonormal bases for the subspace ${\mathcal V}_k$, and let $W_k := Q V_k (V_k^H Q V_k)^{-1}$, $\widetilde{W}_k := Q \widetilde{V}_k (\widetilde{V}_k^H Q \widetilde{V}_k)^{-1}$. Then $$\begin{split}
\eta^{\rm Herm}_{V_k, W_k}(R; B, {\rm i}\omega) & = \eta^{\rm Herm}_{\widetilde{V}_k, \widetilde{W}_k}(R; B, {\rm i}\omega), \\
\widetilde{\eta}^{\rm Herm}_{V_k, W_k}(R; B, {\rm i}\omega) & = \widetilde{\eta}^{\rm Herm}_{\widetilde{V}_k, \widetilde{W}_k}(R; B, {\rm i}\omega)
\end{split}$$ for all $\omega \in {\mathbb R}$.
It suffices to prove $\eta^{\rm Herm}_{V_k, W_k}(R; B, {\rm i}\omega) = \eta^{\rm Herm}_{\widetilde{V}_k, \widetilde{W}_k}(R; B, {\rm i}\omega)$, the other equality follows from this equality immediately. Since the columns of $V_k$ and $\widetilde V_k$ form bases for the same space, there exists a unitary matrix $P$ such that $\widetilde{V}_k = V_k P$. Furthermore, by definition, $$\widetilde{W}_k = Q (V_k P) (P^H V_k^H Q V_k P)^{-1} = Q V_k P P^H (V_k^H Q V_k)^{-1} P = W_k P.$$ The assertion then follows from the following set of equivalences: $$\begin{split}
{\rm i} \omega \in \Lambda \big( (J_{W_k} - R_{W_k}) Q_{V_k} - (B_{W_k} \Delta B^H_{W_k})Q_{V_k} \big)
\;\; \Longleftrightarrow \;\; \\
{\rm det}(W_k^H (J - R)Q V_k - W_k^H B \Delta B^H Q V_k - {\rm i} \omega W_k^H V_k ) = 0 \;\; \Longleftrightarrow \;\; \\{\rm det}(W_k^H \left( (J - R)Q - B \Delta B^H Q - {\rm i} \omega I \right) V_k ) = 0 \;\; \Longleftrightarrow \;\; \\
{\rm det}(P^H W_k^H \left( (J - R)Q - B \Delta B^H Q - {\rm i} \omega I \right) V_k P ) = 0 \;\; \Longleftrightarrow \;\; \\
{\rm det}(\widetilde{W}_k^H (J - R)Q \widetilde{V}_k - \widetilde{W}_k^H B \Delta B^H Q \widetilde{V}_k - {\rm i} \omega \widetilde{W}_k^H \widetilde{V}_k ) = 0 \;\; \Longleftrightarrow \;\; \\
{\rm i} \omega \in \Lambda \big( (J_{\widetilde{W}_k} - R_{\widetilde{W}_k}) Q_{\widetilde{V}_k} - (B_{\widetilde{W}_k} \Delta B^H_{\widetilde{W}_k})Q_{\widetilde{V}_k} \big). \quad\quad\quad
\end{split}$$
In the following we will develop a subspace framework including an Hermite interpolation property for DH systems. For this we first show an auxiliary interpolation result for $B^{H} Q W(\lambda)^{-1} B$, the matrix through which $H_0(\lambda), H_1(\lambda)$ are defined.
\[thm:interpolate\_matrixfuns\] Consider a DH model (\[DH\]) and a reduced model $\dot x_k= (J_k-R_k)Q_k x_k$ with coefficients as in (\[eq:projected\_matrices\]), and let $W(\lambda) = (J - R) Q - \lambda I$ and $W_k(\lambda) := (J_k - R_k) Q_k - \lambda I$. For a given $\widehat{\lambda} \in {\mathbb C}$ such that $W(\widehat{\lambda})$ and $W_k(\widehat{\lambda})$ are invertible, the following assertions hold:
1. If $\: {\rm Im}(W(\widehat{\lambda})^{-1} B) \subseteq {\mathcal V}_k$, then $B^H Q W(\widehat{\lambda})^{-1} B \; = \; B^H_k Q_k W_k(\widehat{\lambda})^{-1} B_k$.
2. Additionally, if $\: {\rm Im}(W(\widehat{\lambda})^{-2} B) \subseteq {\mathcal V}_k$ and the orthonormal basis $V_k$ for ${\mathcal V}_k$ is such that $\:
V_k
=
\left[
\begin{array}{cc}
\widetilde{V}_k & \widehat{V}_k
\end{array}
\right]
$ where the columns of $\:\: \widetilde{V}_k$ form an orthonormal basis for ${\rm Im}(W(\widehat{\lambda})^{-1} B)$, then $\: B^H Q W(\widehat{\lambda})^{-2} B \; = \;$ $B^H_k Q_k W_k(\widehat{\lambda})^{-2} B_k$.
**(i)** If ${\rm Im}(W(\widehat{\lambda})^{-1} B) \subseteq {\mathcal V}_k$ then, since $W_k V_k^H$ is a projector onto ${\rm Im}(Q V_k)$, we obtain $$\begin{split}
B^H Q W(\widehat{\lambda})^{-1} B
\; = \;
B^H Q V_k V_k^H W(\widehat{\lambda})^{-1} B
& \; = \;
B^H W_k V_k^H Q V_k V_k^H W(\widehat{\lambda})^{-1} B \\
& \; = \;
B^H_k Q_k V_k^H W(\widehat{\lambda})^{-1} B.
\end{split}$$ To show that $V_k^H W(\widehat{\lambda})^{-1} B = W_k(\widehat{\lambda})^{-1} B_k$, let $Z := W(\widehat{\lambda})^{-1} B$, and $Z_k$ be such that $V_k Z_k = Z$. (There exists a unique $Z_k$ with this property, because ${\rm Im}(Z) \subseteq {\mathcal V}_k$.) Then $W(\widehat{\lambda}) Z = B$ implies that $ W(\widehat{\lambda}) V_k Z_k = B$, and thus $ W^H_k W(\widehat{\lambda}) V_k Z_k = W^H_k B$. Hence, by Lemma \[thm:reduced\_Wlambda\] we see that $Z_k = W_k(\widehat{\lambda})^{-1} B_k$, implying that $$V_k^{H} W(\widehat{\lambda})^{-1} B \; = \;
V_k^{H} Z \; = \;
V_k^{H} (V_k Z_k) \; = \;
W_k(\widehat{\lambda})^{-1} B_k,$$ as asserted.
**(ii)** Following the steps at the beginning of the proof of part (i), we have $$B^H Q W(\widehat{\lambda})^{-2} B \;\; = \;\; B^H_k Q_k V_k^H W(\widehat{\lambda})^{-2} B.$$ To show that $V_k^H W(\widehat{\lambda})^{-2} B = W_k(\widehat{\lambda})^{-2} B_k$, we exploit that $$\label{eq:intermed_secpow_interpolate}
V_k^H W(\widehat{\lambda})^{-2} B \;\; = \;\;
(V_k^H W(\widehat{\lambda})^{-1} \widetilde{V}_k) (\widetilde{V}_k^H W(\widehat{\lambda})^{-1}B).$$ Now define $Z := W(\widehat{\lambda})^{-1} \widetilde{V}_k$ and $Z_k$ such that $V_k Z_k = Z$ (once again such a $Z_k$ exists uniquely, because ${\rm Im}(Z) \subseteq {\mathcal V}_k$) so that $W(\widehat{\lambda}) Z = \widetilde{V}_k$. Then $
W(\widehat{\lambda}) V_k Z_k = \widetilde{V}_k$ and hence $ W^H_k W(\widehat{\lambda}) V_k Z_k = I_{\widetilde{k},m}$, where $I_{\widetilde{k},m}$ is the matrix consisting of the first $m$ columns of the $\widetilde{k}\times \widetilde{k}$ identity matrix with $\widetilde{k} := {\rm dim} \: {\mathcal V}_k > m$. This implies that $Z_k = W_k(\widehat{\lambda})^{-1} I_{\widetilde{k},m}$, so the following can be deduced about the term inside the first parenthesis on the right-hand side of (\[eq:intermed\_secpow\_interpolate\]): $$V_k^H W(\widehat{\lambda})^{-1} \widetilde{V}_k
\; = \;
V_k^H Z
\; = \;
V_k^H (V_k Z_k)
\; = \;
W_k(\widehat{\lambda})^{-1} I_{\widetilde{k},m}.$$ As for the term inside the second parenthesis on the right-hand side of (\[eq:intermed\_secpow\_interpolate\]), we make use of the following observation: $$\label{eq:intermed2_secpow_interpolate}
V_k^H W(\widehat{\lambda})^{-1}B
\; = \;
\left[
\begin{array}{c}
\widetilde{V}_k^H \\
\widehat{V}_k^H
\end{array}
\right]
W(\widehat{\lambda})^{-1}B
\; = \;
\left[
\begin{array}{c}
\widetilde{V}_k^H W(\widehat{\lambda})^{-1}B \\
0
\end{array}
\right],$$ where the last equality follows, since the columns of $\widetilde{V}_k$ form an orthonormal basis for ${\rm Im}(W(\widehat{\lambda})^{-1}B)$. Putting these observations together in (\[eq:intermed\_secpow\_interpolate\]), we obtain $$\begin{split}
V_k^H W_k (\widehat{\lambda})^{-2}B_k \; & = \;
(V_k^H W(\widehat{\lambda})^{-1} \widetilde{V}_k) (\widetilde{V}_k^H W(\widehat{\lambda})^{-1}B) \\
\; & = \;
W_k(\widehat{\lambda})^{-1} I_{\widetilde{k},m} (\widetilde{V}_k^H W(\widehat{\lambda})^{-1}B) \\
\; & = \;
W_k(\widehat{\lambda})^{-1} I_{\widetilde{k}} (V_k^H W(\widehat{\lambda})^{-1}B) \\
\; & = \;
W_k(\widehat{\lambda})^{-1} I_{\widetilde{k}} ( W_k (\widehat{\lambda})^{-1}B_k )
\; = \; W_k (\widehat{\lambda})^{-2}B_k,
\end{split}$$ where in the third equality we exploit (\[eq:intermed2\_secpow\_interpolate\]), and in the fourth equality we employ that $V_k^H W(\widehat{\lambda})^{-1}B = W_k (\widehat{\lambda})^{-1}B_k$ which is proven in part (i).
After these preparations we can prove our main interpolation result.
\[thm:main\_Hermite\_interpolate\] Consider a DH model (\[DH\]) and a reduced model $\dot x_k= (J_k-R_k)Q_k x_k$ with coefficients as in (\[eq:projected\_matrices\]), and let $W(\lambda) = (J - R) Q - \lambda I$ and $W_k(\lambda) := (J_k - R_k) Q_k - \lambda I$. Suppose that the subspace ${\mathcal V}_k$ is such that $$\label{eq:subspace_inclusion}
{\rm Im}( W({\rm i} \widehat{\omega})^{-1} B ), {\rm Im}( W({\rm i} \widehat{\omega})^{-2} B ) \subseteq {\mathcal V}_k,$$ and $W_k$ is defined as in (\[eq:defn\_Wr\]) in terms of a matrix $V_k$ whose columns form an orthonormal basis for ${\mathcal V}_k$.
1. The quantity $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})$ is finite if and only if $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega})$ is finite. If $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})$ is finite, then $$\label{eq:main_interpolation_kes}
\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})
\; = \;
\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega}).$$
2. Moreover, if $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \omega)$ and $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \omega)$ are differentiable at $\widehat{\omega}$, then we have $$\label{eq:main_interpolation_kes2}
\frac{d\, \widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})}{ d \omega}
\; = \;
\frac{d\, \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega})}{d \omega}.$$
Without loss of generality, we may assume that the matrix $V_k$ is such that $
V_k
=
\left[
\begin{array}{cc}
\widetilde{V}_k & \widehat{V}_k
\end{array}
\right],
$ with the columns of $\widetilde{V}_k$ forming an orthonormal basis for ${\rm Im}( W({\rm i} \widehat{\omega})^{-1} B )$. It suffices to prove the claims for this particular choice of orthonormal basis, because it is established in Lemma \[thm:eta\_basis\_independent\] that the function $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \omega)$, hence its derivative, are independent of the choice of $V_k$ as long as its columns form an orthonormal basis for ${\mathcal V}_k$.
**(i)** By the definitions of $\widetilde{H}_0 (\lambda), \widetilde{H}_{k,0} (\lambda)$, and part (i) of Theorem \[thm:interpolate\_matrixfuns\], we have $$\begin{split}
\widetilde{H}_0 ({\rm i} \widehat{\omega})
& \; = \;
(B^{H} Q W({\rm i} \widehat{\omega})^{-1} B)^H (B^{H} Q W({\rm i} \widehat{\omega})^{-1} B) \\
& \; = \;
(B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-1} B_k)^H (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-1} B_k)
\; = \;
\widetilde{H}_{k,0} ({\rm i} \widehat{\omega}).
\end{split}$$ This also implies that $L({\rm i} \widehat{\omega}) = L_k({\rm i} \widehat{\omega})$ due to the uniqueness of the Cholesky factors of $\widetilde{H}_0 ({\rm i} \widehat{\omega})$, $\widetilde{H}_{k,0} ({\rm i} \widehat{\omega})$. Therefore, $$H_0({\rm i} \widehat{\omega}) = L({\rm i} \widehat{\omega})^{-1} L({\rm i} \widehat{\omega})^{-H}
= L_k({\rm i} \widehat{\omega})^{-1} L_k({\rm i} \widehat{\omega})^{-H}
= H_{k,0} ({\rm i} \widehat{\omega}).$$ Furthermore, $$\begin{split}
\widetilde{H}_1 ({\rm i} \widehat{\omega})
& \; = \;
L({\rm i} \widehat{\omega})^{-1} (B^{H} Q W({\rm i} \widehat{\omega})^{-1} B)^H L({\rm i} \widehat{\omega})^{-H} \\
& \; = \;
L_k({\rm i} \widehat{\omega})^{-1} (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-1} B_k)^H L_k({\rm i} \widehat{\omega})^{-H}
\; = \;
\widetilde{H}_{k,1} ({\rm i} \widehat{\omega})
\end{split}$$ and $$H_1 ({\rm i} \widehat{\omega})
=
{\rm i}
\left(
\widetilde{H}_1({\rm i} \widehat{\omega})
-
\widetilde{H}_1({\rm i} \widehat{\omega})^H
\right)
=
{\rm i}
\left(
\widetilde{H}_{k,1}({\rm i} \widehat{\omega})
-
\widetilde{H}_{k,1}({\rm i} \widehat{\omega})^H
\right)
=
H_{k,1} ({\rm i} \widehat{\omega}).$$ Since $H_1 ({\rm i} \widehat{\omega}) = H_{k,1} ({\rm i} \widehat{\omega})$, it follows that $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})$ is finite if and only if $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega})$ is finite. Additionally, if $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})$ is finite, then $$\begin{split}
\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})
& \; = \;
\max_{t \in {\mathbb R}} \: \lambda_{\min} (H_0({\rm i} \widehat{\omega}) + t H_1 ({\rm i} \widehat{\omega})) \\
& \; = \;
\max_{t \in {\mathbb R}} \: \lambda_{\min} (H_{k,0}({\rm i} \widehat{\omega}) + t H_{k,1} ({\rm i} \widehat{\omega}))
\; = \;
\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega}),
\end{split}$$ completing the proof of (\[eq:main\_interpolation\_kes\]).
**(ii)** Now let us suppose that $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \omega)$ and $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \omega)$ are differentiable at $\widehat{\omega}$. To prove the interpolation property in the derivatives, we benefit from the analytical expressions [@Lancaster1964] $$\label{eq:eta_derivative}
\begin{split}
& \frac{d\, \widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega})}{ d \omega}
\; = \;
v^H \left( \frac{d H_0({\rm i} \widehat{\omega}) }{d \omega} + \widehat{t} \: \frac{d H_1 ({\rm i} \widehat{\omega})}{d \omega} \right) v, \\
& \frac{d\, \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \widehat{\omega})}{ d \omega}
\; = \;
v^H \left( \frac{d H_{k,0}' ({\rm i} \widehat{\omega}) }{d \omega} + \widehat{t} \: \frac{d H_{k,1}' ({\rm i} \widehat{\omega})}{ d \omega} \right) v,
\end{split}$$ where $$\widehat{t} := \operatorname*{arg\,max}_t \lambda_{\min} (H_0({\rm i} \widehat{\omega}) + t H_1 ({\rm i} \widehat{\omega}))
= \operatorname*{arg\,max}_t \lambda_{\min} (H_{k,0}({\rm i} \widehat{\omega}) + t H_{k,1} ({\rm i} \widehat{\omega}))$$ and $v$ is a unit eigenvector corresponding to $\lambda_{\min} (H_0({\rm i} \widehat{\omega}) + \widehat{t} H_1 ({\rm i} \widehat{\omega}))$ $=$ $\lambda_{\min} ($ $H_{k,0}({\rm i} \widehat{\omega}) + \widehat{t} H_{k,1} ({\rm i} \widehat{\omega}))$. Thus, it suffices to prove that $d H_0({\rm i} \widehat{\omega}) / d \omega = d H_{k,0}({\rm i} \widehat{\omega}) / d\omega$ and $d H_1({\rm i} \widehat{\omega}) / d\omega = d H_{k,1}'({\rm i} \widehat{\omega}) / d\omega$ in order to show (\[eq:main\_interpolation\_kes2\]). It follows from parts (i) and (ii) of Theorem \[thm:interpolate\_matrixfuns\] that $$\begin{split}
d \widetilde{H}_0'({\rm i} \widehat{\omega}) / d\omega
& =
-{\rm i} (B^{H} Q W({\rm i} \widehat{\omega})^{-1} B)^H (B^{H} Q W({\rm i} \widehat{\omega})^{-2} B) \hskip 10ex \\
& \hskip 3ex + \; {\rm i} (B^{H} Q W({\rm i} \widehat{\omega})^{-2} B)^H (B^{H} Q W({\rm i} \widehat{\omega})^{-1} B) \\
& = -{\rm i} (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-1} B_k)^H (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-2} B_k) \\
&\hskip 3ex + \; {\rm i} (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-2} B_k)^H (B^{H}_k Q_k W_k({\rm i} \widehat{\omega})^{-1} B_k)
=
d \widetilde{H}_{k,0}'({\rm i} \widehat{\omega}) / d \omega.
\end{split}$$ Now let us determine the derivatives of the Cholesky factors, which at a given $\omega$ satisfy $$\widetilde{H}_0({\rm i} \omega) = L(\omega) L({\rm i} \omega)^H
\quad
\text{and}
\quad
\widetilde{H}_{k,0}({\rm i} \omega) = L_k({\rm i} \omega) L_k({\rm i} \omega)^H .$$ Differentiating these two equations, and setting the derivatives equal to each other at $\widehat{\omega}$ yield $$\begin{split}
L({\rm i} \widehat{\omega}) \left( \frac{d L({\rm i} \widehat{\omega})}{d \omega} \right)^H +
\frac{d L({\rm i} \widehat{\omega})}{d \omega} L({\rm i} \widehat{\omega})^H
& \; = \;
L_k({\rm i} \widehat{\omega}) \left( \frac{d L_k({\rm i} \widehat{\omega})}{d \omega} \right)^H +
\frac{ L_k({\rm i} \widehat{\omega})}{d \omega} L_k({\rm i} \widehat{\omega})^H \\
& \; = \;
L({\rm i} \widehat{\omega}) \left( \frac{d L_k({\rm i} \widehat{\omega})}{d \omega} \right)^H +
\frac{d L_k({\rm i} \widehat{\omega})}{d \omega} L({\rm i} \widehat{\omega})^H,
\end{split}$$ where we have used that $L({\rm i} \widehat{\omega}) = L_k({\rm i} \widehat{\omega})$ as established in part (i). Thus, both $d L({\rm i} \widehat{\omega}) / d\omega$ and $d L_k({\rm i} \widehat{\omega}) / d\omega$ are lower triangular solutions of the matrix equation $$d \widetilde{H}_0({\rm i} \widehat{\omega}) / d\omega = L({\rm i} \widehat{\omega}) X^H + X L({\rm i} \widehat{\omega})^H.$$ This linear matrix equation has a unique lower triangular solution, so $dL({\rm i} \widehat{\omega}) / d\omega = dL_k({\rm i} \widehat{\omega}) / d\omega$. Now, by the definitions of $H_0 ({\rm i} \omega), H_{k,0} ({\rm i} \omega)$, we have $$L({\rm i} \omega) H_0 ({\rm i} \omega) L({\rm i} \omega)^H \;\; = \;\; I
\;\; = \;\; L_k({\rm i} \omega) H_{k,0} ({\rm i} \omega) L_k({\rm i}\omega)^H.$$ Differentiating this equation at $\omega = \widehat{\omega}$ yields $$\begin{split}
\frac{d L({\rm i} \widehat{\omega})}{d \omega} H_0({\rm i} \widehat{\omega}) L({\rm i} \widehat{\omega})^H
\; + \;
L({\rm i} \widehat{\omega}) \frac{d H_0({\rm i} \widehat{\omega})}{d \omega} L({\rm i} \widehat{\omega})^H \hskip 30ex \\
\; + \;\;
L({\rm i} \widehat{\omega}) H_0({\rm i} \widehat{\omega}) \left( \frac{d L({\rm i} \widehat{\omega})}{d \omega} \right)^H
\;\; = \;\;
\frac{d L_k({\rm i} \widehat{\omega})}{d \omega} H_{k,0}({\rm i} \widehat{\omega}) L_k({\rm i} \widehat{\omega})^H
\;\; + \; \hskip 5ex \\
\hskip 8ex L_k({\rm i} \widehat{\omega}) \frac{d H_{k,0}({\rm i} \widehat{\omega})}{d \omega} L_k({\rm i} \widehat{\omega})^H
\; + \;
L_k({\rm i} \widehat{\omega}) H_{k,0}({\rm i} \widehat{\omega}) \left( \frac{d L_k({\rm i} \widehat{\omega})}{d \omega} \right)^H.
\end{split}$$ Using that $L({\rm i} \widehat{\omega}) = L_k({\rm i} \widehat{\omega})$, $dL({\rm i} \widehat{\omega}) / d\omega = dL_k({\rm i} \widehat{\omega}) / d\omega$, and $H_0({\rm i} \widehat{\omega}) = H_{k,0}({\rm i} \widehat{\omega})$, we deduce that $dH_0({\rm i} \widehat{\omega}) / d\omega = dH_{k,0}({\rm i} \widehat{\omega}) / d\omega$. Next we focus on the derivatives $d \widetilde{H}_1({\rm i} \omega) / d\omega, d \widetilde{H}_{k,1}({\rm i} \omega) / d\omega$. In particular, we use that $$L({\rm i} \omega) \widetilde{H}_1({\rm i} \omega) L({\rm i} \omega)^H
\;\; = \;\;
(B^H Q W({\rm i}\omega)^{-1} B)^{H}.$$ Differentiating both sides of the last equation at $\omega = \widehat{\omega}$ gives rise to $$\begin{split}
\frac{d \left\{ L({\rm i} \omega) \widetilde{H}_1({\rm i}\omega)
L({\rm i} \omega)^H \right\}}{d\omega} \bigg|_{\omega = \widehat{\omega}}
\;\; & = \;\;
{\rm i} (B^H Q W({\rm i} \widehat{\omega})^{-2} B)^{H} \\
\;\; & = \;\;
{\rm i} (B^H_k Q_k W_k({\rm i} \widehat{\omega})^{-2} B_k)^{H} \\
\;\; & = \;\;
\frac{d \left\{ L_k({\rm i} \omega) \widetilde{H}_{k,1}({\rm i} \omega) L_k({\rm i} \omega)^H \right\}}{d\omega} \bigg|_{\omega = \widehat{\omega}} \;\;\; ,
\end{split}$$ which in turn implies that $$\begin{split}
\frac{d L({\rm i} \widehat{\omega})}{d \omega} \widetilde{H}_1({\rm i} \widehat{\omega}) L({\rm i} \widehat{\omega})^H
\; + \;
L({\rm i} \widehat{\omega}) \frac{d \widetilde{H}_1({\rm i} \widehat{\omega})}{d \omega} L({\rm i} \widehat{\omega})^H
\hskip 28ex \\
\; + \;\;
L({\rm i} \widehat{\omega}) \widetilde{H}_1({\rm i} \widehat{\omega}) \left( \frac{d L({\rm i} \widehat{\omega})}{d \omega} \right)^H
\;\; = \;\;
\frac{d L_k({\rm i} \widehat{\omega})}{d \omega} \widetilde{H}_{k,1}({\rm i} \widehat{\omega}) L_k({\rm i} \widehat{\omega})^H
\;\; + \; \hskip 5ex \\
L_k({\rm i} \widehat{\omega}) \frac{d \widetilde{H}_{k,1}({\rm i} \widehat{\omega})}{d \omega} L_k({\rm i} \widehat{\omega})^H
\; + \;
L_k({\rm i} \widehat{\omega}) \widetilde{H}_{k,1}({\rm i} \widehat{\omega}) \left( \frac{d L_k({\rm i} \widehat{\omega})}{d \omega} \right)^H .
\end{split}$$ Once again exploiting that $L({\rm i} \widehat{\omega}) = L_k({\rm i} \widehat{\omega})$, $dL({\rm i} \widehat{\omega}) / d\omega = dL_k({\rm i} \widehat{\omega}) / d\omega$, as well as $\widetilde{H}_1({\rm i} \widehat{\omega}) = \widetilde{H}_{k,1}({\rm i} \widehat{\omega})$ in the last equation, we obtain $d \widetilde{H}_1({\rm i} \widehat{\omega}) / d\omega = d \widetilde{H}_{k,1}({\rm i} \widehat{\omega}) / d\omega$ which implies that $$\begin{split}
\frac{d H_1 ({\rm i} \widehat{\omega})}{d\omega}
& \;\; = \;\;
{\rm i}
\left\{
\frac{d \widetilde{H}_1({\rm i} \widehat{\omega})}{d\omega}
-
\left( \frac{d \widetilde{H}_1({\rm i} \widehat{\omega})}{d \omega} \right)^H
\right\} \\
& \;\; = \;\;
{\rm i}
\left\{
\frac{d \widetilde{H}_{k,1}({\rm i} \widehat{\omega})}{d \omega}
-
\left( \frac{d \widetilde{H}_{k,1}({\rm i} \widehat{\omega})}{d \omega} \right)^H
\right\}
\;\; = \;\;
\frac{d H_{k,1} ({\rm i} \widehat{\omega})}{d \omega},
\end{split}$$ and the proof of (\[eq:main\_interpolation\_kes2\]) is complete.
\[rem:rem2\]
The function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \omega)$ is differentiable at $\widehat{\omega}$ whenever $\widetilde{\eta}^{\rm Herm}(R;$ $B, {\rm i} \widehat{\omega})$ is finite (equivalently $H_1({\rm i} \widehat{\omega})$ is indefinite), $\lambda_{\min}(H_0({\rm i} \widehat{\omega}) + \widehat{t} H_1({\rm i} \widehat{\omega}))$ is simple where $\widehat{t} := \operatorname*{arg\,max}_{t \in {\mathbb R}} \lambda_{\min}(H_0({\rm i} \widehat{\omega}) + t H_1({\rm i} \widehat{\omega}))$, and the global minimum of $\lambda_{\min}(H_0({\rm i} \widehat{\omega}) + t H_1({\rm i} \widehat{\omega}))$ over $t$ is attained at a unique $t$. These conditions guarantee also the differentiability of $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \omega)$ at $\widehat{\omega}$ provided the subspace inclusion in (\[eq:subspace\_inclusion\]) holds. This latter differentiability property is due to $H_0({\rm i} \widehat{\omega}) = H_{k,0}({\rm i} \widehat{\omega})$ and $H_1({\rm i} \widehat{\omega}) = H_{k,1}({\rm i} \widehat{\omega})$ from part (i) of Theorem \[thm:main\_Hermite\_interpolate\].
Additionally, when the concave function $g(t) := \lambda_{\min}(H_0({\rm i} \widehat{\omega}) + t H_1({\rm i} \widehat{\omega}))$ attains its maximum, the maximizer is nearly always unique. The function $g(t)$ is the minimum of $m$ real analytic functions [@Kato1995; @Rellich1969] each corresponding to an eigenvalue of $H_0({\rm i} \widehat{\omega}) + t H_1({\rm i} \widehat{\omega})$. If $g(t)$ does not have a unique maximizer, then at least one of these real analytic functions must be constant and equal to $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \widehat{\omega}) = \max_t g(t)$ everywhere. Thus, a simple sufficient condition that ensures the uniqueness of the maximizer is that $H_1({\rm i} \widehat{\omega})$ has full rank, in which case all eigenvalues of $H_0({\rm i} \widehat{\omega}) + t H_1({\rm i} \widehat{\omega})$ blow up (either to $\infty$ or $-\infty$) as $t \rightarrow \infty$ implying that each of the real analytic functions is non-constant.
### The Subspace Framework for $r^{\rm Herm}(R; B)$
The Hermite interpolation result of Theorem \[thm:main\_Hermite\_interpolate\] immediately suggests the subspace framework in Algorithm \[alg:structured\_kadii\_sf\] for the computation of $r^{\rm Herm}(R; B)$. This resembles the structure preserving subspace framework to compute the unstructured stability radii $r(R; B, C)$ $\; = \;$ $r(J; B, C)$, in particular, in the way the subspaces ${\mathcal V}_k, {\mathcal W}_k$ are built. At every iteration, a reduced problem is solved using the ideas in Section \[sec:small\_scale\_st\] and employing Algorithm \[eigopt\]. Letting $\widehat{\omega}$ be the global minimizer of the reduced problem, the subspaces are expanded so that the original function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \omega)$ is Hermite interpolated by its reduced counter-part at $\omega = \widehat{\omega}$.
Assuming that the sequence $\{ \omega_k \}$ converges to a minimizer $\omega_\ast$ of the function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ such that $H_1({\rm i} \omega_\ast)$ is indefinite with full rank and $\lambda_{\min} (H_0({\rm i} \omega_\ast) + t_\ast H_1({\rm i} \omega_\ast))$ is simple where $t_\ast := \operatorname*{arg\,max}_{t \in {\mathbb R}} \lambda_{\min} (H_0({\rm i} \omega_\ast) + t H_1({\rm i} \omega_\ast))$, it can be shown that the sequence $\{ \omega_k \}$ converges at a super-linear rate. Here the analysis in [@Aliyev2018] applies. The conditions that $H_1({\rm i} \omega_\ast)$ is indefinite with full rank, $\lambda_{\min} (H_0({\rm i} \omega_\ast) + t_\ast H_1({\rm i} \omega_\ast))$ is simple, and the interpolation properties $H_0({\rm i} \omega_k) = H_{k,0}({\rm i} \omega_k)$, $H_1({\rm i} \omega_k) = H_{k,1}({\rm i} \omega_k)$ ensure that the full function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i} \omega)$, as well as the reduced function $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i} \omega)$ for all large $k$ are differentiable at all $\omega$ in a ball ${\mathcal B}(\omega_\ast,\delta)$ centered at $\omega_\ast$ and of radius $\delta$. This differentiability property is essential for the applicability of the rate-of-convergence analysis in [@Aliyev2018].
Choose the initial interpolation points $\omega_{1}, \dots , \omega_{j} \in {\mathbb R}.$ $V_j \gets {\rm orth} \begin{bmatrix} W({{\mathrm i}}\omega_1)^{-1}B & W({{\mathrm i}}\omega_1)^{-2}B
& \dots & W({{\mathrm i}}\omega_j)^{-1}B & W({{\mathrm i}}\omega_j)^{-2}B \end{bmatrix}$,\
$\;\;\;\; W_j \gets QV_j(V_j^HQV_j)^{-1} $. \[defn\_init\_subspaces\_structured\] $\; \displaystyle \omega_{k+1} \gets \: {\rm argmin}_{\omega \in {\mathbb R}} \; \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$. \[solve\_keduced\_structured\] $\widehat{V}_{k+1} \gets \begin{bmatrix} W({{\mathrm i}}\omega_{k+1})^{-1}B & W({{\mathrm i}}\omega_{k+1})^{-2}B \end{bmatrix}$. \[defn\_later\_subspaces\_structured\] $V_{k+1} \gets \operatorname{orth}\left(\begin{bmatrix} V_{k} & \widehat{V}_{k+1} \end{bmatrix}\right)
\quad \text{and}\quad W_{k+1} \gets Q V_{k+1}(V_{k+1}^H Q V_{k+1})^{-1}$.
Numerical Experiments {#sec:str_num_exp}
---------------------
In this section we present several numerical tests for Algorithms \[eigopt\] and \[alg:structured\_kadii\_sf\], on synthetic examples and the FE model of the disk brake. The test set-up is similar to the one for the unstructured case in Section \[sec:DH\_numexps\]. In particular, for Algorithm \[alg:structured\_kadii\_sf\], we use the same stopping criteria with the same parameters, but now in terms of $f_k := \operatorname*{arg\,min}_{\omega} \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$, so we terminate when $| f_k - f_{k-1} | \leq \varepsilon | f_k + f_{k-1} |/2$ holds, or when one of the other two conditions stated in Section \[sec:test\_setup\] holds. The initial subspaces are also chosen as described in that section.
### Synthetic Examples
**Small Scale Examples.** We first present numerical results for a small dense random example, where $J, R, Q \in {\mathbb R}^{20\times 20}$ and $B \in {\mathbb R}^{20\times 2}$. These matrices are generated by means of the MATLAB commands employed for the generation of the dense family in Section \[sec:numexp\_syn\]; only here the matrices are $20\times 20$ instead of $800 \times 800$ and $R$ is of rank $5$.
The spectrum of $(J-R)Q$ is depicted on the top in Figure \[fig:structured20by20\_spec\]. Application of Algorithm \[eigopt\] to this example yields $r^{\rm Herm}(R;B) = 0.0501$, and the point that is first reached on the imaginary axis under the smallest perturbation is $1.9794{\rm i}$, i.e., $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ is minimized at $\omega = 1.9794$. At the bottom of Figure \[fig:structured20by20\_spec\], the spectra of matrices of the form $(J - (R + B \Delta B^H))Q$ are plotted for $100000$ randomly chosen Hermitian $\Delta$ such that $\| \Delta \|_2 = 0.0501$. One can notice that some of the eigenvalues (nearly) touch the imaginary axis at $1.9794{\rm i}$ marked with a circle.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![A DH system of order $20$ with random system matrices. **(Top)** A plot of the spectrum for $(J-R)Q$. **(Bottom)** All eigenvalues of all matrices of the form $(J - (R + B \Delta B^H ))Q$ for $100000$ randomly chosen Hermitian $\Delta$ with $\| \Delta \|_2 = r^{\rm Herm}(R; B)$ are displayed. The circle marks $1.9794{\rm i}$, the global minimizer of $\widetilde{\eta}^{\rm Herm}_k(R; B, \lambda)$ over $\lambda \in {\rm i} {\mathbb R}$.[]{data-label="fig:structured20by20_spec"}](structured_spec_20by20.pdf "fig:"){width="9cm"}
![A DH system of order $20$ with random system matrices. **(Top)** A plot of the spectrum for $(J-R)Q$. **(Bottom)** All eigenvalues of all matrices of the form $(J - (R + B \Delta B^H ))Q$ for $100000$ randomly chosen Hermitian $\Delta$ with $\| \Delta \|_2 = r^{\rm Herm}(R; B)$ are displayed. The circle marks $1.9794{\rm i}$, the global minimizer of $\widetilde{\eta}^{\rm Herm}_k(R; B, \lambda)$ over $\lambda \in {\rm i} {\mathbb R}$.[]{data-label="fig:structured20by20_spec"}](structured_spec_20by20_pert.pdf "fig:"){width="9cm"}
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The subspace framework for this example is illustrated in Figure \[fig:structured20by20\_sf\], where the solid, dashed curves correspond to the plots of the full function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, the reduced function $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$, respectively, and the circle represents the minimizer of $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$. On the top row, the framework is initiated with two interpolation points at 0 and near -20; the dashed curve interpolates the solid curve at these points. Then the subspaces are expanded so that the Hermite interpolation property is also satisfied at the minimizer of the dashed curve on the top, leading to the dashed curve at the bottom, which has nearly the same global minimizer as the solid curve. Note that starting from $\omega$ near -14 and for smaller $\omega$ values, the matrix $H_1({\rm i} \omega)$ turns out to be definite for this example, meaning for such $\omega$ values the point ${\rm i} \omega$ is not attainable as an eigenvalue with Hermitian perturbations. In practice, we set the objective to be minimized at such $\omega$ a value considerably larger than the minimal value of the objective, which in this example is 0.1.
**Large Examples.** The remaining synthetic examples are with larger random matrices. We created three sets of matrices $J, R, Q, B$ using the commands at the beginning of Section \[sec:numexp\_syn\] for the generation of the dense family. Each of the three sets consists of four quadruples $J, R, Q \in {\mathbb R}^{n\times n}$, $B \in {\mathbb R}^{n\times 2}$ with the same $n$, specifically $n = 1000, 2000, 4000$ for the first, second, third set. The results obtained by applications of Algorithm \[alg:structured\_kadii\_sf\] to compute $r^{\rm Herm}(R; B)$ are reported in Tables \[table:structured\_1000exs\], \[table:structured\_2000exs\], \[table:structured\_4000exs\] for these sets, respectively. For the first family with $n = 1000$, the subspace framework, i.e., Algorithm \[alg:structured\_kadii\_sf\] already needs less computing time than Algorithm \[eigopt\] except for the last example where quite a few additional subspace iterations have been performed. On this family, the direct application of Algorithm \[eigopt\] and the subspace framework return exactly the same values for $r^{\rm Herm}(R; B)$.
For bigger systems Algorithm \[eigopt\] becomes too computationally expensive, so we do not report results here for the larger dimensions. A remarkable fact we have observed is that the number of subspace iterations to reach the prescribed accuracy is usually small and seems independent of $n$. By the definitions of the structured and unstructured radii, we must have $r^{\rm Herm}(R; B) \geq r(R; B, B^H)$, and the presented radii in the tables are in harmony with this.
All of these examples involve optimization of highly non-convex and non-smooth functions. Figure \[fig:structured1000by10000\_sf\] depicts $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ as a function of $\omega$ with the solid curve for the first example in the first family with $n = 1000$. The same figure also depicts the reduced function $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ with the dashed curve for the same example at termination after 7 subspace iterations. Even though the reduced function $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ represented by the dashed curve in this plot is defined for projected matrices onto $48$ dimensional subspaces, it captures the original function $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ remarkably well near the global minimizer $\omega_\ast = -70.9623$.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ --
![ Progress of Algorithm \[alg:structured\_kadii\_sf\] on a random DH system of order $20$. The solid, dashed curves display $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ as functions of $\omega$, respectively. The square, circle represent the global minima of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$. **(Top Plot)** The reduced function interpolates the full function at two points, at $\omega = 0$ and near $\omega=-20$. **(Bottom Plot)** Plot of the refined reduced function after applying one subspace iteration. []{data-label="fig:structured20by20_sf"}](structured_spec_20by20_init.pdf "fig:"){width="8cm"}
![ Progress of Algorithm \[alg:structured\_kadii\_sf\] on a random DH system of order $20$. The solid, dashed curves display $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ as functions of $\omega$, respectively. The square, circle represent the global minima of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$. **(Top Plot)** The reduced function interpolates the full function at two points, at $\omega = 0$ and near $\omega=-20$. **(Bottom Plot)** Plot of the refined reduced function after applying one subspace iteration. []{data-label="fig:structured20by20_sf"}](structured_spec_20by20_iter1.pdf "fig:"){width="8cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ --
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Application of Algorithm \[alg:structured\_kadii\_sf\] to compute $r^{\rm Herm}(R;B)$ on a dense random DH system of order $1000$. The full and reduced functions $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$ and $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ at termination after $7$ subspace iterations are plotted with the solid and dashed curves, respectively. The point $(\omega_\ast, \widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega_\ast))$ with $\omega_\ast$ denoting the global minimizer of $\widetilde{\eta}^{\rm Herm}_k(R; B, {\rm i}\omega)$ is marked with a circle. []{data-label="fig:structured1000by10000_sf"}](structured_1000_ex1_red.pdf "fig:"){width="8cm"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------------- ----------------- ------------------------------------ ------------------ ------------------------------------ ----------------- ------------------------------------
\[0.5ex\] $\#$ Alg. \[eigopt\] Alg. \[alg:structured\_kadii\_sf\] Alg. \[alg:dti\] Alg. \[alg:structured\_kadii\_sf\] Alg. \[eigopt\] Alg. \[alg:structured\_kadii\_sf\]
\[0.5ex\] 1 0.0129 0.0129 0.0113 7 99.2 66.2
2 0.0101 0.0101 0.0101 1 66.5 7.1
3 0.0141 0.0141 0.0128 5 128.5 72.5
4 0.0096 0.0096 0.0072 20 121.2 247.2
---------------- ----------------- ------------------------------------ ------------------ ------------------------------------ ----------------- ------------------------------------
: Performance of Algorithm \[alg:structured\_kadii\_sf\] to compute $r^{\rm Herm}(R;B)$ on dense random DH systems of order $1000$. For comparison, the results from Algorithm \[eigopt\], as well as the unstructured radii $r(R;B, B^H)$ by Algorithm \[alg:dti\] are also included. The fourth column contains the number of subspace iterations, and the fifth the run-times (in $s$). []{data-label="table:structured_1000exs"}
---------------- ------------------------------------ ------------------ ------------------------------------ ------------------------------------
\[0.5ex\] $\#$ Alg. \[alg:structured\_kadii\_sf\] Alg. \[alg:dti\] Alg. \[alg:structured\_kadii\_sf\] Alg. \[alg:structured\_kadii\_sf\]
\[0.5ex\] 1 0.0102 0.0061 6 149.1
2 0.0040 0.0012 1 17.1
3 0.0125 0.0109 6 175.9
4 0.0101 0.0100 3 126.7
---------------- ------------------------------------ ------------------ ------------------------------------ ------------------------------------
: Performance of Algorithm \[alg:structured\_kadii\_sf\] to compute $r^{\rm Herm}(R;B)$ on dense random DH systems of order $n = 2000$. The fourth column contains the number of subspace iterations, and the fifth the run-times (in $s$). []{data-label="table:structured_2000exs"}
---------------- ------------------------------------ ------------------ ------------------------------------ ------------------------------------
\[0.5ex\] $\#$ Alg. \[alg:structured\_kadii\_sf\] Alg. \[alg:dti\] Alg. \[alg:structured\_kadii\_sf\] Alg. \[alg:structured\_kadii\_sf\]
\[0.5ex\] 1 0.0090 0.0087 2 282.6
2 0.0084 0.0068 3 319.3
3 0.0104 0.0086 2 333.1
4 0.0040 0.0007 46 786.5
---------------- ------------------------------------ ------------------ ------------------------------------ ------------------------------------
: Performance of Algorithm \[alg:structured\_kadii\_sf\] to compute $r^{\rm Herm}(R;B)$ on dense random DH systems of order $n = 4000$. The fourth column contains the number of subspace iterations, and the fifth the run-times (in $s$). []{data-label="table:structured_4000exs"}
### FE Model of a Disk Brake
We also applied our implementation of Algorithm \[alg:structured\_kadii\_sf\] to the FE model of the disk brake described in Section \[sec:numexp\_brake\] which is of the form (\[insta\]), (\[eq:insta\_matrices\]) with $G(\Omega), D(\Omega), K(\Omega), M \in {\mathbb R}^{4669\times 4669}$, and $J, R, Q \in {\mathbb R}^{9338\times 9338}$.
The computed value of the structured radius $r^{\rm Herm}(R; B)$ along with the computed global minimizer $\omega_\ast$ of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, as well as the number of subspace iterations, run-time (in $s$) and the subspace dimension at termination are listed in Table \[tab:brake\_squeal\_structured\] for various values of $\Omega$. It is worth comparing the computed values of $r^{\rm Herm}(R; B)$ in this table with those for the unstructured stability radius $r(R; B, B^T)$ listed in Table \[tab:brake\_squeal\_unstructured\]. The computed structured and the unstructured stability radii are close, though the structured stability radii are slightly larger as expected in theory.
-0.2ex
$\Omega$ $r^{\rm Herm}(R; B)$ $\omega_\ast$ iterations run-time dimension
---------- ---------------------- --------------------- ------------ ---------- -----------
2.5 0.01067 $-1.938\times 10^5$ 2 145.8 72
5 0.01038 $-1.938\times 10^5$ 2 133.5 72
10 0.01026 $-1.938\times 10^5$ 2 132.0 72
50 0.01000 $-1.938\times 10^5$ 2 132.5 72
100 0.00988 $-1.938\times 10^5$ 1 94.4 66
1000 0.00810 $-1.789\times 10^5$ 2 127.7 72
1050 0.00794 $-1.789\times 10^5$ 2 126.4 72
1100 0.00835 $-1.789\times 10^5$ 3 171.2 78
1116 0.01092 $-1.789\times 10^5$ 2 127.4 72
1150 0.00346 $-1.742\times 10^5$ 2 124.3 72
1200 0.00408 $-1.742\times 10^5$ 2 121.2 72
1250 0.00472 $-1.742\times 10^5$ 2 119.1 72
1300 0.00517 $-1.742\times 10^5$ 2 117.1 72
: Structured stability radii $r^{\rm Herm}(R; B)$ computed by Algorithm \[alg:structured\_kadii\_sf\] for the FE model of a disk brake of order $9338$ for several values of $\Omega$. The column $\omega_\ast$ depicts the computed global minimizer of $\widetilde{\eta}^{\rm Herm}(R; B, {\rm i}\omega)$, whereas the last three columns depict the number of subspace iterations, the total run-time (in $s$), and the subspace dimension at termination. []{data-label="tab:brake_squeal_structured"}
Concluding Remarks
==================
We have proposed subspace frameworks to compute the stability radii for large scale dissipative Hamiltonian systems. The frameworks operate on the eigenvalue optimization characterizations of the stability radii derived in [@MehMS16a]. At every iteration, we apply DH structure preserving Petrov-Galerkin projections to small subspaces. This leads to the computation of the corresponding stability radii for the reduced system. We expand the subspaces used in the Petrov-Galerkin projections so that Hermite-interpolation properties between the objective eigenvalue function of the full and the reduced problems are attained at the optimizer of the reduced problem. This strategy results in super-linear convergence with respect to the subspace dimensions. We have illustrated that the frameworks work well in practice on several synthetic examples, and a FE model of a disk brake.
Matlab implementations of the proposed algorithms and subspace frameworks are made publicly available on the web[^1]. Some of the data (including the one associated with the disk brake example) used in the numerical experiments are also available on the same website.
One difficulty is that the proposed frameworks converge only locally. As a remedy for this, we have initiated the subspaces to attain Hermite interpolation at several points on the imaginary axis between the full and initial reduced problems. One potential strategy that is currently investigated is to employ equally spaced interpolation points. Another potential strategy finds the poles closest to these equally spaced points, then employs the imaginary parts of the poles as the initial interpolation points. Another research direction that is currently investigated is the maximization of the stability radii, when $J, R, Q$ depend on parameters in a given parameter set. As an example, for the dissipative Hamiltonian system arising from the FE model of a disk brake, even in the simple setting considered here, $J, R, Q$ depend on the rotation speed $\Omega$.
[^1]: <http://home.ku.edu.tr/~emengi/software/DH-stabradii.html>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Energy efficient real-time task scheduling attracted a lot of attention in the past decade. Most of the time, deterministic execution lengths for tasks were considered, but this model fits less and less with the reality, especially with the increasing number of multimedia applications. It’s why a lot of research is starting to consider stochastic models, where execution times are only known stochastically. However, authors consider that they have a pretty much precise knowledge about the properties of the system, especially regarding to the worst case execution time (or worst case execution cycles, WCEC).
In this work, we try to relax this hypothesis, and assume that the WCEC can vary. We propose miscellaneous methods to react to such a situation, and give many simulation results attesting that with a small effort, we can provide very good results, allowing to keep a low deadline miss rate as well as an energy consumption similar to clairvoyant algorithms.
author:
- |
Vandy [<span style="font-variant:small-caps;">Berten</span>]{}, Chi-Ju [<span style="font-variant:small-caps;">Chang</span>]{}, Tei-Wei [<span style="font-variant:small-caps;">Kuo</span>]{}\
National Taiwan University\
Computer Science and Information Engineering dept.\
{vberten, ktw}@csie.ntu.edu.tw, [email protected]
bibliography:
- 'VaryingWCEC.bib'
nocite: '[@*]'
title: Managing Varying Worst Case Execution Times on DVS Platforms
---
Introduction
============
Motivations
-----------
In the past decade, energy efficient systems have been actively explored. With the tremendous increase of the number of mobile devices, research in this field is still only at its beginning. Moreover, as many of those devices run now multimedia applications, real-time stochastic low-power systems will have a major role to play in the next few years. One of the characteristic of stochastic systems, is that their parameters are likely to change in the time. For instance, a device decoding a movie will need more time to process a sequence in a movie with a lot of color and movement that a dark and quite sequence. What authors usually propose in the literature is to observe continuously the tasks execution time, in order to update their knowledge about the distribution. And once this distribution seems to be too far away from the one which was used by the scheduler, the scheduler is updated. But if the worst case execution time seems to increase, the system cannot afford to wait for collecting enough information in order to update the scheduler: some actions need to be taken as soon as possible, in order to avoid to miss deadline. This is what we want to do in this work: propose an efficient strategy to react to WCEC variation.
Related work
------------
Energy-efficient real-time task scheduling attracted a lot of attention over the past decade. Low-power real-time systems with stochastic or unknown duration have been studied for several years. The problem has first been considered in systems with only one task, or systems in which each task gets a fixed amount of time. Gruian [@Gruian01; @Gruian01b] or Lorch and Smith [@pace01; @pace04] both shown that when intra-task frequency change is available, the more efficient way to save energy is to increase progressively the speed.
Solutions using a discrete set of frequencies and taking speed change overhead into account have also been proposed [@Xu04; @Xu07b]. For inter-task frequency changes, some work has been already undertaken. In [@Mosse00], authors consider a similar model to the one we consider here, even if this model is presented differently. The authors present several dynamic power management techniques, with different aggressiveness level.In [@Aydin01], authors attempt to allow the manager to tune this aggressiveness level, while in [@Xu07b], they propose to adapt automatically this aggressiveness using the distribution of the number of cycles for each task. The same authors have also proposed a strategy taking the number of available speeds into account from the beginning, instead of patching algorithms developed for continuous speed processors [@Xu07]. In [@RTCSA08], we generalize and uniformize the model presented in several of the previous papers, and propose two contributions: first, we gave a general sufficient and necessary condition of schedulability for this model, and second, we presented a new approach to adapt a continuous-speed-based method to a discrete-speed system. Some multiprocessor extensions have been considered in [@Chen07].
Model
-----
The model considered in this document is the same as the one presented in [@RTCSA08] by the same authors. We resume it here for the sake of completeness, but the interested reader is invited to have a look at the given reference.
- We have $N$ *tasks* $\{T_i, i\in [1,\dots,N]\}$ which run on a DVS CPU. They all share the same deadline and period $D$ (also known as the *frame*), do not have offset (periods are then synchronised) and are executed in the order $T_1$, $T_2$, …, $T_N$. The maximum execution number of cycles of $T_i$ is $w_i$;
- The *CPU* can run at $M$ frequencies $f_1<\dots<f_M$;
- We have $N$ *scheduling functions* $S_i(t)$ for $i \in [1,\dots, N]$ and $t \in [0, D]$. This function means that if $T_i$ starts its execution at time $t$, it will run until its end — unless the task is suspended before — at frequency $S_i(t)$, where $S_i(t)\in \{f_1, f_2, ..., f_M\}$. $S_i(t)$ is then a step function (piece-wise constant function), with only $M$ possible values.
This model generalizes several scheduling strategies proposed in the literature, such as [@Xu07; @Xu07b]. Figure \[Sit\] shows an example of such scheduling function set.
A scheduling function can be modelized by a set of points (black dots on Figure \[Sit\]), representing the beginning of the step. $\mid S_i \mid$ is the number of steps of $S_i$. $S_i[k], k\in\{1, \dots, \mid S_i \mid\}$ is one point, with $S_i[k].t$ being its time component, and $S_i[k].f$ the frequency. $S_i$ has then the same value $S_i[k].f$ in the interval $\Big[S_i[k].t, S_i[k+1].t\Big[$ (with $S_i[\mid{}S_i\mid+1].t = \infty$), and we have
$$S_i(t) = S_i[k].f$$ where $k = \max \Big\{j\in \{1, \dots, \mid S_i \mid \} \Big| S_i[j].t \le t\Big\}$
Notice that finding $k$ can be done in $O(\log \mid S_i \mid)$ (by binary search), and, except in the case of very particular models, $\mid S_i \mid \le M$.
Energy and time overhead for frequency changes can easily be taken into account in the model.
### Schedulability Condition
The scheduling functions $S_i(t)$ can be pretty general, but have to respect some constraints in order to ensure the system schedulability and avoid deadline misses. If tasks never need more cycles than their worst case execution cycles (WCEC), or, in other word, if the knowledge about WCEC is correct, we show in [@RTCSA08] that the following (necessary and sufficient) condition ensures that every job in a frame can finish before the end of this frame (if the system is expedient and $T_1$ starts at the beginning of the frame):
$$\label{eq:sched}
S_i(t) \ge \dfrac{w_i}{z_{i+1}-t} ~\forall i \in [1, \dots, N], t \in [0, z_i[,$$
$$\text{where } z_i = D-\frac{1}{f_M}\sum_{k=i}^N w_k,$$
We don’t take frequency change overheads into account, but we have shown in [@RTCSA08] that thoses penalties are easy to integrate.
### Danger Zone
In [@RTCSA08], we define the concept of *Danger Zone*. The danger zone of a task $T_i$ starts at $z_i$, where $z_i$ is such that if this task is not started immediately, we cannot ensure that this task and every subsequent task can all be finished by the deadline (assuming WCEC are correctly known). In other words, if a task starts in its danger zone, and this task and all the subsequent ones use their WCEC, even at the highest frequency, some tasks will miss the (common) deadline. The danger zone of $T_i$ is the range $]z_i, D]$, where
$$\label{eq:DZ}
z_i = D-\frac{1}{f_M} \sum_{k=i}^N w_k.$$
### Notation
In this paper, we will use the following notation:
$$\lceil x \rceil_A = \left\{
\begin{array}{ll}
\min_{y\in A} \{y \big| y\ge x\} & \text{if } x \le \max \{A\} \\
\max_{y\in A} \{y\} & \text{otherwise}
\end{array}
\right.$$
This notation generalizes the classical notation for the ceiling: $\lceil x \rceil = \lceil x \rceil_\mathbb{Z}$. For the sake of readability, $\lceil x \rceil_\mathcal{F}$ will denote $\lceil x \rceil_{\{f_1, \dots, f_M\}}$, or, in other words, the first frequency higher or equal to $x$, or $f_M$ if $x$ is higher than $f_M$.
Table \[tab:symb\] resumes symbols used in this document.
$c_i$ Effective number of cycles of $T_i$
------------------------- ----------------------------------------------------------------------------------
$D$ Deadline
$f_i$ Available frequencies
$M$ Number of available frequencies
$N$ Number of tasks
$\kappa_i(\varepsilon)$ $\mathop{argmin}_\mathcal{K} \{ \mathbb{P}[c_i<\mathcal{K}] \ge 1-\varepsilon\}$
$S_i(t)$ Frequency if $T_i$ starts at $t$
$T_i$ Task number $i$
$w_i$ Worst case execution cycles (WCEC) of $T_i$
$z_i$ Beginning of $T_i$ danger zone
: Symbols used in this document.\[tab:symb\]
Varying WCEC
------------
Let us assume that we have a set of functions $S_i(\cdot)$, supposed the be as efficient as possible, thanks to any algorithm such as [@RTCSA08] or [@Xu07]. Those functions have been computed knowing the execution length distribution, as well as the worst case execution time $w_i$.
It is realistic to consider that the distribution can vary according to the time. As long as $w_i$ does not change, there is no problem: from time to time, when the current distribution is too far away from the distribution used for the computation of $S_i(\cdot)$, those functions are rebuild. The computation of those functions can for instance be done during the slack time of any frame, or on another CPU. Meanwhile, the “old” functions can still be used without risking deadline violation. In the worst case, deprecated functions can make the system consuming more energy than an up-to-date set of functions.
But the WCEC can also vary. If a $w_i$ decreases, again, this does not have any consequence on the schedulability. If this $w_i$ increases, this means that suddenly, an instance of $T_i$ runs over the maximal time it was supposed to run. Here is the aim of this work: study miscellaneous possible reactions to this situation, allowing to reduce as much as possible deadline misses or task killings (which in some cases are intrinsically unavoidably), before the system can rebuild its functions.
The strategies we propose here are composed of two phases:
- First, we have to decide what to do with tasks overpassing their $w_i$ in the current frame;
- Then, we have to adapt the function $S_i(\cdot)$ in order to guarantee the schedulability of the next frame taking into account the new $w_i$ (if other tasks won’t be longer than expected. Otherwise, again, the schedulability is not guaranteed). Of course, we do not completely rebuild $S_i(\cdot)$ which would take too much time. Instead, we change them slightly, improving schedulability, but possibly reducing power consumption efficiency.
In this work, we will assume that we cannot tolerate that at time $D$, some tasks of the current frame are still running. This means that we prefere to kill some tasks instead of missing deadlines. We need then a mechanism allowing to kill a task if still running by some given time that can be provided at the beginning of this task execution.
No preemption mechanism
=======================
We first assume that tasks cannot be preempted: tasks can then be killed, but if at the end of the frame, it happens that there is some time left, the killed task cannot be resumed. We consider that a situation becomes problematic once a task enters (or does not end before) the danger zone of the next task. We have mainly three solutions:
- We do not want any subsequent task to enter the next danger zone. We kill then the running task just before the danger zone of the next task;
- We let the task running without any control. At time $D$, we kill any task that would be still running;
- An hybrid solution.
Kill at danger zone
-------------------
In this case, we consider that a task running longer than expected is “guilty”, and its then killed as soon as it could make any task to be not able to meet its deadline. When task $T_i$ is started, a timeout is set (and disabled when the task finishes) on $z_{i+1}$ (see Equation \[eq:DZ\]), the start of $T_{i+1}$ danger zone (or $D$ if $i=N$).
The main disadvantage is that the danger zone is usually very pessimistic. There is then a high probability (depending upon the computation cycles distribution) that the system kills $T_i$, but some time is still available after $T_N$ (except, of course, if $i=N$).
Another disadvantage of this technique is that, as the next task $T_{i+1}$ starts at the limit of its danger zone, it will be killed as soon as it uses more cycles than $w_{i+1}$. So killing a task “cancels” the laxity of the next one.
Kill at D
---------
Here, we accept a task to enter in the next danger zone. Of course, if a task starts during its danger zone, it should run as fast as possible, then, $$S_i(t) = f_M ~\forall t > z_i $$
At the beginning of a frame, at timeout is set on $D$, and if a task is still running at this time, this task is killed, and any subsequent task is just dropped. The main advantage of this technique is that it allows to use the totality of the available time. Then the “time overflow” has to be huge before a task is killed.
However, the last task of a frame is always expected to run at a speed which allows it to use the totality of the remaining time if it uses its WCEC. Indeed, if the frequency cannot be changed in the middle of a task, this is usually the speed giving the minimum energy consumption. Then, the laxity for the tasks at the beginning of a frame is larger than the laxity at the end, which could be considered as unfair.
Another disadvantage is that if a task is killed or not run, the last one does obviously not reach its end, even if it does not require more than $w_N$. Then, the larger $i$, the larger the probability to be killed or not run. Again, this is unfair.
One possibility is to put at the end of the frame tasks that are less important, or even optional.
Hybrid solutions
----------------
An hybrid solution is to give more space to the last tasks but still starting some tasks in their danger zone expecting them to run less than their WCEC. We allow any task to enter the danger zone of the next task, but the large $i$, the less intrusion is allowed. In a strict solution, $T_i$ would be killed at $z_{i+1}$. We want to define a set of points $\{{\tilde{z}}_i\}$ such as $T_i$ is killed at ${\tilde{z}}_{i+1}$, with the following properties:
- ${\tilde{z}}_i \ge z_i$: we are less strict than just killing at the next danger zone;
- ${\tilde{z}}_i < {\tilde{z}}_{i+1}$: when a task is killed, the next one needs to have some time left;
- ${\tilde{z}}_{N+1} = D$: $T_N$ is killed at $D$ if still running (because of firm deadlines);
We propose two methods allowing this: the first one allows a task $T_i$ to use a fixed ratio of the time interval $[z_{i+1}, D]$, the second one takes the job length distribution into account.
The first one defines then $${\tilde{z}}_i = z_i + (D-z_i) \delta_i$$ for some $0\le \delta_i \le 1$. We can see easily that $\delta=1$ corresponds to “Kill at D”, and $\delta=0$ at “kill at Danger Zone”.
### Percentile approach {#sect:vareps}
An alternative hybrid solution is to consider that any of the subsequent tasks will not need more than its $1-\varepsilon$ percentile. If $\kappa_i(\varepsilon)$ $(\in [0, w_i]) = \mathop{argmin}_\mathcal{K} \{ \mathbb{P}[c_i<\mathcal{K}] \ge 1-\varepsilon\}$, ($\kappa_i(\varepsilon)$ is then the length which is not overpassed by $(1-\varepsilon)\times 100\%$ of jobs) where $c_i$ is the effective number of used cycles, then we only accept $T_i$ to run up to
$${\tilde{z}}_{i+1} = D - \frac{1}{f_M} \left(\sum_{j=i+1}^N \kappa_j \right).$$
In the following, for clarity, $\kappa_i$ stands for $\kappa_i(\varepsilon)$. For instance, we set $\varepsilon$ to $5\%$, and, using the $T_{k}$ number of cycles distribution, we compute $\kappa_k$ such as in $95\%$ of cases, $T_k$ uses less than $\kappa_k$ cycles. Then we behave as if $\kappa_k$ was the WCEC of $T_k$ for each $k>i$.
The tolerance $\varepsilon$ does not have to be the same for every task: we could set $\varepsilon=0$ for some important tasks, and $\varepsilon_i<\varepsilon_j$ if $i$ has an higher priority than $j$.
In order to improve the fairness between $T_N$ and other tasks, we can also consider a decreasing sequence of $\varepsilon$. However, this technique could be considered as unfair: first tasks have much more laxity than last tasks. A way of working around this would be to consider the $1-\varepsilon$ percentile only for some fixed number of tasks, say $K$, and using the following value: $${\tilde{z}}_{i+1} = D - \frac{1}{f_M} \left(\sum_{j=i+1}^{\min\{i+K, N\}} \kappa_j + \sum_{j=i+K+1}^N w_j \right).$$
Available preemption mechanism
==============================
If tasks can be preempted, we can then suspend a task as soon as it enters the next danger zone (or later, at ${\tilde{z}}_{i+1}$), and resume it:
- Either after the end or the preemption of $T_N$. Of course, there is no guaranty that $T_N$ finishes before $D$, but we can assume that in general, it is highly likely;
- Or as soon as some slack is available, which means that a tasks $T_i$ ends before ${\tilde{z}}_{i+1}$.
There are in this case several questions:
- At which speed the resumed task should be run?
- What if several tasks have been suspended?
Frequency for resumed task (resume at the end)
----------------------------------------------
The safest method would be to run a resumed task at speed $f_M$, but it will consume a lot of energy. In some cases, we might have more information. First, we could know the global worst case execution cycle number $W_i$, such as $w_i\le W_i$. In this case, we can choose the lowest frequency allowing to run the remaining part of $T_i$ assuming it could use $W_i$ in total. Then, if $T_i$ is the only task to run at time $t$, if $c_i$ is the number of cycles already consumed, we can choose
$$f = \left\lceil \dfrac{W_i-c_i}{D-t} \right\rceil_\mathcal{F}.$$
If $f_M < \frac{W_i-c_i}{D-t}$ (then $f=f_M$), this means that even at $f_M$, we cannot guarantee to finish the task before $D$. But this does not mean that we are going to miss the deadline, because the task can of course require less cycles than $W_i$. In such a situation, if not missing deadline is more important than energy consumption, $f_M$ is the best frequency.
In some situations, we could for instance assume that the WCEC does not vary too abruptly. In other terms, if $w_i$ is the current WCEC, a task will never need more that $w_i \cdot (1+\alpha)$, for some $\alpha>0$, or at least we accept to kill a task if this condition is not met. In that case, a similar frequency selection can be used: $$f = \left\lceil \dfrac{w_i \cdot (1+\alpha)-c_i}{D-t} \right\rceil_\mathcal{F}$$
Frequency for resumed task (resume at the first slack)
------------------------------------------------------
Again, the most conservative way is to resume tasks at their maximal speed. This conservative method is of course the safest, but consumes more energy. The disadvantage of resuming tasks as soon as we have some available slack is that we don’t really know the available total slack. We only know the time we have before the next tasks start, but some slack might be available later. So several heuristics might be considered:
- Resume tasks at the current speed;
- Resume tasks at $f_M$
- According to the “emergency” (current available slack, number of tasks to be run, probability to find some slack later, …), choose a frequency.
Frequency for other tasks
-------------------------
If we choose to resume tasks after $T_N$, and if a task has been suspended, a first method would be to just “forget” about this task before the end of $T_N$, which means choose the frequency for subsequence tasks without taking this suspended task into account. But in order to reduce the power consumption, the system should select the smallest frequency allowing $T_N$ to finish just in time if is uses its WCEC. If the execution number of cycles variance is rather small, the slack time could be pretty much tiny. Especially because if the user has the choice, it is in general better to put tasks with smaller variance at the end, because this allows to finish the last task very close to the deadline, and then reduce the idle time, and have a better repartition.
A solution would be to increase the speed of tasks as soon as some tasks are waiting in the “suspended tasks queue”. The most conservative way is that we stop using $S$ function when tasks are suspended, and always use $f_M$. A more optimistic way would be to increase the speed that $S$ would have chosen, for instance the next available frequency, or + the number of suspended tasks.
Several tasks
-------------
If several tasks have been interrupted and should be resumed, the situation is slightly more complex, and some strategy needs to be defined. Let $R$ be the set of tasks to resume.
First, if we do not have any information about the maximal number of remaining cycles, the safest way is to run tasks at $f_M$. We also have to choose in what order tasks from $R$ are resumed. The simplest way is to resume them in the order given by indices. But this give some fairness problems: tasks with a higher index have a higher likelyhood to be killed. This could fit the user requirements if tasks are sorted according to their priority, but not in case of uniform priorities.
In order to improve the fairness, a simple method consists in resuming tasks randomly, possibly with some criteria according to the user needs.
If we have for any task in $R$ the knowledge of its global maximum $W_i$ (with possibly $W_i=w_i \cdot (1+\alpha)$), we can then use the frequency $$f =
\left\lceil
\dfrac{\sum_{i \in R} (W_i-c_i)}{D-t}
\right\rceil_\mathcal{F}.$$
Of course, for better energy consumption, this frequency has to be recomputed before resuming any task of $R$, because some previous resumed tasks could have required less time than their expected maximum.
### Multiple rounds
In the previous section, we only consider situations where resumed tasks are not suspended later. But we can also consider that tasks can be interrupted or resumed several times. We then have to wonder if we would better to finish the maximum number of tasks, or to progress in every task in parallel. It might be interesting to consider two families of tasks:
- If a not finished tasks does not give any benefit, the time spent in running such tasks is just wasted. In this case, it would be probably better to run a task up to its completion;
- If a not finished can give some feedback (for instance with lost of quality and/or precision, as it is the case in the Imprecise Computation Model [@Leung07]), then we could try to improve the fairness instead of the number of finished tasks.
In the first case, we want to maximize to number of finished tasks. If we do not have any information about the remaining time, then any order might be equivalent if there are not priority between tasks. If we have some information about the remaining time, then we could heuristically resume tasks sorted by then smallest remaining time first.
In the second case, we could prefer to improve the fairness by running tasks in parallel. Of course, we are only interested by the fairness at time $D$, then we do not need “real” parallelism. For instance, we can allocate to each task in $R$ the same amount of time. Then, if $T_N$ ends at $t'$, every task gets $(D-t')/\parallel R\parallel$ (or any weighted repartition). If a task reaches its allocated amount of time, it is re-suspended. If some tasks needed less than their allocated time, a second round of resumes can be performed. This second round contains necessarily less tasks, as it is because some tasks end before their allocated time that we still have some left time. Notice that the worst number of preemptions occurs when at each round, one task finishes earlier than the time it receives. So if $r=\mid R\mid$, the number of preemptions is $(r-1)\times(r-2)\times\dots\times2 = (r-1)\times r/2$.
In order to avoid to have too much switching times, we could possibly decide (arbitrarily) to run only one task if the remaining time goes bellow some bound.
Preemption and intra-task frequency changes
-------------------------------------------
If a preemption mechanism is available, we can use it in order to increase the frequency of a task. For tasks respecting their WCEC, we assume that scheduling functions do not give any information allowing to optimally change the frequency during execution. So if a task does not need more than its WCEC, we do not consider changing its speed in this work.
But if some task does not respect its expected delay, we can the interrupt it, and resume it immediately at a higher frequency. This allows to be safer, especially if we do not have information about the global WCEC. Several policies can be considered, the safest being to use $f_M$ as soon as a WCEC is not respected. But this strategy is certainly the most energy consuming.
We can also consider to take into account the laxity we have, that is, the remaining time before the next danger zone. We can for instance increase the frequency as the laxity diminishes.
Adaptation of Si(.)
===================
If $T_i$ needed $c_i >w_i$ cycles in the previous frame, then $c_i$ might be considered as the new WCEC (the case where $T_i$ has been killed after $c_i$ cycles is considered later). Of course, it is usually not realistic to rebuild the whole set of scheduling functions $S_i(\cdot)$ before the next frame. We choose then to adapt the scheduling function in order to guarantee the schedulability, where the WCECs are now $w_i^\text{new} = \max\{c_i, w_i\}$.
Let first consider that only one task $T_j$ overran its WCEC. The scheduling has now to guarantee that even if $T_j$ requires again $c_j$ cycles, it will end before $z_{j+1}$. We know that $S_j(\cdot)$ guarantees that if $T_j$ requires $w_j$ cycles, it will end before $z_{j+1}$. So we need to build a set of functions $S'_i(t)$ such as (see Equation \[eq:sched\]):
- If $i < j$: $$\label{eq:Sit1}
S'_i(t) \ge \frac{w_i}{D - t - \frac{1}{f_M} \left(c_j + \sum\limits_{\begin{subarray}{c}k=i+1\\ k\ne j\end{subarray} }^N w_k \right)}$$
knowing that $$S_i(t) \ge \frac{w_i}{D - t - \frac{1}{f_M} \left(w_j + \sum\limits_{\begin{subarray}{c}k=i+1\\ k\ne j\end{subarray} }^N w_k \right)};$$
- If $i = j$: $$\label{eq:Sit2}
S'_j(t) \ge \frac{c_j}{z_{j+1} - t}\text{~~ knowing that ~~} S_j(t) \ge \frac{w_j}{z_{j+1} - t};$$
- If $i>j$: $$\label{eq:Sit3}
S'_i(t) \ge \frac{w_i}{z_{i+1} - t}\text{~~~ knowing that ~~~} S_i(t) \ge \frac{w_i}{z_{i+1}- t}.$$
For $i>j$, we can simply take $S'_i(t) = S_i(t)$. The lower bound of $S_i(t)$ does not depend on tasks running before $T_i$. We will now propose two adaptation methods for $i\le j$.
Using Schedulability Condition
------------------------------
A first adaptation consists in using the property that we want a function $S'_i(t)$ close to $S_i(t)$, but respecting the schedulability condition $S'_i(t) \ge \dfrac{c_j}{z_{i+1}-t}$. We propose then to simply use $S_i(t)$ as long as it respects the condition, and to use this condition when required. Or, more formally,
$$\begin{aligned}
S'_i(t) &=& \left\lceil \max\left\{S_i(t), \frac{c_j}{z_{i+1}-t}\right\} \right\rceil_\mathcal{F} \\
&=& \max\left\{S_i(t), \left\lceil \frac{c_j}{z_{i+1}-t}\right\rceil_\mathcal{F} \right\} \end{aligned}$$
Then, if a task $T_j$ took $c_j>w_j$ cycles in this frame, we make the following changes for the next frame:
- $w_j \gets c_j$;
- Functions $S_i(\cdot)$ are changed in the following way:
$$S'_i(t) \gets \left\{
\begin{array}{ll}
\max\left\{S_i(t), \left\lceil \dfrac{c_j}{z_{i+1}-t}\right\rceil_\mathcal{F} \right\} & \text{if $i\le j$,} \\ \\
S_i(t) & \text{if $i>j$,}
\end{array}
\right.$$
Adaptation of $S_i$ for $i \in [1, \dots, j]$ can be done using Algorithm \[alg:adaptSSched\].
\[alg:adaptSSched\]
Horizontal shift
----------------
Another adaptation can be done using some properties of the function. For $i<j$, we propose to define $S'_i(t) = S_i\left(t+\dfrac{c_j-w_j}{f_M}\right)$.
This corresponds to a left shift of the amount of time required by the “new cycles”, or ($c_j-w_j$). Let us now check the schedulability condition of $S'_i$ for $i\le j$.
We have $$\begin{aligned}
S'_i(t) &=& S_i\left(t+\dfrac{c_j-w_j}{f_M}\right) \\
&\ge& \frac{w_i}{D - t - \dfrac{1}{f_M} \left(c_j-w_j+w_j + \sum\limits_{\begin{subarray}{c}k=i+1\\ k\ne j\end{subarray} }^N w_k \right)}\\
&=& \frac{w_i}{D - t - \dfrac{1}{f_M} \left(c_j+ \sum\limits_{\begin{subarray}{c}k=i+1\\ k\ne j\end{subarray} }^N w_k \right)}\end{aligned}$$ which is the condition .
Combining the previous adaptation for $i=j$ and this one for $i<j$, $S_i$ can be adapted in the following way:
$$S'_i(t) \gets \left\{
\begin{array}{ll}
S_i\left(t+\dfrac{c_j-w_j}{f_M}\right) & \text{if $i<j$,}\\ \\
\max\left\{S_i(t), \left\lceil \dfrac{c_j}{z_{i+1}-t}\right\rceil_\mathcal{F} \right\} & \text{if $i=j$,} \\ \\
S_i(t) & \text{if $i>j$,}
\end{array}
\right.$$
This adaptation can be done in $O(j \times M+M \log M)$ (if $\mid S_i \mid \le M \forall i$), both numbers being usually very small, with Algorithm \[alg:adaptS\_Shift\]. Notice that the logarithmic complexity is because of the “ceiling” (we need to find the smallest frequency higher than a value). But in practice, we don’t have to consider all frequencies, because the new $S_j[k].f$ is obviously higher than the previous one. And if $c_j$ is not too different from $w_j$, the new frequency will simply be the next one, or the next to next one.
\[alg:adaptS\_Shift\]
If several tasks overpass their WCEC in the last frame, we can just consider that this last frame is equivalent to as much frames as the number of tasks having overpassed their WCEC, and in each of those frames, only one task overpassed its WCEC. We can then just apply the transformation given hereabove successively, once for each task.
Remark that if a task has been killed after using $c_i$, it may happen that the new WCEC is actually larger that $c_i$. So we may consider a value higher than $c_i$ as the new WCEC. However, in the next frame, there is a non null probability that $T_i$ still has some laxity after $c_i$ cycles in the next frame. This makes that, stochastically certainly, the system will converge to the new correct WCEC, possibly after having been killed again.
Adapting Killing/Suspend Time
-----------------------------
Let ${\tilde{z}}_i$ be the time at which $T_{i-1}$ is killed (or suspended). According to the relationship between ${\tilde{z}}_i$ and $z_i$, ${\tilde{z}}_i$ must also be adapted when some WCEC changes. Let $z'_i$ (resp. ${\tilde{z}}'_i$) be the danger zone (resp. kill/suspend time) after the change. If $j$ is the index of the task increasing its WCEC to $c_j$, we have
$$z_i = D - \frac{1}{f_M} \sum_{k=i}^N w_i.$$
and
$$\begin{aligned}
z'_i &=& D - \frac{1}{f_M} \sum_{\substack{k=i \\i\ne j}}^N w_i- \frac{1}{f_M} c_j \\
&=& \left\{
\begin{array}{ll}
z_i - \dfrac{c_j-w_j}{f_M} & \text{if }i\le j \\
z_i & \text{otherwise}
\end{array} \right.\end{aligned}$$
Then, the new danger zone can be obtained by subtracting $\dfrac{c_j-w_j}{f_M}$ when $i\le j$.
If ${\tilde{z}}_i = z_i + (D-z_i) \delta_i$ (which includes “kill/suspend at danger zone” and “kill/suspend at $D$”), we have, if $i\le j$:
$$\begin{aligned}
{\tilde{z}}'_i &=& z'_i + (D-z'_i) \delta_i \\
&=& z_i - \frac{c_j-w_j}{f_M} + \left(D - \left(z_i - \frac{c_j-w_j}{f_M}\right)\right) \delta_i \\
&=& z_i + \left(D - z_i \right) \delta_i - \frac{c_j-w_j}{f_M} + \frac{c_j-w_j}{f_M}\delta_i \\
&=& {\tilde{z}}_i - (1-\delta_i) \frac{c_j-w_j}{f_M}\end{aligned}$$
The new kill/suspend time is then obtained by subtracting $(1-\delta_i) \dfrac{c_j-w_j}{f_M}$ for $i \le j$.
The percentile approach is a little bit more difficult to solve, because this method is based on the task length distribution, and we assume that the known distribution is obsolete when we need to adapt $S$ functions, and therefore we should not use it anymore. We need then to adapt the kill/suspend time according to some heuristic.
We propose two transformations for $\kappa_j$ (we assume $c_j > w_j$):
1. $\kappa'_j = \kappa \dfrac{c_j}{w_j}$
2. $\kappa'_j = \kappa + (c_j - w_j)$
The first adaptation assumes that the whole distribution is stretched from $[0, w_j]$ to $[0, c_j]$. The second adaptation assumes that the distribution is shifted upwards with a shift of $(c_j-w_j)$. We consider the generic general percentile approach:
$${\tilde{z}}_i = D - \frac{1}{f_M} \left(\sum_{k=i+1}^{\min\{i+K-1, N\}} \kappa_k + \sum_{k=i+K}^N w_k \right).$$
The first adaptation gives:
$${\tilde{z}}'_i = \left\{\begin{array}{l}
D - \frac{1}{f_M} \left(\sum\limits_{k=i}^{m_i^K} \kappa_k + \kappa_j(\frac{c_j}{w_j} -1)+ \sum\limits_{k=i+K}^N w_k \right) \\ \\
\hfill \text{if }j\le m_i^K \\ \\
D - \frac{1}{f_M} \left(\sum\limits_{k=i}^{m_i^K} \kappa_k + \sum\limits_{k=i+K}^N w_k + (c_j-w_j)\right)\\ \\
\hfill \text{otherwise}
\end{array}
\right.$$ where $m_i^K = \min\{i+K-1, N\}$, or
$${\tilde{z}}'_i = \left\{\begin{array}{ll}
{\tilde{z}}_i - \dfrac{\kappa_j}{f_M} (\frac{c_j}{w_j} -1) & \text{if }j\le m_i^K \\ \\
{\tilde{z}}_i - \dfrac{c_j-w_j}{f_M} & \text{otherwise}
\end{array}
\right.$$
If we consider the second adaptation, we can easily get:
$${\tilde{z}}'_i = {\tilde{z}}_i - \dfrac{c_j-w_j}{f_M}$$
WCEC diminuing
--------------
If the WCEC for some job has not been observed for a long time, we could consider to reduce it. Let $c_j$ ($<w_j$) be the “new” worst case execution cycle of $T_j$. While there is no schedulability necessity to adapt the scheduling, we could try to take this information into account in order to reduce the consumption. Again, we want to have simple changes before recomputing correctly scheduling functions.
Of course, we still need to ensure the schedulability (according to the new WCEC set). For instance, we could use an adaption very close to the one we did before:
$$S'_i(t) \gets \left\{
\begin{array}{ll}
S_i\left(t-\dfrac{w_j-c_j}{f_M}\right) & \text{if $i<j$,}\\ \\
\left\lceil S_j(t) \dfrac{c_j}{w_j} \right\rceil_\mathcal{F} & \text{if $i=j$,} \\ \\
S_i(t) & \text{if $i>j$,}
\end{array}
\right.$$
In the case $i<j$, we want to do a right shift. The schedulability of the case $i=j$ is very easy to show. The schedulability for other cases can be proved in a very similar way as before. Algorithm \[alg:adaptS\_Shift\] can be used with a slight adaptation.
Simulation Results
==================
Scenario
--------
{width="\lw"} {width="\lw"} {width="\lw"} {width="\lw"}
In the following simulation results, we will present two different scenarios. This first one is very simple and smooth, the second one uses data coming from experimental measurement on video decoding platforms. For the first scenario, we assume 4 tasks, using a normal distribution for execution length. We consider two phases: the first one, for which characteristics are known by the system, and which contains roughly 160 frames, and the second phase (40 frames), in which three out of the four tasks change their behaviour, and increase their average and worst execution time. This 200-frames scenario is of course run several hundreds of times to obtain usable statistics. Figure \[fig:scen\_norm\] shows graphically an example of actual job execution number of cycles.
The reason we use such a short scenario (200 frames) is that the “critical interval” where jobs are killed is most of the time very short: if we consider a scenario close to the saturation, only a very few number of jobs need to be killed before the system knows the new worst case execution times, and does not have to kill jobs any more (unless the load is too high and the schedulability cannot be guaranteed anymore). And this number of killed tasks only depends of the way the transition of the two phases happens, and not of the length of those two phases.
{width="\lw"} {width="\lw"} {width="\lw"} {width="\lw"}
The second scenario uses data collected on experimental environment in the Computer Science and Information Engineering department of the National Taiwan University, Taipei. We consider 9 tasks, and assume that they are distributed according to the workload presented in [@RTCSA08]. Again, we consider two phases, but much longer: the first phase contains 20.000 frames, the second one 4000. See Figure \[fig:scen\_DIV\]. We refer this scenario as the *DIV workload*.
As the second scenario contains a huge amount of frames, the effect of the transition (number of killed job/lost cycles before the S-functions are up-to-date) is very small. In order to highlight this effect, we also present a short version of this scenario, keeping only 250 frames amongst 20.000.
For every following figure but Fig. \[fig:scen\_DIV\] and \[fig:scen\_norm\], the horizontal axis represent the frame length. This means that to produce such a plot, we simulated the behaviour of a system with a large frame length (or deadline), measure some metric, and start the same experiment with a shorter frame length. Large value of frame length corresponds then to a low load, and small value to a high load.
Fairness
--------
Measuring the fairness experimentally is not very easy, because for most cases, killing a job should be something quite rare. We first proposed to measure the laxity of a job as:
$$\mathcal{L}_i = \frac{1}{\text{nb instances of }T_i} \sum_{\text{instance $k$ of $T_i$}} \frac{e_k}{r_k}$$ where $e_i$ is the number of cycles that instance $i$ actually run, and $r_i$ is the number of cycle that instance $i$ required. The lower $\mathcal{L}_i$, the higher the number of lost cycles. We define the fairness as:
$$\frac{\min_i\{\mathcal{L}_i\}}{\max_i\{\mathcal{L}_i\}}$$
The drawback of this measurement is that a strategy which kills a very few jobs but always the same (for instance the last one) will have a fairness very close to $1$, while a strategy which kills more often, but different jobs will have higher fairness. However, intuitively, and fair strategy should be a strategy which, when jobs are killed, each task has a similar probability to be the victim. So we propose then
$$\mathcal{L}_i = \frac{1}{\text{nb killed of }T_i} \sum_{\text{killed instance $k$ of $T_i$}} \frac{e_k}{r_k}$$ Notice that at low loads, the number of killed jobs is usually very small, then the fairness is computed on a small number of jobs, which gives pretty much erratic values.
No preemption mechanism
-----------------------
![Fairness, without preemption mechanism, with four tasks (normally distributed length). The closer to 1, the fairer.\[fig:norm\_nopreempt\_fairness\]](Figs/fairness){width="\linewidth"}
![Killing rate, without preemption mechanism, with four tasks (normally distributed length)\[fig:norm\_nopreempt\_kill\]](Figs/killrate){width="\linewidth"}
![Energy consumptions, without preemption mechanism, with four tasks, and normally distributed length\[fig:norm\_nopreempt\_energy\]](Figs/energy){width="\linewidth"}
In this first set of simulations, using the workload corresponding to Figure \[fig:scen\_norm\], we assume that we do not have any preemption mechanism, and that we are then only allowed to kill a job and not to resume it. We show in Figures \[fig:norm\_nopreempt\_fairness\], \[fig:norm\_nopreempt\_kill\], and \[fig:norm\_nopreempt\_energy\] the influence of the factor $\delta$ — which gives the laxity we authorise between the danger zone of the next frame, and the end of the frame — on the fairness, the killing rate and the energy consumption.
It is not surprising that if no flexibility is given (jobs are killed at the next danger zone, $\delta=0$), the fairness (Fig. \[fig:norm\_nopreempt\_fairness\]) is much better (values close to 1) than if we give a large flexibility ($\delta=1$), because in this later case, last jobs are more likely to be killed than first jobs. On the other hand, being more “rigid” increases the number of killed jobs at the end of any frame, because we waste some free time that would be available at the end. So choosing an appropriate value of delta should be done carefully, according to the kind of workload we consider, but also depending whether we have to give more importance to fairness of killing rate. In the example we give here, $\delta=0.2$ seems to be a good trade-off: the fairness (Fig. \[fig:norm\_nopreempt\_fairness\]) if very close to $\delta=0$, but we are comparable to $\delta=1$ regarding to the killing rate (Fig. \[fig:norm\_nopreempt\_kill\]). From the energy point of view (Fig. \[fig:norm\_nopreempt\_energy\]), we don’t see any significant difference between different $\delta$’s.
Remark that on the right side of the energy plot (Fig. \[fig:norm\_nopreempt\_energy\]), we see a very huge difference between $\delta$’s. But at this load, a huge number of jobs intrinsically need to be killed, because the frame length is no small to allow tasks to finish. So in most case, designer are not really interested in the system behaviour at those loads.
Preemption mechanism
--------------------
First, we compare the difference between having or not a preemption mechanism (Figure \[prempt-nopreempt\]). Pretty much obviously, we observe that the killing rate is lowered when preemption/resuming mechanism is available. We show for instance a simulation with the normal distribution and $\delta=0.2$. Notice that if we consider high $\delta$, we don’t see any difference between using or not a preemption mechanism, because if $\delta$ is high enough, jobs are almost never suspended, and only the last one in the frame is killed.
![Comparison between using or not using preemption mechanism, with four tasks (normally distributed length), and $\delta=0.2$.\[prempt-nopreempt\]](Figs/prempt-nopreempt){width="\linewidth"}
In Figures \[prempt-energy-S\], \[prempt-lost-S\] and \[prempt-fairness-S\], we show respectively the energy consumption, average number of cycles, and fairness for various $\delta$ factors, when preemption mechanism is available, and using the normal distribution workloads. We can see in this set of simulation that there is no need to let jobs running in the danger zone: we are better to suspend them as soon as they enter the next danger zone ($\delta=0$), and resume them when some slack time become available.
![Relative energy consumption, with respect to $\delta=0$, with four tasks (normally distributed length), with preemption mechanism (resume at slack)\[prempt-energy-S\]](Figs/prempt-energy-S){width="\linewidth"}
![Killing rate, with four tasks (normally distributed length), and preemption mechanism (resume at slack)\[prempt-lost-S\]](Figs/prempt-killrate-S){width="\linewidth"}
![Fairness, with four tasks (normally distributed length), and preemption mechanism (resume at slack)\[prempt-fairness-S\]](Figs/prempt-fairness-S){width="\linewidth"}
Realistic workload
------------------
In the few next plots, we will show some results where a more realistic workload (see Figure \[fig:scen\_DIV\]) has been used for simulations. We aim here at comparing our adaptation method to two other scenarios: first, a simple case where the scheduling function is not adapted when the workload changes, then an idealistic situation where we know in advance when the change occurs, and we also know the distribution of the two phases.
In the first figure (Figure \[DIV-energy\]), we observe that our adaptation method saves much more energy that no adaptation at all. The very low energy consumption we can see for the no adaptation scenario at high load comes from the high rate of killed jobs for this method. Indeed, all of those lost cycles do not consume energy at all.
In Figure \[DIV-killrate\], we show the average killing rate for those three scenarios. However, if we consider the same workload as before, we could not see any difference between our dynamic method and the “2 phases” method, because our adaptation only needs to kill a few jobs, and this number of jobs does not depends upon the length of the simulation. It’s why in this plot, we present measurements done on a much shorter workload ($\pm$ 250 frames). Please notice that the killing rate in the “no adaptation” scenario was not affected significantly by the length of the simulation, as long as we keep the same ratio between the first and the second phase. We observe in this simulation that the killing rate in our adaptation can be very close to zero, which means that even if we do not have a good knowledge of the distribution (and its worst case execution time), we can still, with very small effort, avoid to kill most jobs that would have been killed if we needed to collect a new distribution before adapting the $S$-functions.
![Energy consumption, with 9 tasks (DIV workload). We compare our dynamic adaptation mechanism with a clairvoyant method (Two phases), and a simple method which does not adapt its information when the length raises up.\[DIV-energy\]](Figs/DIV-energy){width="\linewidth"}
![Killing rate, with 9 tasks (DIV workload). Same comparison as for Figure\[DIV-energy\].\[DIV-killrate\]](Figs/DIV-killrate){width="\linewidth"}
Miscellaneous experiments
-------------------------
We have also performed many experiments we do not present in details here, by lack of space. Here are a few conclusions we have drawn:
- We almost have not observed any difference between “resume at end” and “resume at slack”. Slack can be used very often, but when slack used, we already entered danger zone in most cases, and then often the frequency is $f_{max}$ up to the end.
- The two adaptation methods we proposed (using schedulability condition or horizontal shift) do not show significant difference, whatever the metric we consider.
- Idem for the adaptations we proposed for the danger zone: they both show the same performance or fairness.
Conclusion
==========
In this paper, we have shown that with a small effort, we can efficiently manage varying WCEC on DVS frame-based systems. We provide several algorithms and methods allowing to first have an efficient behaviour as soon as some task overpasses its WCEC, and secondly, adapt the scheduling functions to improve the schedulability. We provide several proofs showing the correctness of our algorithms, and present many simulation results attesting the performance of the proposed methods. Through those simulations, we shown that we can be very close to a clairvoyant algorithm, both from the killing rate point of view, and from the energy consumption point of view.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.